Length-Conditioned Majority Voting (LCMV) improves LLM reasoning by exploiting the empirical bias that longer chain-of-thought traces are less likely to be correct. Instead of filtering out long responses, it learns a logistic prior over correctness conditioned on CoT length and adapts it at test time using only unlabeled data and pseudo-labels from standard majority voting. The method is model-agnostic, adds negligible computational overhead, and is evaluated across 11 models and 8 benchmarks, making it a pragmatic drop-in replacement for standard MV.

Multi-party floor control is the next bottleneck for full-duplex voice agents: systems must decide not just when speech ends, but whether overlapping listeners are backchanneling or genuinely taking the floor. MultiPartyTurnBench introduces 240 synthetically generated, channel-separated conversations spanning two to five speakers and six interaction types, evaluating End-of-Turn (EOT) and Interruption (INT) detection with exact causal annotations. The results reveal a stark capability asymmetry: turn completion is largely tractable, while selective interruption handling remains near chance even for state-of-the-art commercial agents.

Theory-of-Mind reasoning requires connecting an agent's perception to their belief and then to their action, yet large language models often fail to compose these inferences even when each part is known. This paper proposes rule-based reinforcement learning: using Group Relative Policy Optimization to train models to emit concise if–then rules that bridge perception, belief, and action before answering. The authors show that this approach outperforms free-form chain-of-thought reasoning and supervised fine-tuning across four ToM benchmarks, while using roughly half the tokens, and that rules from larger models improve downstream GPT-4o-mini accuracy.

Reinforcement learning for language agents suffers from sparse trajectory-level rewards that do not directly supervise environment dynamics. DGS-WM addresses this by introducing Decision-Guided Selective World Modeling: instead of forcing the policy to predict entire environment observations as auxiliary targets, it selects only observation units that support the subsequent policy response and discounts already-mastered targets via confidence-adaptive weighting. This selective co-training yields consistent gains over full-observation counterparts across QA and web-shopping benchmarks.

The paper asks how much meaning language models can actually share through language alone. To measure this, the authors introduce two tools — an epsilon-estimation algorithm that finds near-optimal prompts and estimates how much performance is left on the table, and a reconstruction bottleneck that scores how well a receiver model recovers a sender model's concept from labeled examples alone. They then apply observational scaling laws across 1.2k qualitative concepts in eight domains, classifying each as already articulable ('reaches'), potentially articulable with stronger models ('rising'), or permanently tacit ('bounded').

Multimorbid prescribing is a high-stakes coordination problem: patients with multiple chronic conditions receive fragmented, single-disease treatment plans that frequently trigger adverse drug events. PRISMATIC addresses this by treating prescription generation as a multi-agent synthesis task rather than monolithic text generation. The system deploys six specialized agents—ranging from information extraction to safety and rationality critics—on a shared blackboard, coordinated by an inner iterative-refinement loop (DFPC) and an outer long-term weight-adaptation loop (DSWA). The work matters because it reframes clinical AI safety as an explicit multi-objective coordination problem, moving beyond end-to-end LLM output toward modular, auditable reasoning.

The paper tackles the token inefficiency of visible reasoning traces in Chinese LLMs, which are typically generated in modern vernacular Chinese. It proposes Wenyan Thinking, a method that rewrites intermediate reasoning into Classical Chinese (wenyanwen) to produce more compact traces while keeping final answers short and benchmark-compatible. The work matters because it treats linguistic register itself as a controllable compression axis for reasoning, offering a practical way to improve accuracy under tight generation budgets without architectural changes.

Reinforcement learning with verifiable rewards suffers from coarse scalar supervision and moving-target instability in self-distillation. FSPO eliminates the persistent teacher model by storing sparse historical next-token distributions as fixed function-space targets, combining them with privileged disagreement correction and reliability-adaptive grounding. Across scientific reasoning and tool-use benchmarks, it achieves roughly 4% higher average accuracy than SDPO under a 5-hour budget while cutting peak memory from about 70 GB to 56 GB on Olmo3-7B-Instruct.

As large language models proliferate, identifying which model family generated a given text must remain reliable even when the domain or model version changes at test time. This paper proposes ATLAS, a generative framework that learns source-conditioned text distributions through lightweight adapter tuning and attributes text by comparing clipped token-level likelihoods. A new benchmark, MAABench, demonstrates that this likelihood-based approach outperforms direct discriminative classifiers under distribution shifts, while also enabling identification of distillation teachers.

The paper tackles the months-long lag in confirming US overdose deaths by predicting suspected overdoses directly from death investigation reports available days after a fatality. Its core idea is an expert-guided pipeline that extracts verbatim evidence spans—feature-guided rationales—to support classification, comparing them against templated verbalizations and against joint self-rationalizing models. The best configuration improves over a deployed rule-based system by roughly 15% relative macro-F1 while preserving expert oversight, offering a plausible path to faster public-health surveillance.

This paper tackles the problem that LLM agents often return replies which a human requester cannot interpret because they rely on private working context unseen by the user. The authors define reply usability as the ability of a requester to understand and act on a reply from only what they have seen ($R$), and measure it with an LLM-based reader model that flags information gaps. Across eight current models, 54% of 480 replies fail, with even the best model failing 30% of the time. They show that appending explicit requirements or rewriting with a secondary model substantially lowers failures, making usability an actionable target for agent design.

Typed decision models return probabilities over fixed answer choices instead of generating free text, making them a natural fit for tabular prediction tasks. This paper evaluates four such models against three large language models on row classification, fact verification, and anomaly detection, subjecting every request to 21 content-preserving rewrites. The authors find that decision models match LLM accuracy at far lower cost and are far more robust to how questions are phrased, yet no model escapes fragility when the table layout is altered.

Simulated students let developers test tutoring systems before real learners are available, but LLM role-play suffers from competence leakage because the model substitutes its own knowledge for the learner's. HACHIMI-II counters this by recasting simulation as third-person prediction: it builds a student model, infers a fine-grained learner state for each step, and compiles that state into a deterministic behavior harness that constrains generation. On five datasets, this cuts the ability-distribution gap by 2.4–8.4× and more than doubles error-structure fidelity over the strongest baseline.

Existing retrieval-augmented MT applies retrieval uniformly, yet LLM parametric knowledge suffices for many inputs and external evidence is only conditionally useful. This paper proposes SRAT, an agentic RL framework that treats reasoning, retrieval, and translation as learnable actions optimized via GRPO for final translation quality. It learns when to retrieve, what to query, and how to integrate evidence, supported by two retriever-alignment strategies and evaluated across idiom, knowledge-intensive, and document-level settings. The work is important because it challenges the default assumption that retrieval is always beneficial and shows that adaptive, quality-aligned evidence acquisition consistently outperforms fixed pipelines.

Data augmentation for document-level NLP risks breaking inter-sentential coherence and invalidating sentence-aligned labels. This paper proposes an LLM-driven discourse parsing framework that preserves sentence order, count, and rhetorical relations $R_{i,j}$ while replacing key phrases $\mathbf{K}_i$ to generate augmented documents $\mathbf{D}'$. By keeping the $n$-sentence structure fixed, labels such as summary indices or QA spans can be transferred without expensive LLM relabeling.

Language models frequently derail on multi-step reasoning because a single early error propagates through the rest of the generation. REWIND attacks this by treating token-level entropy as a diagnostic: it traces backward from high-entropy peaks to the sharpest local gradient jumps, uses those positions as cut points for suffix regeneration, and filters proposals with a likelihood-based acceptance rule. The result is a targeted, test-time repair loop that requires no fine-tuning or external verifier and that the authors show improves accuracy across multiple model families and reasoning benchmarks.

Existing agent benchmarks generally assume cooperative, emotionally stable users, overlooking a major source of real-world failure. This paper introduces τ-Emotion-Bench, a ten-domain benchmark that evaluates tool-calling agents under controlled emotion-conditioned user interactions, together with Tracer, a pipeline that collects training trajectories through an interaction-aware harness where dissatisfied users can terminate inefficient exchanges. It matters because it reveals large reliability gaps in frontier models and shows that small models can be trained to substantially improve interaction consistency with far less data than prior methods.

Legal reasoning is not merely the mechanical application of rules to facts; it is shaped by a judge's underlying legal philosophy. This paper introduces LePhilBench, the first benchmark to test whether LLMs can both understand jurisprudential tendencies and apply them to predict judicial outcomes. It pairs jurisprudential profiling with philosophy-aware judgment prediction and grounds the tasks in a dataset of 220 annotated U.S. Supreme Court opinions, offering a rigorous new lens on models' legal reasoning capabilities.

Multilingual speech technologies still fail to reach most of the world's languages because existing corpora are limited to a few hundred well-resourced languages and narrow domains like parliamentary proceedings or Bible readings. ELARSpeech addresses this gap by curating 30,202 hours of speech across 675 underrepresented languages from the Endangered Languages Archive, yielding both labeled and unlabeled subsets with rich acoustic and communicative diversity. It matters because it provides a large-scale, publicly available testbed that reveals existing self-supervised speech models struggle on the true linguistic long tail despite massive pretraining.

KoAGE investigates whether large language models can know, learn from context, and coin Korean neologisms. The authors introduce NeoTrace, an automated pipeline that collects and dates 4,271 neologisms across six genres, and propose a cutoff-aware benchmark that separates lexical recall from contextual inference and generation. By comparing model outputs with attested human coinages under original materials and simulated community settings, the paper offers a temporally grounded view of how models adapt to evolving vocabulary.