Papers

Long-horizon agents have drawn growing attention for their ability to work autonomously – from hours to days – on economically valuable tasks, a capability essential to automated discovery and recursive self-improvement (RSI). Yet open-source models still struggle on long-horizon tasks, largely for lack of both training data and a working recipe, and scaling both is hard: long, complex tasks are difficult to curate and verify, and their rewards are sparse. We address both. First, we introduce Tev-Gym, the first long-horizon agent training dataset with verifiable, dense, progressive rewards – 1,170 verified task environments across 7 domains. It is built on a latent-world-first approach that first constructs a latent world: a complete and correct hidden specification of the task domain, workflow, and expected state after each step. Each task is then progressively rendered from this specification into an agent-facing environment, enabling controllable complexity and a natural progress metric. Second, we turn this metric into a training recipe: via progress-potential reward shaping (PPRS), we convert progress into a dense per-step reward that improves training efficiency and makes otherwise-intractable tasks trainable. Training with this recipe yields Tev-27B, which achieves substantial gains on the Agents' Last Exam computing and math sciences domains (ALE-CM), matching GPT-6 Astra and Claude Opus 5 at Pass@5. Despite training only on ALE-related data, it generalizes to the Terminal-Bench (TB) 3.0 and 4.0 subsets, surpassing GLM 5.2 and Gemini 3.7 Flash, and reaches Claude Opus 4.8-level performance on middle-horizon tasks such as TB 2.1.

Reasoning models achieve strong performance through long chains of thought (CoT), but often spend more tokens than needed to solve a problem. Existing ways to shorten the CoT require additional training for each model, add a fixed steering vector to the model's hidden states, or use decoding heuristics such as stopping early or banning words like "wait", which can reduce accuracy. We propose DoubtLess, a training-free decoding method in which a much smaller reasoning model steers the large one towards brevity. The small model reads the large model's CoT so far twice, once prompted to be brief (the brief thinker) and once prompted to be verbose (the verbose thinker), and the difference between its two logit vectors, scaled by a strength, is added to the large model's logits at every step of the CoT. On seven benchmarks of competition mathematics, graduate-level science and competitive programming, DoubtLess shortens the CoT of Qwen3-32B by 31% with unchanged accuracy, outperforming five inference-time baselines. We also show that DoubtLess reduces thinking tokens per turn by 20% in a self-debugging coding agent with unchanged task performance. It also generalizes to two other model families with unchanged accuracy. When combined with training-based compression, it further shortens the CoT of a model already trained for brevity. Ablations and behavioral analysis show that DoubtLess reduces repeated checking, with most token savings coming after the correct answer first appears in the CoT. Our code is available at https://github.com/leoluolol/DoubtLess.

How can we build an agent that thinks beyond human language, yet still acts through it in the external environment? Modern LLM agents still reason and interact through text, while latent reasoning offers an efficient internal medium but lacks native tool access and reasoning transparency. We bridge these two directions with Hybrid Reasoning Agents (HybridAgents), which dynamically navigate between latent reasoning and text-based interaction throughout a unified generation process. Architecturally, a lightweight HybridLink enables seamless transitions between text and latent modes, while the agent's continuous working memory preserves lossless hybrid context across multiple turns. These designs preserve tool-integrated agentic capabilities while making latent thoughts directly inspectable. Building on this architecture, we further develop dedicated post-training algorithms for hybrid reasoning: HybridSFT stabilizes tool-integrated hybrid generation by teaching the backbone to coordinate with HybridLink, then HybridRL scales reasoning by learning when to think in latent space and when to act through text. We evaluate HybridAgent at 3 model scales across 11 benchmarks covering mathematics, science, medicine, search, and agentic reasoning. Compared with frontier LLM Agents and agentic frameworks spanning both text-based and latent methods, HybridAgent consistently improves performance by 10.4%, while delivering 35.5%-54.7% token reduction, 1.8x-2.6x faster inference, and 2.7x-3.7x cost savings. Code and data are available at https://github.com/togethercomputer/HybridAgent.

Earth observation (EO) models are still built by hand. For each new task, experts design a training recipe and tune it against a benchmark. We show that this loop can be handed to an agent. Any verifiable task can be cast as a harness and a verifier, where the harness defines what the agent may use and the verifier defines what counts as success. Every EO task can be made verifiable, through labeled benchmarks, independent sensors, physical models or the repeated observation of the same place. Given only these two components, a general-purpose coding agent builds the EO model and improves it recursively on its own outputs. Settings long treated as separate research problems, from learning with a handful of labels to adapting across domains, become choices of harness and verifier. On six EO benchmarks spanning optical, SAR and multispectral time-series imagery, with outputs that range from land-cover maps and ship instances to canopy heights and answers about large images, agent-built models surpass the state of the art on every one, and on land-cover mapping, they do so from 11% of the training labels. This establishes a new paradigm for EO, in which people design harnesses and verifiers and agents build the models, so EO models improve as agents do.

With the rapid development of multimodal large language models (MLLMs), evaluating their video understanding capabilities has become increasingly important. Existing video understanding benchmarks cover only part of the evaluation dimensions, limiting systematic diagnosis of omni-modal video understanding. We introduce OmniCompass, a holistic benchmark for omni-modal video grounding and understanding. OmniCompass defines two evaluation tracks: OmniCompass-VL assesses vision-language models on tasks requiring visual evidence, while OmniCompass-AV assesses omni-modal models on tasks requiring joint audio-visual evidence. Each track includes three complementary tasks: temporal grounding (TG), spatial grounding (SG), and question answering (QA), assessing when a referenced target appears, where it is, and what happens. QA covers five perception and three reasoning categories in OmniCompass-VL, and three perception and two reasoning categories in OmniCompass-AV. OmniCompass-VL contains 820 videos and 2,460 instances across domains and three duration types: short, medium, and long. A narrative- and audio-rich subset forms OmniCompass-AV with 240 videos and 720 instances. Each instance is evaluated under two paired settings: text-only and text+image, specifying the same target through text and visual-language guidance. All instances are manually annotated with multi-round quality verification. Evaluations of 50 MLLMs show that open-source models still lag behind proprietary counterparts, particularly in audio-visual understanding. Precise temporal and spatial grounding remains challenging even for leading models, and their predictions still exhibit inconsistencies across the paired target specifications.

Large language models (LLMs) generally rely on sampling temperature to regulate the amount of randomness in their outputs. Mainstream methods for selecting temperature typically involve fixed, heuristic global settings and context or token-level calibration, rather than a geometrically interpretable representation of the communicative scenario. In this work, we introduce a simple and interpretable temperature selection framework, the Scenario-Actor, which adapts decoding to the inferred situational characteristics. Scenario-Actor extracts scenario-relevant structures from Transformer hidden states and determines the temperature from scenario embedding, a structured geometric representation of the communicative setting. Empirical results show that on Llama-3.1-8B-Instruct, Scenario-Actor achieves strong goal-completion performance in interactive social scenarios, significant gains on open-ended instruction following, and competitive performance on multi-turn conversation compared to multiple baselines.

Pre-trained Large Language Model (LLMs) are effective on many tasks in zero-shot and few-shot settings. For machine translation of low-resource languages, the obvious options are few-shot prompting and fine-tuning on a small parallel set. We ask: what if moderate amounts of low-resource monolingual data are also available? We argue that, rather than continued pretraining, a more effective use of this data is the pre-LLM era technique of back-translation (BT). The current fine-tuned LLM translates monolingual sentences in each language, and the resulting pairs train it in the reverse direction. Updates on the parallel set are interleaved for stable learning. With 1K parallel pairs on Odia and Swahili, BT beats continued pretraining into the low-resource language by 7.7 to 16.4 XCOMET-XXL points in domain. With beam-decoded intermediate translations, it also beats a model with 10 time more parallel data. In-context examples retrieved from the parallel data further improve BT. We also analyze the source, quality and amount of monolingual text. Across five low-resource languages, with Hindi as a high-resource control, and two 3B backbones, BT improves translation into the low-resource language by 16 to 173% XCOMET-XXL over parallel-only fine-tuning. Blind human evaluation on Odia and Sanskrit agrees with the automatic metrics.

Harness self-evolution revises an agent's software scaffolding, including prompts, memory, workflows, and executable components, while keeping model weights fixed. We formalize a feedback-driven process in which accepted revisions persist across runs. The framework separates update targets from proposal and admission mechanisms. It distinguishes changes to interpreted state from revisions to tools and code, repository-level control logic, and plugin composition. We characterize search through evolutionary optimization and examine its reliability through continual PAC-learning. Under explicit assumptions, the analysis yields an expected generalization bound, a retention bound conditional on historical regression budgets, and a high-probability admission rule using fresh, independent validation data. We examine how persistent harness revisions affect knowledge sharing and organizational workflows, with implications for scientific reliability, physical safety, and institutional accountability. Cases from prior work illustrate model–harness co-design, distributed execution, and plugin reuse. Reliable deployment requires continual validation as task distributions and evaluators change, with independent safeguards for learning, runtime privileges, and consequential actions. We provide a comprehensive 66-page discussion of related work, theory, algorithms, evaluation, applications, and the broader impact of Harness RSI, together with continuously updated code and supplementary materials, at https://github.com/zgcagi/Harness-RSI.

Agentic reinforcement learning (RL) enables large language models (LLMs) to learn from multi-turn interactions with verifiable outcomes. Textual skills can improve exploration by distilling prior interactions into reusable guidance. However, existing skill-based RL approaches often either assume skill injection is beneficial or rely on indirect feedback or rollout-intensive evaluation to assess its utility. Our pilot studies show that rollout-based utility estimation for skill credit assignment is costly and unstable under limited sampling, while skill benefits vary across candidates and depend on the current policy's competence on each query. To address these challenges, we introduce SkillScaffold, a policy-adaptive training scaffold that combines low-cost skill-utility estimation with selective skill injection. By reusing outcome-verified training trajectories, SkillScaffold generates candidate skills and estimates their utility through outcome-aware likelihood comparisons, without candidate-specific environment rollouts. The resulting utility guides whether and which skill to inject, retaining skill-free execution when no candidate is estimated to help. Joint training enables the Generator to produce skills tailored to the evolving Executor. Scaffolded RL encourages skill internalization for skill-free inference. Experiments on ALFWorld, WebShop, ScienceWorld, and $\tau^2$-Bench demonstrate gains over strong agentic-RL and skill-based RL baselines across both short- and long-horizon tasks. Ablations and further analyses support the importance of utility-based selection and abstention. Our code is available at https://github.com/Vaidurya00/SkillScaffold.

Frontier language-model agents can now improve robots on their own: they write control code, test it, and keep what works as skills and memories for the next task. We introduce SEAR (Self-Evolving Agentic Robotics), a framework and benchmark built around two questions. Is the agent really learning, or just trying more often? And does what it learns still help on new tasks, in a new simulator, and on a real robot? The framework runs an agent in three modes: it writes code that becomes the policy (code-as-policy), acts live in the control loop (agent-as-policy), or works as a researcher that fine-tunes a vision-language-action model (auto-research). The benchmark adds controls that separate retained experience from repeated attempts, familiar tasks, privileged simulator state, larger budgets, and luck. Across seven frontier language models, 330 simulated and five physical tasks, and over four years of agent runtime, we highlight three findings. First, the learning curve of a robotic agent depends both on the overall time budget and the task difficulty relative to model capability. Harder tasks make longer sessions worthwhile; on easier tasks, re-rolling short attempts does just as well. Second, control code and shared routines transfer across tasks and simulators, while task-specific solutions do not. On a real robot, we observe a performance gap between agent-designed practice and novel human-designed scenes. Third, different models evolve more effectively in different ways. Some do best acting live, others writing code first. And a text-only model keeps pace with a comparable multimodal one: perception helps with error assessment, but agentic reasoning and tool use can matter more. We release the benchmark and codebase to support future research.

LLM ensemble gains are often measured against a single baseline, but the value attributed to additional candidates can change with the comparison used. We study an LLM aggregator under four matched conditions: Direct (no candidates), Single-slot (one candidate in a multi-slot interface), Duplicate (one candidate repeated), and Multi-candidate (separately generated candidates). Under either non-diverse control, overall candidate gain (CIV) splits algebraically into an additional-candidate contrast (DSV) and a candidate-context contrast (CCV); the split, but not CIV, depends on the chosen control. On MultiHopRAG model diversity (k=2), the repetition gap Rep=-8.33 pp [-13.25, -3.75] quantifies the difference between gains measured against Duplicate and Single-slot. On a matched FRAMES stronger-panel extension (n=200, k=3, aggregator output limit 4096), paired Delta Rep=+9.67 pp [+6.00, +13.50]; within panels, Rep changes sign (original -6.33; stronger +3.33). In FRAMES evidence aggregation, candidate-context contrasts remain strongly negative under both non-diverse controls. Our results show why ensemble evaluations should state the reference control before attributing gains to additional candidates.

Long-horizon agents accumulate progress in persistent task state, yet repeatedly process interaction histories describing how that progress was achieved. This raises a natural question: can long-horizon context instead be organized around reusable progress rather than the history that produced it? In this paper, we introduce CarryOn, an object-oriented context runtime that organizes continuation around reusable task objects. Rather than treating compressed histories or reconstructed accounts of progress as the primary basis for continuation, CarryOn builds a context view from continuation carriers, complementary execution context, and a live tail of recent interactions. A core challenge is keeping this reusable basis valid after its surrounding history is retired: how can CarryOn update the basis without reconstructing the full representation at every fold? CarryOn maintains carrier records across folds, revising affected records against bound evidence and applicable environment checks while carrying forward unaffected records and jointly updating execution context. Experiments on OSWorld 2.0 and BrowseComp-Plus show improved gains of 5.8% and 7.0% while reducing total token usage by 51.9% and 22.9%, respectively. Project page: https://ethannju.github.io/carryon

Large language model multi-agent systems extend individual agents' capabilities through collaboration across specialized roles, tools, and workflows. Yet collaboration makes failures harder to diagnose: one agent may misinterpret a correct result while others carry the mistake forward. Failure attribution therefore seeks to identify the responsible agent and decisive error step. Recent methods trace dependencies, but knowing where information came from is not the same as knowing where it went wrong. We introduce BLADE (Blame Localization via Audited Dependency Evidence), a training-free framework that moves beyond tracing information to auditing its use. Using Jev for structured probabilistic judgments, BLADE uses dependency probabilities to select relevant source-decision handoffs for auditing. At these handoffs, it examines whether receiving decisions use their sources faithfully and within task requirements, distinguishing error introduction, propagation, and repair. These findings inform a global diagnosis that considers all execution steps. Without reference answers, BLADE achieves the best or tied-best step accuracy among eight strategies under a shared Jev decision interface on 1,228 failed runs across five Who&When and Who&When Pro subsets, outperforming All-at-once by 2.2-11.1 percentage points. Making responsibility evidence explicit offers a path toward more interpretable and reliable agentic systems.

Rubric-based reward models (RMs) evaluate responses against explicit criteria. However, existing methods often derive criteria from observed response candidates, introducing response-induced bias that may limit rubric generalizability. Their judges also struggle with mechanical constraints when relying on reading alone, and commonly used pairwise judges provide only relative preferences, complicating their direct use for policy optimization. We propose ReVeR, a pointwise rubric RM that derives hard rules and graded principles by reasoning about the instruction alone and evaluates responses independently through tool-augmented judging. Both components are trained through label-blind distillation, in which preference labels only filter a teacher's outputs, yielding ReVeR-Traces, a large-scale dataset of rubric-reasoning and judging trajectories. Experiments demonstrate that ReVeR outperforms pairwise rubric RMs in average preference accuracy, most clearly on instruction following, as well as in test-time scaling and downstream policy optimization. We further find that generator RL improves preference accuracy while reducing rubric fidelity, and that higher preference accuracy does not necessarily translate into better downstream policies. These findings highlight that rubric fidelity, preference accuracy, and downstream utility are distinct properties, shedding light on future RM research.

Evaluation for automatic crisis reporting should measure what practitioners need, not what is convenient to compare. Through LLM-assisted, author-approved content analysis, we derive a 15-slot information schema from professional humanitarian situation reports using qualitative content analysis, then evaluate 121 human and 261 LLM-generated reports across five disasters and three hazard types. Under google/gemini-3.7-flash on Cyclone Idai, the best machine arm is nominally above the empirical human ceiling (0.827 vs. 0.821), but the difference is statistically indistinguishable from zero (95% bootstrap CI $[-.049,.059]$; permutation $p=.852$), and the ordering does not replicate under openai/gpt-5.6-luna. On a blind 100-item Idai sample, the author agrees with the Gemini judge at 0.910 accuracy and unweighted Cohen's $\kappa=.799$. All four judges clear the plan's raw-agreement criterion of .7; only Gemini reaches .7 on chance-corrected $\kappa$. Across the three primary events, machine reports reproduce 5.4-11.0% of same-day official figures. Across those same events, raw exact-number matching links 14.3-19.6% of official figures and 85.4-90.1% of machine figures to the social stream; manual audit confirms 45/50 declared absences, 4/25 official matches, and 24/25 machine matches semantically. The issuing-provenance gap is positive in all 12 observed event-judge cells: report date, organisation, series, and authority were not supplied at professional-report coverage in any cell. On the three primary events, ROUGE-L is uninformative, BERTScore is negatively associated with completeness after multiplicity correction, and a generic LLM judge changes sign.

Reinforcement learning (RL) from unit tests is the common recipe for post-training code models, but a unit test only asks whether the final program passes. We analyze what this unit-test-only reward teaches a model during debugging and find that, while it raises pass rates, it also teaches the model to edit more and to hallucinate bugs in code that is already correct. We introduce PreciseCoder, a training algorithm that augments the unit-test reward with a signal for semantically verified, localized edits, which correlates with smaller edits and higher pass rates. On 4B, 9B, and 27B models, PreciseCoder reduces the average number of edit lines from 7-15to 3 at a pass rate on par with unit-test-only training. Trained only on single-line bugs, PreciseCoder can generalize to repository-level program repair, where it resolves more issues with smaller patches. Its precision also transfers from weaker to stronger models: given the reasoning trace of a 9B PreciseCoder model, frontier models such as GPT-5.6 Sol and Qwen3.8-Max gain 4.6-12.4 points in edit precision with no additional training. Linguistic analysis of the thinking traces shows that our reward encourages the model to reason about individual lines, name the buggy ones, and form consistent bug hypotheses.

Natural Language Autoencoders (NLAs) are proposed as oversight tools for language models: an Activation Verbaliser (AV) writes an English explanation of an activation vector, and an Activation Reconstructor (AR) maps the explanation back to a vector. Their value for oversight depends on a human monitor being able to tell a true account of an activation from a false one, given the Av's explanation. We formalise NLA training and NLA oversight deployment as two different referential games from Emergent Communication: training is a generative referential game, while oversight is a discriminative, multi-modal, partially-observable referential game played over a many-to-many relation between the monitored language model's activations and the accounts that are true of them. The gap between the two games predicts that AV's explanations will identify which activation they describe without discriminating its true account from closely matched false ones, since training never asks explanations to rule anything out. Using the NLA's own AR as a maximally-sympathetic surrogate monitor model, we experimentally confirm this theoretical insight on one released checkpoint: true accounts beat accounts of other activations in all 320 comparisons, but beat fluent rivals differing in a single claim only 60.2% of the time (std.err. 6.7%, chance at 50%, $n = 35$). A meaning-preserving rewrite by an uninvolved model costs only 2.1% of the distance to another activation's account, so the failure is not due to the NLA pair having developed private conventions in the AV's explanations. Our rival accounts test only transcript-level truths, which the monitored language model's conversation settles, so these results bound the easiest case and leave claims about the monitored model's own processing untested. We state conjectures that would make both results consequences of the framework.

Code-switching (CS), the alternation between languages within an utterance, is common in multilingual speech but underrepresented in pretraining corpora. Increasing CS coverage requires finding relevant utterances in large speech collections so that they can be included in training data. One way to find such utterances is to transcribe the speech and detect language alternation in the text, but this depends on obtaining accurate multilingual transcriptions. We propose a lightweight method that detects CS directly from speech. It uses frame-level language information in pretrained speech encoders to determine whether both languages are present within an utterance. Specifically, we identify language-selective hidden-state dimensions using monolingual recordings and aggregate their frame-level activations into a CS prediction. With a small CS set used to tune a decision rule, we show that the Whisper-small encoder achieves a mean F1 of $0.798$ across 14 language pairs in CS-FLEURS, compared with $0.776$ for a CNN and $0.751$ for a HuBERT linear baseline trained for monolingual language identification. Our findings suggest that pretrained speech encoders already encode the information needed to identify CS utterances, and they provide a lightweight tool for building more representative multilingual corpora.

Large language models rely on prompt context to perform in-context learning, follow formats, and adapt to local evidence. This ability, however, also raises a robustness concern: models may reuse irrelevant tokens from earlier context when answering a later query. Prior work identifies this behavior as contextual entrainment and finds attention heads whose ablation reduces it, arguing that these heads are largely task- or relation-specific. But are the mechanisms behind contextual entrainment truly isolated by relation, or do relation-conditioned searches recover reusable components shared across relations? Using factual-distractor relation-completion prompts across 19 lexical-semantic subtypes, we identify attention heads whose ablation reduces entrainment and test how the resulting head sets transfer recovery ability across exact relations, semantic families, and mixed relation distributions. Same-family transfer exceeds different-family transfer, family residuals weaken once globally shared heads are removed, and subtype-specific searches repeatedly rediscover many global heads. Clean-completion and context-use probes further suggest that this shared intervention can reduce harmful copying while preserving ordinary relation completion and simple in-context abilities. Our findings provide a more precise perspective on contextual entrainment and its mitigation: harmful context following is not best addressed by searching for isolated relation-specific circuits, but by using heterogeneous attack prompts to reveal reusable intervention targets whose effects are shaped by semantic relation structure.

Biomedical language models (LMs) provide strong representations for clinical tasks, but they often fail to capture the hierarchical structure among medical concepts. This limitation is especially pronounced for rare and fine-grained clinical codes in long-tailed vocabularies, where hierarchy-agnostic LM representations become less informative as concept frequency decreases. We introduce Semantic Medical Identifier (SMI), a training-free framework for constructing ontology-grounded Semantic IDs and hierarchy-aware medical concept representations. For each diagnosis, procedure, or medication concept, SMI defines its Semantic ID as the coarse-to-fine path through an expert-defined medical ontology. SMI then aggregates frozen biomedical LM embeddings across the corresponding ontology levels to form hierarchy-aware medical concept representations that preserve coarse-to-fine clinical semantics. The base SMI uses uniform aggregation across ontology levels, while a level-weighted variant selects global level weights on validation data. Our theoretical analysis shows that SMI yields more informative medical concept representations than hierarchy-agnostic LM embeddings, particularly for rare and fine-grained clinical concepts. Experiments show that SMI captures meaningful hierarchical structure and improves clinical relation prediction. On MIMIC-IV, SMI improves mortality prediction, readmission prediction, drug recommendation, and rare drug prediction. On eICU, SMI yields more robust patient representations under cross-hospital distribution shift.

Recent advances in zero-shot voice cloning systems make it possible to reproduce a recognizable voice from a short recording and attach it to arbitrary virtual characters. Prior work suggests that familiar vocal identity can carry over associated personality impressions to new targets such as robots or virtual characters. However, these works have mainly examined familiarity and personalization effects rather than directly testing whether a cloned familiar voice systematically transfers personality impressions to an embodied virtual character. Such transference raises the concern that cloned voices could be used to manipulate people into engaging with virtual characters as capable social actors. Across two studies, we investigated whether personality impressions transfer when a virtual character speaks with the cloned voice of a familiar person. Study 1 examined this question in a real-world university setting using the cloned voice of a lecturer familiar to the participating students. Study 2 examined the same question online using voices of social media content creators and members of their communities. Drawing on the concepts of identity activation, trait transfer, and person reconstruction, we show that evidence of personality transfer was limited in both studies. Observed effects were selective and did not generalize across traits or study contexts. In the online study in particular, recognizable voice identity cues activated expectations about the speaker but did not reconstruct the broader person impression. Qualitative responses further indicated that judgments were frequently shaped by perceived artificiality and by whether voice and appearance fit together. Our findings moderate common concerns that cloned voices straightforwardly manipulate people into perceiving artificial characters as capable social actors.

Language model systems repeatedly make test-time decisions by comparing similarity: they retrieve exemplars, rank candidates, and revise responses. Yet similarity is not unique: lexical, semantic, structural, and stylistic resemblance can point to different candidates, and the task-aligned notion can change across instances. We introduce BASIS (Bayesian Axis Selection via Interactive Signals), a training-free Bayesian method that uses weak test-time feedback to infer which criterion in a supplied, task-specific library governs the current decision. It maintains a joint posterior over candidate targets and criteria, updates it from scalar scores, pairwise preferences, or verifier outputs, and selects probes by expected information gain. In a finite noisy model, we characterize when the target-criterion pair is recoverable and show why fixed-criterion and fixed-mixture rules can fail under criterion shift. Across concept recovery, reasoning exemplar selection, and response refinement, BASIS improves six-probe concept-recovery Success@1 from 56.0% to 68.1%, reasoning accuracy from 72.3% to 73.1%, and refinement win rate from 61.4% to 63.8% over tuned static mixtures. It also exceeds compute-matched best-of-N selection on reasoning (72.1%) and refinement (61.2%). Gains are largest under criterion mismatch and with larger candidate pools, and matched controls indicate that adaptive criterion weighting explains much of the benefit, with joint target-criterion inference adding further value. BASIS selects only among supplied criteria and does not discover new ones.

Chemical language models (CLMs) are pretrained on large corpora of molecules with the aim of learning representations that transfer to diverse downstream tasks. Here, we systematically study the effects of pretraining using a series of such CLMs across 64 molecular property prediction datasets. We find that the downstream gains of pretraining in CLMs do not scale and performance saturates below gradient-boosting baselines. Across the regimes evaluated, only supervised pretraining on traditional physico-chemical descriptors consistently improves downstream performance for the CLMs tested whereas some forms of pretraining result in models worse than a randomly initialized encoder. While models can resolve molecular structure well, they do so no better than a randomly initialized model, suggesting that, at best, CLMs learn descriptor-like properties in addition to low-level token statistics during pretraining, but surprisingly little chemistry beyond what classical molecular features already capture.

Combining pretrained models is a common strategy for building AI systems, whether by aligning a vision encoder with a language model or merging fine-tuned checkpoints. However, finding compatible pairs typically requires costly empirical search, as compatibility is known only after alignment or merging. We propose COMPASS (COMPatibility ASsessment Score), a theoretically motivated diagnostic that estimates whether a model pair will be compatible prior to alignment or merging. Building on the observation that representational convergence is most robust at the level of local neighborhoods, COMPASS decomposes model compatibility into two cheap-to-compute axes: a pair-specific ease, defined as the $k$-NN overlap between the two encoders, and a paradigm-level ceiling, capturing the maximum overlap their training paradigms allow. Together, these two axes provide a ranking of candidate model pairs and prescribe how to act on each pair: (i) prioritize pairs that already agree and whose paradigms are compatible, (ii) invest additional compute on pairs with compatible paradigms but low initial agreement, (iii) switch the pretraining paradigm when it cannot support sufficient agreement, or (iv) deprioritize pairs that score poorly on both axes. We apply COMPASS to $72$ cross-modal alignment experiments and $392$ task-arithmetic model merges and show that compatibility estimates by COMPASS strongly correlate with downstream success, achieving average Pearson correlation of $0.91$ on cross-modal alignment tasks and $0.70$ on model merging.

Interactions happen universally and spontaneously in daily lives. Studying involvement in interactions helps us analyze who can influence who and has a wide potential in applications. Specifically, we consider conversational involvement in free-standing conversations. We study two operationalizations of an FCG: F-formations and conversation floors, and propose that current assumption that F-formations capture single conversations is insufficient. To illustrate how different implementations could impact the detected F-formations, we employ three analyses. First, we propose that speaking status can be used for estimating conversation floors, and leveraged it to analyze floors within F-formations. Second, we analyze spatial characteristics of the detection results by applying different feature sets on chosen baselines, and find that groups resulting from head features are more likely to be subsumed by groups detected using other features. Finally, we do a joint analysis which combines the spatial features and floor patterns in the conversation. We observe that certain pipelines can detect groups with single floors from upper body features, while retaining F-formations with lower body features. We conclude that a more detailed theory and framework is needed to holistically demonstrate conversational involvement within interactions.

Searching text-rich knowledge graphs requires following relations while examining textual evidence. Predefined graph interfaces constrain what evidence agents can access, even when they write code to combine tool calls or process retrieved subgraphs. We propose direct knowledge graph interaction (DKGI), a harness in which language model agents create graph access tools directly over complete node and edge files. The agent writes Python code to follow relations, combine node sets, and select textual evidence, interleaving execution with shell search and file inspection. This lets agents process large node sets in bulk while returning selected text for the model to assess. Execution feedback guides tool creation and revision, while the evolving tools determine what evidence the agent gathers next. On benchmark datasets spanning three domains, training-free DKGI exceeds the strongest published baseline by at least 6.7 Hit@1 points on every graph, even with a minimal harness. Controlled ablations show the benefits of multi-turn programming over shell-only access and predefined graph tools, while trajectories show agents revising code to extend searches or address errors. These findings highlight the value of allowing tool creation and tool use to co-evolve during graph search. Our code and data are available at https://github.com/leopoldwhite/DKGI.

Self-distillation is an emerging approach that trains a model to imitate its own predictions when conditioned on privileged information, such as correct answers or environmental feedback. Recent methods have reported better sample efficiency and generalization than reinforcement learning (RL), but it remains unclear whether these gains extend beyond in-distribution evaluations and whether they stem from privileged distillation. In this paper, we revisit two representative methods, Self-Distillation Policy Optimization (SDPO) and On-Policy Self-Distillation (OPSD), across code, science, tool use, and math reasoning tasks. We find that their reported gains do not consistently generalize to unseen problems and are largely explained by overfitting or transferring existing reasoning behavior, rather than effectively using privileged information. Moreover, self-distillation gradients show weak alignment with policy gradients that directly optimize expected reward, questioning whether they provide an effective direction for policy improvement. Our experiments across Qwen3 models (1.7B-14B) and OLMo-3-7B suggest that effectively leveraging privileged information to improve reasoning remains an open challenge. Our code is available at https://github.com/sleepypandu/does-sd-improve-reasoning

Legal question answering requires answers that are not only plausible, but also grounded in legally valid and citable evidence. Standard retrieval-augmented generation often relies on one-shot semantic matching, which is insufficient when legal evidence must be selected according to authority, applicability, and diversity across heterogeneous sources. We introduce Legal-R1, an agentic retrieval framework for evidence-based legal reasoning. Legal-R1 decomposes legal reasoning into a multi-agent workflow, where a Lawyer agent learns when to retrieve, which legal source to query, what evidence to seek, and when the collected evidence is sufficient. A Legal Assistant agent retrieves from normative legal sources and precedent cases with authority-aware and diversity-aware selection, while a Judge agent generates the final answer from the complete evidence trajectory. We post-train only the Lawyer agent with masked supervised fine-tuning and trajectory-aware reinforcement learning. Experiments on multiple Chinese legal benchmarks show that Legal-R1 improves over general-purpose, legal-specific, and retrieval-augmented baselines.

Dialogue and collaborative assistants repeatedly judge what a partner knows, believes, or intends. Typed decision models offer a potential alternative to generative large language models for bounded Theory of Mind (ToM) selection. We present Jev-ToM, comparing five typed decision models and seven generative baselines on 11,117 English text questions from nine ToM benchmarks across six capability groups. TypeSafe Jev reaches 73.2% source-balanced accuracy including native controls and ranks first or second across all six capabilities, leading beliefs, goals/desires/intentions, and emotions. Other typed models show distinct strengths across tasks. Separate resource comparisons suggest that Jev could reduce the cost and waiting time of repeated ToM selection. Confidence-based selection remains benchmark-dependent, and supplied reasons can repair errors or overturn correct answers depending on their attachment. Candidate expansion reveals distinct selection and output-validity problems. Related-question cues account for most of Jev's downstream gains, with task-dependent effects from predicted states. The findings motivate testing Jev as a ToM selection component in assistants, with task-specific confidence calibration, candidate design, and output validation as integration priorities.

Large language models increasingly tackle tasks that require multi-step reasoning, where errors in intermediate claims can affect the final answer. Self-consistency (SC) improves reasoning by voting over sampled final answers, but its vote leaves intermediate disagreements unresolved. To address this problem, we introduce DDRG (Disagreement-Driven Verification Guided by Reasoning Graphs), a framework for selective correction from a frozen bank of reasoning traces. Our insight is that graph alignment can expose claim-level disagreements and provide correction evidence without expanding the sampled trace bank. DDRG builds source-grounded claim graphs and sends localized conflicts to an answer-blind verifier. A verifier-supported refutation that reaches the majority answer's supporting path triggers one focused repair; otherwise, DDRG returns the exact SC vote. Experimental results across five benchmarks show that native zero-shot DDRG improves macro-average accuracy over direct answering by 8.65 and 15.83 percentage points for Qwen2.5-7B and Llama3.1-8B. It also exceeds CoT, the evaluated GoT and GoV question-answering adaptations, and SC on this macro metric. DDRG links local verification to selective correction without enlarging the initial candidate bank.