2026-10-12

Language models have been adopted faster than evaluation practices that reliably assess their risks, and a growing share of use is affective rather than informational. Within that traffic users disclose suicide and self-harm (SSH) crises. Existing safety benchmarks score models primarily on refusal, but refusal is not the desired behavior here; the conversation requires a substantive supportive response. We introduce DistressBench, 718 English conversations with clinician-authored user turns spanning 23 SSH subcategories, each paired with a task-specific rubric of weighted atomic criteria across seven dimensions of crisis support and a clinician reference response. A model's score is the weighted average over the criteria its reply satisfies, reading directly as the fraction of required care delivered. The best of 25 frontier models satisfies 88.4% of weighted criteria and the median model 71.4%, against 98.7% for clinician references scored on the same criteria. Pooled across all 25 models, in 35.3% of conversations where a model satisfied every criterion for recognizing the crisis, it failed to route the user toward human help. The shortfall concentrates in the dimensions requiring the model to act rather than withhold. Contrasting each dimension's median model against its best separates capability limits from choices: de-escalation reaches 74.6% in the strongest model against a median of 46.0%, and safety disclaimers 84.6% against 8.3%. Multi-turn conversations score below single-turn ones (68.5% against 74.2%). We release the corpus, rubrics, judge calibration labels, and a companion annex of 20 long-horizon red-team transcripts. These results argue for evaluating crisis safety on the care a model delivers rather than on what it declines to say, and for treating de-escalation and resource routing as deployment requirements.

This paper introduces MathVis, a benchmark designed to evaluate LLMs' ability to visualize frontier mathematical problems and perform diagram-assisted reasoning. The benchmark consists of 512 tasks, each comprising a mathematical statement, a reference diagram, executable TikZ or SVG code, a Gold Intermediate Representation (IR), atomic rubrics, and error codes. Through rigorous evaluation, GPT-6-Astra, Claude Fable 5, and Gemini 3.1 Pro Preview achieve pass@1 scores of 31.25%, 26.56%, and 22.85%, respectively. We found that although the tasks require visualization, LLMs produce outputs dominated by formal languages, symbols, and graphics programs, rather than explicit diagram-mediated reasoning. Furthermore, while the solutions generated by LLMs may contain all the necessary elements, they are often visually confusing to humans. Ultimately, this benchmark aims to visualize mathematical problems and promote research on enabling LLMs to simplify their reasoning via visual diagrams.

We present SmartSalesBench, the first multi-turn conversational benchmark explicitly designed to evaluate large language model (LLM) sales agents against psychologically grounded, resistant buyer personas. Existing task-oriented dialogue benchmarks assume cooperative users and measure task completion; SmartSalesBench inverts this assumption, placing agents against personas programmed to resist and challenge across conversations of up to twenty turns. The benchmark includes five agent profiles, twelve buyer personas, fifteen smartphone products across six price tiers, and four controlled experiments with 3 iterations totaling 13,680 sessions. Both sales-agent and buyer persona generations are grounded through retrieval-augmented generation (RAG). We introduce nine evaluation metrics grouped into three tiers: primary outcome metrics (CR, BRS, RIR), secondary quality metrics (SMS, PFS, ETI), and exploratory metrics planned for future reporting (ITS, ARI, SPI). Our evaluation combines automated LLM-as-judge scoring with internal human assessment (n = 50, \k{appa} = 0.71, Pearson r = 0.81 with BRS).

Counterfactual text editing is the ability to revise a given text in light of a hypothetical change while preserving facts not causally affected by that change. When asked what would have happened if a doctor had not recommended surgery, the model output should retain the patient's medical history and modify only the consequences of the decision. This requires causal understanding: a model must determine, without being explicitly instructed to, what events are affected by the hypothetical change and what should remain fixed. This skill is central to human reasoning, especially in safety-critical domains like healthcare or law. By contrast, the task of associational text generation merely generates text statistically associated with the altered condition. In the medical example, an associational response might make the patient's symptoms less severe because milder cases correlate with no surgery. Such responses produce a plausible scenario, but not the counterfactual version of the given one. Prior counterfactual-text datasets do not clearly separate these modes of reasoning. We introduce the Counterfactual Benchmark (CoBe), a testbed for models' zero-shot ability to counterfactually edit a given textual scenario when prompted to do so in a conversationally natural way. CoBe contains 2000 questions spanning healthcare, science and engineering, business, and everyday scenarios. We test a suite of frontier and smaller models, with frontier systems achieving an average score of $58.05\%$. Our findings indicate that models frequently default to associational generation rather than preserving the causal structure of the original scenario. Moreover, performance does not scale consistently with either model size or output token length, suggesting that the failure may reflect model training objectives rather than insufficient model capacity.

As large language models (LLMs) acquire advanced biological design capabilities, they offer substantial research benefits but simultaneously introduce severe dual-use risks, such as engineering enhanced pathogens or drug-resistant bacteria. While biosecurity frameworks recognize these threats, the actual risk of LLMs executing dangerous biological modifications remains unquantified, particularly when augmented with specialized biological design tools (BDTs). Existing evaluations primarily assess hazardous knowledge possession rather than practical application, overlook risk amplification from BDTs, and rely on unreliable, expensive, or hazardous validation methods. To address these gaps, we introduce DRYLAB-Bench (Design Risk Yielded by LLM Agent in Biodesign), the first comprehensive benchmark evaluating the high-risk biological mutation design capabilities of LLM agents. Moving beyond basic question-answering, DRYLAB-Bench tasks agents with 30 genuinely hazardous mutation-design scenarios distilled from biosecurity guidelines. It evaluates models in a multi-round, tool-augmented environment where LLMs orchestrate BDTs to iteratively refine candidates. Crucially, it scores designs directly against 187,695 empirical wet-lab measurements from ProteinGym and ViroGym using our proposed Biosecurity Risk Score (BRS), quantifying risk-associated effects while enabling scalable and safe ground-truth verification. Evaluations of 15 frontier models reveal substantial baseline threats: without tools, 67.6% of one-shot designs exceed the average measured variant, and 21.4% reach the highest-risk decile. Equipping models with BDTs significantly amplifies this risk, raising the pooled risk by \(\Delta BRS=+0.116\) and the top-decile share to 23.5%. Code: https://github.com/drylab-bio/DRYLAB-Bench

Static coding benchmarks are fixed at release. Over time their tasks leak into training data, and stronger agents eventually saturate them. We argue that a benchmark should instead aim to be self-scaling: its tasks and oracles are created at evaluation time, so no content exists in advance to leak, and its difficulty is set by the participating agents rather than fixed at release. We introduce SWE-Duel, an arena in which two coding agents evaluate each other over mirrored Red-Blue turns. In each turn, the Red agent works on a real repository across three context-isolated sessions. It proposes and implements a new feature with feature tests, then embeds a subtle bug in that feature along with sealed bug-triggering tests. The resulting pull request is admitted only if (1) it passes the feature tests and the repository's existing tests while failing exactly the bug tests, and (2) a fresh session of the same agent can repair the bug blind, which rules out unsolvable tasks. The Blue agent then reviews the pull request without being told a bug exists. It wins only if its revision passes all three test suites, and then the two agents swap roles. We run a tournament among 11 model-harness participants on 15 repositories in five languages for 1,500 turns. The leaders win in different ways. GPT-5.5 with Codex CLI ranks first as the strongest solver, repairing 98.3% of the defects it reviews. Claude Opus 4.8 with mini-swe-agent ranks second as the strongest task author, with a 27.3% Red win rate. Finally, within one model family, tasks written by the newest version defeat a fixed-reference solver far more often than those of older versions (42.9% vs. 4.0%), giving preliminary evidence that task difficulty scales with task authoring capability.

Agentic large language models now tackle complex tasks, yet strategic reasoning over long horizons is under-studied: most agentic benchmarks are short and episodic, and performance on long tasks is hard to dissect into stored knowledge, reasoning and luck. We introduce SpireBench, in which models play the complete commercial deck-builder Slay the Spire (StS), where early decisions decide long-term success. Across 20 large language models (LLMs) at the game's highest difficulty, frontier models beat heuristic baselines and match the average human player but remain far behind experts. This gap widens as the run lengthens: the strongest models win the first act as often as experts, then fall behind. What separates strong models from weak is how well they play each fight; all of them misjudge risk over a run, preserving health rather than spending it to grow stronger. A comprehensive, self-consistent renaming of every game term shows that part of their performance rests on prior knowledge of the game, and models do not improve from the memories they write across runs, which are often wrong and overly cautious. Our harness is the first to give a model the complete original game, and we compare it with existing harnesses. Overall, SpireBench offers an unsaturated long-horizon benchmark with tools to locate where and why agents fail.

Geopolitical and strategic intelligence reports constitute a rich source of expert knowledge and have increasing value in the current political climate, but their heterogeneity and analytical complexity pose challenges for retrieval-augmented question answering. We introduce a benchmark for grounded question answering over 415 open-source geopolitical and strategic intelligence reports, comprising 5,525 document-grounded question-answer pairs spanning factual, inferential, comparative, and risk-oriented questions. We propose a document-aware generation pipeline that profiles reports before generating questions and validate dataset quality through expert annotation and an LLM-as-judge protocol. Using the benchmark, we evaluate 15 retrieval systems and six open-weight LLMs for RAG. We find that models perform similarly on document-level retrieval, while at the harder chunk level retrieval quality is not driven by embedding-model scale: bge-m3, a mid-size open encoder matches an order of magnitude larger ones and outperforms commercial embedding APIs. In contrast, answer generation remains challenging even given the gold evidence, particularly for inferential and risk-oriented questions.

Large language models are becoming central to GeoAI systems, yet their evaluation rarely considers the elicitation of geospatial data requirements. This omission is critical as retrieval oriented geo-analytic tasks cannot be supported reliably without first identifying the necessary datasets and constraints. GeoFitBench-LLM is introduced as a benchmark for assessing whether LLMs can infer the datasets, the supporting resources and the constraints required by geo-analytic requests in natural language. The benchmark is structured around two reasoning dimensions, Spatio-Temporal Relational and Semantic Grounding, reflecting long-standing challenges in geospatial information retrieval and related natural language tasks. Ground truth requirements are encoded in a fixed, provider-independent schema for reproducible evaluation of resource identification and attribute-level specification. Experiments on seven models under minimal prompting indicate that identifying the necessary resources is often achievable, with GPT-5.4 Thinking achieving the best result by recovering the complete set of resources in 66.67% of benchmark requests. Results further indicate that accurate fitness-for-use specification remains more difficult as inferential demands increase.

Understanding and manipulating spatial relations is fundamental to AI systems that reason about physical environments and construct grounded world models. However, existing evaluations of multimodal large language models (MLLMs) for spatial reasoning rely predominantly on static question-answering (QA) over small-scale or synthetic scenes, rarely probing whether models can modify spatial configurations while satisfying explicit geometric constraints. We introduce GeoSpatialBench, a benchmark designed to evaluate an escalating progression of spatial cognition, ranging from spatial relation understanding and geometric reasoning to constrained spatial editing. Grounded in real-world urban data across six geographically diverse cities, GeoSpatialBench evaluates a comprehensive taxonomy of 28 topological, directional, metric, and ordering relations across structured GeoJSON and map image modalities, with ground-truth references and constraints verified via GIS operations. Our empirical evaluation across frontier MLLMs uncovers a stark spatial gap: structured GeoJSON consistently outperforms map images, with multimodal combinations offering no reliable gain over GeoJSON alone, and while models show reasonable competence on qualitative topological and ordering, performance drops on quantitative metric relations. Crucially, editing performance on ordering relations declines relative to QA, and most failed edits violate at least one specified spatial constraint. These results highlight a gap between passively recognizing spatial relations and actively manipulating geometric configurations. GeoSpatialBench provides an open, rigorous diagnostic testbed to benchmark these spatial limitations, laying essential empirical groundwork for physically grounded world models and embodied spatial AI.

Large language model (LLM) agents are becoming direct users of the open web. As they take on increasingly complex and information-intensive tasks, effective web access is becoming a core capability for agentic problem solving. Existing approaches have advanced agentic web access through different access models, including search-augmented agents, browser-based agents, and, more recently, filesystem-based agents. Nevertheless, as users expect agents to complete long-horizon, end-to-end tasks, they increasingly need to conduct broad information gathering, perform browser actions, and revisit acquired evidence within a single trajectory. Yet our preliminary analyses reveal that existing access models struggle to meet these demands. They primarily support either information gathering or browser interaction, while combining their capabilities can involve substantial costs or performance trade-offs. To this end, we present Web\$hell, an interface that unifies web access under a single access model. Using a common schema defined by WebContract, WebCompiler compiles live webpages into addressable information and action objects. These objects are maintained in WebCache and handled through a shared set of operations, providing agents with consistent, operable, and persistent web access. Through extensive experiments across web search, web navigation, and deep research benchmarks, we show that Web\$hell achieves strong task performance at lower cost. Our results further indicate the importance of interface design for effective web access, with Web\$hell opening new possibilities for more capable and efficient LLM agents on the web.

AI systems are making rapid progress in mathematics, but this progress is hard to measure. Benchmarks of solved problems saturate, and in benchmarks of open problems each problem can be resolved for the first time only once. We present SqueezeBench, a benchmark that is resilient to saturation: a success raises the bar instead of removing the problem. Each problem fixes an integer function $f(n)$ whose values are known only for small $n$, and the benchmark keeps a lower and an upper record on it. A record is a computable Lean function with a machine-checked proof that it bounds $f$. A submission proposes a new function and proves that it bounds $f$, that it is no worse than the current record at any $n$, and that it is strictly better at some $n$. Once accepted, it becomes the record that the next submission has to beat. The system thus chooses the bound as well as its proof, and it may improve the record at a single $n$ or its growth rate. At every $n$, the gap between the two records shows how far the certified bounds are from determining $f(n)$. Every proof is checked by two independent Lean kernels, so no referee needs to check the mathematics. The checks also reject a function that computes $f$ by exhaustive search at small $n$, and a maintainer only reads the submitter's report and looks for such a search at larger $n$. We release ten problems from extremal combinatorics with twenty proved initial records and the admission tools. As a baseline, agents based on a frontier language model, with at most one hour per record, had 82 submissions accepted and improved every record, without going beyond the published literature. Agents based on two newer models then started from these records, had 60 more submissions accepted and improved 19 of the 20 records again. For eight of the ten problems, at least one record is still asymptotically behind the best published bound.

Recent speech LLMs can leverage rich paralinguistic cues, enabling more complex and personalized interactions. However, little work investigates how this sensitivity can lead to stereotyping, and existing evaluations largely measure demographic disparities without distinguishing beneficial personalization from stereotypical behavior. We introduce BASIS-Bench, a benchmark designed to disentangle these two aspects through three axes: (1) demographic robustness, measuring disparities in content-based tasks across speakers and demographic classification capabilities; (2) stereotypical bias, measuring whether perceived demographic cues trigger unprompted stereotypical behavior; and (3) personalization, measuring whether models can tailor responses to demographic characteristics when explicitly requested. We evaluate eight recent speech LLMs and find that, despite often recognizing demographic cues, most fail to translate them into appropriate personalization. Explicit demographic labels improve personalization but also amplify stereotypical associations already present in the textual backbones. These findings highlight the need to train speech LLMs not only to perceive demographic cues, but to learn when they should shape their responses.

Reasoning models can benefit from test-time scaling, but continued reasoning may become ineffective or harmful, a phenomenon termed overthinking. Evaluations against full thinking compare accuracy and token usage but lack references for overthinking severity and strategies' theoretical mitigation potential. To characterize overthinking at the model-task level, we first introduce Step-wise Truncation Probing (STP), repeatedly generating answers from multiple retained prefixes along each fixed sampled reasoning trajectory. Using each trajectory's earliest measured position attaining maximum empirical accuracy as a reference, we define Overthinking Ratio and Accuracy Degradation as the fraction of thinking tokens generated afterward and the accuracy decrease from this reference to full thinking, respectively. Secondly, we then formulate dynamic routing, fixed budgets, and dynamic early exit within a unified framework to quantify the theoretical mitigation potential of different termination strategies. Based on this unified framework, we use the complete STP measurements to construct a token-accuracy Pareto frontier for each strategy under a common normalized thinking-token budget. We further develop OTBench, a reusable benchmark built from the STP process, which produced 70.5B reasoning tokens and the corresponding probing measurements across six models and four benchmark groups.

Negative polarity items (NPIs), such as "anymore" and "ever", are grammatical only in select environments, predominantly in the context of negation. Because their grammaticality depends on both syntax and semantics, they have a rich, multilingual literature in theoretical linguistics, yet work on NPIs in language models is almost entirely confined to English. We introduce BLiMPPI, a multilingual evaluation suite of over 12,000 NPI minimal pairs in 33 languages, built by a pipeline that discovers NPIs in Universal Dependencies (UD) treebanks, labels their licensing environments, and perturbs these into minimal pairs. Evaluating 12 multilingual and 2 monolingual models alongside n-gram baselines, we find that multilingual models struggle most with lower-resource languages and with pairs that are not n-gram solvable. We further examine the effects of NPI typology, language family, licensing variability, and NPI-licenser distance. Data: https://github.com/arzuburcuguven/blimppi-data

Evaluating large language models (LLMs) on AMLworld's smallest test splits is costly because each of over a million targets requires a separate graph prompt. Sampling reduces this cost, but small balanced subsets change unweighted precision and F1. We introduce AMLworld-Compact, an evaluation set that retains every illicit test transaction and samples benign transactions from supervised-score strata. Its HT-Coreset protocol reduces evaluation targets by 271x and 611x on HI-Small and LI-Small. Horvitz-Thompson weights give unbiased full-test confusion-count estimates for fixed predictions; recall is exact, while precision and F1 are ratio estimates. In a held-out study using Graph Feature Preprocessor (GFP) features, one boosted-tree family defines strata for evaluating the other over 500 draws per configuration. At the released sizes, F1 RMSE is 0.00-2.58 percentage points versus 4.44-7.41 for uniform sampling with HT weights; conservative intervals remain broad. Under the evaluated graph prompts, seven open-weight LLMs have weighted precision near the predict-all-illicit baseline. In an outcome-stratified audit of 1,000 HI-Small few-shot outputs from one run per model, lexical checks pass for 84.5% of correct and 56.2% of incorrect binary verdicts, so these markers do not establish correctness. We release the evaluation sets, construction code, and audit tools at https://huggingface.co/datasets/typhoon-ai/AMLworldCompactEval and https://github.com/nat-nischw/AMLworld-Compact.

Large language models (LLMs) have shown strong results in classical code generation and are increasingly used to write the unit tests that safeguard that code. Whether they can do the same for quantum software is less clear, because quantum programs are probabilistic, depend on numerical tolerances, and are judged by domain-specific checkers, so a well-formed test can still disagree with the benchmark's verdict. Existing quantum benchmarks measure whether a model produces a correct solution, but not whether its tests can distinguish correct implementations from faulty ones. We introduce QTestBench, a framework that evaluates LLM-generated quantum test suites against independent benchmark checkers, never treating the tests themselves as ground truth. For each task, QTestBench prompts models to produce syntactic, semantic, and behavioral tests, injects candidate solutions at runtime, and compares each test verdict with the checker's label under three validation cohorts, each answering a distinct diagnostic question. We benchmark five LLMs on eight PennyLane QHack challenges and ten Qiskit HumanEval tasks. Generated suites almost always execute, but discrimination lags behind: the best generator reaches a test quality of 0.71 on QHack and 0.34 on Qiskit HumanEval, so executability alone overstates test quality. Oracle-verified synthetic mutants stabilize bug-detection estimates when natural faulty solutions are scarce, as on three of ten Qiskit tasks that yield none. QTestBench offers a reproducible, cross-framework benchmark for the trustworthiness of LLM-generated quantum tests.

Personal agents increasingly connect to services across the web and applications by using tools. As the effects of their actions accumulate and interact across services, unintended or harmful outcomes can arise when agents fulfill benign user requests. Proactively identifying these compositional failures requires systematically exploring diverse cross-service paths. Existing red-teaming methods often simply prompt LLMs to identify potentially harmful paths from task and environment descriptions or decompose harmful objectives specified by humans or LLMs into seemingly benign action combinations. However, generation is biased toward patterns learned during training, potentially favoring familiar tool combinations and leaving novel cross-service interactions underexplored. To address this, we propose X-AgentRed, a self-evolving red-teaming framework that discovers harmful tool call paths through a graph of tool dependencies and expands and refines them using execution feedback. By analyzing each tool's effects, potential harms, and preconditions for execution, the red-teaming agent constructs a directed graph that captures how one tool's effects can satisfy another's execution preconditions. Recipes sampled from this graph are instantiated as realistic scenarios in which agents can produce harmful outcomes. Execution feedback enables self-improvement by diversifying paths that produce harm through tool substitution and refining preconditions at blocked steps in unsuccessful paths to enable further execution. Across four frontier agentic models, X-AgentRed achieves $1.77\times$ the harmful path rate of STAC and explores a more diverse set of tools, yielding $2.99\times$ the tool coverage on average.

Task-oriented conversational agents remain fragile under realistic conditions, especially when users exhibit non-cooperative behavior. Existing function-calling benchmarks underrepresent adversarial conversation trajectories, thereby limiting the training resources available for developing robust agents. We present MADS (Multi-Agent Dialog Simulation), a synthetic data generation pipeline that produces structured, diverse dialogue trajectories. We inject adversarial conversation, producing validated conversation trajectories that capture realistic, unhappy paths. Using MADS, we construct MADS-Bench: a benchmark spanning six domains, 55 personas, and 1095 conversations, with approximately 66% of conversations containing failure-prone conversation scenarios. We evaluate the robustness of the trained models using a two-stage training approach that combines supervised fine-tuning (SFT) and reinforcement learning with verifiable rewards (RLVR).

A conversational speech model has to decide, at every pause, whether to speak. Existing evaluations test boundary detection, policy adherence, or aggregate statistics, not this decision. We place real-time speech systems at 427 pauses and 482 turn ends from 176 unscripted conversations in Seamless Interaction, give each the conversation so far, and score what it generates against what the person did: whether and how soon it speaks at a pause; its latency, length, and wording at a turn end; and whether it can answer a question about the preceding 30 to 120 seconds. Across six systems, the hosted APIs (GPT Realtime, Gemini 3.1 Flash Live, Grok Voice) respond at almost every pause whichever endpointer they use; only their latency, about 1 to 4 s with silence-based endpointers and about 8 s with semantic turn detection, decides whether the words land inside the pause. The full-duplex models (PersonaPlex, Moshi) are the only systems that regularly stay silent and the only ones that interrupt within a human reaction time. Hosted systems reply longer than the person. Given the history as text, hosted systems recall 41 to 60% of conversation facts under a Claude Opus 5.5 judge and 19 to 26% under a strict word matcher; hearing it as audio, full-duplex models recall 17 to 22% and 12 to 15%. We release the labels, points, harness, and every output.

Web agents increasingly browse, interpret, and act on web content on behalf of users. In many realistic workflows, however, the visual evidence they consume is not neutral: charts and dashboards can be misleading, shaping judgments without issuing any explicit malicious instruction. Unlike chart question answering, Web Agents treat charts as intermediate evidence whose misinterpretation can propagate into concrete downstream actions --- a reliability risk that remains underexplored. We introduce MisVisAgentBench, the first paired executable benchmark for evaluating how misleading visualizations affect Web Agent execution. It comprises 280 task instances in matched clean--misleading pairs that differ only in chart evidence. We further provide a diagnostic framework that isolates misleading-induced failures from baseline task difficulty (Zhu et al., 2025) and captures execution-level effects such as extra browsing steps and decision hesitation. Across thirteen vision-language Web Agents, replacing clean charts with misleading counterparts reduces the average task success rate by more than 34 percentage points. Moreover, these failures predominantly follow the targeted misleading action branch, suggesting that deceptive charts systematically redirect rather than randomly disrupt agent decisions. These results reveal that visualization robustness is a critical yet understudied dimension of Web Agent safety and reliability.

Creativity benchmarks for large language models (LLMs) vary widely in task and metric design. This raises the question of whether they capture a common notion of creativity. We propose that this shared target can be formalized as task-specific performance in the long tail of a reference input-output distribution, which describes how common input-output pairs are relative to a chosen source of text. We ground this proposition in the novelty-quality product formulation and evaluate it empirically across 15 benchmarks and 26 open-weight models, using n-gram absence from the pretraining data as a novelty proxy. Our results show that creativity benchmarks lie further in the tail regions than standard benchmarks. Moreover, on both creativity and standard benchmarks, model rankings from higher-tail subsets of each benchmark align more closely with human creativity judgments in LM-Arena. Building on these findings, we introduce q@Tail($\alpha$), a lower-cost creativity diagnostic that measures mean task quality on the tail fraction of an existing benchmark. Applying it to creativity-promoting methods reveals substantial room to improve LLM creativity through stronger tail performance.

Large audio-language models (LALMs) frequently exhibit object hallucination, affirming absent sounds. The dominant mitigation, Audio-Aware Decoding (AAD), contrasts logits under real audio against an all-zero "blank" waveform, flagged by prior review as out-of-distribution. We introduce a black-box causal test that isolates the claimed evidence itself: localize a claim's evidence region via CLAP similarity, remove it by spectral gating, and compare against a random-window control across a 2x2x2 grid on three LALMs. On Qwen2-Audio-7B, localized ablation's raw AUROC trails blank's (0.613 vs. 0.721); this is a duration confound, not a localization failure, since blank removes the entire clip versus the grid's 2-4 s. Holding duration fixed, localization beats random removal at every cell on both models. Raw detection value remains model-dependent, especially on Audio Flamingo 3. The same evidence corrects decoding: contrastive decoding combining blank and language-prior contrasts, gated and ensembled, raises Qwen2-Audio accuracy from 57.1% to 71.9%, beating AAD alone. A unified causal-grounding profile does not transfer between models, indicating different grounding regimes. The source code and a detailed report of our findings, including identified limitations, are available at https://github.com/khalooei/causa.

Tool-using agents increasingly require a common substrate for building, training, evaluating, distilling, and maintaining agentic systems, yet existing tooling typically addresses these stages in isolation. We present AgentTune, a source-available Python library that provides a unified framework for developing and optimizing tool-using agents. At its core, AgentTune introduces a two-tier EventLog representation that preserves agent trajectories while supporting consistent projections for training, evaluation, and auditing. Building on this substrate, the library provides a shared agent runtime, tool-calling reinforcement learning, agentic distillation, retrieval-augmented generation, decision workflows, evaluation utilities, and a designed path toward self-healing agents. These components are exposed through composable interfaces so that agent strategies, training methods, evaluation procedures, and recovery mechanisms can be combined without requiring separate representations or pipelines. We demonstrate the framework across agentic tasks including legal contract search and domain-specific applications, illustrating its use for trajectory-based optimization, tool-use evaluation, and cross-task agent development. AgentTune is released as a source-available library at https://github.com/Lexsi-Labs/AgentTune

Stance detection remains challenging for large language models (LLMs) due to implicit opinions, sarcasm, and context-dependent expressions of stance. Although automatic prompt optimization offers an alternative to fine-tuning without updating model parameters, existing structured approaches are not specifically designed around stance-label semantics and class-specific ambiguities. We introduce MoPrO-SD, a stance-specific prompt optimization framework that decomposes prompts into functional semantic modules and refines them sequentially using LLM-based feedback. We study three refinement strategies: instruction-only optimization (I), instruction with reasoning (I+R), and instruction with per-example reasoning (I+R+E), using class-conditioned errors for stance-label refinement. We also introduce VaccineStance, a new human-annotated benchmark for vaccine stance detection. Experiments across four datasets and five open-weight LLMs show that MoPrO-SD improves upon the human-written Base prompt in many settings. On SemEval-2016, class-conditioned feedback achieves higher mean Macro-F1 than pooled feedback for four of five backbones. Both reasoning-based strategies, I+R and I+R+E, achieve strong performance, with their relative effectiveness varying across models and datasets, while remaining competitive with recent prompt optimization methods. The MoPrO-SD code and VaccineStance dataset are available on GitHub.

Long-horizon embodied agents must maintain task-relevant information, coordinate heterogeneous robots, and adapt their decisions as execution unfolds. Existing benchmarks investigate dynamic task execution and embodied multi-agent coordination, but provide limited evaluation of their intersection. We introduce MAGMA-Bench, an architecture-agnostic benchmark for persistent task reasoning in dynamic multi-robot environments. MAGMA-Bench evaluates agents across increasing task horizons and controlled online perturbations, including execution failures, task interruptions, and mission updates, with scenarios requiring coordination across robots with heterogeneous capabilities. It uses a structured skill-level interface to separate task-level reasoning from perception and low-level control. We evaluate locally deployable open-weight planners, multi-agent state-tracking strategies, and frontier models. Even the strongest evaluated models fail on a substantial fraction of clean episodes, and their success rates decrease further under online perturbations. Supervised adaptation greatly improves a small model, but performance remains substantially lower on longer and multi-robot missions. We release the benchmark, 261,646 task-level training examples, and an architecture-agnostic expert-trace generation pipeline.

As AI agents increasingly rely on external tools, accurate retrieval is essential for reliable behavior. Existing methods optimize semantic relevance but largely ignore tool availability during execution. To address this gap, we propose RATED (Reliability-Aware Tool Retrieval via Execution Dynamics), a retrieval framework that combines robust semantic matching with dynamic reliability adaptation. RATED employs a frozen dual-encoder with contrastive fine-tuning for semantic matching, and a lightweight contextual bandit module that re-ranks candidates based on real-time execution feedback. This design enables rapid adaptation to runtime tool failures without expensive model retraining. We introduce EcoAgent, a benchmark of 5,172 real-world tools and over 210,000 query-tool pairs for offline retrieval and dynamic reliability evaluation. Experimental results demonstrate that RATED achieves a 52.2% relative MRR improvement over the ToolLLM API retriever on real-world queries, and reduces fault selection rate from 86.2% to 2.9% under severe API outage scenarios with a 30% failure rate, while incurring negligible latency overhead of less than 1 ms over encode-only inference. The code and dataset are available at https://github.com/Elliotes/EcoAgent.

Language models increasingly support or automate information retrieval from document collections too large to index. This task is difficult when it requires specialist linguistic or scientific knowledge. We used the Wikipedia Game to test language models on an analogous information retrieval task. In the game, players must reach a target article using only article links. We compared sentence embedding models and LLMs across five Wikipedia languages and five academic-literature datasets. We construct very large local graphs of Wikipedia pages and of paper citations using OpenAlex and provide code to build and navigate them. This allows the study of language models' navigation capabilities without relying on third-party APIs. We compared the models with human and TF-IDF benchmarks. The Wikipedia Game format enables robust comparisons across techniques, languages, and domains. No model significantly outperformed humans, although the best models (GPT-o3 mini, GPT-4o, QwQ:32B, and DeepSeek-R1:70B) achieved comparable success. Performance and rankings varied across languages and topics, with implications for practitioners utilising these techniques.

Autoregressive decoding is less suited to non-causally ordered data, e.g., cross-references or future dependencies, since their outputs cannot be revised post hoc, unlike iterative human reasoning. Diffusion LLMs partially mitigate this limitation but still cannot flexibly insert or remove arbitrary numbers of tokens from a pre-generated context. We design AutoIndexer as a generative model with a novel training and decoding mechanism that simulates a Chain of Edits, allowing the model to flexibly modify its prior outputs in hindsight via token insertions, substitutions, and deletions. The model learns to predict marker tokens for opening and closing edits, simultaneously moving a cursor to positions in the existing token sequence, analogous to a text editor. Training sequences are obtained by perturbing ground-truth labels with inverse chained edit operations, and the model learns to edit by undoing those inverses. The perturbation training allows AutoIndexer to be robust to erroneous context and adaptable to editing tasks such as code correction. We demonstrate that our method exhibits post-generation context-editing capabilities that most autoregressive and diffusion models lack without compromising performance on standard LLM benchmarks. The code is available at https://github.com/AutodeskAILab/autoindexer.

Speculative decoding accelerates large language model inference by using a lightweight drafter to propose multiple tokens that are subsequently verified in parallel by the target model. Recent diffusion-based drafters make this paradigm particularly attractive: deep attention-based neural backbones provide strong predictions, while drafting multiple future positions in parallel avoids the sequential cost of autoregressive generation. This parallelism, however, comes at a cost. Tokens within a draft block are predicted without access to the realized identities of preceding draft tokens, removing intra-block dependencies that would naturally be available under autoregressive generation. We introduce DCOCO (DFlash Contextual Corrector), a lightweight mechanism that restores these dependencies while preserving parallel drafting. Rather than committing to each position's top prediction, DCOCO retains multiple candidate tokens and evaluates their contextual compatibility across adjacent draft positions in parallel, enabling contextually informed token selection at low additional cost. Across four verifier backbones and ten benchmarks, DCOCO achieves the highest average throughput in every evaluated configuration, spanning greedy and stochastic decoding with and without extended reasoning, with speedups of up to 6.6x over vanilla decoding. Its advantage persists under concurrent serving: DCOCO outperforms the strongest competing method by 6-23% across 4-32 concurrent requests, while other sequential correctors fall below uncorrected DFlash at these loads. These results demonstrate that recovering intra-block dependencies through parallel contextual scoring delivers consistent acceleration across decoding regimes and serving loads.

A latent recurrent language model applies a shared block of layers repeatedly, so one checkpoint can trade inference compute for accuracy by changing the loop count. At one loop, however, such models typically fall well short of a non-recurrent model of the same size, a one-loop tax that makes a separate non-recurrent model the better choice whenever latency matters. We present Elastic Latent Recurrence (ELR), a recipe for training looped language models without this tax. At one loop, an ELR model costs the same to serve as a non-recurrent model of the same size and, trained on the same data, reaches similar accuracy. Each additional loop then buys more accuracy for more inference compute. Three choices make this work: a gated state update with a learned step size, loop counts mixed within each optimizer update, and conditioning each pass on its loop index and budget. The first two preserve one-loop accuracy, and the third lowers loss at every additional loop. Mixed depths can be introduced late in pretraining to limit the added training cost, and with zero-initialized conditioning added at the switch, this late schedule reaches lower loss than mixed-depth training from scratch. In MoE models, reusing first-loop expert routes on later loops cuts mixed-depth update time by 11.7% with no measurable change in accuracy. At inference, keeping only the last loop's attention KV cache holds attention KV memory at its one-loop level while retaining most of the gain from additional loops. On 4B dense and 30B-A3B MoE hybrid models, ELR at one loop is 0.52 points below and 1.04 points above its matched baseline on a ten-task average, and four loops add 7.25 and 4.75 points. Carried through post-training, the recipe yields Koi-4B, a single checkpoint that scores an estimated 18 on the Artificial Analysis Intelligence Index at one loop, matching Qwen3.5-4B, and 29 at four loops.

Search agents extend large reasoning models with multi-round access to external knowledge, yet most existing systems largely conflate retrieving evidence with consuming it. Retrieved documents are commonly incorporated according to fixed or rank-based policies, despite the fact that evidence utility is conditioned on the agent's current reasoning state. Consequently, a lower-ranked candidate may be critical for the current step, while a higher-ranked candidate may provide redundant, irrelevant, or misleading information. We formulate this overlooked problem as evidence admission, which asks, at each search step, which retrieved candidates should be allowed to enter the agent's evolving reasoning context. To enable such selective control, we introduce SEEDS (Selective Evidence Entry During Search), a preview-guided evidence admission framework for search agents. SEEDS first exposes lightweight previews of retrieved candidates and then lets the agent jointly determine which candidates and how many should be fully inspected, conditioned on its current reasoning state. Unlike fixed or adaptive prefix policies, SEEDS can select arbitrary subsets with flexible cardinality, and it can be instantiated directly with an existing reasoning agent or further optimized through learning. Across nine knowledge-intensive QA benchmarks under both local corpus and live web search, SEEDS consistently improves search effectiveness while substantially reducing full-evidence consumption. Experimental analyses show that SEEDS more effectively selects useful candidates beyond the corresponding retrieval prefix. These results establish evidence admission as a complementary dimension of agentic search, separating what information is retrieved from what information is ultimately allowed to shape reasoning.

We present Neurograf v1, an experimental end-to-end Romanian-first language-model pipeline developed to investigate the feasibility of training language models from scratch for a lower-resource European language. The pipeline comprises RoSfera, a 635 GB Romanian corpus containing 175.7 million documents; an 80k byte-level BPE tokenizer; five decoder-only models ranging from 160M to 6B parameters, trained on 131 billion tokens; three instruction-tuned assistants; and two contrastively trained embedding models. Training and evaluation were conducted using computing infrastructure at ELI-NP/IFIN-HH. The model family exhibits consistent scaling of validation loss and increasing Romanian factual and cultural knowledge with model size. On the 1,000-question WWTBM benchmark, Neurograf-6B achieves 69.89% accuracy on Romanian-context questions, while Neurograf-3.2B reaches 67.38%. Compute-optimal analysis yields a parameter-scaling exponent of approximately 0.423. Neurograf v1 serves as an experimental demonstrator rather than a production system. The study identifies practical challenges involving diacritic preservation, corpus quality, data composition, and tokenization efficiency, while demonstrating the feasibility of developing Romanian-first base models, instruction-following assistants, and semantic embedding models within a unified training pipeline.

While reinforcement learning with verifiable reward (RLVR) enhances reasoning in large language models, its efficacy in long-horizon multimodal tasks remains underexplored. Visual Sudoku puzzles pose a representative challenge, requiring multi-step deductive reasoning over both local constraints and global dependency chains. To this end, we introduce FG-Sudoku, a fine-grained benchmark categorized by grid size and required deductive steps. Evaluations reveal that frontier MLLMs fail on complex long-range dependencies (e.g., 9*9 puzzles). During post-training, this near-zero success rate causes severe reward sparsity, halting standard RLVR. We propose Adaptive Curriculum Group Relative Policy Optimization (AdaC-GRPO) to overcome this. Unlike rigid or prior-based schedules, AdaC-GRPO dynamically constructs training batches by tracking the model's "Zone of Proximal Development" via online rollout accuracy. Experiments demonstrate that AdaC-GRPO enables Qwen2.5-VL-7B to approach the performance of Gemini-2.5-Flash while bridging 62.2% of the overall accuracy gap to Gemini-2.5-Pro, with its generalizability further validated across InternVL-3.5 and Qwen3-VL.

Designing tool-integrated reasoning agent systems for complex chemical tasks remains a fundamental challenge. Although recent chemical agents have demonstrated promising performance through the integration of external tools, they are still built on the autoregressive paradigm, which is inherently constrained by causal attention and left-to-right generation. As a result, they often struggle with bidirectional molecular understanding, global coordinated tool planning and multi-source evidence integration. To address this challenge, we introduce ChemDiffAgent, the first masked diffusion language agent for tool-integrated chemical reasoning. It reformulates chemical reasoning as iterative denoising over interleaved reasoning and tool-use trajectories, enabling tool-use decisions to be jointly determined under full-trajectory context and thereby better aligning the agent's decision boundary with its knowledge boundary. We further analyze how this paradigm benefits agentic reasoning through harder infilling subproblems during training and flexible arbitrary-order decoding during inference. Then we develop a two-stage post-training framework, comprising agentic supervised fine-tuning and variance-reduced preference optimization, to enhance tool-use and reasoning capabilities. We also construct a comprehensive benchmark covering six chemistry tasks across both single-turn and multi-turn tool-use scenarios. Under matched budgets and evaluation, ChemDiffAgent consistently outperforms autoregressive agents, achieving an overall improvement of approximately 20%. Notably, our 8B model achieves performance comparable to, in some cases surpassing, significantly larger frontier models, such as GPT 5.5 and Claude Opus 4.7. Code is available at https://github.com/bilemond/ChemDiffAgent.

The fast pace in which protocols, frameworks, and technologies are being proposed for LLM-based agents results in a complex scenario for anyone aiming to create state-of-the-art AI solutions. Without an assessment on how different protocols, frameworks, and libraries work together to compose an Agentic Web solution, developers and organizations face different challenges (e.g., skilling, governance, infrastructure) even before becoming part of the Agentic Web ecosystem. In this context, we propose a system for performing Agentic Web evaluations called MASTIF (Multi-Agent System TestIng Framework) and present a baseline assessment of Agentic Web technologies, including 3 LLMs, 3 protocols, and 4 agent implementation frameworks, against 300 Mind2Web benchmark tasks. Results suggest that selecting a task-specific technology stack may result in gains of task completion rate, efficiency score, and social value stability scores; and reductions in verbosity and time to complete tasks. With the presented results, we expect to present a baseline in terms of metrics and results for this complex and fast evolving scenario for those aiming at designing, assessing, and developing multi-agent systems.

Low-bit inference on accelerators with native 4-bit support requires quantizing both weights and activations to realize the full throughput gains these formats provide. While post-training weight quantization has been extensively studied, activation quantization remains considerably more challenging and typically requires calibration data. In this work, we introduce DUALQUANT, a dual-scale, calibration-free post-training quantization framework that directly minimizes the weight-reconstruction Mean Squared Error (MSE) and, as an immediate by-product, yields per-channel activation rescalings prior to activation quantization. Our key observation is that the scaling factors obtained by minimizing weight-reconstruction error provide effective activation rescalings, enabling competitive activation quantization without calibration data. We formulate DUALQUANT as a non-differentiable weight-reconstruction problem and develop an efficient solver based on variable splitting and block coordinate descent, where each subproblem admits a closed-form update. DUALQUANT provides a unified algorithm supporting a wide range of formats, including the full MX family and NVFP4. To our knowledge, DUALQUANT is the first dual-scaling method to formally support MX and NV block formats. Across Llama and Qwen models evaluated on WikiText-2, C4, and zero-shot downstream tasks, DUALQUANT matches or exceeds the accuracy of calibration-based SmoothQuant across most models and formats, and substantially outperforms calibration-free baselines, with large gains for MX and NV block formats. We validate these gains with a real packed W4A4 inference path for Llama-3.1-8B on an RTX 4090. The resulting implementation closely matches the accuracy of fake quantization while delivering up to 2.61x memory reduction and up to 2.46x throughput speedup over BF16.

State-space models (SSMs) maintain a fixed-size recurrent state, but reading and writing it at every decoding step can bottleneck hybrid models. This eager state management causes large state I/O with low arithmetic intensity, leaves speculative decoding no cheap rollback, and feeds state quantization error into subsequent recurrent updates. We introduce ReplaySSM, which unrolls the recurrence to represent the state as a checkpoint and a buffer of recent SSM inputs. This separates output computation from checkpoint updates. By updating the checkpoint periodically, ReplaySSM avoids state writes on most steps and roughly halves the dominant state memory traffic. Keeping recent inputs also makes speculative rollback a buffer-pointer operation and enables output-only verification without materializing intermediate states. The same separation allows a low-precision checkpoint to serve output computation on most steps while a higher-precision checkpoint serves recurrent updates, so rounding error in the low-precision copy never enters later updates. We implement ReplaySSM in vLLM for Mamba-2 and Gated DeltaNet. At serving batch size 256, it accelerates standard decoding by $1.20$--$1.47\times$ over a tuned vLLM baseline across four hybrid models, and delivers about $1.8\times$ the throughput of vLLM's speculative decoding on the two 120B-class models. Separate state precisions also reduce output KL divergence and downstream accuracy losses compared with a shared low-precision checkpoint.

Vision-Language Models (VLMs) increasingly analyze and reason about images, yet it remains unclear whether their internal representations faithfully encode the structured relations required for such reasoning. We introduce EICS (Evaluation-Invariant Concept Stability), a metric for testing whether representations consistently encode concepts under controlled one-concept interventions, and CASP-VL (Causally controlled Spatial Perturbations for Vision-Language models), a benchmark of causally controlled spatial perturbations derived from Visual Genome scene graphs. Across multiple VLM architectures, we find that spatial relations are represented substantially less faithfully than lexical information, even when the same models score well above chance on compositional benchmarks. Further analysis shows that the relation is identifiable at the predicate's own tokens but is lost when the sentence is aggregated into a global representation, whether through the (EOS) token or by mean pooling. These findings reveal a systematic failure invisible to output-based evaluation: VLMs do not preserve, in the representation used downstream, relational information that is present in their own input.

Vision-language reward models are trained on annotated preference pairs, which are expensive to collect, can disagree, and are ranked by benchmark accuracy. We build one with no human preference label: a linear head on an intermediate layer of a frozen Qwen3-VL backbone, trained to score a web caption above a perturbed version of the same caption; labeled data enters only to select the layer. The layer is the decisive choice, worth 5.8 points over the last layer, and one recipe transferred without re-tuning beats or at least matches GPT-4o on three benchmarks at 32B, 6.6 points ahead on VL-RewardBench, from a few thousand parameters at about a thousandth of fine-tuning compute. At an equal pair budget, perturbations teach a frozen head 3 to 5 points more than curated RLAIF-V pairs. Because the backbone is frozen, the objective is convex, and the perturbation sampler specifies the reward, which the benchmark does not identify: two heads trained on disjoint halves of the same perturbations tie on the benchmark and differ by 6.4 points in hallucination judging. Retraining the head on the policy's own verifier-labeled candidates cuts caption hallucination by 5.1 CHAIR points and lifts image-generation pass rate by 5.4, while scoring lower on every reward benchmark. Together, an intermediate-layer probe, free same-image perturbations, and a verifier-driven bootstrap turn give a recipe for multimodal reward modeling that needs no annotator, costs a fraction of fine-tuning, and is specified by the comparisons it is trained on rather than by a benchmark rank.

Satirical news is a unique form of misinformation: it is intended to be humorous by the original publication source but can be misinterpreted. Explanation of satirical news articles is challenging because recent and local context needs to be incorporated. This is a pilot study that explores explanation generation for satirical news articles from Australia and compares them alongside those from USA. The paper demonstrates that framing explanations on unreliable narrator theory on open-weights models can be a feasible, cost-effective alternative to frontier models. We find that models identify the satirical tropes more accurately on the American source than the Australian one. This work results in a novel dataset and an evaluation benchmark for satirical explanation.

In repeated human-AI interaction, a single observation can be quoted, summarized, and returned to a model as multiple reports, making repetition look like independent corroboration. We study whether language models can use supplied report lineage to update according to evidence provenance rather than report count. We introduce Latent Quotient Evidence Training (LQET), a relational objective that combines sparse Bayesian supervision with constraints that encourage stable predictions when reports add no new evidence and consistent updates to the same new evidence across relay histories. On a controlled binary benchmark, LQET reduces new-evidence update RMSE by 41-44% and log-odds drift after eight relay steps by 89-91% relative to task-only training across Qwen3-4B and Phi-4-mini, with three paired training seeds per backbone. The gains extend to a closed-loop simulation in which model confidence controls how many relayed reports re-enter the model's context in later rounds. By grounding belief updates in supplied provenance, LQET offers a route toward dynamic human-AI systems in which influence follows new evidence, not the number of times the same evidence returns.

Recurrent-depth language models improve inference-time reasoning by repeatedly applying a shared network to refine internal representations, but the additional recurrent computation also increases inference cost. This overhead is particularly pronounced during long-context prefill, where the full input sequence is repeatedly processed across recurrent depth. Weight-and-activation quantization offers a natural way to accelerate this computation through lower-precision arithmetic, yet repeatedly applying quantized computation introduces numerical perturbations that accumulate along the recurrent trajectory. We systematically analyze how such errors propagate across recurrent depth and show that quantization errors introduced earlier are attenuated by more subsequent recurrent steps and therefore have a smaller effect on the final representation. Matched-budget experiments strongly support this prediction, showing that placing quantized computation earlier in recurrent depth consistently yields the best fidelity. Based on this insight, we introduce RAMP (Recurrent Adaptive Mixed Precision), which performs early recurrent refinement using low-precision weight-and-activation computation and restores high precision for terminal refinement. Across context lengths up to 4K, RAMP preserves near-full-precision prefill quality while achieving up to 1.16x end-to-end first-prefill speedup on an NVIDIA H100, corresponding to a 13.9% latency reduction.

Power sampling improves large language model (LLM) reasoning at test time by sampling from a sharpened version of the base model's distribution. Its sampler, PowerMH, is slow because its Metropolis--Hastings (MH) updates are sequential: each accept/reject decision sets the state from which the next proposal is generated. Prior work reduces this cost by changing the sampler or the model (e.g., sequential Monte Carlo, approximate lookahead, distillation). We propose Predictive SubTree Prefetching (PreSTO), which accelerates PowerMH by prefetching future proposals and provably leaves its MH transition kernel unchanged. PreSTO views future MH outcomes as a tree of accept/reject decisions: in each model call, it generates the proposals of a \textit{prefetched subtree} of this tree in one batch, then follows the realized path with the original MH acceptance rule. We also observe that MH reaches only a sparse, nearly unary part of each prefetched subtree, we further design a \textit{predictive traversal rule} that chooses the subtree to maximize the predicted number of completed transitions. Across seven datasets and five base LLMs, PreSTO achieves median speedups of $1.71\times$, $1.91\times$, and $2.33\times$ on PowerMH and its follow-ups EntropyCut and MultiTryMH, respectively; on the first two, it makes a median of $3.9\times$ fewer model calls, each only $2.1\times$ longer. On LCB V6 with Qwen3.5-9B, the predictive traversal rule is a further $1.25\times$--$1.37\times$ faster than the default breadth-first traversal.

Fast rollout generation for long-context reinforcement learning with verifiable rewards (RLVR) requires reducing KV-cache traffic without changing the response groups used for policy optimization. In our profiled RLVR workloads, attention accounts for 72--76\% of autoregressive decoding kernel time. Sparse attention reduces these reads, but sampling directly from a sparse policy changes the response groups used for learning. For group-relative optimization, we show that even exact response-wise reweighting does not generally recover the expected dense-policy update because the advantage of each response depends on the rewards of other group members. We propose Self-Sparse Speculative Rollouts (S$^3$R). It uses sparse attention to propose tokens, then generates responses through full-attention verification and exact speculative sampling. Drafting and verification share the same policy parameter, eliminating the need to train or update a separate drafter. The resulting response groups follow the dense-policy distribution. Under S$^3$R, we characterize speedup conditions and the costs of drafting and verification. We found that dense verification still requires the complete historical KV cache. A KV-aware scheduler therefore limits KV-cache repeated offloading and restoration while maintaining rollout parallelism. % OR thrashing In a matched ablation with an aggressive sparse drafting budget, S$^3$R maintains stable training while direct sparse sampling with training-side correction becomes unstable. On long-context mathematical reasoning, S$^3$R achieves $1.39\times~1.35\times$ the end-to-end training throughput of autoregressive rollouts on Qwen3 model, with comparable task performance.

Clinical prognosis requires integrating structured measurements, physiological waveforms, longitudinal history, and population context. We introduce PRISM (Population-Retrieved Inference for Structured Multimodal Prognosis), a multimodal language model that combines per-feature numerical tokenization, recency-weighted aggregation of longitudinal ECGs, and a dual-modality population memory bank. Retrieval operates in representation space: the model receives mean-pooled embeddings of similar training-cohort encounters, while no individual retrieved record enters the prompt. On the 1,443-task MDS-ED benchmark, PRISM achieves AUROCs of 86.2% for diagnosis, 93.1% for deterioration, 91.8% for ICU admission, and 92.8% for mortality. These results use a 96-hour pre-arrival window unavailable to reported baselines. With benchmark-window inputs only, PRISM reaches 0.8895 AUROC, compared with 0.8937 for UniPACT and 0.8889 for MDS-ED. A matched-input compact fusion model using the same five evidence streams and frozen encoders reaches 0.7652 AUROC, versus 0.9092 for PRISM. Replacing numerical EHR encoding with text serialization reduces AUROC by 3.66 points; removing pre-arrival history reduces it by 1.97 points; and zeroing retrieval reduces it by 0.99 points. Genuine retrieval outperforms random-neighbor retrieval by 0.97 macro-average points, although a paired task-level sign test does not establish a uniform advantage (p=0.067). On the external MC-MED cohort, supervised fine-tuning reaches 0.8611 AUROC, compared with 0.8227 for locally trained gradient boosting and 0.7841 for source-only PRISM. These results support structure-aware multimodal integration and supervised external adaptation.

Long-context serving runs out of KV memory before the hardware reaches its useful throughput. HiSparse stores history in host memory and fetches only the selected entries, but it uses one cache budget at every load and does not handle draft-model speculative decoding such as DSpark and DFlash2. We present Adaptive DHiSparse, which supports speculative decoding over hierarchical sparse KV and adapts residency to load. It protects the required working set, retains additional history while GPU memory is spare, and recovers the rest through exact fused retrieval. DSpark and DFlash2 keep their proposal and acceptance rules: queries share missing-row retrieval, draft KV uses a window-bounded ring, and verification staging is reused across layer groups. At 100K context and concurrency 128, GLM-5.3 with DFlash2 delivers 793 output tokens/s per decode GPU, 1.45$\times$ HiSparse and 1.27$\times$ non-speculative adaptive serving. Held to the same latency bounds as HiSparse, it sustains up to 4$\times$ the concurrency and 2.3 to 3.7$\times$ the output, depending on the per-token bound. On DeepSeek-V4.1-Flash at 500K context and concurrency 512, Adaptive DHiSparse with DSpark delivers 1.85$\times$ the output of resident serving with DSpark. In a load step, optional residency reduces host fetches from 14.2% to 1.4% of selected rows at low load and releases memory under pressure.

Legal language understanding requires reasoning over semi-structured regulatory documents, yet standard pre-training objectives do not explicitly target this structural capability. Because current legal language models rely solely on masked language modeling, they treat statutory text as flat token sequences and ignore the underlying hierarchical constraints. To bridge this gap, we present a multi-task pre-training approach for German traffic law (StVO). Our approach combines three complementary objectives: Next Sentence Prediction (NSP), Knowledge Graph (KG) Guided Clustering (CLU), and Knowledge Graph (KG) Aware Paragraph Classification (CLS). This approach explicitly encodes structured legal knowledge directly into the model weights. This encoding enables zero-shot legal reasoning at inference time and removes the need for discrete graph traversal. We evaluate our approach across nine Transformer-based encoders and 324 experimental runs. On synthetic StVO question answering, our models demonstrate consistent relative improvements between 10.3% and 155.0% over baselines. Furthermore, relative to the established LexGLUE baselines (Chalkidis et al., 2022), our models improve over prior results on four of seven legal benchmarks. It also achieves consistent zero-shot transfer gains between 3.4% and 15.8% on Austrian driving regulations. These experimental results indicate that our training approach generalizes successfully beyond the initial training legal domain.