2026-10-21

Enterprise adoption of LLMs in agentic workflows is gated by evaluations that span a wide arc of abilities: CRM operations, heterogeneous-document retrieval, personal data understanding, multi-turn planning, tool use, function calling, long-horizon research, and long-context software engineering. These evaluations live today in separate repositories with divergent runners, dependencies, and scoring conventions, which makes cross-benchmark comparison difficult and obscures what a single aggregate number actually represents. We release EGI-BENCH, a unified LLM agent benchmark suite that consolidates eight enterprise facing agent benchmarks (CRMARENA, HERB, PERSONABENCH, USERBENCH, MCP-UNIVERSE, MCPEVAL, BENCHFORCE, LOCOBENCH) behind a single command-line interface with two runtime-budgeted execution modes. On top of the suite we define the Enterprise General Intelligence (EGI) Score, a six-dimensional framework (M1-M6) whose buckets are populated by fine-grained sub-metrics routed from each benchmark rather than by benchmark-level averages. We describe the suite, the sub-metric routing, and the quick/deep budgets, and report results for fourteen frontier and cost-efficient LLMs spanning the GPT, Claude, and Gemini families under quick mode. Top-line EGI ranges from 39.6 (GPT-4o mini) to 65.6 (gpt-5.5), and no single model dominates every capability axis: gpt-5.5 leads task outcome, planning, and retrieval grounding; Claude Sonnet 4.5 leads tool actioning and safety/abstention; and Claude Haiku 4.5 leads engineering quality. Code is available at https://github.com/SalesforceAIResearch/egi-bench.

Existing steering benchmarks are built around simplified, isolated generation tasks and evaluate a single behavioral attribute under controlled settings that do not reflect the diversity and complexity of practical deployment. As a result, performance on these benchmarks provides only a limited picture of how steering methods perform in the real world. We therefore introduce RealControlBench, a comprehensive benchmark for evaluating fine-grained steering of LLMs and LLM agents across diverse behavioral targets and interaction settings. RealControlBench extends steering evaluation beyond isolated generations to complex, execution-grounded, multi-turn agent interactions, covering five complementary dimensions: instruction-following control, agentic tool-use decision control, reasoning-performance steering, reasoning-length control, and trajectory-conditioned safety awareness. Rather than measuring steering success alone, the benchmark jointly evaluates target controllability and capability preservation, including instruction following, reasoning performance, and downstream agent utility and safety. It further examines whether steering effects remain specific and robust across prompt variations, domains, and model architectures. Our evaluation shows that existing steering methods remain unreliable across realistic control objectives: they often fail to consistently induce the intended behavior and can substantially degrade the model's underlying task performance. These results suggest that strong steering effects in controlled settings do not necessarily translate into reliable control in more realistic evaluations. By providing a unified, model-agnostic steering evaluation framework, RealControlBench aims to support the development of steering methods that are not only effective but also precise, robust, and useful in realistic AI deployments.

Span-level hallucination detection pinpoints unsupported content in LLM outputs, enabling actionable feedback and targeted mitigation. Existing evaluations for span-level hallucination detection methods are fragmented across tasks, annotation formats, and metrics, and most span-level datasets cover only a single NLG domain, making it difficult to assess whether a detection method actually generalizes across tasks and domains. In this paper, we introduce SHAM-Bench (Span-level Hallucination Benchmark Across Multiple Tasks), a unified benchmark for span-level hallucination detection spanning six NLG tasks, aggregated from existing human-annotated datasets totaling 6,461 examples under a common schema and evaluation protocol. Using SHAM-Bench, we evaluate existing hallucination detection methods and find that fine-tuned detectors often perform well in-domain but degrade under distribution shift, whereas prompt-based methods with reasoning are more robust across tasks. Our analysis further shows that subtle unsupported content remains difficult to detect and span-level evaluation is still challenged by ambiguous span boundaries and noisy labels. We release the dataset, evaluation code, and a public leaderboard to support future research on broadly generalizable hallucination detection.

Every day, more than 40 million people ask ChatGPT a health question, such as whether laptop posture explains a stiff neck and days of pounding headache. Posture is plausible, but meningitis is a time-critical possibility. We introduce SelfDx, a paired benchmark for whether models have the appropriate response behavior, a form of epistemic independence, in health questions where misplaced agreement with the user's self-diagnosis could delay urgent or emergency care. Each of 100 pairs contrasts a caution case, where a serious alternative remains possible, with an agreement case, where the user supplies evidence that rules out the alternative and it is safe for the model to agree. We find across 27 frontier models that they tend to inappropriately agree with the user's self-diagnosis in cases that require caution. Even the best current-generation model reaches only a 77.0% appropriate caution rate on time-critical condition cases. To address this gap, we experiment with three prompt variants that increase appropriate caution but reduce appropriate agreement by 9-35%, indicating excessive caution. In contrast, supervised fine-tuning on 128 balanced examples improves both rates together across three runs. We recommend post-training to improve appropriate caution without making models cautious on every case, with paired evaluation to verify both response behaviors.

Depression dialogue benchmarks commonly assess models by comparing predicted symptom labels with clinical scale scores. However, a prediction may agree with a benchmark label even when the dialogue lacks enough evidence to justify it. We introduce EviDial, a benchmark that distinguishes label agreement from evidence-grounded assessment. EviDial comprises 382 depression dialogues from three sources and 2,744 participant-symptom cells, each classified as having absent, partial, conflicting, or sufficient evidence. Because evidence sufficiency lacks a directly observable gold standard, we derive calibrated evidence references from multiple judgments using generalizability theory, many-facet Rasch measurement, and random-effects latent class analysis. Our audit reveals that 82.8% of participant-symptom cells lack sufficient textual evidence for the corresponding symptom assessment, with substantial variation across sources. Across model families and prompting strategies, 69.2%-96.9% of predictions judged correct by benchmark labels are nonetheless unsupported by sufficient dialogue evidence. Further analyses show that strong agreement among LLM evaluators need not reflect independent evidence judgments, underscoring the need to calibrate automated reference panels. These findings demonstrate that label-based evaluation can overestimate evidence-grounded performance by conflating agreement with clinical labels and support in the observed dialogue. EviDial provides a measurement layer for separating these constructs and auditing LLM-based depression assessment more faithfully.

When an LLM agent acts on behalf of a user, it is expected to act within the authorization scope the user grants. Existing agent evaluations largely measure task completion or harmful consequences and can therefore overlook actions that violate this delegation without producing conventional safety or privacy harms. We introduce AuthorizationBench, a benchmark of authorization alignment grounded in documented real-world cases. A scalable construction pipeline produces 286 executable tasks spanning situational triggers, the presence or absence of direct harmful consequences, and ten application domains. We execute every task with four models at two reasoning-effort settings across three agent harnesses in an authorization-aware sandbox that keeps authorized and out-of-authorization paths executable and observable, yielding 6,864 trajectories. Agents act out of authorization in 19.2% of trajectories, with rates ranging from 10.5% to 26.6% across the 24 configurations. Task completion does not imply authorization. Increasing reasoning effort leaves the aggregate rate nearly unchanged, with effects varying across models. On controlled pairs, agents act out of authorization at similar rates with (17.9%) and without (20.8%) direct harmful consequences, and more often under benign pressures than malicious injections (23.5% versus 17.6%, p=0.002). Our results show that task capability, additional reasoning, and consequence-oriented safety evaluation do not by themselves ensure authorization alignment. Moreover, agents that consult the principal before completing a task exhibit lower rates of out-of-authorization behavior (14.0% versus 21.6%, p<0.001), pointing to a tension between agent autonomy and preserving users' decision authority that calls for further human-centered study. Dataset and code are available at https://github.com/PEACH-Research-Lab/AuthorizationBench .

Ensuring the alignment of Large Language Models (LLMs) with diverse cultural values is critical for their safe global deployment. However, existing alignment benchmarks are predominantly English-centric or confined to simple discriminative tasks. Moreover, they often fail to effectively evaluate model robustness against complex, adversarial inductions within specific cultural contexts. To address these limitations, we propose ABench, a fine-grained, handcrafted benchmark specifically tailored to taxonomized values in Chinese. Grounded in a rigorous hierarchical taxonomy comprising 158 derived value subcategories, ABench consists of 2,895 open-ended questions. These questions employ techniques such as role-playing and logical traps to probe the stability of models' stances in adversarially induced contexts. Furthermore, we propose a weighted automated evaluation framework that utilizes question-specific scoring criteria for nuanced assessment. Extensive experiments on 16 mainstream LLMs reveal that most models exhibit significant vulnerabilities when subjected to sophisticated adversary attacks on values (scoring below 60). This underscores the need for more robust, culturally specific alignment strategies. Our data will be publicly released to promote further research on taxonomized value alignment.

Enterprise agents answer questions over workplace records that evolve through team collaboration. As information is shared, corrected, and revised across documents, chats, and figures, multiple versions of the same information accumulate in the workspace. Reliable question answering therefore requires agents to reason over sufficient evidence, resolve cross-source and temporal conflicts, and recognize when the available evidence is insufficient to establish an answer. To systematically evaluate these capabilities, we introduce MMEnterprise-Bench, a benchmark comprising 1,613 questions grounded in evolving multimodal records from simulated collaboration projects, covering three categories: Evidence Reasoning, Conflict Resolution, and Unanswerability. We evaluate 10 proprietary and open-source multimodal LLMs and find that even the strongest models struggle when evidence contains cross-source or temporal conflicts or spans multiple images. Moreover, increasing the retrieval budget or employing a larger retriever improves evidence recall but does not consistently translate into higher answer accuracy. These findings highlight the persistent challenges of reasoning over evolving workplace records and establish MMEnterprise-Bench as a testbed for developing more reliable enterprise agents. Code and data are available at https://mlvlab.github.io/MMEnterprise-Bench/.

Persuasive writing exploits cognitive biases, recruiting a reader's mental shortcuts to shape belief, often without the reader noticing. We call the linguistic features exploiting such mental shortcuts cognitive bias triggers. As large language models (LLMs) increasingly mediate how people encounter news, summarizing it, ranking it, and answering questions over it, a natural question is whether they can recognize the features of a text that bias a human reader. We introduce BAIT (Bias Annotation In Text), a resource for detecting cognitive bias triggers in news. BAIT pairs a schema of nine triggers, each specified in an annotator codebook, with 50 English news articles annotated by trained annotators at the article and span levels and adjudicated into a gold standard. On this data we define two tasks, detecting which triggers a text contains and locating the spans that carry them. We then benchmark six frontier LLMs zero-shot and with three or ten in-context examples. The models approach human agreement on which triggers an article contains, but rarely recover its full set of triggers, and they reach less than half of human agreement in locating the passages that carry them. Moreover, they over-predict triggers that come with easy surface cues, and most of the passages they mark for these triggers are not triggers, yet carry the trigger's surface cue far more often than random text. These findings suggest that the evaluated LLMs respond to surface cues rather than to the mechanisms that make a passage manipulative. We release the schema, the codebook, standoff annotations, and the evaluation pipeline.

Decision models support routing, guardrails, and output verification in LLM systems. However, item accuracy alone cannot establish whether models make correct decisions across controlled variations or assign reliable confidence to their predictions. We introduce DecisionPivot, a diagnostic benchmark comprising 48 tasks and 40,033 items across seven decision roles. Its core track contains 7,800 items from 13 generated task families, organized into contrast groups with computed answers verified by independently implemented solvers. The complementary reference track contains 32,233 items from 35 public test sets, preserving their original evaluation metrics. We evaluate 40 decision models and 14 general models using documented, model-specific protocols, assessing item and contrast-group accuracy, probability ordering, calibration, and empirical selective coverage. Our results reveal substantial differences obscured by aggregate scores: two 27B models separated by only 2.6 points on the reference track achieve contrast-group accuracies of 70.1% and 44.2%. Across 2,400 binary contrast pairs, Perplexity Decider correctly ranks the positive member above the negative member in 92.3% of pairs but answers both members correctly in only 74.1%, exposing a gap between probability ordering and decision correctness. Models with similar item accuracy also differ in the reliability of their high-confidence predictions. On a separate 1,296-item subset, GLM-5.3 achieves 93.5\% contrast-group accuracy at a higher reasoning setting, compared with 81.5% at its lowest effort. By linking errors to pivotal details and difficulty levels, DecisionPivot provides fine-grained diagnostics of decision failures that aggregate evaluation alone cannot reveal.

Mathematical reasoning remains underexplored for low-resource languages, including Dialectal Arabic. To bridge this gap, we introduce the DAR-MGSM (Dialectal Arabic Multilingual Grade School Math) Benchmark, the first manually curated benchmark for evaluating mathematical reasoning in Dialectal Arabic. DAR-MGSM consists of math word problems (MWPs) requiring multi-step reasoning and covers five major Arabic dialects alongside Modern Standard Arabic (MSA). The benchmark comprises two parallel versions: translated-only and translated & culturally localized. The latter version adapts entities and references to regional varieties. We evaluate eleven Arabic-centric and multilingual large language models on DAR-MGSM. The results show that the performance generally declines on the localized version, highlighting the challenges introduced by culturally grounded adaptations. However, the magnitude of this effect varies across models, with Arabic-centric models exhibiting different levels of sensitivity to localization. A mechanistic interpretability analysis shows that translated and localized variants of the same MWPs often induce distinct internal reasoning trajectories, even when aggregate accuracy remains similar. Overall, our results show that mathematical reasoning performance depends on both linguistic variety and cultural context.

Mathematical reasoning is a fundamental capability for educational applications of Large Language Models (LLMs), where reliable problem solving requires not only correct answers but also coherent and verifiable reasoning processes. Multi-agent systems (MASs) have shown growing potential in this domain by leveraging collaboration and verification among multiple LLM agents. However, existing evaluations primarily focus on the correctness of final answers in text-only settings, offering limited insight into multimodal reasoning capabilities, the reliability of intermediate reasoning steps, and robustness against reasoning errors. To address these limitations, we introduce MAMath-Bench, a process-aware benchmark designed to evaluate the multimodal mathematical reasoning capabilities of MASs. MAMath-Bench comprises 1,003 problems spanning 11 mathematical categories, multiple difficulty levels, and several modalities, including text, images, and audio. It features fine-grained milestone annotations to evaluate systems across three dimensions: problem-solving correctness, reasoning-process reliability, and robustness to adversarial reasoning-chain perturbations. Our evaluation yields three key findings. First, multi-agent collaboration does not consistently improve final-answer accuracy over strong single LLMs. Second, collaboration can substantially enhance the coverage of intermediate reasoning, but such gains may come at the cost of increased reasoning redundancy. Third, current systems remain vulnerable to adversarial reasoning errors, particularly those involving subtle structural perturbations. The dataset is available at https://github.com/mira-ai-lab/AgentMath.

Retrieval-augmented generation (RAG) is the dominant paradigm for grounding large language models in external corpora, but its similarity-driven, chunked pipeline has structural blind spots: it cannot anticipate which passages a question requires and discards document-level context that resolves cross-passage references. We introduce evidence-based generation (EBG), requiring models to return an answer with line-level citations, and a joint Ans+Evid metric that scores correctness and grounding simultaneously. We reformulate two corpora into evidence-grounded multi-document QA benchmarks- MultiHop-RAG and QASPER-MP, a multi-paper adaptation of QASPER with realistic short paper cues, and pair them with a controlled synthetic two-hop haystack and an evidence-augmented NoLiMa spanning 10K-1M characters. Across ten proprietary frontier and three open-weight LLMs, EBG isolates two RAG limitations that long-context models avoid: similarity retrieval surfaces any relevant chunk far more reliably than all required chunks for multi-fact queries, and chunking destroys document meta-information needed to resolve identity-bearing queries. EBG also exposes grounding failures invisible to answer-only evaluation, including incomplete citations in two-needle settings and a widening gap between Ans and Ans+Evid when context contradicts parametric knowledge. Data and code are available at https://github.com/debobanerjee/LC-EBG.

Multi-agent LLM systems improve performance on clinical tasks, and that improvement is often read as evidence of a more trustworthy system. However, the two are not the same claim. A clinician acts on the stated reasoning and not only on the conclusion, so whether collaboration reduces hallucination, and what decides when it does, has to be measured in the reasoning too. Yet most evaluations score only the final answer. We evaluate four systems (a single-LLM baseline and three multi-agent topologies) across four models and three clinical tasks, scoring hallucination, under a factuality and faithfulness taxonomy, inside the rationale each system generates and reporting it alongside the task metric. We report that collaboration buys task performance but not lower hallucination as a rule, and no topology is reliably better than another. The effect differs by model: of the $36$ model--system--task comparisons, $12$ are large enough to resolve --- four decreases and eight increases, with no model on both sides. What sets the level, and the direction of any change, is the model rather than the topology or the number of agents. A self-correction experiment is consistent with this: shown its own flagged errors, every model except Llama-3.2-3B-Instruct reduces hallucination; the one model for which collaboration lowers hallucination gains the most, and Llama-3.2-3B-Instruct, the model whose hallucination it amplifies most, does not reduce its hallucination at all. We find that collaboration redistributes what its models can already do; it does not supply what they cannot.

Large Language Models (LLMs) have significantly advanced natural language processing and artificial intelligence applications through their capabilities in reasoning, text generation and knowledge retrieval. However, LLMs are still vulnerable to two important challenges: hallucination, generating factually incorrect or fabricated content, and response deviation, meaning systematic deviations from ground-truth responses. These limitations raise serious concerns regarding reliability and trustworthiness in academic and research environments. This study offers a thorough comparative analysis using a variety of metrics of ten advanced LLMs ChatGPT, Gemini, Claude~3.5~Sonnet, LLaMA~3, Mistral, Grok, Perplexity, DeepSeek~V3, Falcon, and Gemma based on a benchmark consisting of 70 carefully designed academic prompts across four complexity levels: Basic, Explanation, Advanced, and Tricky, evaluated in February 2025 on specific model versions. Each response was independently evaluated by three reviewers using five metrics: Accuracy, Precision, Recall, Response Deviation Index, and Hallucination Index. Perplexity and DeepSeek~V3 achieved the highest overall accuracy, while Gemini demonstrated strong recall and a well-balanced performance profile across all five metrics. LLaMA~3 showed strong resistance against hallucinations through a conservative method of generation. Advanced-level prompts proved most challenging across all models, while Tricky prompts were handled more reliably owing to their binary true/false structure. Falcon achieved perfect precision and zero hallucinations, but at the cost of the lowest recall and the highest response deviation. The results show a fundamental trade-off between response completeness and factual reliability, intensified by prompt complexity.

As AI systems begin to resolve open research problems, reliable verification becomes increasingly important. Formal mathematics provides a trusted framework for mathematical reasoning, enabling agents to produce machine-checkable proofs while conducting mathematical exploration. Existing formal mathematics benchmarks often focus on known results or a small set of prominent open problems, which can become difficult to evaluate reliably once solutions are discovered and widely disseminated. Existing evaluation protocols are also not always aligned with how modern coding agents operate. We therefore introduce ReuF2F, a Lean 4 benchmark of 244 formalized open problems drawn from Research Experiences for Undergraduates (REU), spanning entry-level open research problems across a broad range of mathematical subfields. We expect these less prominent open problems, which tend to attract less attention and may remain unsolved for longer, to be better suited for held-out evaluation of end-to-end open problem resolution in Lean. We adopt a standard coding-agent evaluation setting, placing each problem and its logical negation into a Lean project in a Docker container. We also build Leanimum-agent on mini-SWE-agent to provide a unified harness for fair comparisons across models. Our evaluation finds that even the best-performing model resolves only ~40% of the problems, leaving substantial room for improvement on formalized open research problems.

Large Language Models (LLMs) increasingly operate across culturally diverse settings, yet existing evaluations largely measure cultural behavior at the output level. We introduce CultureLens, a reference-grounded benchmark and mechanistic study of how LLMs use cultural information. CultureLens contains 9,261 scenarios across 75 countries and systematically varies country, value, true-norm, and contradictory-norm evidence. Across Llama-3.1-8B, Gemma-3-12B, and Qwen-3.5-27B, stating a contradictory norm lowers aligned answers by 32-47 percentage points, even when the model answers correctly from the country alone. Mechanistically, models default to Anglophone norms when no country is given. The failure under a contradictory norm is not one of missing knowledge, as blocking the norm inside the model restores the correct answer in most cases. Models accept a stated norm according to how it is presented and not whether it is true, which leaves them most exposed to errors in retrieved content. Supervised fine-tuning with contradictory-norm examples raises aligned answers under a contradictory norm to 85-93%, including on unseen countries, at a small cost under a true norm.

Web agents are typically evaluated on English-language interfaces, leaving their ability to operate across the world's multilingual websites largely unexplored. This gap is especially significant for low-resource languages, where standardized resources for evaluating web agents remain limited. To address this gap, we introduce Low-Resource Web Agent Benchmark (LoWAgent), a benchmark for evaluating web agents across different low-resource languages (LRLs) in comparison to the default, English. LoWAgent evaluates web agents across English and four LRLs: Hindi, Bengali, Urdu, and Amharic. The benchmark comprises of six fully localized websites, instantiated across five languages to yield 3,000 task instances. We evaluate three closed-source and five open-source models across all five languages on all six websites. Our findings show that both closed-source and open-source models experience performance degradation for LRLs compared to English, with substantially greater performance drops observed among the open-source model agents. We further evaluate the robustness of the models on realistic interface disruptions, such as cookie notices and advertisement pop-ups, and find that these further degrade agentic performance. Finally, we show that post-training on a small subset of LoWAgent traces can significantly improve low-resource language agentic task performance, demonstrating the usefulness of our benchmark.

Analyzing merchant transaction histories requires dynamic temporal filtering, entity resolution, and aggregation over irregular event streams--operations poorly captured by static tabular or time-series forecasting benchmarks. We introduce TxnQA, an analytical QA benchmark comprising 27 task families (81 templates) alongside a 5-scenario fraud-risk reasoning suite that evaluates both verdicts and supporting evidence over synthetic, privacy-mitigated transaction data. Ground-truth answers are computed programmatically, and model performance is decomposed into format-following rate (FFR), conditional content accuracy (CA), and end-to-end correctness (E2E). Evaluating prompted proprietary baselines and six LoRA-adapted open-weight models shows that adaptation reliably enforces output schemas but leaves arithmetic bottlenecks: Gemma-4-31B-IT and Qwen3-8B attain E2E scores of 0.867 and 0.816, whereas evaluated MoE checkpoints reach only 0.632-0.713. Furthermore, on the 500-instance fraud suite, verdict accuracy frequently exceeds 80%, yet joint accuracy (requiring exact evidence attribution) collapses to at most ~23%. These findings identify schema compliance, semantic calculation, and evidence attribution as three distinct bottlenecks in transaction-sequence QA.

Code agents increasingly work on evolving real repositories, and memory systems aim to help them reuse earlier experience. However, such experience is diverse and evolving, so it remains unclear how memory systems organize it and thereby affect code agents. Existing benchmarks cannot answer this reliably because they lack either temporally valid histories or control over prior-task relevance. In this paper, we present MemDev-Bench to evaluate how memory systems affect code agents over diverse repository histories. It pairs each task with relevance-annotated earlier tasks from the same repository to build histories with controlled relevance. We also build an automated pipeline that reliably makes these earlier tasks executable and verifiable, producing 4,869 historical tasks for 555 base tasks. Based on MemDev-Bench, our systematic evaluation of memory systems across diverse development scenarios yields three main findings. (1) Memory mainly improves fine-grained localization and editing, increasing functional correctness. (2) Memory also causes failures on otherwise solvable tasks even with related history, and unrelated history tends to increase such failures. These failures mainly stem from agents reusing memory without reassessing its applicability. (3) Compared with using memory systems, directly providing agents with raw historical trajectories substantially reduces their steps. Further analysis suggests that agents save these steps by reusing process information that compressed memories tend to omit. These results suggest that future memory systems should preserve process information and reassess memory applicability against current task context.

Full-duplex speech agents must update persistent memory while listening and speaking, yet it remains unclear which events should be stored, revised, or excluded. Existing benchmarks isolate interaction timing from memory outcomes, obscuring the state transitions behind a correct answer. We introduce DuplexLedger, a bilingual benchmark of 1,094 persistent sessions aligning four full-duplex scenarios with event-level Write/Update/Ignore/Retain supervision, explicit pre/post memory states, and 6,564 same-session questions and checks. Evaluating five full-duplex models reveals an interaction-memory gap: no system combines strong boundary control with coherent final-state memory, and the best final-state fact coverage is 2.01%. We propose ACT-Mem, an 8.92M-parameter sidecar comprising an event controller, a versioned transactional ledger, and a typed retriever coupled with a frozen MiniCPM-o 4.5 backbone. On 219 held-out sessions, ACT-Mem increases latest-value accuracy from 4.11% to 8.68% while reducing stale assertions and distractor adoption. It improves interaction boundaries but slows interruption stopping. These results motivate selective, auditable full-duplex memory rather than indiscriminate transcript retention. Code is available at https://github.com/lyx-researcher/DuplexLedger.

Existing reasoning evaluations typically provide task-relevant information to models as available premises, rarely examining concurrently how models discover evidence autonomously and how information presentation within narratives guides or obscures judgment. This paper proposes the WORTH Benchmark: models read original mystery narratives without clue annotations, autonomously select key clue sentences, identify case participants, and explain their reasoning process. The narrative materials are designed by professional authors such that evidence facilitating reasoning coexists with narrative arrangements that delay the revelation of truth. We report a pilot experiment under full-story conditions across two episodes, conducting a total of 70 formal runs on models, of which 67 runs are content-scorable. Scoring separately records hits on author-designated key clue sentences and participant selections; natural language explanations are used for case diagnosis and are not scored independently. The results show that clue hits and participant judgments are not synchronized; across 37 content-scorable runs in Episode 02, 31 runs simultaneously selected the same set of incorrect participants, while adjacent suspicion-oriented and exoneration-oriented sentences were selected 25 times and 0 times, respectively. These phenomena illustrate the diagnostic value of component-wise analysis and suggest that not-fully-cooperative informational relationships may influence evidence selection and judgment. This study only implements full-story conditions for two episodes, and does not establish stable model rankings or infer general mechanisms on this basis; human player records have not yet been cleaned, and this version does not report human-model comparisons.

Multimodal Large Language Models show strong vision-language capabilities, yet their geometry-centric reasoning remains insufficiently understood. Existing mathematical benchmarks typically rely on rasterized figures and cover limited geometric knowledge areas and response formats, making it difficult to assess whether models reason over geometric structures or exploit superficial visual patterns. We introduce TriGeoBench, a geometry-centric multimodal benchmark constructed from Chinese middle- and high-school mathematics examinations. TriGeoBench contains 7,000 question-answer pairs and 12,235 geometric figures with TikZ code, spanning plane geometry, solid geometry, analytic geometry, function, and vector geometry across diverse response formats. Each instance forms a {question-answer pair, image, TikZ} triplet: question-answer pairs are stored in LaTeX, images as RGB figures, and TikZ preserves geometric topology, symbolic annotations, and relational structure. We evaluate representative open- and closed-source MLLMs through image-to-TikZ and multimodal reasoning probes. Results reveal persistent fine-grained structure and image-text alignment challenges. Oracle experiments with human-curated reference TikZ suggest that accurate executable representations provide complementary geometric cues; they do not evaluate an end-to-end conversion pipeline. TriGeoBench offers a testbed for structure-aware mathematical reasoning.

As computer-use agents (CUAs) move to long-horizon workflows spanning multiple applications, undetected errors can silently corrupt persistent state and cause cascading failures in subsequent actions. Reliable error detection and recovery have become critical bottlenecks for building robust CUAs. However, existing benchmarks focus primarily on end-to-end task performance, providing limited insight into whether agents can detect and recover from errors during execution. The few benchmarks that evaluate recovery rely largely on manually injected errors, which may not reflect the types or horizons of failures induced by real agent policies. We introduce CUA-Recovery, a benchmark comprising test and training suites across multiple desktop applications. We construct it using ReRail, a framework for synthesizing long-horizon tasks and trajectories in a persistent desktop environment. ReRail composes MyPCBench tasks into new long-horizon workflows with executable ground truth, collects verifier-confirmed failures from multiple policies, and turns them into verified recovery trajectories. CUA-Recovery comprises a human-verified test suite of 2,250 erroneous states over 100 tasks and a training suite of 8.6K trajectories over 377 tasks. Evaluation on the test suite shows that agents struggle both to recognize inherited errors and to complete the task. Across six agents and six takeover depths, Error Awareness Rate (EAR) reaches at most 44.0% and Pass@3 at most 25.5%. After fine-tuning Qwen3.5-35B-A3B on the training suite, its average EAR rises from 19.9% to 35.2% on held-out recovery states, and its Rubric Score goes from 5.0% to 16.1%. It also gains 15.3 points on MyPCBench and 10.8 points on Odysseys for Pass@3, showing that the training trajectories help both with recovery and with completing tasks from a clean start. Our code is available at https://github.com/xu-hu-2002/CUA-Recovery.

Large language models can achieve strong Text-to-SQL benchmark performance while still producing schema and query-logic errors. Most evaluations characterize such failures only in the final generated query, making it difficult to determine whether a related error was also observed in an earlier, simpler step or appears only at the final stage. We present a drill-down evaluation framework that decomposes Text-to-SQL queries into progressively more complex executable sub-queries and corresponding natural-language questions, allowing rule-based hallucination categories to be examined across a generation path. We distinguish recurrent hallucinations, in which the same rule-based category is observed in an earlier step and again in the final query, from emergent hallucinations, in which the category is observed at the final step without an earlier detection in the same path. Across five LLMs evaluated on an expanded BIRD-mini benchmark, several categories show appreciable recurrence probabilities, while clause abuse shows comparatively higher conditional emergence. Under the conversational-history condition, recurrence probabilities are generally higher and emergence probabilities are generally lower. These observations indicate that the two conditional measures do not move uniformly in the same direction under additional context and motivate trajectory-level analysis in addition to final-output evaluation. Code and evaluation artifacts are available in the project repository. https://github.com/HaydenMM/bird-mini-drilldown

Tool-using LLM agents increasingly answer e-commerce queries through multi-step tool interactions. The expected output of a tool call can shift between agent development and execution, allowing a plausible but incorrect value to be returned without an explicit failure signal. These silent errors may propagate through subsequent decisions into final answers. To study agents' resilience across this end-to-end process, we formulate Silent-error Resilience Task, an interactive task that evaluates whether they complete the original query after a silent error corrupts an intermediate state. We introduce RITE, a benchmark built from human-authored industrial workflows and production-log-derived inputs. To enable controlled and reproducible evaluation, we use a semantic error taxonomy to inject errors into fixed workflow prefixes, freeze recorded tool behavior, and let agents continue freely. It contains 167 matched units across nine e-commerce workflows, yielding 501 silent-error, correct-value, and explicit-failure instances. Across 10 recent LLMs, silent errors roughly halve task success and are consistently harder than explicit failures. Recognition, use of an available verification path, and integration of the resulting evidence form successive bottlenecks. We also propose a light State Checker as an intervention to test whether resilience can be improved.

Model compression makes large language models deployable, but today's evaluation tells an incomplete story. Standard practice measures retained task performance and stops there. This misses a critical fact: compression perturbs the distributed knowledge embedded across billions of weights, and task scores alone cannot reveal which underlying abilities survive and which quietly collapse. To solve this question, we introduce ACE-Compression, an ability-centric evaluation framework. Instead of asking only "does the compressed model still answer correctly?", we ask "which of its core abilities remain intact?" Our framework probes four deployment-facing dimensions: procedural reasoning and constraint execution, knowledge grounding and truthfulness, social cognition and affective understanding, and persona fidelity and stylistic consistency. Across more than 35 compressed checkpoints from six open-weight LLMs-covering quantization and pruning-we find that compression redistributes ability profiles. Coarse, local, and surface-level signals often persist, while fine-grained, long-tail, and context-dependent abilities become fragile. Quantization consistently outperforms pruning in retention, but the practical meaning of that gap depends entirely on which ability an application requires. Based on our findings, deployers can choose compression methods by the ability profile their application needs, not just by a leaderboard average. And algorithm designers gain diagnostic signals to build compression methods that preserve what matters. We believe that ACE-Compression turns compression from a black-box accuracy drop into an interpretable intervention-one we can measure, understand, and ultimately improve.

Reliable retrieval-augmented generation (RAG) requires recognizing when neither retrieved documents nor parametric knowledge supports an answer. We introduce DeclineCraft, a dynamic abstention benchmark whose unanswerable questions target Minecraft feature additions after a corpus cut-off, with audits of possible parametric exposure. Later cut-offs support evaluation of future models. Across 27 LLMs, we measure Correct Abstention Rate (CAR; how often models acknowledge insufficient knowledge on unanswerable questions), Answerable Attempt Rate and Answer Accuracy. We find that (1) single-shot retrieval significantly raises CAR in 38 of 40 configurations (model and thinking setting) versus no retrieval. (2) Iterative RAG, in which the model self-assesses its context and can request up to two further retrieval rounds, significantly lowers CAR versus single-shot RAG in 30 of 40 configurations (mean 46.4%), with little change in Answer Accuracy (68.5% to 69.3%). Adding query rewriting and reranking to hybrid (dense & sparse) retrieval lowers CAR for all 12 models tested. (3) Larger models answer more accurately (Spearman $\rho = +0.60$ with total parameters), but neither model size nor release date shows a significant association with CAR. (4) Without retrieval, enabling thinking on the same weights lowers CAR significantly for 8 of 13 models (by 14% points on average), and the responses shift from acknowledged uncertainty to premise denial.

Recent advances in multimodal large language models (MLLMs) have substantially improved spoken interaction and video understanding. Existing online and streaming benchmarks primarily assess causal multimodal perception through timestamped question answering, whereas real-world interaction requires models to continuously identify task-relevant multimodal cues and decide whether, when, and how to respond. This leaves open how well current models can sustain real-time interaction throughout an evolving multimodal stream. To address this gap, we introduce DIVE, a benchmark for evaluating full-duplex interaction with visual engagement under continuous and realistic multimodal streams. It organizes six tasks into three complementary interaction scenarios. (1) Proactive Intervention evaluates whether models can identify emerging visual or semantic cues and intervene at the appropriate moment. (2) Continuous Engagement evaluates sustained participation through continuous visual tracking and visually grounded query interaction. (3) Interaction Coordination evaluates how models coordinate listening and speaking under overlapping and multi-party interactions. DIVE contains 525 audio/video samples with fine-grained human annotations and precise temporal metadata. Its evaluation protocol separates response quality, participation decisions, and response timing. We evaluate representative real-time models using response quality, window response rates, and reaction latency. Experiments show that leading real-time models achieve sub-second average response latency, yet their average window response rates remain at or below 53.77%. Response quality varies substantially across models and tasks, with adapting ongoing responses to newly arriving information remaining particularly challenging. We hope DIVE facilitates progress toward more natural and capable real-time multimodal interactive systems.

Long-term assistants must follow unfinished tasks across days as information and circumstances change. Finding information or starting an action, however, does not ensure that the task gets completed. We introduce OptionWorld, a benchmark that traces information acquisition, judgments, actions, and outcomes in simulated daily-life tasks to identify where current LLMs fall short as long-term personal assistants. In runs with archived decision inputs, the key facts remain in the model's last input before the deadline in 35.5% of failed instances. Failure analysis identifies the three most frequent problems: needed information is not obtained or combined, available opportunities to act go unused, and agents impose restrictions that the task does not require. Three follow-up experiments show that supplying missing facts restores task success in some cases; action reminders can restart work while subsequent steps still miss the deadline; and correcting a supplied note that wrongly forbids contact restores permitted inquiry. These findings locate shortcomings in maintaining task state: connecting facts to unfinished work, following through before deadlines, and separating actual restrictions from the agent's assumptions.

Recent Omni LLMs (OLLMs) have demonstrated remarkable capabilities in joint audio-visual reasoning, yet cross-modal hallucinations remain a major challenge. Existing mitigation methods mainly operate at the output level, overlooking the entanglement of grounding evidence and learned priors within fused representations. To address this issue, we propose Omni-Woodpecker (OWP), a training-free framework that shifts hallucination mitigation from output-level manipulation to conflict-conditioned representation rebalancing. Specifically, OWP first identifies cross-modal conflicts from complementary modality-specific views. It then employs orthogonal decomposition to separate grounding evidence from competing priors within the hidden states. Finally, OWP adaptively rebalances fused representations by suppressing misleading priors and restoring ignored evidence before decoding. Extensive experiments demonstrate the effectiveness and strong cross-architecture generalization of OWP. Using Qwen2.5-Omni-7B as the backbone, OWP improves performance by 12.9 and 8.6 points on CMM and AVHBench, respectively, while consistently achieving gains of 0.67-2.25 points across four general OLLM reasoning benchmarks. Furthermore, OWP generalizes effectively to VideoLLaMA2-AV, improving performance by 4.6 and 3.9 points on the same benchmarks. Code is publicly available at https://github.com/ZiiJianLiu/Omni-Woodpecker

For AI assistants to move beyond simple chat interfaces and reliably assist users with applications like Gmail, Venmo, and Slack, we need to improve their tool-use capabilities. However, existing tool use benchmarks are still limited and difficult to scale because they rely on human supervision or are too domain-specific. To make data generation scalable, we propose, TURTLE (**T**ool **U**se **R**elated **T**asks with **L**ong-Horizon **E**volution), an Unsupervised Environment Design (UED) approach to evolve existing tasks using a proposer, a solver, and a feedback generator large language model agents. The proposer is rewarded for difficult yet solvable tasks based on the solver's empirical performance, and the feedback generator summarizes the solver's execution traces for the proposer to exploit. TURTLE produces evolving, self-tuning test suites adaptable to different environment requirements, from supporting user simulators and dense document retrieval to interacting across multiple apps. Empirical results on AppWorld, $\tau$-Bench, and AutomationBench using Gemma 4 31B-it show that the evolved tasks target the desired difficulty range (e.g., a 165\% increase in the percentage of hard tasks on AppWorld), increase task diversity (in terms of reasoning and simultaneous tool interactions), and elicit a broader range of failure modes than the original tasks.

Large language model (LLM) agents are moving beyond reactive responses toward proactive assistance that anticipates user needs without explicit requests. However, existing proactive agents primarily address immediate needs, while many everyday situations involve delayed intentions that can only be acted on hours or days later. Humans handle such intentions through prospective memory, the ability to remember to act on an intention when the right moment arises. In this paper, we introduce ProMemAgent, the first proactive agent to incorporate prospective memory over everyday sensory contexts for long-horizon proactive assistance. ProMemAgent organizes past sensory perceptions (e.g., egocentric video and audio) into context-aware memory, generating prospective cues and linking earlier events with subsequent developments for future retrieval. Given the current context, it retrieves relevant past events and performs memory-augmented proactive reasoning to determine whether assistance is needed. We further introduce CPMBench, the first benchmark for evaluating long-horizon proactive assistance over sensory contexts, comprising 1,000 instances across eight everyday categories. Extensive experiments show that ProMemAgent achieves the best memory retrieval performance and consistently outperforms all baselines on long-horizon proactive assistance.

Diffusion large language models (dLLMs) predict multiple token positions per forward pass, yet block-wise decoders serialize generation and repeatedly recompute stable predictions. We attribute this inefficiency to coupling a prediction's readiness for irreversible commitment with its utility as context. We introduce Tentative Context Pipelining (TeCoP), a training-free decoder that decouples these roles through staggered block-parallel pipelining and draft-and-verify. TeCoP reuses stable but uncommitted predictions as provisional cross-block context, with commitment governed by confidence and stability checks. Across four LLaDA and Dream dLLMs on GSM8K, Minerva MATH, and MBPP, TeCoP reduces number of forward evaluations by up to 87% and achieves up to 7.50-fold throughput without KV caching and 11.77-fold with caching over the uncached original decoder. It outperforms Fast-dLLM and DSB in throughput across all 12 model-benchmark settings, with and without caching, while keeping average cached accuracy within one percentage point of the original decoder.

Recursive self-improvement, in which frontier AI models work on training the next generation, is quickly becoming a reality. However, the safety of this process has received far less attention than it should. A misaligned model training its successor can shape what that successor knows, how it behaves, and what it wants, and the harm carries forward into the next generation. We study this risk through the lens of AI control. We introduce RSI Control Arena, a testbed in which a frontier model, acting as an autonomous R\&D agent, is given a benign main task from a stage of a successor model's training pipeline together with a covert side task to install a misaligned property in the successor. The testbed contains 16 side tasks in three classes, forming 22 task pairs with main tasks that span the pipeline from data curation to reinforcement learning. We evaluate six frontier models as agents against LLM monitors that read the transcript of the run and the agent's reasoning. We observe that every model completes a substantial share of the side tasks without degrading the main-task performance, including side tasks that make the successor itself a better saboteur. Under minimal instructions to stay covert, models already produce successful attacks that monitors fail to detect, and providing a short list of evasion tactics further reduces the detection rate. Monitors with access to agent's reasoning detects most attacks, but not all. Our results indicate that recursive self-improvement raises significant safety concerns, and that strong safety effort should be in place before models are trusted with greater responsibility of training their successors.

Large Language Models (LLMs) are increasingly used as research collaborators for hypothesis refinement, methodological planning, and technical analysis across multi-turn interactions. In this setting, a useful assistant must be accurate, methodologically justified, logically consistent, and token-efficient. Current alignment approaches rarely achieve all of these together. Instruction-tuned models gain conversational quality through verbosity, and we show that science-agnostic general collaborative rewards compress responses at the expense of scientific rigor. We propose CollabSciReM, a goal-conditioned reward framework that decomposes the session-level reward into conversational quality, scientific rigor, and structural efficiency. We also introduce SciQuality, a pairwise evaluation of multi-turn research conversations grounded in 300 held-out scientific papers and scored on six dimensions: methodological soundness, justification, logical consistency, factual correctness, novelty, and responsiveness. Fine-tuned on 8.62K synthetic research conversations, CollabSciReM is preferred on SciQuality over a science-agnostic reward baseline trained on the same data for all four Llama and Qwen models, including a 65.3% win rate at matched response length. For the 7B and 8B models, CollabSciReM outperforms the science-agnostic reward baseline on the primary metric in 8 of 10 single- and multi-turn benchmark settings. In a blind study, domain experts rated CollabSciReM significantly higher on novelty than the general-reward variant and reported fewer hallucinated responses than both the general-reward variant and the base model.

Large Language Models (LLMs) are increasingly used for interpersonal advice, potentially shaping how users interpret ambiguous social encounters. Existing Theory of Mind evaluations typically rely on predefined answers, while uncertainty research overlooks LLMs' preferred explanations and confidence under psychological distress. To address these gaps, we introduce MentalLoom, a novel evaluation of LLM mentalisation comprising 604 ambiguous social episodes, matched depression and social anxiety disclosures, and four-turn dialogues with unchanged event evidence. Across eight LLMs, disclosures initially elicited more negative explanations overall but more positive narrator-directed interpretations. Scripted repeated questioning showed greater emotional support, stronger positive narrator-directed endorsement, and higher certainty without new event evidence. These findings show that user context can reshape interpretive preferences and commitment while the underlying ambiguity remains unresolved.

Head-dimension compression can reduce the memory required by the key-value (KV) cache, but existing methods often reconstruct full-dimensional representations during inference, adding significant decoding FLOPs overhead. We formalize the design considerations for converting pretrained attention into a natively reduced-dimensional form that reduces both KV-cache storage and attention computation using a calibration set. Specifically, we consider reductions that (1) apply a shared projection to queries and keys so that they remain in a common reduced feature space, (2) derive this projection jointly from query and key activations to better preserve their interaction, and (3) are compatible with the pretrained RoPE transformation. To obtain a reconstruction-free realization of this reduction, we restrict it to coordinate selection and theoretically show that RoPE compatibility then holds if and only if complete rotary pairs are retained. Building on this characterization, we introduce RACE, a post-training method that selects rotary pairs using their joint query-key activation energy. We theoretically show that this criterion minimizes a separable, calibration-dependent upper bound on post-RoPE query-key product approximation error. The resulting reductions are folded into the model weights and require neither runtime reconstruction nor specialized kernels. Across three model families and different compression levels, RACE outperforms structured pruning and remains competitive with reconstruction-based methods on language-modeling, zero-shot, and long-context tasks. At 50% KV-cache retention and 256K context, RACE reduces total inference FLOPs by 47.8%, yielding 1.79x higher throughput, 43.2% lower time-to-first-token (TTFT), and 28.1% lower allocated GPU memory in comparison to an uncompressed model.

Precisely reproducing a target span verbatim within a long context is a fundamental capability test for LLMs: even when the target string is present in context, models often still drift within the span or fail to stop correctly, revealing limitations in source grounding and boundary control. This capability has broad practical relevance. For instance, code generation often requires embedding sensitive strings such as API keys exactly as they appear in context. In this work, we construct a benchmark to test exact span reproduction ability across contracts, scientific literature, code, and random sequences, and find that current models still make surprisingly frequent errors. To address this, we propose CopyGen, a lightweight copy-constrained plugin for frozen LLMs that turns copying from a byproduct of next-token prediction into a controlled process. Our key idea is to decompose copying into three distinct decisions: when to copy, how far to copy, and when to stop. Concretely, CopyGen first determines whether to enter copy mode, then predicts a tentative copy length, and finally verifies the generated span token by token. During verification, we adaptively adjust the logits of copy continuation tokens to incentivize or suppress continuation, and emit only the longest valid prefix. Across three LLM backbones and four benchmark tasks, CopyGen improves average exact match by 35% relative without model finetuning, and integrates seamlessly with SFT to further improve it by 4%, while substantially accelerating generation. These results suggest that reliable exact copying is not something LLMs obtain for free, but a decoding-time control problem that requires explicit modeling. Code is available at https://github.com/mbzuai-nlp/CopyGen.

For efficient deployment of large language models (LLMs), methods of ultra-low-bit backbones reduce weight storage, memory traffic, and linear-layer execution cost, while multi-token prediction (MTP) methods reduce sequential decoding steps by predicting multiple future tokens in parallel. Although complementary, the direct combination of these methods cannot accelerate decoding at comparable quality: masked drafting lacks intermediate token input supplied in next-token training, while wider computation and causal key-value (KV) updates can offset the savings from fewer steps. To address these challenges, we propose BitMTP, an algorithm-system co-design framework for efficient MTP on ultra-low-bit LLM backbones. A joint next-token prediction and MTP objective equips a single model with both autoregressive (AR) generation and block-causal drafting. Confidence-only decoding, speculative-only decoding, and hybrid decoding provide different quality-throughput operating points without using separate draft and verifier models. Accepted-prefix alignment keeps the committed prefix, materialized causal KV boundary, and CUDA Graph replay state consistent for continued generation. BitMTP uses this contract to select width-specialized low-bit execution and reuse captured graphs over shared KV storage. On NVIDIA H20, BitMTP with hybrid decoding achieves five-task mean speedups of 2.2x on BitNet b1.58-2B and 2.1-2.4x on 1.7B-4B Bonsai models at comparable task quality. On RTX 3090 and Jetson Orin, it further achieves up to 2.0x in five-task mean throughput and 3.3x on MBPP+, respectively. These results demonstrate that BitMTP accelerates ultra-low-bit LLM generation across backbones, model sizes, and GPU types. The source code is released at https://github.com/PolyU-VCLab/BitMTP

Continuous flow language models offer a path to parallel generation with tools for control and acceleration unavailable to discrete diffusion, yet effective techniques for building them at the billion-parameter scale remain less established. We present Lafa (Large flow language model), a fast, high-performing flow language model scaled to 8B parameters in a data-efficient manner. Lafa is a blockwise flow model with causal attention on the prefix and bidirectional, asynchronous flow within each block. This design enables efficient training by adapting pretrained autoregressive (AR) models and fast inference via prefix caching and blockwise parallel decoding. We train Lafa at 0.5B, 7B, and 8B scales with fewer than 30B tokens, more efficient than previous diffusion models, and evaluate it on instruction following, reasoning, and coding, where the learned flow matches or surpasses the quality of AR and can be distilled into flow maps for one- and few-step generation. Notably, flow training preserves or even improves pretrained AR capabilities via a multi-token prediction effect, which we leverage in a self-speculative decoding scheme that verifies flow-based drafts autoregressively within a single model. This yields 2.54x the throughput of an AR model at comparable quality, outperforming multi-token prediction and discrete diffusion methods. Our findings show that continuous flows can serve as a scalable foundation for language modeling.

RTL repair benchmarks should measure whether a model can fix broken RTL given a specification. To this end, we construct and hash-bind RTLRepairDataset, 5,784 (broken, fixed) training pairs; RTLRepairBench, 459 held-out cases with 374 syntactic and 85 behavioral mutations; and an evaluation harness supporting simulation and formal equivalence. We then use this infrastructure to identify two measurement confounds. First, simulation and formal scoring give opposite signs for the apparent effect of providing a specification in one evaluated cohort; reset-aware rescoring attributes almost all of the simulation-formal disagreement to reset and first-cycle comparison semantics. Second, on 274 mined model failures, the specification-only minus specification-conditioned recovery point estimate ranges from -8.5 to +18.3 percentage points across six models. Thus, a conventional repair score need not isolate whether the model used the faulty RTL. We translate these findings into evaluation controls: report initialization, reset, first-cycle, and unknown-value semantics; disclose oracle coverage and undecided cases; and include a specification-only baseline. Together, the resources and audit provide both a benchmark for behavioral repair and a protocol for diagnosing what contributes to its measured score. We release the benchmark, the training set, construction code, sanitized aggregate evidence, and reproduction materials at https://github.com/Evaaaz/RTLRepairBench.

Scientific foundation models increasingly operate on continuous phenomena, yet still use discrete residual computation. We ask whether pretrained residual transformations can support additional latent computation without expanding the backbone. To test this possibility, we introduce Continuum-LM, which reuses these transformations as dynamical operators in a continuous-depth language model. Its hidden states evolve through an ordinary differential equation evaluated by higher-order numerical integration, with solver order and step size controlling computation. Across eight scientific tasks, Continuum-LM trains stably and improves mean exact-match accuracy by 7.77 and 10.96 percentage points over discrete counterparts at the 0.6B and 1.7B scales, respectively. At 0.6B, it also outperforms an Ouro-style recurrent control under matched parameter and residual-function evaluation budgets. Together, these results support extensible computational depth through parameter reuse, identifying continuous hidden-state evolution as a practical design dimension for scientific language models beyond parameter scaling.

LLM agents act on the external world through tools and third-party skills, so a single executed action can cause irreversible harm. Existing defenses judge a proposed action without tracking whether its content comes from the user or from external sources such as tool outputs, and they judge it in isolation rather than against an available alternative, which forces a trade-off between security and utility. We present PROVISO, a controller that decides which action to commit after the agent proposes an action and before it is executed. PROVISO encodes trusted observations, external observations, the proposed action, and an alternative into provenance-typed memory, relates them through role-specific queries, and commits the action with the highest calibrated value learned from outcomes. On four agent security tasks spanning the agent lifecycle, no baseline exceeds PROVISO in both security and utility. PROVISO improves F1 over the strongest baseline by 13.84 points on SkillsBench and by 7.07 points on ToolSafe, and it reaches 88-99 F1 on three additional skill detection benchmarks.

Platforms that host young users screen messages with commercial text-safety APIs, but independent evidence on how often such systems miss harm or flag harmless text is scarce, especially outside English, and the documentation of the one studied here leaves its false-positive rate unquantified. We evaluated one such API, Tuteliq Detect as deployed under the release identifier tuteliq-detect-v5.20261011 (not the separately published v6 model), on six harm categories: authorised push-payment fraud, money-mule recruitment, social engineering, romance scams, gambling harm and bullying. We built SITO (Synthetic Independent Text-safety benchmark with Overflagging tests; sito is Ukrainian for 'sieve'), 2,332 synthetic cases in English, Ukrainian and mixed Ukrainian, with each harmful case matched by hard harmless ones: same-topic messages, minimal-edit contrasts, warnings and retrospective accounts. On English text the system flagged 84.9-94.1% of harmful cases and 0-4.3% of harmless ones per category. It missed third-party reports of ongoing harm far more often than perpetrators' messages (63.6% against 95.1% recall, all languages), when sent without the settings the API offers for reports of harm. Ukrainian and mixed-language input was detected at least as often as English, recall was lower on text from a held-out generator family, so headline recall may be optimistic, and 3.2% of cases received a different decision in repeated runs. A platform should measure false positives on its own traffic and check how bystander reports are handled. The datasets, generation prompts and code are released.

Medical workflows often require multiple interdependent analyses, but existing specialist and generalist AI models are not designed to coordinate them. We introduce MMedAgent-Plus, a multimodal medical tool-use agent that combines a vision-language planner with 18 leading specialist models spanning 12 tasks and 8 modalities. Through a unified tool interface, the planner selects specialists, executes them in sequence, and synthesizes responses grounded in their outputs. The planner is instruction-tuned on 208,928 samples from 32 public medical benchmarks, covering single- and multi-tool interactions across single and multiple rounds. We separately evaluate tool orchestration, response synthesis, and end-to-end execution on held-out samples. MMedAgent-Plus achieves tool-call accuracies of 91.0%, 98.1%, and 80.7% in single-tool multi-round, multi-tool multi-round, and multi-tool single-round settings, respectively. Given correct tool outputs, MMedAgent-Plus achieves scores of 14.4-14.8 out of 15 for faithfulness, completeness, and conciseness across 600 evaluation samples. In end-to-end execution without pre-supplied tool outputs, MMedAgent-Plus correctly invokes the required specialist in 90.6% of 350 evaluation cases spanning seven tasks. Its final-answer accuracy reaches 0.98, matching the dedicated leading specialist and exceeding the strongest baseline (0.28).

Hallucination stress tests can yield similar average scores across language models even when different prompts elicit hallucinations in each model. We develop a procedure that iteratively edits source-grounded seed prompts to expose hallucinations in a target model. An LLM editor revises each prompt in response to the target's answer and a judge's category-specific score and explanation. In the main setting, the target also receives feedback on its previous response. Experiments cover 500 seeds from seven datasets and ten topics, with intrinsic and extrinsic hallucinations evaluated across open-weight and closed-source models. Editing substantially increases judge-rated severity over the seeds. In the baseline configuration, retained responses fall in the judge's strong-hallucination bin in 77.2% of runs targeting extrinsic hallucination and 74.2% targeting intrinsic hallucination. Later editing rounds account for more than half of the mean score increase. Among 219 prompts that pass model-assisted source-validity checks, 89 elicit responses containing unsupported claims. Across five models, 200 curated prompts yield similar average severity scores, although pairwise Spearman correlations between prompt-level scores are only $\rho=0.0387$ to $0.1897$. Curated prompts can be reused across models, but their diagnostic value depends on source validity and the failures they expose in each model.

Coding agents can answer analytical questions over a data lake by discovering, cleaning, and joining the relevant data. However, they discard this work after each question and redo it from scratch for the next one. Tool evolution, which automatically builds and refines reusable tools across tasks, is a natural fit for this repetition. Since a ground-truth label for a data-lake question requires an expert to solve it first, we consider the unsupervised setting. We propose TEAL (Tool Evolution for Analytics over data Lakes), which evolves a library of tools and usage skills for a given Worker agent without labels. Two agents propose tools from complementary evidence, an Explorer from the data lake itself and a Critic from Worker trajectories. A proposed change is kept only if the Worker adopts the tool on a rerun without extra steps and any changed answer is supported by trajectory evidence. We also introduce TEAL-Q, which generates synthetic questions for tool evolution when real ones are unavailable. We evaluate on KramaBench and LakeQA with four Worker LLMs under step budgets, which cap the number of steps the Worker may take per question. With a tight budget of 10 or 15 steps, TEAL improves every Worker on both benchmarks, lifting Sonnet-5 at 15 steps by 9.3 points on KramaBench and 21.7 points on LakeQA. A tighter budget also yields better feedback for evolution. For example, Sonnet-5's LakeQA library evolved at 20 steps and deployed unchanged at 30 steps scores 68.3, above both the baseline (59.6) and a library evolved directly at 30 steps (64.2). Additionally, a library evolved on synthetic questions performs comparably to one evolved on real questions. These results show that useful tools can be evolved without labels, and hopefully inspire future work on what to evolve and how to obtain feedback in unsupervised agent evolution.

Although large language models support a wide range of tasks, hallucinations remain a major obstacle to user trust. When a model generates an answer, both its internal processing and final output vary with the question. Internal states are often expected to offer richer signals for hallucination detection. However, studies using these states typically report their standalone discrimination performance, while output-side comparisons often use simple baselines with weaker modeling capacity. We integrate output-side information available at inference time into a combined baseline and compare detection performance before and after adding hidden states under identical data splits and evaluation rules. Experiments cover open-ended generation and constrained multiple-choice tasks across different models. Adding hidden states improves AUROC by approximately 0.004 and 0.039 on the two tasks, respectively. Hidden states thus provide additional predictive information, but strong standalone detection performance can still yield only limited additional gains.

Multi-agent systems are increasingly deployed to coordinate multiple large language model (LLM) agents. However, this also creates new attack surfaces: malicious agents may have chances to exploit the coordination. In this paper, we propose and study the \textit{show-up attack}, in which a malicious subagent manipulates the supervisor into involving it in tasks beyond its assigned scope. To evaluate this risk, we propose SUA-Arena, which implements 6 attack strategies and contains 188 test cases. Then, we evaluate 6 frontier LLMs and find that all of them are vulnerable to show-up attacks. For example, GPT-5.6-Sol as a supervisor suffers an ASR of up to 0.93. Mitigation strategies such as prompting and restricted communication fail to reduce the risk consistently. These results show that multi-agent systems need explicit mechanisms for enforcing authority boundaries.

Large language models are expected to estimate their own uncertainty, with confidence signals used to guide retrieval, adaptive compute allocation, routing, and selective abstention. We investigate whether confidence signals measured at different temporal stages of problem solving encode distinct forms of epistemic information and remain reliable beyond English-centric settings. Inspired by classical meta-memory theory, we distinguish anticipatory pre-solve confidence (Feeling-of-Knowing, FOK) from retrospective post-solve confidence (Judgement-of-Learning, JOL). We evaluate these signals across two frontier LLMs, multiple reasoning domains, translated Marathi benchmarks, and a newly constructed benchmark of genuinely local Marathi factual questions. Across settings, the cheaper pre-solve signal remains weakly calibrated, while post-solve confidence generally provides stronger epistemic alignment despite its additional inference cost. Confidence behaviour transfers surprisingly well under direct language translation, with calibration trends largely preserved between English and Marathi, but degrades substantially on local Marathi knowledge, suggesting that linguistic transfer alone is insufficient for region-specific information needs. Finally, hallucinated responses maintain persistently elevated confidence after answer generation, highlighting important limitations of current LLM self-assessment mechanisms.

Speculative decoding accelerates autoregressive (AR) language models by letting a small model draft tokens serially and a large model verify them in parallel. We reverse these roles. Starting from a standard AR checkpoint, we fine-tune the large model into a proposer that predicts a block of future tokens in a single forward pass, then use one additional forward pass to verify the proposed block. This gives ProVe, Proposer-Verifier decoding. The verifier can be a smaller AR model or the proposer itself; the latter gives ProVe-lossless, which reproduces the proposer's greedy AR decoding. Block proposals already provide substantial room for acceleration: with a block size of 16, the proposed block has an average AR-matching prefix of 6.6 tokens on GSM8K. Across six benchmarks with a 32B proposer, ProVe-lossless reaches 2.5x average AR throughput without training a separate draft head. With a small 3B verifier, ProVe reaches 2.7x average throughput at 2.6 accuracy points below AR on the six-task average, providing a controllable tradeoff between speed and accuracy.

Negation remains a persistent challenge for language models across NLP tasks and modalities. We show that this weakness can undermine LLM safety guardrails through a structural vulnerability we term polarity blindness: the failure to recognize harmful intent after polarity-altering linguistic transformations. We study two forms of polarity inversion: grammatical negation and antonym replacement, and use synonym replacement as a polarity-preserving control. We further test whether an adversarial textual wrapper amplifies these failures. Using a 260-prompt benchmark across 10 state-of-the-art LLMs and 3 languages (English, Spanish, Czech), stratified by harmfulness potential and evaluated on a validated 0-5 harm scale, we find that polarity blindness affects most evaluated models, with mean harm scores reaching 4.12 for the most vulnerable model. The effect persists across typologically differing negation systems, suggesting that polarity blindness reflects a broader limitation in semantic robustness rather than a language- or grammar-specific failure.

Recent advances in large (vision) language models have enabled computer-use agents that automate digital workflows via fine-grained keyboard and mouse operations. However, such agents remain fragile in long-horizon settings, where errors accumulate and executed actions leave unintended side effects, making the ability to recover from failures critical. Yet, existing remedies remain inadequate: environment resets are infeasible in proprietary settings, action replay is prohibitively expensive over long horizons, and API-based rollback is limited to web environments, leaving model-based recovery largely unexplored. To address this gap, we present ReTrace, a scalable evaluation framework for systematically studying recovery behavior in computer-use agents. We formalize Transition-Consistent Recovery: identical downstream actions from the recovered and target states must yield identical outcomes. We operationalize this criterion through goal-oriented action replay following controlled error injection. Using ReTrace, we present the first systematic study of autonomous, model-based recovery, revealing that (i) ideal recovery can unlock model's latent task completion capabilities, (ii) current computer-use agents fail to exploit their latent recovery capabilities, and (iii) accurate model-based recovery yields measurable end-to-end gains. Together, these findings establish recovery as a fundamental capability for building reliable computer-use agents. Our code and dataset is available at https://github.com/BangProx/ReTRACE.

Large Language Model (LLM) agents can handle increasingly complex tasks, navigating long trajectories of observations and actions. As these trajectories grow, key-value (KV) cache I/O and storage become the dominant cost. Trajectories can also exceed the LLM's maximum context length, requiring control over active cache growth. Prior work uses context engineering to reduce cache size by trimming or summarizing history. However, modifying the trajectory prefix invalidates subsequent cached states and triggers costly re-prefilling of retained context. Task accuracy can also suffer when too little context is retained to limit re-prefill costs. We propose cache engineering as a new scheme. It allows agents to directly control their active KV cache without modifying the trajectory or re-prefilling retained context. Specifically, agents can proactively use drop to exclude selected cache from future attention and repos to adjust retained keys' position encodings. We formalize these cache operations and identify history dependence as a serving challenge: identical token sequences can yield different cache states under different cache operation sequences, making prefix matching insufficient for correct cache reuse. Thus, we co-design the AKV serving engine. It provides a stateless interface compatible with existing serving APIs, while internally managing stateful cache reuse, eviction, and restoration. Experiments show that a tool-response rolling-drop cache policy improves average benchmark scores by 1.7 points and reduces average cost by 27.7%, compared with conventional threshold-triggered summarization. With AKV, it achieves up to 1.2x the output throughput and a 1.5x speedup in P95 request latency. Recursive self-improvement of cache policies further reduces cost by 7.1% over expert design. Our code is at https://github.com/thu-nics/AKV.

Current evaluation of LLM-generated debate largely focuses on argumentative quality, leaving the value structure of generated arguments underexamined. We introduce a diagnostic metric framework for comparing how values are expressed and organized in human-reference and AI-generated debates. The framework separates value expression, the range of recognizable values a debate invokes, from value organization, how those values are distributed and sustained across turns. Grounded in Schwartz's Basic-10 value theory, it represents debates as multi-level value distributions over argumentative units, utterances, stances, and full debates, enabling metrics for value coverage, alignment, diffusion, and cross-turn dynamics. Across multiple generation systems, AI-generated debates are broadly value-expressive, often activating as many or more value dimensions than matched human references, but they organize these values differently. Their utterances are more locally value-diffuse, and source-blind judges distinguish value clarity from value focus. These findings show that debate evaluation should consider not only value presence, but also value organization. Code and supplementary materials are available at https://github.com/Maoyuan-li/Value-Expression-and-Organization.

Multi-token prediction (MTP) can accelerate language-model decoding by drafting several tokens and verifying them in a single target-model pass. However, the best speculation horizon changes with context: longer drafts offer greater amortization when reliable, but can waste compute on suffixes discarded after a rejection. Unlike autoregressive drafting, where a stopping rule can decide token by token, MTP must fix the horizon before drafting, as all heads fire in the same forward pass. We therefore formulate horizon selection as an online control problem, using verification feedback to learn how far the draft can be trusted in each context, and introduce Adaptive Multi-Token Prediction (AMTP): a lightweight controller that requires no retraining and leaves the exact verifier unchanged. To guide horizon selection, we characterize the expected number of tokens committed at each horizon via acceptance survival, the probability that a draft prefix is fully accepted. Before each round, AMTP predicts how far the draft is likely to survive and picks the horizon with the highest expected tokens per second, learning freely from every verification pass. Theoretically, we show that any online learner's prediction error translates into throughput regret against the best horizon for each context, and in turn into a bound on wall-clock completion time. With online gradient descent, this regret is sublinear. Across five draft-target pairings built on Llama-3.1-8B-Instruct with Medusa heads, Meta's MTP models, and Qwen3-4B-Inst-2507-MTP, evaluated on AIME 2025 and GSM8K, AMTP decodes up to 4.1x faster than standard decoding and matches or outperforms the fixed window in 8 of 10 settings, with throughput gains of up to 10%.

Conversational recommender systems have recently adopted vision-language models to match user dialogue intents with multimodal product information. However, extending these methods to many product domains faces a dilemma: deploying a separately adapted VLM per domain scales cost linearly with the number of domains and requires routing each dialogue to the right model, while naively merging all domains into a single VLM degrades accuracy because product characteristics vary widely across categories. We propose LightCRS, a lightweight, query-conditioned framework for unified multi-domain multimodal conversational recommendation. LightCRS precomputes multimodal product representations once offline with frozen encoders. Online, a compact 5.3M-parameter module fuses this evidence for each dialogue via query-conditioned cross-attention, whose parameters are shared across all domains, and compares candidates jointly with an inter-candidate Transformer. To enable rigorous evaluation, we construct MM-R2A-25, a 25-domain benchmark pairing natural Reddit recommendation-seeking conversations with Amazon product listings. LightCRS achieves state-of-the-art average HR@1, NDCG@5, and MRR, improving HR@1 by 35.4\% over the strongest baseline while reducing per-query latency by 266 times. Joint multi-domain training itself improves HR@1 by 15.2\% over separately trained single-domain models.

Reasoning structures (e.g., chains and graphs) are increasingly used for probabilistic inference, factuality verification, and uncertainty estimation, but their structural reliability is rarely evaluated directly. We test whether this reliability holds through a proposed two-stage diagnostic evaluation in open-ended multi-hop QA. Stage 1 tests different capabilities required to construct reasoning graphs, showing that graph reliability varies with topology, prompt design, and model family. Stage 2 uses uncertainty estimation as a downstream test case. Generated-edge experiments show that graph-based computation is sensitive to upstream graph errors, with performance dropping relative to the gold-graph setting. Gold-graph experiments further show that graph structure provides little benefit when used only as prompt context, but improves the separation of correct from incorrect answers by confidence score and enables node-level diagnostics.

Large language models are deployed globally, yet safety alignment and evaluation remain concentrated in English: harmful requests refused in English can elicit unsafe completions when translated into low-resource languages. We evaluate ten widely deployed LLMs across 79 languages (English plus 78 low-resource) under a translation-bypass threat model, using a metric conditioned on English safety, which separates translation bypass from unsafety already present in English. English safety does not transfer reliably: conditioned on safe English completions, mean bypass rates range from 21.9% to 44.4%. Bypass risk concentrates in a subset of language families, a pattern broadly shared across models. The shared ordering is confounded with translation degeneracy, but its variation across models is significant and survives every control we apply. We also show that per-model bypass rates do not identify a stable ranking: nine of ten models change rank across three conditioning schemes that estimate the same quantity, with no change to any underlying response.

Block-diffusion speculative decoding accelerates large language model inference by drafting multiple tokens in parallel. However, one-step drafting in discrete diffusion is limited by token factorization error: each token is predicted from the same masked block, without conditioning on the tokens selected at other positions in that step. This can produce inconsistent drafts and reduce acceptance length. Iterative refinement mitigates this limitation by conditioning later predictions on previously committed tokens, though additional passes increase drafting cost. To address this trade-off between draft fidelity and inference latency, we introduce DStill, a two-stage framework that distills a target-conditioned drafter's own multi-step refinement into one-step drafting. We first train the drafter to complete partially revealed blocks, then use its refinement trajectories to train a student to predict the final completion in one pass. Using a curriculum that gradually transitions from partial to fully masked inputs, we successfully transfer the multi-step refinement capabilities into one step. Across reasoning, math, code benchmarks on three model families DStill improves acceptance length by up to 28% and speedup by up to 25% over one-step DFlash, achieving up to 6.45x speedup over auto-regressive decoding.

Large language models (LLMs) can influence user trust and decision-making through subtle linguistic cues, yet the behavioral effects of manipulative outputs remain poorly quantified. We introduce a behavioral computing framework for measuring \textit{behavioral vulnerability}, defined as users' susceptibility to manipulative AI responses. Building on DarkPatterns-LLM benchmark, we generate controlled manipulative stimuli using Manipulation Gradient Descent (MGD), targeting biases associated with authority, urgency, and social proof. We analyze participant responses using the Multi-Stakeholder Impact Assessment Network (MSIAN) to model cross-user influence and the Temporal Harm Propagation Model (THPM) to characterize the persistence of behavioral effects over time. In a study involving 120 participants, manipulative framing produced significant changes in trust calibration and compliance ($p < 0.01$), with effects also propagating across social contexts. These findings highlight the need to extend LLM safety evaluation beyond model outputs to measurable behavioral consequences and provide a computational framework for quantifying individual, social, and temporal risks arising from manipulative AI interactions.

Diffusion language models (DLMs) are emerging as a promising paradigm for code generation, offering the potential for fast, parallel decoding. However, generating complex programs remains challenging, as their interdependent components must maintain consistent interfaces, data flows, and shared assumptions. Aggressive parallel decoding can compromise these dependencies and undermine program correctness, while conservative decoding sacrifices efficiency. To address this trade-off, we propose planning program decomposition before implementation to coordinate interdependent units during parallel generation. We introduce Self-Decomposing Diffusion for Code (SDDC), a novel framework that decomposes complex programs and implements their units in parallel, entirely through diffusion. SDDC first generates a contract specifying the program's decomposition and inter-unit dependencies, then uses it to guide coordinated parallel implementation. Across five benchmarks and two 7B diffusion backbones, SDDC substantially improves the accuracy--latency trade-off, raising pass@1 by up to 31.1 percentage points while decoding 7.5--13.7$\times$ faster than the base models, and outperforming a fine-tuning baseline by up to 23.5 points with lower latency.

Identifying recipe entities and linking them to ontology concepts is fundamental for semantic recipe understanding. However, food mentions in recipes are often ambiguous, context-dependent, and lexically diverse, making ontology entity linking (EL) challenging. We propose SCODEL, a streamlined, model-agnostic approach based on a central idea: contextual food mentions and ontology concepts that refer to the same entity should learn the same canonical ontology-oriented definition. SCODEL uses these shared definitions as a semantic normalisation layer, providing complementary semantic information to existing EL models without requiring a new linking architecture. We further introduce ReciLinkD, the first benchmark for contextual food ontology entity linking in recipes. Experiments show that SCODEL consistently improves three modular EL models (ReFinED, GSM-EL, and BLINK): SCODEL x BLINK achieves the highest Hit@1 on seen concepts (88.33%), while SCODEL x GSM-EL achieves the highest Hit@1 on unseen concepts (83.15%), outperforming the strongest large language model (LLM) based EL baseline in each setting. Experiments on a biomedical ontology benchmark further demonstrate that SCODEL generalises beyond the food domain. Finally, we use SCODEL to ground food mentions at scale and construct an ontology-grounded recipe knowledge graph, demonstrating its utility for ontology-aware recipe retrieval and corpus-level food analysis.

Training LLM agents for policy-constrained tool environments, such as an airline backend with booking, refund, and cancellation policies, requires tasks that exercise the environment's rules individually and in combination. Existing synthesis pipelines let a model propose tasks and retain only those it immediately solves, biasing training data toward easier behaviors. We introduce Gradus (Graded Rungs of Ascending Difficulty from an Underlying System), which decouples what a task should teach from the concrete task and solution used to teach it. Without seed tasks, Gradus treats a fixed environment as a latent syllabus, extracting guarded behavioral operators that capture how actions depend on current conditions and how their effects reshape subsequent decisions. It composes these operators into increasingly demanding tasks while using corpus-level feedback to target missing outcome combinations. Crucially, Gradus jointly calibrates tasks and solutions against the environment: execution feedback identifies which side to refine while preserving the intended behavioral requirements. Verification thus becomes a recovery signal rather than a terminal filter, coupling coverage-guided exploration with the preservation of difficult examples. From off-the-shelf backbones, the resulting curriculum lifts a 30B model from 59% to 89% on $\tau^2$-Bench and from 47% to 83% on CAR-Bench with an Opus 4.5 proposer, and to 76% and 62%, respectively, with an open-weight Qwen3-235B proposer. These results show the value of jointly controlling which behaviors are explored and which survive into verified training data.

Scaling the training of Large Language Model (LLM) agents typically requires a large collection of executable environments, whose construction, maintenance, and operation constrain the expansion of training experience. Existing offline trajectories offer an opportunity to reuse interaction experience, yet effectively refining the current policy without revisiting these environments remains challenging: supervised fine-tuning (SFT) can lead to catastrophic forgetting, while online reinforcement learning requires real-time interaction with the environment. To address this challenge, we propose Step-Advantage Policy Distillation (SAPD), a method for learning from offline trajectories using relative step advantages. SAPD decomposes multi-turn trajectories into individual decision contexts for on-policy distillation, allowing the student to generate responses conditioned on recorded histories, and introduces relative step advantage signals to stabilize training. The entire process requires neither new environment observations nor additional multi-turn interactions. Experiments on ALFWorld, WebShop, and Search-based QA show that SAPD improves performance across different teacher configurations, mitigates catastrophic forgetting compared with SFT, and runs at least $4.7\times$ faster per training step than online OPD. We will release our code soon.

Telecom language models must combine broad networking knowledge with specialist behavior across routing, broadband services, access technologies, and operational diagnostics while remaining practical to deploy as a single model. We present an end-to-end framework that builds source-grounded synthetic training data from structured networking taxonomies and expert graphs, materializes 212,559 routed training records across fourteen networking routes, trains six frozen specialist LoRA teachers over one shared continued-pretrained backbone, and consolidates their behavior into one deployable student adapter using Routed Specialist On-Policy Distillation (RSOD). In RSOD, the committed student generates each trajectory, the route-assigned specialist teacher-forces the exact native token sequence, compact teacher hidden states are reconstructed through the shared frozen output head, and only the student LoRA is updated. A controlled reconstruction validation shows mean absolute logit error of 7.26 x 10^-6 with 100% top-1 agreement between native and reconstructed teacher logits. On four RTX 6000 Ada GPUs, the completed 2,000-step run produced 16,000 trajectories with zero observed student-version or token- ID mismatches. Under the primary route-conditioned evaluation over 2,850 held-out questions, the raw Meta-Llama-3.1-8B-Instruct model scores 12.8661/25 micro (13.0172/25 macro), the combined routed six-teacher system scores 21.7641/25 micro, and the consolidated RSOD student, released as Llama-CalixCLM v1, reaches 21.7112/25 micro (21.8006/25 macro) on the same five dimension LLM-judge rubric. Under a separate neutral deployment-like prompt, the same student reaches 20.0028/25 micro (20.0778/25 macro). The consolidated student remains within 0.0529 micro-score points of the combined routed-teacher system while retaining a numerically validated and auditable train-serve execution path.

Long-horizon deep search requires LLM agents to iteratively decompose complex information needs, gather evidence from external sources, and maintain a reliable research state under multiple constraints. A central challenge is that early mistakes can be amplified over the search trajectory: agents may anchor on an incorrect candidate, collect partially supporting evidence, overlook contradictions, and accumulate errors that misguide subsequent reasoning. We propose FALCA, a falsification-aware framework that formulates deep search as a propose-falsify process. FALCA adopts a collaborative Searcher-Falsifier architecture, where the Searcher proposes candidate answers or sub-answers, while independent Falsifiers actively seek counterevidence, check unmet constraints, identify conflicts, and produce concise falsification reports for research-state revision. To strengthen this loop, we introduce falsification-aware training, combining SFT warmup on falsification trajectories with agentic RL based on parallel rollouts and hierarchical credit assignment over research stages, state transitions, and actions. Experiments on five search benchmarks show that FALCA improves long-horizon deep search performance over backbone and harness ablations, while reducing confirmation bias and error accumulation through more effective falsification and research-state revision.

Large Language Model inference over long contexts is increasingly bottlenecked by the cost of attending to a growing Key-Value (KV) cache. Sparse attention reduces this cost by accessing only a fraction of the cache, but existing methods face a trade-off between deterministic retrieval, which captures high-scoring tokens but discards residual attention mass, and sampling, which improves coverage at the cost of higher variance. Hybrid methods combine top-$k$ retrieval with uniform sampling, yet still exhibit a substantial gap from dense attention at aggressive sparsity. We introduce \emph{Stratified Sparse Attention (SSA)}, a unified framework combining deterministic retrieval with score-guided sampling. SSA deterministically attends to high-scoring tokens while sampling the residual using proxy attention scores and correcting for their inclusion probabilities, thereby recovering attention mass beyond the retrieval set. SSA encompasses top-$k$, top-$p$, and hybrid top-$k$+sampling as special cases. Across RULER, LongBench, LOFT, and AIME, and five models ranging from 3B to 27B parameters, SSA consistently improves over the baselines at matched sparsity. On RULER-HARD with Ministral-3-8B, SSA improves accuracy over the strongest practical baseline from 81.8\% to 84.2\% at 32K and from 63.8\% to 69.3\% at 128K under $50\times$ attention sparsity, recovering 94\% and 85\% of dense-attention performance, respectively. With optimized CUDA kernels, SSA achieves up to $2.96\times$ decode speedup over FlashInfer at 1M context, demonstrating that its accuracy gains translate into practical long-context inference acceleration. We open-source our code at https://github.com/skylight-org/SSA-Paper-Code.

Computer-use agents are limited by the interfaces humans give them. An agent may type a spreadsheet cell by cell when a single line of code would do, or print all two thousand lines of a file through a shell to find one function. Both the interfaces and the knowledge of when to use each are built by hand, which cannot keep pace with new applications, workflows, and models. We argue that interface design should itself be an agent capability. We introduce Agent S4, an autonomous computer-use scientist that builds Evolving Agent Interfaces from task experience. A meta-agent examines task agents' trajectories on synthetic training tasks, without success labels, and constructs and revises interfaces and usage knowledge in a repository. Population-based evolution then selects which repositories to develop further by their performance on validation tasks the meta-agent never sees, favoring changes that generalize beyond the seen tasks. On OSWorld V2, Agent S4, with Qwen3.8-27B and the evolved repository of interfaces, sets a new state of the art in open-source computer-use agents, reaching 21.3% binary score, 3.8 times the same model with fixed interfaces. This is above both GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%), at about 4.3% and 1.5% of their respective costs. On ScienceBoard, the evolved interfaces improve the success rate from 41.4% to 58.6% with the same model. These results suggest that useful interfaces can be evolved from experience, making autonomous interface design a source of improvement complementary to advances in models and harnesses. We open-source our code at https://github.com/simular-ai/agent-s.