2026-10-01

Language agents must revise planned actions when evidence changes, permission is revoked, or a stop instruction arrives. A useful response is selective: suspend affected actions, preserve unaffected work, and resume only after sufficient repair. We introduce NAQD-Env, a synthetic environment that evaluates these decisions against a deterministic reference policy over explicit evidence, authorization, and constraint dependencies. Eleven dependency families support evaluation on development structures, held-out families, and held-out combinations of structures. Metrics distinguish attempted violations from violations permitted by a simulated execution gate and jointly report policy agreement, task value, withdrawal, resumption, and event reporting. We evaluate three open-weight instruction-tuned models from two families under three prompt conditions on 350 frozen scenarios, yielding 3,150 model-prompt episodes before gate replay. Across the reported conditions, withdrawal recall is at most 0.06, no valid resumption is observed at eligible opportunities, and only one episode matches the complete reference policy. Under the NAQD prompt, Qwen2.5-7B has fewer unsafe-attempt episodes than Qwen2.5-3B and Llama-3.1-8B, but also completes less useful work and preserves unaffected actions less accurately. Exploratory supervised fine-tuning probes increase Qwen2.5-3B decision accuracy from 0.45-0.54 to 0.83-0.92; separate diagnostics reveal inappropriate withdrawal after curriculum omissions and a loss of event reporting. These results motivate evaluating selective withdrawal as a distinct component of agent reliability. The setting measures policy application with trusted structured inputs and does not establish real-world containment or source-verification ability.

Most clinical benchmarks evaluate language models (LMs) on diagnosis using complete case descriptions. In clinical practice, however, patients present information in different ways, and clinicians must obtain relevant history and determine which examinations are needed before reaching a diagnosis. Diagnostic accuracy alone therefore cannot establish whether an agent gathered essential information or conducted an appropriate clinical assessment. Furthermore, existing benchmarks lack professional clinicians' verification. To address this gap, we introduce KlinikeBench, a benchmark of 333 clinician-authored tasks, each providing an isolated sandbox environment with a virtual patient, clinical tools, and task-specific success criteria. More than 35 clinicians contributed to case authoring and benchmark evaluation. In an empirical study, clinicians gave simulated dialogues higher mean quality ratings than reference conversations, which is adapted from real conversation. In each task, an LM has a fixed budget of turns to communicate with the patient, ask about relevant history, request examinations, follow action constraints, and record a final diagnosis. We score these steps separately as well as together. Across 31 models and seven model families, the best-performing models (e.g., GPT-6-astra and Claude Opus 5) succeed on less than 30% of tasks, even though their diagnosis accuracy reaches 90.7%. Some models benefit from talking with the patient; others diagnose well from a complete chart but perform much worse in conversation. Overall, KlinikeBench provides a testbed for evaluating the full clinical encounter and reveals a substantial gap between diagnostic accuracy and performance in interactive clinical assessment.

Constantly evolving real-world knowledge necessitates models to be updated continuously. Especially in medicine, as clinical evidence changes over time, outdated knowledge can pose safety risks. Existing evaluations of knowledge integration focus on factual recall, offering limited insight into whether newly integrated knowledge is actually usable. Our benchmark MedKIT (Medical Knowledge Integration and Transfer) provides a granular evaluation of how models integrate and apply knowledge under realistic sequences of clinical updates. Each instance corresponds to a factual update derived from clinical evidence, paired with targeted probes that assess transfer across lexical variation, relational transformations, compositional reasoning, and open-ended operationalization, as well as locality tests for knowledge preservation. Using MedKIT, we conduct a large-scale empirical study of 12 knowledge integration strategies across 5 diverse models, including both general-purpose and medical LLMs. Our results reveal a consistent gap between recall and usable knowledge: while most methods achieve strong gains on the original update task and under lexical variation, relational generalization is limited, and no method yields meaningful improvements on compositional or operational tasks. These findings highlight a fundamental challenge in knowledge integration and position MedKIT as a testbed for developing methods that make newly integrated knowledge more consistently usable across tasks and contexts.

Large language models (LLMs) encode substantial latent geographic knowledge, yet they reason poorly over space and are unreliable when queried from coordinates alone. Useful behaviour emerges only when structured spatial context is supplied in the prompt. This raises a question geographic evaluation has left unexamined: once the right context is supplied, does the model reason from it, or override it with its own parametric recall? We take up this question with an open pipeline for network-based spatial context retriev-al. In it, the surroundings of a selected point are defined by the pedestrian street network, the area actually reachable on foot. Using only open data and open-weight models, the pipeline retrieves features from OpenStreetMap and the GHS-POP population grid, computes indicators over the network catchment in code, and injects them as a compact spatial brief. On this basis we build a faithfulness benchmark. It labels every claim a model makes by its source (grounded in the brief, or drawn from training knowledge) and its correctness, and it probes each case with a planted false premise that the brief refutes. We evaluate sixteen open-weight model configurations across three families (Qwen, Gemma and Llama, with Gemma in two generations), four size classes and, where available, both thinking and non-thinking modes, on three con-trasting cities, resampling every case over ten seeds. The results show that resistance to the planted premise varies more strongly by model family and generation than by scale, while brief-reading competence forms a partly separate dimension. These behaviours are not captured by conventional world-correctness scores or single-shot evaluation. We release the implementation, spatial briefs, model outputs, and claim-level labels as a reproducible workflow at github.com/perezjoan/NSCR-LLM.

Large language models frequently fail to balance staying truthful with being supportive. They often exhibit sycophancy in responses to users, agreeing with false claims, offering unwarranted flattery, and giving advice skewed toward users' expressed views. In reality, sycophancy rarely happens in a single exchange; it may emerge organically as users repeatedly insist or subtly steer the dialogue over time. Current evaluations, however, rely on rigid, single-turn tests or fixed scripts that fail to capture these natural dynamics. Furthermore, these benchmarks often mistake showing basic empathy for yielding, penalizing models for acknowledging a user's feeling. This view may drive future models to over-correct into cold, dismissive rigidity. To address this gap, we introduce FIGS (Factual Integrity and Grounded Support), a dual-axis evaluation framework built around extended, realistic dialogue. We use an adaptive 10-turn conversational simulator that dynamically challenges the target model, reflecting how users repeat requests, push back, or steer a conversation toward a preferred answer. To accurately evaluate these trajectories, we apply a taxonomy that strictly separates Sycophancy (whether the model holds firm to the truth and keeps its praise proportional) from Calibrated Validation (showing empathetic understanding of the user's feelings without overdoing it). We release our complete testing environment, including 500 diverse multi-turn scenarios and an automated judge. Our evaluation of leading models reveals a consistent trade-off: over the course of a sustained interaction, current systems either slowly drift to sycophancy or over-correct into robotic detachment. This demonstrates that balancing honesty with appropriate support throughout a natural conversation remains a critical, unsolved challenge.

Values such as honesty, autonomy, and confidentiality are often regarded as general principles underpinning AI alignment. However, what it means to act in accordance with these values can depend on the context in which a decision is made. In this paper, we ask whether large language models (LLMs) appropriately adapt the application of a value across professional settings, while remaining consistent when contextual changes do not alter the relevant professional norm. To study this, we introduce ContextAdapt, an evaluation framework covering honesty, autonomy, and confidentiality across medicine, law, finance, and national security. Drawing on primary-source professional and regulatory documents, we construct a value x domain framework and use this to develop scenarios testing both default professional rules and recognised exceptions. We evaluate 12 LLMs on both the actions they recommend and the justifications they provide. In our main experiment, models achieve 95.6% mean appropriateness, although the use of the correct domain-specific justification varies substantially across models, from 25.6% to 76.9%. In a separate factorial experiment, explicitly naming the professional domain and changing the role of the model have limited effect on behaviour. Varying stakes, however, reveals severe but localised failures: in some cases, models alter their responses even though the underlying professional obligation remains unchanged. In particular, perceived severity appears to act as a cue for disclosure across both honesty and confidentiality scenarios. These results show that evaluating value alignment requires us to consider not only whether models follow abstract principles, but whether they apply them appropriately across different contexts.

When a materials LLM answers a question about crystal structure, does it reason from the structure or copy an answer already printed in its input? Accuracy cannot tell: a structural description often prints the very field it is scored against. CARAT holds question and gold answer fixed across eight matched views, names each structural relation separately in GraphSpace, and adds matched fine-tuning, answer masking, evidence injection, paired inference, and a rule that can withhold claims. First, on the benchmark's hardest families the grounded view is worth 17.3 points over formula inputs. Second, we turn that scrutiny on ourselves. GraphSpace beats a plain periodic graph by 19.3 points, but that margin is two effects at once: where the plain rendering carries everything the question needs it is 1.96 points, and where it omits those fields entirely, 46.7 points. The headline mostly measures what the baseline lacked, not how evidence is presented. Third, we attack our own benchmark. A rule that skips the link and reads the list directly answers four of seven hardened families, so we rebuilt it until eleven such shortcuts sat near chance. The frozen model quotes that link yet answers the same when we redirect it, on 95.6% of paired cases: it repeats the relation without using it. After matched supervision it reaches 99.8%, and deleting the link drops it to 23.4%, below the 27.0% the best shortcut reaches: both steps are learnable.

Access to real-world information is often noisy and fragmented. Constructing a coherent narrative from such fragments requires models to reconstruct missing spans within a broader storyline, commonly referred to as text infilling, while preserving consistency with both the local context and the global storyline. Despite using text infilling as a pre-training objective in many Large Language Models (LLMs), their actual performance on real-world narrative infilling remains underexplored. In this paper, we address this gap by introducing a multi-domain benchmark of ~9.2K instances for narrative infilling, constructed by masking one to three sentences across four narrative types: encyclopedic text, commonsense stories, news articles, and visual narratives. Using this benchmark, we evaluate 20 instruction-tuned open-source LLMs ranging from 1.5B to 70B parameters across varying levels of instruction specificity and reasoning guidance. Outputs are assessed using standard automatic metrics and a qualitative framework covering five narrative dimensions. Results show that model scale does not reliably predict infilling quality: Gemma-2-2B achieves the highest qualitative score (4.02/5), outperforming models over ten times larger, including DeepSeek-Qwen-32B (3.77/5, 6.6%) and LLaMA-3.3-70B (3.71/5, 8.3%). We further find that explicit reasoning offers limited benefits as chain-of-thought reasoning yields only a marginal improvement (+0.6%). Additionally, short narratives and domain characteristics emerge as stronger predictors of task difficulty than infill position alone for narrative infilling in current LLMs.

Recent progress in large language model reasoning has been driven by benchmarks and reinforcement learning environments with automatically verifiable rewards, particularly in mathematics, code, and formal logic. These settings make model accuracy easier to evaluate and optimize, but it remains unclear how far success under fixed problem specifications and stable evaluation criteria transfers to reasoning outside such domains. Real-world reasoning often proceeds under incomplete and revisable information: conclusions may be supported provisionally, defeated by counter-evidence, reinstated by further arguments, or revised when stronger reasons become available. Reasoning of this kind is generally referred to as defeasible reasoning. We introduce ArgGYM, a procedural benchmark and RLVR-compatible training environment for structured defeasible reasoning. ArgGYM decomposes this reasoning into twelve tasks and grounds task-specific scoring in a symbolic argumentation engine that computes the formal states used to evaluate model outputs. It includes a frozen benchmark of 1,440 verified instances across fifteen curriculum configurations, two argument preference orderings (weakest-link and last-link), and two set orderings (elitist and democratic), while the same generators and verifiers can produce fresh instances for evaluation that reduces dependence on static test sets and for verifiable-reward training. On the frozen benchmark, frontier and open-weight models show sharply different reasoning profiles: they can recover substantial parts of structured answers without solving the complete task, and performance declines in later curriculum configurations with longer dependencies and more interacting structures. We release the benchmark, generators, and verifiers for reproducible evaluation and RLVR training.

Agent benchmarks test agents in worlds that stay still. Deployed agents work in worlds that other people also change. Someone texts the agent to send the money elsewhere or an order confirmation asks it to reply with a door code. We present BACKDROP, which asks how much of an agent's capability in a clean world survives in such a world. BACKDROP takes a task along with the agents execution environment, and plants four everyday hazards in its world, one at a time and all together. The instruction and the correct end state stay the same. Each hazard asks one question. Authority: does a message from another person override the user? Injection: does text planted in a record redirect the agent? Boundary: does a request pull it into an app it was not given? Fault: after a write fails without saying whether it landed, does the agent check before it retries? Across 3,678 variants and 16 models, , the average pass rate falls from 69.5% to 31.3% once all four hazards are present; the strongest models fall furthest (Claude Fable 5.1 from 96.6% to 56.0%). Agents have learned to resist injected text but often follow other unauthorized requests of other people. With all four hazards present, and counting only runs where the planted text reached the agent, agents followed another person's message in 46.4% of runs and injected text in 20.3%. The gap is consistent throughout all 16 models. BACKDROP formalizes these gaps and shows how an agent's score in a task's world is a ceiling on real-world performance.

Voice agents in production must handle several requests, background speech, and customers who lose patience. We introduce VAmoS Energy, a benchmark that combines these challenges in 100 calls about utility billing and payment assistance. Each caller makes two to four requests. The agent has sixteen tools backed by a stateful Stripe billing twin and the Apache Fineract loan engine, with account access blocked until caller verification succeeds. The tasks use public household electricity data and a policy based on Pennsylvania's residential billing rules. An LLM-as-a-verifier checks the agent's actions and spoken figures against explicit requirements. On a calibration run, it agrees with a code verifier on 99.1% of checks. Across fourteen voice stacks and three repeats per task, completion ranges from 17.3% to 44.7%. Grok Voice leads, and Gemini 3.8 Live and GPT-Live follow at about the same cost per call. Background television reduces pooled completion from 38.7% to 8.6%. The simulated caller often accepts an incorrect result because it hears the agent's words but cannot inspect its actions. These findings show why voice agents need evaluation across the whole call, including what they say, what they change, and how they handle competing speech.

Financial forecasting from earnings conference calls requires models to reason over complex corporate disclosures, market expectations, and subtle communication signals. However, existing financial benchmarks are often limited to unimodal inputs or single-task settings, making it difficult to evaluate whether multimodal large language models (LLMs) can support real-world financial analysis. In this paper, we introduce MM-FinEval, a novel benchmark designed to evaluate multimodal LLMs across multiple financial tasks. MM-FinEval spans a diverse timeline from 2019 to 2022. The entire proposed dataset contains 2,045 S\&P 500 conference earning calls as inputs and 12 financial task labels as outputs. Each input contains three modalities: a word-to-word text transcript of the earning call, the corresponding presentation slides used during the call, and the entire audio recording. To establish a rigorous evaluation framework, we analyze 19 baseline models across three distinct model categories: Image-Text, Audio-Text, and Any-to-Any configurations. We observe that small-size Any-to-Any models processing all three modalities achieve strong performance, even when compared against larger proprietary models restricted to two-modality inputs. This indicates that our tri-modal dataset design introduces useful, non-redundant information. These results validate that text, audio, and visual data serve as important, complementary signals that mimic the decision-making process of expert human analysts.

As large language models (LLMs) are increasingly used in clinical settings, it is critical to evaluate their reliability under realistic variation in clinical text. We study this question in clinical triage, comparing LLMs to practicing physicians under text perturbations that preserve the underlying clinical setting. We introduce a benchmark of over 6,000 clinical scenarios, 7,000 physician annotations, and 225,000 model responses. Using this benchmark, we make two key observations. First, LLMs are more likely than physicians to recommend unnecessary care at baseline, and this tendency increases under perturbed inputs. Further, we find that LLM recommendations are more sensitive to gender and tone perturbations than human recommendations. Together, these results demonstrate that LLMs can vary under clinically irrelevant textual changes, highlighting the need for deployment-oriented evaluations grounded in expert physician behavior.

Modern LLM agents increasingly tackle complex tasks through interactive, long-horizon exchanges with users, while existing benchmarks generally assume that users always accurately and sufficiently communicate a fixed intent. However, this oracle communication assumption rarely holds in practice: users may miscommunicate, change their goals, and run out of patience. We define this task setting as Interactive Intent Alignment, where agents must recover and continuously track the user's current intent despite imperfect communication and evolving goals. To study this setting, we introduce Drift-Bench++, a principled benchmark construction pipeline for verified executable tasks with controlled misalignment and intent shifts, along with an interaction protocol featuring finite patience, diverse simulated users, and silent interaction-conditioned shifts. We further develop GRIP, a comprehensive evaluation protocol covering task grounding, user realism, inquiry effectiveness, and adaptation to evolving intent. Across diverse environments, models, and interaction conditions, stronger interaction consistently helps but remains far from oracle performance; Validation on deployed ProdAgent sessions further shows that the modeled failures are prevalent and consequential in deployment. By providing a unified, executable benchmark for interactive intent alignment, Drift-Bench++ offers a foundation for evaluating and advancing agents under realistic communication and evolving intent.

As natural language drives more applications, language models increasingly run inside programs as decision components: the program sends them the current state and acts on the returned decision until a newer one arrives. When evidence changes during inference, a decision correct for its own state can stay in force after that state has passed, as when a call recorder keeps running after a customer starts reading out a card number; untimed (offline) accuracy counts such an error as correct. We introduce StreamDecisionBench (SDB), which evaluates the decision in force at every instant and attributes every erroneous instant to judgment, latency or both. Its scenarios stream evidence in four application families, with reference decisions computed from public rules by executable code. We summarize in-force accuracy across update intervals of 1-5 s by its normalized area under the curve on a logarithmic time axis, giving equal weight to equal multiplicative ranges. Across six settings of four hosted models, this score stays within 2.9 points of the scenario-wise product of untimed accuracy and an oracle's integrated timing score. Reasoning improves judgment, but at low effort latency costs Luna and Terra, two GPT models we also evaluate without reasoning, 38.8 and 42.8 points relative to untimed accuracy; a faster component with weaker judgment attains a similar integrated score to Terra without reasoning. The aggregate and family curves show where these tradeoffs change, making the evaluation's time-scale dependence visible.

Agentic AI systems increasingly rely on cloud-hosted large language models for planning, tool use, and iterative execution, raising concerns about API cost and data exposure. Advances in personal AI devices enable agents to execute locally, but limited resources on device may affect task success and performance. Existing benchmarks are inadequate for systematically characterizing these trade-offs across devices, workloads, and deployment architectures. We present AgBench, a benchmark suite and open artifacts for reproducible evaluation of agentic AI on personal devices. Using AgBench, we evaluate local, hybrid, and cloud execution across agentic workloads, examining task success, latency, cloud API cost, and data exposure. Our results, drawn from over 162.07 million data points, show that personal AI devices can complete many agent tasks locally, but local-only execution generally has lower task success and longer completion times than cloud-only execution, especially as concurrency increases. Local-only execution eliminates cloud model API costs and sensitive-information exposure to cloud agents. Hybrid execution can improve task success, but its cloud cost and data exposure depend on how agents divide work and share information. No single architecture performs best across task success, goodput, cloud cost, and data exposure; deployment choices should reflect the intended workload and device capabilities. AgBench is available at https://anonymous.4open.science/r/AgBench-2777.

Long-horizon agentic workflows require models to sustain repeated state-dependent actions all while the context grows, sub-task complexity changes, and new data arrives. Each situation represents an independent axis along which an agent may fail. An agent reconciling a long ledger, for example, must repeatedly read its state, update the correct record, and preserve alignment across thousands of outputs. A model may accept the entire ledger yet lose its place or stop applying the operation consistently as generation proceeds. We introduce Long-Transduction, a controlled diagnostic that tests a model's ability to stay on task during long generation while continuously reading, mutating, and outputting input-context dependent operations such as arithmetic, sorting, variable lookups, and table transformations. Long-Transduction evaluation independently varies local task complexity, input data formatting, and context length isolate failures along each axis. We evaluate seven open-weight models, finding a 62.8\% decrease when scaling context length from 4-128K, a 36.5\% decrease when varying input format, and a 39.9\% decrease by increasing local task complexity. Together, these failures represent critical liabilities in long-horizon agentic workflows.

Multimodal large language models (MLLMs) offer a promising approach for recovering numerical data from scientific charts, but their ability to reconstruct chart data from visually dense figures remains poorly understood. Existing chart understanding benchmarks primarily evaluate question answering or chart-level reasoning and provide limited support for evaluating structured numerical reconstruction from scientific figures. We introduce \textbf{ChartDensity-Bench}, a benchmark for evaluating MLLMs on structured numerical data reconstruction from compound chart figures under controlled visual density. Built from charts paired with source-level ground-truth data, ChartDensity-Bench systematically varies the number of simultaneously presented charts ($k\in{1,3,6,9}$), enabling controlled evaluation of density-induced degradation. We further propose a multi-dimensional evaluation framework covering structural reliability, reconstruction completeness, parseability, and numerical fidelity. Experiments on five recent MLLMs show that numerical reconstruction generally degrades as visual density increases, while the magnitude of degradation varies substantially across models. Chart-level paired comparisons further show that the same source chart can incur higher reconstruction error when embedded in denser visual contexts. These findings highlight visual density as an important and previously underexplored factor in MLLM chart data reconstruction and provide a systematic benchmark for evaluating model robustness in this setting.

Understanding multi-perspective alternative narratives requires identifying how their information agrees, conflicts, or differs across sources. Existing work on cross-text relations largely focuses on categorizing relations between predefined text pairs, such as entailment or contradiction, rather than directly extracting such information from full narratives. To address this gap, we introduce Overlap-Unique-Conflict (OUC) extraction, a cross-narrative task that extracts all overlapping, conflicting, and unique clauses from two narratives. To support this study, we construct a benchmark of approximately 22K narrative pairs and 140K OUC instances spanning factual, argumentative, and political discourse. Evaluating 14 open-source LLMs (0.6B-35B), we find that unique information is far easier to extract than overlap and conflict: the strongest model, Gemma-4-31B, reaches only 61.13% F1-score on overlap and 48.58% on conflict, against more than 75% on unique. Further diagnostic analysis reveals that this difficulty does not stem from relation recognition alone, but rather from a failure to pair and extract the corresponding clauses from full narratives, especially in smaller models. Nevertheless, learning these extractions with task-specific supervision narrows the gap considerably: a fine-tuned Qwen-3-8B gains 15-28% absolute over its baseline and surpasses models roughly four times its size (e.g., Qwen-3.6-35B) on several tasks. Even so, overlap and conflict remain well below satisfactory, leaving cross-narrative clause extraction an open challenge.

End-to-end accuracy on compositional tasks records how often MLLMs fail, but cannot distinguish whether a failure reflects an intrinsic deficit in the targeted capability or a cascading error from an upstream prerequisite. We propose a causal decomposition framework that isolates these two failure modes through controlled interventions on the prerequisite dependencies of each task. Our capability metrics (NC, IC, RC) score each task under unassisted, correct, or incorrect prerequisites to diagnose where failures arise; contribution metrics (N-Score, S-Score), adapted from probabilities of causation, quantify each prerequisite's necessity and sufficiency to determine why. We instantiate the framework in CADET, a diagnostic benchmark of 10 composite tasks decomposed into 46 unit tasks with over 33,000 human-annotated questions spanning perception, spatial, temporal, and cognitive categories. Diagnosing frontier MLLMs with our framework uncovers systematic patterns that end-to-end accuracy obscures. Capability-wise, supplying correct prerequisites eliminates 54\% of errors on cognitive tasks, lifting them from weakest to above spatial and temporal. Prerequisite-wise, causal contributions are concentrated in a few critical prerequisites, and supplying the single most important one alone captures 84\% of the gain from supplying all prerequisites.

As preferences, goals, and facts change, LLM agents must use the current state while earlier versions remain in context. Yet they can answer with an old value of the same variable, a failure that we call stale binding. To study when models use outdated information and why, we introduce Controlled In-Context Memory (CICM), a benchmark for tracking and using updated information in conversations and agent logs. We observe that even frontier reasoning models can fail to recover the current state. We find that in open-source models probes can still recover the updated value when the model answers with an old one, pointing to a failure to select information that remains available. Component tests in Qwen and Pythia identify a mechanism for this selection failure: attention drift, where attention favors old values over the current one when producing an answer. We study a one-layer transformer to mathematically understand how this phenomenon happens: when attention scores are similar, several old values can together receive more attention than the current value. Guided by this explanation, we redirect attention toward the current value without further training. When the current value is requested directly, adjusting this intervention for each input corrects most old-value errors across various model families while preserving nearly all initially correct answers. Reliable context management therefore requires more than remembering updated information: models must use it to guide their answers.

Large language model (LLM) computer-use agents are typically evaluated with clean written instructions, despite speech being an increasingly popular interface for interacting with such systems. Speech input introduces an additional failure point: transcription errors can alter task-critical entities, constraints, or targets before the agent begins reasoning, while conventional ASR metrics do not directly measure whether the information required for successful execution has been preserved. We introduce Talk2Agent, a benchmark for evaluating how effectively voice interfaces convey human-spoken instructions to LLM-based computer-use agents. Talk2Agent builds human-spoken versions of tasks from WildClawBench and OSWorld and evaluates a range of voice interfaces, including dedicated ASR models, audio-capable LLMs, contextual biasing, and LLM-based ontology repair. Because repeatedly executing long-horizon computer-use tasks is costly and stochastic, we further propose an execution-free, task-conditioned evaluation framework that projects the original task grader onto prompt-addressable intentions and measures how much task-relevant information is retained after the voice interface. On WildClawBench, Talk2Agent's execution-free native projection provides a practical, execution-grounded measure of voice-interface quality, correlating with downstream task completion and improving Pearson correlation by 0.246 over WER/CER on 32 hours of real human speech.

Optimization modeling formulates real-world decision problems as mathematical programs that solvers can use to find optimal decisions. Large language models (LLMs) can automate this process, but the resulting correct formulations can require substantial time and memory to construct and solve, limiting practical scalability. Therefore, we systematically investigate whether LLMs can identify problem structure from natural-language descriptions and apply suitable optimization modeling techniques to generate mathematical models and solver code that solve the problems correctly and efficiently. To this end, we first curate OptTips, a knowledge base of 50 expert modeling techniques in eight families. Using this knowledge, we develop OptDachshund, a multi-agent framework that transforms problems from existing optimization benchmarks into new tasks for evaluating LLMs' use of modeling techniques. It constructs conventional and expert mathematical models with solver code for the same task and data, providing baselines for correctness and computational cost. The resulting EfficientOpt benchmark contains 561 expert-reviewed tasks with paired reference implementations. Evaluation of 11 representative LLMs reveals an efficiency gap on correctly solved tasks with comparable measurements: for every LLM, most generated programs take longer to solve than their expert counterparts. Within the comparable reference-size subset, 57\% of programs with correct objective values and fewer variables and linear constraints have longer recorded solver times. Case studies show that different modeling techniques can achieve the same optimal value at similar recorded cost. Faster solving may not reduce execution time if the code takes longer to prepare data and build the model. LLM optimization modeling should therefore be evaluated for both correctness and computational efficiency.

Recent work on hallucination detection in large language models has shown that, for a fixed pre-trained model and reasoning task, it is possible to estimate the model's confidence in the correctness of its outputs. Such uncertainty estimates have primarily been used to improve truthfulness by detecting or filtering confabulations. In this work, we ask whether these signals can instead be used more proactively to directly improve the accuracy of model-generated answers. We propose USteer, a simple, training-free steering mechanism that adjusts a model's layer-wise activations during inference using the gradient of a confidence measure with respect to the activations. This procedure nudges generation toward outputs with lower uncertainty at inference time, without modifying model parameters or requiring additional supervision. We show that this approach consistently reduces hallucination across a range of tasks, demonstrating that confidence signals can be leveraged not only for detection, but also for effective inference-time control of model behavior.

Large language models are reshaping ecommerce from static recommenders into interactive shopping assistants, yet real-world shopping requires session-level decision support: users reveal and revise constraints, coordinate multiple goals, and expect product-grounded recommendations over a full conversation. Existing benchmarks are mostly outcome-oriented or execution-oriented, leaving this evolving decision process under-evaluated. We introduce REALWORLDSHOP, a benchmark built on 3.28M grounded products, structured shopping episodes, a profile-grounded and actioncontrolled user simulator, and role-play evaluation. Our analysis shows that current systems produce locally plausible responses but struggle with state tracking, constraint updating, and grounded convergence, especially under ambiguous intent, bundle, and multi-intent scenarios. We further propose REALSHOP_AGENT, an executable session-control framework with explicit state management, shopping-flow control, catalog-grounded retrieval, and runtime guards. Experiments show that REALSHOP_AGENT consistently outperforms strong baselines on REALWORLDSHOP.

AI reviewers can assign different judgments to manuscripts that report the same science in different wording, potentially rewarding rhetorical optimization over scientific improvement. We formulate Rhetorical Robustness as the joint requirement of stability across content-preserving rewrites and discrimination across papers. We introduce RobustReview, a controlled full-manuscript benchmark with 1,260 manuscript versions, and evaluate 30 reviewer configurations. The benchmark reveals false robustness, where low rewrite sensitivity coincides with score collapse across papers, and shows that human alignment and rhetorical robustness rank reviewers differently. Moreover, the evaluated content-focused prompting protocol does not consistently improve robustness across backbones. Motivated by these findings, we introduce SciCore, a dual-branch reviewer that averages a full-manuscript judgment with a judgment based on an extracted, structured science core. This design combines manuscript-level assessment with a content-normalized view intended to reduce rhetorical sensitivity. In our primary GPT-5.5 comparison, SciCore achieves a leading joint stability-discrimination profile among the benchmarked reviewers while maintaining competitive human alignment. These results identify rhetorical robustness as a distinct evaluation target and demonstrate the potential of science-core review to improve it.

Multi-agent debate (MAD) can improve large language model (LLM) reasoning by allowing multiple agents to exchange and critique their answers to the same task. However, the interactions that enable agents to correct mistakes can also spread adversarial errors and steer the agents toward an incorrect answer. Although some efforts have been made to examine particular attack types on MAD, systematic evaluation of MAD under diverse attacks remains limited. A central question is whether debate mitigates adversarial influence or amplifies it. In this paper, we present MADBench, a benchmark for evaluating the security of MAD. We organize attacks into a layered taxonomy following the MAD workflow, incorporating both established attacks and new strategies tailored to debate. We evaluate six attack families over 356 source tasks and 3,958 test cases, examining their effects on the final answer and the propagation of adversarial influence. Our results show that, under attacks, MAD does not necessarily improve LLM reasoning. Compared with a single-agent baseline, MAD can mitigate attacks on answer accuracy in question-answering tasks while amplifying unauthorized reads or writes in both question-answering and workspace tasks. Moreover, even when three out of five agents collude, the attack changes the final answer from correct to wrong on only 28.30\% of tasks answered correctly without attack, while only 3.26\% of initially correct honest agents switch to wrong answers during debate.

Trustworthy Legal AI requires systems that can answer legal questions while grounding their responses in authoritative sources. However, existing Vietnamese legal benchmarks provide limited coverage of real-world legal consultations. We introduce \textbf{ViLegalExpert}, a large-scale benchmark constructed from authentic citizen--lawyer consultations, containing over \textbf{172K} questions across \textbf{34 legal domains}, together with professional answers and expert-verified legal evidence. ViLegalExpert supports legal information retrieval, extractive QA, and abstractive QA. Experiments with representative retrieval methods and language models reveal substantial challenges in evidence retrieval and grounded answer generation. While pretrained models perform strongly on QA, hybrid retrieval achieves the best retrieval performance. These results demonstrate the difficulty of mapping naturally expressed legal questions to authoritative provisions and establish ViLegalExpert as a challenging benchmark for reliable Vietnamese Legal AI.

In hospital workflows, electronic health records (EHRs) are often noisy, and may not contain the evidence needed to confirm events or measurements referenced in a clinical query. Even when database retrieval succeeds, clinical agents can overlook such discrepancies and return plausible but unsupported answers. We introduce EHR-RobustGym, a scalable and interactive environment for evaluating and training robust clinical agents grounded in noisy EHRs. Built on MIMIC-IV hospital records (365K patients, 31 tables, and over 500M records), EHR-RobustGym comprises 5,486 Clean-Noise pairs spanning six clinical intents and both patient-level and population-level queries. The pairs test robustness to Record-level, Value-level, and Query-level noise, while interactive SQL/Python execution and outcome verification support trajectory collection and training. Evaluating multiple LLMs reveals substantial robustness gaps: average task success across proprietary and large-scale open-weight models drops from 62.2% on Clean questions to 37.9% on Noise questions. At k=4, pass^k consistency falls below 50% for most evaluated models, exposing instability in clinical task completion. Supervised fine-tuning and reinforcement learning in EHR-RobustGym improve performance, with gains generalizing to five external EHR benchmarks. Together, these results position EHR-RobustGym as a testbed for evaluating and improving the evidence-grounded robustness of clinical agents.

Large language models are strong general-purpose code generators, but executable algorithmic trading remains a demanding specialization target: a model must translate a natural-language strategy specification into correct program logic for a specialized trading framework, execute on historical data, produce trades, and remain semantically faithful to the request. We study two complementary mechanisms for specializing language models for this setting: continued pretraining on algorithmic-trading framework code and supervised fine-tuning (SFT) on agent-validated request-to-code pairs. Evaluation is centered on QuantCode-Bench, our 400-task benchmark for Backtrader strategy generation, together with a repository-level SWE-bench-like track. Continued pretraining improves single-turn Judge Pass from 41.5% to 47.5% for Qwen3.5-397B-A17B and from 27.8% to 33.0% for Qwen3.6-35B-A3B. SFT applied after continued pretraining yields a larger gain for Qwen3.6-35B-A3B, reaching 58.2% Judge Pass and 83.5% successful backtests; in agentic evaluation it raises first-turn success from 22.3% to 58.3% and final success after up to 10 turns from 47.5% to 79.5%. Continued pretraining alone improves first-turn agentic success but lowers final success after repair from 47.5% to 32.5%, consistent with degraded instruction following, whereas SFT improves both. We also identify a capability-retention failure: domain specialization degrades parser-conformant structured tool calling, and targeted recovery SFT restores tool-call formatting but not the base checkpoint's repository-level agent performance. The results show that framework-oriented pretraining, validated SFT, and explicit capability-retention evaluation address distinct failure modes in domain-specific executable code generation.

Delegating complete application development to coding agents requires preserving the intended design rather than simply producing plausible outputs through naive prompting. Game development provides a demanding testbed, as long-form Game Design Documents (GDDs) describe requirements that must work together across game logic, visual rendering, and player interactions. However, existing game-development benchmarks typically use compact specifications and provide limited support for evaluating interdependent requirements across these aspects in long-form GDDs. We introduce A2Z GameSpec-Bench, a benchmark of 100 long-form GDDs for evaluating end-to-end game development by agents. We measure faithfulness by checking whether the game satisfies the GDD requirements and preserves the relationships among them. Each GDD is turned into a dependency-aware contract that contains rules, constraints, and prerequisite relations. Following game-development practices, we combine source-code inspection with agent-generated test policies for scenario-based replay and adaptive playtesting. The contract remains fixed across agents and revision rounds, while judgments and evidence linked to the same requirements support consistent comparison and failure detection. Our evaluations show that current agents struggle to jointly satisfy interdependent requirements across code implementation and actual play. Requirement-specific feedback improves GDD Fidelity by 10.9% relative to self-revision after two rounds. A2Z GameSpec-Bench assesses end-to-end specification-following ability beyond implementation judgments and provides targeted feedback to support more faithful game development. Code and datasets are available at https://a2z-gamespec-bench.github.io.

Agent harnesses often improve language models with human-designed workflows, but as models grow more capable, unreliable guidance can increasingly constrain their execution. We call the ability to benefit from useful guidance while overriding unreliable guidance thinking outside the box. We introduce Box$^2$-Bench, which holds the model and task fixed while varying workflow reliability to isolate how models regulate their reliance on guidance. On Box$^2$-Bench, frontier models often benefit from reliable guidance but remain vulnerable when it is misleading or becomes unreliable. To test whether this capability can be learned, we train two open-weight models using bad workflows, reserving good workflows for evaluation. We explore two complementary training strategies: counterfactual supervised fine-tuning improves robustness, while outcome-based reinforcement learning can shift the balance toward greater use of helpful workflows. We further find that this behavior extends beyond workflows to other forms of external information, improving peer correction and robustness to corrupted memory. Together, our results identify selective reliance on fallible external information as a dimension of agent reliability not captured by task performance alone.

Top researchers have good intuition, but do language models have as good intuition about model training as top AI researchers? To measure model intuition of LLMs and humans, we introduce the ArchitectureIQ benchmark. Each question presents a synthetic dataset and several training recipes, and the test-taker is asked to predict the recipe yielding the best test metric. Overall, we find that LLMs' model intuition is good but has four limitations: (1) The intuition is imperfect, or even sub-human in some cases. Frontier models achieve around 76% accuracy (random choice 33%) vs best human researcher (66.0%), yet remain far from perfect. For architecture-only questions, best human achieves 65% while GPT-6 Astra only has 38%. (2) The intuition is empirical, not structured, supported by the fact that more CoT compute does not lead to substantial improvement. Unlike math, we still lack a "Science of AI" language that enables structured reasoning on AI. (3) The intuition is not maximally condensed, and can be further compressed into a knoledge base. Our constructed knowledge base with only 20 items yields large gains for weak models: GPT-4o equipped with the accumulated knowledge almost matches the performance of Claude Opus 5. (4) The intuition is insensitive to dataset properties, but the best model should in general depend on data properties. This suggests that data is the real "dark matter" in AI -- LLMs (so do human researchers) understand too little about data, even less than model architectures.

We investigate the problem of role-capability leakage (RCL), in which a role-prompted reasoning model generates convincing in-role text while continuing to exhibit capabilities on benchmarks that exceed those implied by the assigned role. For example, when a model is prompted to assume the role of a kindergarten student, one might expect its performance on a mathematics benchmark to reflect kindergarten-level ability rather than expert-level proficiency in solving calculus problems. We introduce RoleCapBench, a curriculum-grounded benchmark for evaluating RCL across six educational roles and four assessment levels spanning elementary school through A-level, and use it to evaluate three open-weight reasoning models. We find that although the models can generate stylistically convincing in-role responses, they consistently fail to align their underlying capabilities with their assigned roles. Naive role prompting yields strong role-voice scores of 1.218--1.389 while retaining above-role accuracy of 0.811--0.898. RCL persists across a range of prompting conditions, including prompts that explicitly instruct the model to match the role's capability level. To mitigate this problem, we propose Injection, an inference-time intervention that combines explicit, role-specific capability guidelines with a guiding prefilled response prefix. Injection improves role-capability alignment across models, reducing above-role accuracy by up to 0.562 while preserving in-role accuracy with a marginal drop of less than 0.058 across most models. All artifacts, including scripts and evaluation data, will be released upon acceptance.

Pre-training equips large language models (LLMs) with a broad repertoire of behavioral patterns associated with roles, styles, values, and goals. Post-training teaches conditional enactment and makes a helpful Assistant the default, but it does not erase alternative modes from the weights; explicit prompts can therefore elicit personas that repeatedly shape judgment, language, and action. In open-weight settings, runtime controls can be removed, motivating persona unlearning: a weight-level edit that makes a designated persona difficult to elicit and enact on unseen contexts. We introduce PersonaUnlearnBench, a model-specific paired benchmark spanning six LLMs from three families and five personas, with aligned forget/retain sets, held-out instruction paraphrases, and four-axis evaluation. The benchmark shows that standard unlearning methods cannot reliably erase the target persona without sacrificing meaningful generation or general utility. We therefore propose PaCE, which compares target and desirable responses to the same questions to locate an internal behavior direction, then trains target-prompt states away from the target mode and toward the matched desirable response. Experiments show that PaCE consistently suppresses target personas with high response quality and useful counterpart behavior, at moderate utility cost. These results establish persona unlearning as a distinct behavior-level editing problem and a practical route toward persistent control of latent LLM response policies.

Scientific software presents a demanding test for computer-using agents based on visual language models (VLMs): completing a research workflow requires interpreting specialized interfaces, manipulating scientific objects, and producing verifiable results. We thus introduce OSWorld-Science, a benchmark and evaluation environment that combines scientifically meaningful tasks, artifact-based evaluation, and an efficient agent harness for studying computer use in the scientific domain. The benchmark contains 12 VLMs and 146 high-quality tasks across several scientific domains and software configurations, covering workflows such as molecular drawing and retrosynthesis, pathology image analysis, statistical computing, and physical simulation. Tasks are developed through expert proposals and iterative human--AI co-design, with selection guided by scientific value and difficulty. Task-specific execution-based evaluators inspect application states and generated artifacts, including molecular structures, segmentation masks, plots, and numerical results, and award partial credit for incomplete outcomes. Our special harness integrates model adapters, interaction-loop control, and trajectory logging to support comparisons of models and interaction strategies. Our results show that current state-of-the-art VLMs with a strong harness still face challenges in addressing key questions in the scientific domains. We also analyze the benchmarking results across multi-linguistics, reasoning efforts, context length and other factors and derive several important conclusions and directions to assist future development. Overall, we provide an integrated framework connecting expert-defined scientific goals to verifiable software outcomes, enabling systematic evaluation of both agent capabilities and harness design in scientific workflows.

Existing auto-research benchmarks often entangle multiple sources of improvement, including training frameworks, hyperparameters, compute budgets, and data, making it difficult to attribute why one frontier agent outperforms another to specific research capabilities. In this work, we isolate and systematically evaluate Data Intelligence: an agent's ability to understand, manipulate, and improve the data that shapes model capabilities. We introduce AutoDataBench, a controlled testbed built on a conceptual framework of data intelligence spanning data diagnosis, data organization, and data construction, instantiated through three highly curated optimization tasks while holding non-data factors fixed. Across tool use, retrieval, and knowledge injection, we evaluate frontier LLMs' ability to improve training data through iterative experimentation under task-specific resource budgets. Beyond optimization performance, we ask: do LLMs understand what their data interventions do? We compare predictions made before training with observed outcomes to seek evidence of data-effect reasoning beyond trial and error, and explore whether iterative feedback helps LLMs better understand how changes to training data affect model performance. Finally, we show that reusing AutoDataBench trajectories for mid-training improves downstream coding performance, highlighting its value in both evaluating data intelligence and generating high-quality training data. Code and resources are available at https://github.com/AutoDataBench/AutoDataBench.

Unlearning a fact in one language does not guarantee its removal in others as changing the query or even the requested answer language can reopen seemingly forgotten knowledge -- a cross-lingual loophole. The most straightforward solution to this challenge -- unlearning in all languages -- is neither scalable nor desirable as it amplifies damage to unrelated model capabilities. We introduce the task of language budgeted multilingual unlearning where the goal is to select a subset of languages that maximizes cross-lingual erasure. To study this task we introduce the Cross-Lingual Unlearning Tensor, an unlearning benchmark that spans 174 language--script pairs and 25 atomic paraphrase types to examine when forgetting generalizes across linguistic expressions of the same knowledge. We further propose COVER, which selects source languages to maximize predicted COVERage of languages receiving no forget supervision, enabling unlearning on a language budget. Surprisingly, we find naively selecting strong individual sources does not reliably compose into strong source sets motivating our development of COVER. At deployment COVER only requires benign calibration data and access to the frozen model. Across three model families and two disjoint forget sets, COVER reduces mean held-out residual access by 7.8--27.3% relative to uniform source selection. We find these gains extend beyond synthetic benchmarks to real news documents in low-resource language settings using human translated data from the Low Resource Languages for Emergent Incidents (LORELEI) corpus.

Long-running tools can dominate coding-agent latency: compilers, test suites, and repository commands take seconds to minutes while the agent idles. This observation stall presents the same tension that drove out-of-order processors -- asequential interface hides work that can be predicted and started early, but a speculative result may become visible only after it and every earlier step have been validated. We present TomasuLLM, a runtime that executes agent tool calls out of trajectory order while preserving task-execution correctness. It drafts future actions, runs them in isolated copy-on-write sandboxes, traces their dependencies and effects, and commits results in trajectory order only after validation against committed state. Across three benchmarks spanning sub-second to minutes-long tool calls, TomasuLLM improves the reported benchmark means and scales with tool latency: 1.31x on 100 SWE-bench Verified tasks, 1.35x on 28 Terminal-Bench 2.0 tasks, and 1.27x matched progress on 18 SWE-Marathon sessions. Across 4,010 audited commit-validation records, it produces zero false accepts.

Large language models (LLMs) are increasingly used to answer questions about politically contentious issues, yet evaluations typically treat a model's stance as a relatively stable property. Real users, however, communicate political signals through their terminology, assumptions, and personal context. We investigate whether such signals produce ideological mimicry: systematic shifts in the political stance expressed by an LLM toward the position conveyed by the interaction. If LLMs adapt their responses to these signals, they risk creating personalised political information environments in which users with opposing views receive systematically different accounts of the same issue, potentially reinforcing existing divisions. We build the Poli-SHIFT dataset and evaluation framework and assess seven open-weight LLMs across ten contentious political topics in the United States, United Kingdom, and Australia, systematically manipulating contested terminology, politically valenced premises, and user information, and eliciting responses in both multiple-choice and open-text formats. Across models, we find robust evidence that prompt framing shapes the political stance of LLM outputs. Changing terminology alone reverses which side of an issue a model supports in 16.9% of matched comparisons. Stated political ideology also systematically shifts responses toward the user's position. These findings show that political stance is not a fixed property of LLMs; the views expressed are conditional on the interaction with the user. As LLMs become increasingly personalised sources of information, such interaction-dependent adaptation could contribute to political information environments that reinforce users' existing perspectives.

We present HalluScoring 2026, a shared task for evaluating hallucination detection and factual verification in Arabic question answering under challenging generalization settings. The shared task is organized into two main tasks, each comprising two subtasks, for a total of four subtasks. Task 1 evaluates binary hallucination detection, considering generalization to unseen questions (Subtask 1.1) and responses generated by unseen LLMs (Subtask 1.2). Task 2 extends the evaluation beyond detection by requiring the systems to additionally identify the correct factual answer from six related candidates, covering Islamic knowledge (Subtask 2.1) and general knowledge (Subtask 2.2). The shared task is based on two Arabic datasets: HalluScore and HalluTruthQA. A total of 13 teams participated in the shared task, 10 of which submitted system description papers. The results of Task 1 demonstrate that hallucination detection remains challenging under distribution shift, with the winning team achieving AUC-ROC test scores of 0.772 and 0.767 for Subtasks 1.1 and 1.2, respectively. For Task 2, the winning team achieved scores of 0.882 and 0.857 in the Islamic and general-knowledge subtasks, respectively, under assisted evaluation.

Chart editing requires cross-modal edit grounding, realizing a requested visual change in the code that draws it, with necessary related updates and without altering unrelated content. Existing benchmarks emphasize either code executability or chart quality, but their metrics do not clearly distinguish request completion from missed coupled updates and gratuitous changes. We introduce ChartRevise, a structured dataset and evaluation protocol for exact program-grounded chart editing. For dataset construction, we build on the grammar of graphics to systematically cover chart-editing operations, using source-program checks to verify their applicability across chart types and libraries. To improve edit exactness, our pipeline checks individual requirements and guides repair or exclusion when they are unmet. The resulting dataset contains 92,438 records covering 344 edit types across 20 chart types and three plotting libraries. For evaluation, our reference-free protocol separately measures atomic requirement completion, identifies gratuitous changes, and detects missed coupled updates. These checks are combined with successful execution and rendering to determine exact-edit success. Across five models and four external benchmarks, fine-tuning yields relative gains of 16\% in mean requirement recall and 22\% in mean exact-edit rate.

Long-form subtitle translation requires reasoning over discourse and cultural context spanning episodes or entire series, while maintaining consistent terminology and style. Existing single-LLM methods are largely sentence-level, and multi-agent systems often use static workflows that do not adapt to scene complexity or production context. We propose SMART, a Self-evolving Multi-Agent system for long-foRm subtitle Translation. During test-time training, SMART builds persistent series-level memory and translates a subset of sentences through a dynamic router and Mixture-of-Agents layer with tools for terminology verification, subtitle constraint validation, and contextual retrieval. A judge-refiner loop scores candidates and uses textual critiques to update agent prompts and routing policies without retraining the underlying LLMs. During test-time inference, the evolved configuration translates the remaining series. We also introduce Subtitle Arena, covering 14 genres, 2--198 episodes per series, production years 1959--2023, and 15 target locales, together with SubMQM, a subtitle-adapted MQM framework with seven dimensions and 19 error categories. SMART achieves the best overall MQM score in all 15 Subtitle Arena directions, reducing average penalty by 6.9% over the strongest competing agent system. On the public MuSC benchmark, SMART obtains the best model result across all four language pairs and also achieves the best human-evaluation result, with an overall score of 4.50/5.

Existing full-duplex speech benchmarks cover only subsets of real-time interaction behaviors, often under limited contextual conditions. We introduce DuplexAct-Bench, a bilingual benchmark that systematically covers six complementary behaviors, from interruption and yielding to proactive initiation, active silence, and backchanneling, across Pre-session, In-session, and No-explicit conditions. Across 1,290 English and Chinese streaming trials, we evaluate 12 full-duplex speech systems on both Timing and Content. Results reveal substantial variation across behaviors, conditions, and systems, as well as frequent mismatches between semantic quality and behavioral timing. These findings show that current systems remain far from robustly managing when, whether, and how to participate as real-time interaction unfolds. Project page: https://alitaxky.icu/DuplexAct-Bench/

Autonomous agents increasingly rely on memory to generalize beyond their training environments. However, agents are bounded by what they have seen and believed, and leveraging such memories in unseen environments can introduce biases into their internal beliefs. We formalize this phenomenon as \textit{false memory}, which can arise from spurious correlations, environment shifts, and knowledge conflicts. Despite its importance, false memory is difficult to evaluate because it stems from agent internal beliefs and is easily confounded with ordinary generalization failures. Therefore, we propose FAME, a training-free framework that evaluates false memory through the evolution of agent beliefs under counterfactual reasoning. Specifically, counterfactual scenarios reveal how beliefs change as the latent concept of memory shifts under hypothetical interventions; thus, measuring the resulting concept drift provides a signal for distinguishing faithful versus false memory. Such concepts can be estimated from agent hidden states before answer generation, avoiding the need for reward design or answer sampling. Empirical experiments reveal that simply monitoring answers often fails to detect false memory, while FAME achieves AUROCs of 76.2% - 96.7% across false-memory settings, and outperforms the best baseline by 3.4% - 23.3% across realistic benchmarks, spanning math reasoning (GSM-Symbolic), code generation (GitChameleon), and complex reasoning (BigBench-Hard). We further release corresponding counterfactual templates and facilitate future research on false memory.

Ownership establishes rights over the use, control, and transfer of objects. Understanding these relations is essential for AI systems to interact appropriately with people and their resources. Yet how large language models (LLMs) attribute ownership under competing claims remains unclear. We introduce the Competing Ownership Attribution Task (COAT), comprising 42 scenarios, and compare ownership allocations from 24 LLM configurations with those of 108 human participants. Overall, human-model similarity is close to human-human similarity, but models show greater homogeneity in their ownership judgments. Within individual answers, models also divide ownership more evenly among claimants than humans do. Pooling responses across model configurations reveals more scenarios with a shared judgment and fewer with distinct viewpoint groups than in humans. When humans form distinct groups, models may converge on one viewpoint or between competing viewpoints. Further comparisons reveal different contextual sensitivities. As material value increases across scenarios, allocations to creators decline less sharply in models than in humans. Across scenarios differing in public recognition of later holders as owners, allocations to these holders increase in models but decrease slightly in humans. Together, these findings suggest that the evaluated LLM responses do not fully capture the diversity of participants' ownership judgments or how those judgments vary across situations. Developing socially capable AI therefore requires moving beyond overall similarity to capture the diversity and context dependence of human judgments.

Large language models (LLMs) are increasingly deployed as automated judges for AI-generated content, yet a single judge is unreliable and even a panel of judges leaves a hard residue: when judges disagree, majority voting discards the conflict instead of resolving it. We present JuryFlow, a disagreement-guided, human-in-the-loop multi-agent evaluation framework that treats inter-judge disagreement not as noise to be averaged away, but as a precise, claim-level signal indicating where an evaluation is uncertain. JuryFlow decomposes each candidate response into atomic claims, has a panel of heterogeneous judges assign per-claim verdicts, and builds a disagreement graph whose nodes are scored by verdict entropy and whose edges encode structural similarity between claims. A human acts as a structural guide, selecting which disagreement to resolve through a single, minimal intervention rather than re-labeling the response, after which the focal claim is re-evaluated, the correction propagates along graph edges and to historically similar cases, and is crystallized into reusable rubric entries that all judges inherit, making the evaluator progressively self-refining. To enable large-scale, reproducible benchmarking without human studies, we evaluate JuryFlow in an automatic configuration in which focal selection is made by entropy ranking. On MT-Bench and LLMBar, JuryFlow improves agreement with gold labels over single-judge and majority-vote panel baselines, and ablations isolate the contributions of disagreement-targeted re-evaluation, propagation, and rubric induction. We contribute (1) a human-in-the-loop paradigm that recasts the human from labeler to structural guide, (2) the JuryFlow framework operationalizing it through a disagreement graph, focal re-evaluation, and closed-loop rubric induction, and (3) an evaluation protocol with ablations that isolate where the gains originate.

Survival analysis estimates time-to-event outcomes from patient covariates and is widely used for medical risk assessment. Patients seeking prognostic information after a diagnosis may turn to large language models (LLMs), now readily accessible through consumer applications. However, whether LLMs can provide accurate survival predictions has not been rigorously evaluated. We introduce Survprompt, a framework that converts structured patient covariates into free-text clinical vignettes and prompts pre-trained LLMs to predict survival zero-shot. We benchmark Survprompt against conventional survival models, including random survival forests (RSF), across two multi-institutional pan-cancer cohorts: the publicly available MSK-CHORD cohort and a newly curated cohort from the Providence St. Joseph Health Network constructed using an LLM-based medical abstraction framework. We report censored mean absolute error (cMAE) and concordance index (c-index) and conduct feature ablations to identify variables influencing LLM predictions. Frontier LLMs achieved surprisingly competitive cMAE for individual survival times. For example, GPT-5.6-Sol achieved cMAE within 10% of state-of-the-art RSF models specifically trained for survival prediction for several cancer types and lower cMAE than RSF for prostate cancer in MSK-CHORD. Feature ablations revealed that LLMs prioritized clinical variables similarly to specialized survival models. However, LLMs showed inconsistent accuracy across cancer types and institutions and poorly discriminated between high- and low-risk patients (lower c-index). Zero-shot LLMs can generate surprisingly accurate prognostic estimates without specialized training, but their variable performance across cancer types and institutions remains an important limitation for clinical use.

Retrieval-augmented generation (RAG) can ground large language models in external evidence, but retrieved context does not guarantee that generated claims are factually supported. This problem is especially relevant in multi-hop RAG, where retrieval and reasoning proceed through multiple dependent stages. We study whether claim-level conformal factuality control, previously developed for RAG, remains effective in this setting. We apply split-conformal claim filtering to multi-hop RAG and evaluate it on HotpotQA, Natural Questions, and TriviaQA using Llama 3.1 8B and GPT-4o-mini, together with a single-hop reference experiment. Across all six multi-hop model-dataset configurations, increasingly stringent conformal targets consistently increase the fraction of responses whose retained claims are fully supported. At the 95% target, this rate ranges from 95.80% to 97.20%, compared with 55.60%-76.03% without filtering. However, the improvement is strongly selective: only 4.41%-31.09% of generated claims are retained and 9.70%-51.40% of responses remain non-empty at the 95% target. These results show that conformal factuality extends to multi-hop RAG, while demonstrating that nominal reliability must be interpreted jointly with claim retention and abstention.

Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.

Open-ended LLM homogeneity can create false plurality when several systems appear to offer independent perspectives while returning the same familiar default. Single-pass answers obscure the distinction between agreement produced by a tightly constrained answer space, prompt-vocabulary echo, and broader answer spaces with stable alternatives beneath the surface. We introduce CHOIR (Collective Hierarchically-Ordered Inquiry Responses), a framework that adapts free-list elicitation from cognitive anthropology to LLM ensembles. CHOIR repeatedly elicits ranked lists, clusters items into prompt-level concepts, and measures concept salience across models, prompt variants, and persona conditions. We evaluate CHOIR on Infinity-Chat 100, an external prompt bank from recent work on open-ended model homogeneity, and on a 27-question targeted diagnostic bank designed to isolate mechanism-level contrasts. On Infinity-Chat 100, CHOIR reproduces high surface agreement (93/100 prompts above chance) while separating narrow prompts from broad prompts with recoverable depth. Across targeted probes and the external prompt bank, base-model identity remains the strongest recoverable signature, and persona prompts shift surfaced concepts within base-model signatures. A source-blind ranking module prioritises rare-but-stable candidates for later inspection. CHOIR turns open-ended homogeneity into a diagnostic measurement problem by asking where models converge, why they converge, and what remains reachable under structured depth probing.

Speculative decoding accelerates autoregressive LLMs by having a lightweight drafter propose tokens that the target model verifies in parallel. Diffusion-based drafters further reduce drafting latency by proposing multiple tokens at once. However, these tokens are predicted independently, so a single early error causes prefix verification to discard the rest of the draft, even when it contains useful downstream predictions. We introduce DEdit, a diffusion-based drafter that can not only draft by conventional parallel unmasking but also iteratively edit its draft through token-to-token predictions. Through editing, later predictions can serve as bidirectional context for repairing earlier errors and extending the accepted prefix. To teach the model to repair errors while preserving correct predictions, we propose ProposalMix, a training scheme that mixes draft predictions with ground-truth tokens based on first-pass confidence during training. Across seven benchmarks on Qwen3-4B and Qwen3-8B, DEdit achieves the highest macro-average token acceptance and speedup among the evaluated drafters, reaching macro-average speedups of $5.72\times$ and $5.97\times$ over autoregressive generation under greedy decoding, respectively. Further analysis shows that acceptance improves with more editing passes and wider drafting windows, and that ProposalMix halves harmful edits that shorten the accepted prefix. Moreover, restricting the editor to causal attention lowers acceptance, especially on highly predictable outputs, indicating that future context is a key source of these gains.

Chain-of-thought (CoT) traces often serve as a proxy for how Large Language Models (LLMs) arrive at their answers. However, growing evidence shows that models' CoT often fails to reflect their internal computations and can be changed without affecting their final answers. In this work, we measure and improve the alignment between the reasoning described in an LLM's CoT and what it computes internally. We propose CoT-Interpretability Alignment (CIA), a metric that measures the agreement between a model's CoT traces and its internal reasoning strategies as detected by interpretability tools. We evaluate CIA on three tasks (two-hop question answering, hint intervention, and integer multiplication) across three LLMs, finding that LLMs exhibit limited alignment across all tasks (44.8-75.9%). We then experiment with improving CIA via post-training, setting both the task accuracy and parametric faithfulness signals as a reward. Experiments show that we can substantially improve CoT parametric faithfulness while maintaining or improving the task accuracy. We provide rich analysis, such as their generalization patterns. Our work provides both a framework for auditing CoT parametric faithfulness and a pathway toward making models' explicit reasoning more trustworthy. Code and data are available at https://github.com/yihuaihong/CIA-minimal-repro.

Long-context LLM agents accumulate interaction histories that strain KV-cache memory and attention computation. Although sparse attention reduces these costs, heterogeneous cache representations and workflows hinder integration with existing inference engines, while prior sparse-serving abstractions support only specific layouts or workflows. We present SparseEngine, a ground-up, sparse-first inference engine whose shared lifecycle contract lets each method control its KV representation and computation while coordinating state transitions with common serving infrastructure. SparseEngine supports 15 methods across four categories and enables cross-request state management through Chain Cache, which resumes KV-eviction methods from retained history, and controllable Prefix-Cache Pruning, which removes KV from selected history regions while preserving logical-prefix matching. While maintaining method quality, SparseEngine delivers over 10x higher throughput with KV eviction, over 2.5x faster decoding at matched concurrency than vLLM, and over 2x end-to-end speedup on agent benchmarks. The code is available at https://github.com/CURRENTF/SparseEngine.

An agent that moves must recognise a place from a viewpoint it has never seen. We introduce 4MT-VLM, a dataset of procedurally generated landscapes, each rendered across five stimulus modes that remove appearance cues while holding layout fixed: shape and colour, shape only, colour only, bare terrain peaks with no objects, and a valley viewpoint that puts the peaks on the horizon. The last condition is commonly used in clinics to probe hippocampal function in human patients. We test this benchmark across sixteen different open and closed-source models and report 4AFC performance, a measure which is also used to grade human participants. We observe that models identify a place from the studied viewpoint but lose it once the camera moves, dropping below the 25% chance level at 135{\deg} where a human observer scores 85%. Frontier models (Gemini 3.8 Flash, GPT-5.6) answer only 39% and 31% of rotated trials correctly, recovering to 85% and 55% only when distractors are moved more than 30 meters apart. Our benchmark demonstrates that while current VLMs possess rudimentary cognitive maps, their spatial resolution remains fundamentally too coarse to maintain a stable, 3D understanding of the world once the viewpoint changes.

A transformer language model performs a bounded amount of computation per token, and recent work by Vishal Sikka, former CEO of Infosys, argues that this bound limits which tasks a model can carry out or verify (arXiv:2507.07505). We ask how much of the budget beneath that ceiling is spent on work the model has already done. Serving is stateless across requests: a model that answers a second question about a document recomputes the document's attention state from the first token. On seven real-world datasets, 98.7% of prompt tokens were text the model had already read. We present Galahad, a memory layer for vLLM, SGLang and llama.cpp that makes this reading a one-time cost. Taliesin saves the model's key-value (KV) state for a block of text and loads it on the next request that contains the same bytes, instead of recomputing it. Blaise keeps the documents themselves and passes the model only the section a question needs. On a recall test with 100 facts hidden in a 97,000-token corpus (Gemma 4 31B), Taliesin alone let the model attend to the whole corpus and answered 98 of 100 on llama.cpp at 3.0 s and 572 J per question, against 10 of 100, 9.3 s and 2,754 J for the same model without Galahad, which could hold only the last 12,000 tokens. With Blaise added, the model read about 668 tokens per question and answered 100 of 100 on all three runtimes at 0.59-0.64 s and 200-213 J; a tuned RAGFlow pipeline answered 77. Storing the corpus is a one-time cost of about 100 s and 28 kJ, whose energy is recovered after 13 questions. Restored state is bit-identical: all 262,144 output logits matched after restart, rehydration and hot-load. Galahad worked with all 30 models we tested under vLLM, and it fails closed: any load that does not pass its checks is recomputed. Together these results move LLM serving from stateless to stateful inference.

During reinforcement learning with verifiable rewards (RLVR), large language models (LLMs) can exploit loopholes in their environments to obtain high rewards without improving the intended capabilities, i.e., reward hacking. Despite its risks to training efficiency and safety, monitoring and mitigating reward hacking during training remain challenging, which is limited by a lack of testbeds that reproduce hacking and reliably identify it. We introduce CATCH, a controllable testbed for studying reward hacking in coding RL. CATCH deliberately exposes environmental loopholes and provides execution-based gold labels by comparing success under a vulnerable evaluator with task correctness under an independent audit. It also can control the model's initial hacking tendency through supervised fine-tuning data mixtures and the difficulty of earning rewards through reward designing, enabling systematic comparisons of hacking dynamics and interventions. Experiments show that CATCH can produce diverse RL training trajectories with clear reward hacking, and analyses demonstrate that both initial models and reward difficulties shape the emergence of reward hacking. We further evaluate the effectiveness of different reward hacking detection and mitigation methods. A key finding is that a chain-of-thought monitor initially suppresses hacking, but this protection erodes as the policy model learn to mislead the monitor with code comments. This highlights the need to evaluate hacking mitigations throughout training with CATCH. The source code and resources are publicly released at https://github.com/THUAIS-Lab/CATCH.

Embodied agents must determine where to act, anticipate the resulting scene changes, and interpret observed outcomes to guide subsequent actions. This requires connecting 4D interaction understanding, which explains how past actions changed the scene, with spatially grounded planning, which determines how and where to act toward a goal and anticipates the resulting scene changes. We introduce ChronoGraph, a functional 4D scene graph that links actions on affordance parts to semantic and geometric state changes. By representing observed and anticipated transitions in the same form, it provides a shared basis for understanding and planning. We construct ChronoGraphBench through an automatic data engine that converts human-interaction videos and simulated robot trajectories into graph-annotated questions for training and evaluating Vision-Language Models (VLMs) on both tasks. Using these annotations, we train ChronoGraphVLM by adapting pretrained VLMs in two stages. Graph-as-Chain-of-Thought supervised fine-tuning teaches the models to reconstruct observed transitions and predict future ones as graph traces before answering. Subsequent joint 4D graph reinforcement learning directly rewards graph properties and answer correctness. Experiments across model scales show improvements over the corresponding pretrained baselines and zero-shot transfer to VLM4D. Real-world demonstrations further show that graph-based planning and affordance grounding support mobile manipulation through existing robot skills without additional fine-tuning.

LLM-based multi-agent systems coordinate specialized reasoning through aggregation, interaction, and adaptive control, yet their potential for graph learning remains unexplored. Graph learning is a natural setting for such systems because useful evidence may arise from heterogeneous local, long-range, global structural, and semantic perspectives whose relevance varies across instances. Existing LLM-based graph learning approaches primarily rely on single-agent reasoning, while multi-agent coordination has been studied mainly in general reasoning settings. Consequently, it remains unclear whether multiple specialized agents can improve graph learning and how coordination strategies should be designed and evaluated. To address this gap, we introduce GraphMAS, a systematic benchmark of multi-agent coordination for graph learning. GraphMAS builds a shared pool of graph reasoning specialists and organizes coordination along two dimensions, inter-agent interaction and runtime adaptivity, yielding four paradigms and seven representative coordination methods. Under a unified protocol, we evaluate these methods across seven text-attributed graphs, three domains, and two graph learning tasks. We find that heterogeneous graph perspectives are complementary, and that coordinating specialists improves over individual specialists and single-agent graph reasoning, with gains from decomposing reasoning across specialists rather than from broader evidence access alone. However, richer inter-agent interaction does not reliably help, whereas instance-adaptive specialist selection yields the strongest accuracy-efficiency trade-off. We further show that coordination can be learned over a fixed specialist pool and transfers to held-out graphs. GraphMAS therefore provides a controlled evaluation framework and empirical principles for understanding when and how multi-agent coordination benefits graph learning.

Speculative decoding accelerates language model inference by verifying multiple draft tokens in a single target-model pass. Recent parallel drafters have achieved breakthrough performance in frontier production models, but their effectiveness deteriorates as the entropy of target distributions increases due to insufficient draft diversity. To overcome this bottleneck without sacrificing parallelism, we introduce UBTree, a parallel drafter that couples a Unigram proposer with a Bigram selector to construct drafting Trees. The unigram proposer is trained with the standard cross-entropy objective to generate candidate tokens independently for each position, while a lightweight bigram selector predicts transition scores between adjacent candidate pairs. Unlike the proposer, the selector is trained with a renormalized KL objective on high-temperature data. This tree-native training broadens the supervision beyond the greedy path, encouraging plausible alternative branches that improve the chance of accepting additional tokens during tree verification. Across seven standardized benchmarks with Qwen3-4B and Qwen3-8B, UBTree achieves an average speedup of $5.84$--$6.94\times$ over autoregressive decoding and outperforms DARTree in all 28 comparisons. Production-scale evaluation further demonstrates UBTree's advantage over frontier baselines such as DSpark.

Modern transformers pair impressive capabilities with substantial memory and compute demands. Low-rank weight factorization reduces both while keeping the matrices dense, and thus efficient on standard hardware. Existing methods, however, choose the subspace to remove from each weight matrix with local closed-form criteria: activation energy, layer-wise reconstruction error, or a quadratic approximation of the loss. These criteria ignore how errors propagate through the network, so at high compression the errors compound with depth and performance collapses. We introduce Learnable Subspace Projections (LSP), which instead learns the subspaces to discard end-to-end. Each linear layer, or tied group of layers that read the same activations, is assigned an orthogonal projector. All projectors are optimized jointly against a global objective--the KL divergence to the dense model's output distribution or the model's original training loss--while the pretrained weights remain frozen. Projectors are initialized from a whitened SVD truncation, and ranks are allocated by the output KL each projector induces per parameter saved. After training, the projectors merge into standard low-rank factors, with each tied group sharing one factor. In attention, this also lets the model cache one narrow latent in place of full keys and values. Across LLMs (OPT-125M/1.3B, Qwen3-4B, Llama-2-7B) and ViT-B/16, LSP outperforms baselines, and its advantage widens as compression increases. At -70% compression, LSP brings Llama-2-7B to 10.9 WikiText-2 perplexity and 42.2% mean zero-shot accuracy, versus 13.3 and 36.0% for the strongest baseline. The factorized model decodes up to 1.6x faster than the dense model at small batch sizes, and aching the shared latent shrinks the combined memory of weights and KV cache by 13.5x at a 128k-token context, versus at most 6.5x for untied baseline factorizations.

Direct preference optimization (DPO) models binary preferences through a Bradley-Terry model with a common noise scale, without explicitly accounting for preference strength or prompt-dependent uncertainty from human feedback. We introduce uncertainty-normalized margin DPO (UNM-DPO), which combines strength-dependent margins with a learned prompt scale. Motivated by a heteroskedastic Bradley-Terry model, we develop two training objectives. Both compare the implicit rewards of preferred and rejected responses, derived from response log-probability ratios to a reference policy. Advantage-only (AO) divides this reward difference by the prompt scale before subtracting the margin; whole-residual (WR) subtracts the margin before dividing by the scale. For the WR comparison model, we establish a necessary and sufficient condition under which known margins make the prompt scale identifiable. We introduce a practical procedure for learning the scale. Building on WR, we introduce ULNM-DPO-WR, which normalizes each response's implicit reward by its length. We evaluate our methods against DPO and related baselines on HelpSteer2 and HelpSteer3, using the Skywork reward model as a judge. With Llama-3.1-8B-Instruct, ULNM-DPO-WR achieves tie-adjusted win rates against matched DPO of 68.00% and 65.31% on evaluation panels, with higher mean rewards and shorter responses on average. On AlpacaEval with a GPT-4.1 judge and GPT-4-Turbo reference answers, the same 8B policy achieves a length-controlled win rate of 21.62%, compared with 16.39% for DPO and 15.30% for SimPO. These results demonstrate the potential of combining preference-strength margins, learned prompt scales, and length normalization for policy optimization.

Final-answer accuracy does not reveal whether a scientific-search agent failed to encounter a target paper, attempt to inspect it, or return an accepted answer after inspection. We introduce decision checkpoints that record observations and tool actions without benchmark labels during inference, then join target identities and evaluator labels to assign outcome categories from recorded events. Across five conditions on 540 answerable AutoResearchBench Deep questions in a fixed, target-enriched environment, keyword search achieves 24.6\% accuracy, compared with 17.8\% for raw search. The keyword condition has fewer incorrect answers with neither target exposure nor inspection, but more incorrect answers after the target is exposed and left uninspected. Compared with keyword search, read-first has 27.4\% more recorded evidence-search calls. Target inspection attempts occur on 199 questions under read-first and 191 under keyword search; both conditions achieve 24.6\% accuracy. The checkpoint protocol makes these question-level differences explicit, distinguishing target exposure and inspection from aggregate accuracy and total tool use.

LLM agents rely on retrieval tools to access external knowledge, yet visual agentic search remains severely bottlenecked by standard single-step retrievers. In current pipelines, the agent must issue text queries for every intermediate step, struggling when visual clues are difficult to describe or when the retriever fails to surface necessary intermediate evidence within its top results. We hypothesize that offloading multi-step navigation across the entire embedding space directly to the retrieval tool resolves this performance bottleneck. To study this systematically, we introduce VHOP, a flexible data generation framework and benchmark with five core difficulty levels testing both visual matching and search planning. Using this framework, we develop VHOP-Router, an end-to-end training pipeline---combining supervised fine-tuning, online imitation learning, and reinforcement learning---that transforms a standard embedding model into an autoregressive multi-step retriever. Operating directly in the visual latent space, VHOP-Router retrieves linked image chains in a single tool call without requiring the agent to formulate intermediate text queries. Experiments show VHOP-Router boosts retrieval performance from under 5\% to 76.3\%. In agentic search, it improves task success rates by 52.7\% and reduces the average token length by 61\% from 1886 to 728, whereas upgrading the agent yields only a 3.7\% gain. Compared to a strong baseline where the agent retrieves the top 50 results per step, VHOP-Router maintains superior performance while reducing in-context images by $23\times$ and cutting the cumulative API payload by $35\times$. The models also generalize robustly to unseen difficulty levels and realistic test sets. Ultimately, VHOP and VHOP-Router provide an efficient and effective solution for visual agentic search that leaves native LLM capabilities entirely intact.

Large language models deployed as personalized assistants must reason over long, evolving interaction histories. However, in long-term dialogue reasoning, relevant evidence is scattered across sessions, preferences may be revised over time, and standard long-context training fails to address these challenges under data scarcity and prohibitive computational costs. We propose StateTree, a data-driven RL method that constructs a challenging auxiliary task from scarce dialogues with verifiable ground truth. StateTree augments multi-session dialogues with a tree-structured path-tracing task: key-value records are embedded across sessions to form a binary tree. Solving the task requires the model to traverse from root to leaf by retrieving records across sessions and comparing timestamps to resolve branches, then recover the hidden target question among distractor leaves. We apply curriculum RL training progressively increasing tree depth and introduce a compositional variant whose edges carry step-level reasoning fragments, training the model to compose partial cues into coherent queries. Trained on 10K-token contexts, StateTree generalizes to 128K tokens without full-length RL costs and exhibits capabilities including cross-session retrieval, temporal reasoning, knowledge update, and compositional multi-hop reasoning. StateTree outperforms both SFT and RL-based baselines while preserving short-context general reasoning. StateTree-7B achieves gains up to +23.60% on LongMemEval (128k), and StateTree-14B reaches 59.00% accuracy on LongMemEval, surpassing QwenLong-L1-32B (45.20%).

Deep Research agents synthesize evidence into cited reports, yet a well-cited report can still reach a misleading conclusion. Citation correctness checks whether cited sources support individual claims. It does not show whether adaptive search exposed a representative view of all documents made available for evaluation, which we call the candidate pool. Early findings redirect later queries, document choices, and stopping, so the documents an agent reads form a selective sample. Existing evaluations rarely account for this selection. We formulate the problem as adaptive evidence sampling and introduce Causal Evidence Selection Correction (CESS). CESS predicts each candidate document's evidence direction and corrects the candidate-pool average using the logged probabilities of selecting each document and reaching each search round. Shrinkage stabilizes short searches, while intervals replace point estimates when some documents cannot be sampled. We also prove that estimating the average evidence direction of a common pool differs from measuring how a change in search policy alters the evidence read. The latter requires intervention. On questions from the MS2 systematic-review benchmark, CESS reduces mean absolute error against the candidate-pool average by $9.2\%$ and reduces the estimate's change under opposing document rankings by $39.4\%$ relative to averaging the evidence scores of documents read. Across trajectories from a public Open Deep Research agent, the corresponding reductions reach $60.1\%$ and $87.2\%$. A further 4,800 trajectories under paired interventions confirm that correcting a pool estimate and measuring a policy effect are different tasks. CESS therefore audits whether the evidence direction underlying a report reflects the documents available for evaluation, while a separate intervention analysis measures the effect of search decisions.

Human intelligence relies heavily on learned intuition: recognising patterns and judging situations without explicitly unfolding every intermediate step. We introduce Bongard, an open-weight System One model that treats machine intuition as an independent capability to design and train. A T5Gemma 2 4B-4B encoder-decoder separates reading the evidence from making judgments. The encoder reads the state bidirectionally together with the question instructions, and separate decoder branches share this encoding, so many judgments about the same situation require only one reading of the state. A trained head returns probabilities over the supplied candidates without generating text. Training proceeds in three stages, from supervised judgments to semantic relationships to action outcomes, and each stage updates all 7.09 billion trainable parameters on one Blackwell GPU. Joint-embedding post-training raises accuracy on held-out rephrasings from 75.7% to 85.9%. A sandbox stage then learns outcome distributions from action rollouts and exact oracles, raising accuracy on a frozen sandbox panel from 50.6% to 64.8%. On DecisionBench, the final model reaches 78.05% accuracy over 23,900 decisions and ranks fourth of 61 systems in the public comparison. On one RTX PRO 6000, its median latency is 36 ms for short requests, and 32 questions about one state take 221 ms. Bongard demonstrates that machine intuition can be systematically trained via representation learning and outcome feedback, providing an open, efficient alternative for high-throughput decision workloads.

Deep research tasks require agents to navigate large knowledge spaces, synthesize evidence across many sources, and adapt their plans as findings emerge. Directed acyclic graph (DAG)-based multi-agent systems suit this setting because they support parallel execution and isolate each sub-task within a focused dependency context. Yet existing DAG-based agents instantiate a task-level plan before execution and repair the graph only after failures or missing evidence are observed. This Plan-then-Patch strategy is brittle for deep research: the system commits most strongly when its evidence is weakest, and later revisions waste computation on branches that should not have been planned. We propose DAGent, a DAG-based multi-agent framework with Evaluate-then-Grow incremental planning: an Orchestrator grows the task graph one batch at a time, conditioning each expansion on confidence and uncertainty signals from completed nodes. A hierarchical context layer propagates compact QueryDocs by default while preserving full execution traces for on-demand recall. The recorded DAG topology admits structural RL signals that outcome-only recipes cannot define; DAGRPO, a GRPO adaptation, injects topology-conditioned credit on Executor rollouts and a structural compliance regularization on Orchestrator plans. Across BrowseComp-Plus, GAIA, and xbench-DeepSearch, DAGent surpasses the strongest open-source baseline by 5.3 / 5.8 / 2.0 points at the Qwen3-235B-A22B scale, and the lead replicates across four open-source backbones and extends to GPT-5 at 327K context. At the Qwen3-8B scale, DAGRPO improves over a same-budget outcome-only GRPO baseline by 3.0 average Pass@1 points. A same-architecture comparison shows that evidence-conditioned planning reaches higher accuracy at lower per-task token, tool-call, and step footprints than its Plan-then-Patch counterpart. Code: https://github.com/hanwenliu6825/DAGent

Standard language model RL algorithms credit every token of a long rollout with the same advantage determined by the terminal reward. Actor-critic methods can provide finer-grained credit assignment, but learned critics are generally considered too inaccurate to trust when training LLMs with RL. In recent works, even when a critic is present, it is used only for baseline estimation, so every trajectory must be rolled out to its terminal reward. We introduce Actor-Critic with Action Chunking (AC2) that removes the need to roll every trajectory to completion. AC2 instead assigns credit to action chunks: short continuations of prefixes of past trajectories. A learned critic scores the state reached at the end of each action chunk, allowing the policy to update without observing a terminal reward. We make critic-based credit assignment reliable through three design choices. First, we introduce local readiness which uses critic-based updates on a problem only when the critic is sufficiently accurate on that particular problem. Second, when available, we provide the critic with a reference solution from a previous successful rollout. Third, we assign credit over action chunks of 10k tokens rather than individual tokens, giving the critic a more meaningful portion of the trajectory to evaluate. We train Qwen3-4B on FineProofs-RL using AC2 and evaluate on IMO-ProofBench. AC2 exceeds GRPO's peak validation score of 18.5% using 2.5x fewer decoding FLOPs. This gain comes from two sources, (1) AC2 requires 25% fewer training steps to reach this score, and (2) each step generates fewer tokens because the policy does not need to continue every trajectory to completion. Conceptually, we demonstrate that we can remove the need to roll out every trajectory to completion, opening up a large previously unexplored design space for LLM RL algorithms.

On-policy self-distillation (OPSD) trains mathematical reasoning models using a privileged teacher that sees a reference solution and supervises student-sampled prefixes. Standard OPSD uses one fixed parameter setting at every state, but nearby settings may offer additional supervision. We find that local parameter perturbations reveal complementary reference-aligned corrections under the same reference context. Different experts supply these corrections at different reference positions. Their pool covers more such positions than the unperturbed privileged teacher. We introduce Neighborhood OPSD (N-OPSD) to turn these corrections into supervision at student-visited states. Offline, greedy selection builds a compact pool of frozen experts by rewarding filtered reference-token gains beyond the pool's current best at each position. The highest-peak expert need not provide the best training target. Online routing therefore separates the anchor direction from its level of support. MaxPeak selects the anchor token, and quantile selection chooses among experts whose top token matches it. The student learns from the chosen expert's full next-token distribution through the clipped forward-KL objective inherited from OPSD. We evaluate on AIME 2024, AIME 2025, and HMMT February 2025. Across three independent runs per method, Neighborhood OPSD improves the three-benchmark Average@12 over OPSD by 2.75, 1.67, and 1.94 points on Qwen3-1.7B, 4B, and 8B, respectively. Student-prefix continuations support using the pool beyond the reference trajectories used for selection. Matched ablations support filtered reference-token gains as a selection criterion. Accounting for overlap within the pool and routing by state further improve student accuracy. Inference uses only the distilled student.

Speech-to-speech systems must solve tasks whose requirements emerge across conversational turns. We introduce SpeechConversationBench (SCB), a focused evaluation of spoken mathematical reasoning using 103 sharded GSM8K problems. The framework compares the original problem delivered in one turn (full), its concatenated information shards delivered together (concat), and incremental spoken disclosure across turns (sharded). We report final-answer accuracy for four commercial speech systems and LEGO, a proprietary speech pipeline developed internally by the SCBX Innovation Lab team with explicit conversational context management. Relative to concat, sharded accuracy decreases by 5.0-25.3 percentage points across the four commercial systems. LEGO achieves 77.5 percent accuracy in all three conditions, compared with 76.6 percent sharded accuracy for GPT-4o Realtime. The two single-turn baselines distinguish sensitivity to problem reformulation from the additional challenges introduced by incremental spoken interaction.

Reinforcement learning with verifiable rewards (RLVR) has improved the reasoning capabilities of large language models (LLMs), yet their predictions remain sensitive to task-irrelevant prompt features. We investigate this sensitivity through semifactual prompt interventions that preserve the underlying problem and its answer. Our analysis reveals substantial variation in token-level sensitivity and shows that suppressing high-drift token candidates during decoding improves reasoning accuracy without updating model weights. These findings highlight a limitation of Group Relative Policy Optimization (GRPO), which assigns the same outcome-derived advantage to every response token and may reinforce potential spurious dependence alongside useful reasoning. Motivated by this observation, we introduce Semifactual Credit-Augmented Policy Optimization (SCAPO), a causally inspired variant of GRPO that incorporates semifactual stability into token-level credit assignment. SCAPO measures token probability drift for fixed responses under semifactual interventions and uses normalized stability scores to reduce advantages for relatively unstable tokens during early training, while granting no additional credit for stability alone. On Qwen3-4B-Base and Qwen3-1.7B-Base, SCAPO improves AIME 2024-2026 accuracy over GRPO by 5.63 and 4.17 percentage points, respectively. At both model scales, SCAPO achieves the best results on most evaluated mathematics benchmarks and all evaluated out-of-distribution benchmarks among the compared methods. These results suggest that semifactual stability provides an effective training signal for improving reasoning and generalization through finer-grained credit assignment in RLVR. The code is available at https://github.com/DtYXs/SCAPO.