Zhiheng Qian, Junhao Du, Jing Peng, Hanqi Li, Kunyang Peng, Feng Xue, Shuai Wang, Kai Yu
Written transcripts of the same speech can differ in ways that are not uniquely determined by the speech signal. We formulate evaluation-side transcript canonicalization (ETC), which maps multiple surface forms to a single canonical form when their differences are not determined by speech. To our knowledge, ETC-Bench is the first and most comprehensive multilingual benchmark for evaluation-side transcript canonicalization, covering 181 phenomenon types across 12 languages. Each instance is a minimal pair that shares a sentence context and differs only in a target span exhibiting one phenomenon. To evaluate both canonicalization success and semantic preservation, we propose Semantic-preserving Exact-match Normalization Accuracy (SENA), which combines exact agreement between independently canonicalized pair members with a pair-relative BERTScore constraint against information loss. We further introduce the Taxonomy-guided Agent Canonicalization Skill (TACS), which provides LLM agents with phenomenon descriptions and explicit canonicalization contracts. Experiments with five baseline configurations and three TACS-based agents show that the agents outperform the baseline systems on shared languages. ETC-Bench provides a systematic testbed for evaluation-side transcript canonicalization across languages.
Haolin Chen, Deon Metelski, Leon Qi, Tao Xia, Joonyul Lee, Steve Brown, Kevin Riley, Frank Wang 等 33 位作者
End-to-end automation of realistic healthcare operations stresses three capabilities underrepresented in current benchmarks: policy density, decisions must be grounded in a large library of medical, insurance, and operational rules; Multi-role composition: a single task requires the agent to play multiple roles with handoffs; and multilateral interaction: intermediate workflow steps are multi-turn dialogs, such as peer-to-peer review and patient outreach. We introduce $χ$-Bench, a benchmark of long-horizon healthcare workflows across three domains: provider prior authorization, payer utilization management, and care management. Each task hands the agent a clinical case in a high-fidelity simulator of 20 healthcare apps exposed via 87 MCP tools, which it must drive to a terminal status through tool calls and writing the role's artifacts, guided by a 1,290+ document managed-care operations handbook skill. Across 30 agent harness/models configurations, the best agent resolves only 28.0% of tasks, no agent clears 20% on strict pass^3, and executing all tasks in a single session slumps the performance to 3.8%. These results raise the hypothesis that similar gaps are likely to surface in other policy-dense, role-composed, irreversible enterprise domains.
View- and target-specific echocardiography AI models capture useful but complementary information and may not achieve sufficient predictive performance in isolation. A learned ensemble is a natural way to combine these outputs, but heterogeneous models typically require each output to be converted into a binary decision, probability, confidence score, or fixed-length feature vector. We developed a supervised large language model (LLM)-based late-fusion module that directly aggregates view-specific predicted finding statements and predicted measurements in language space and learns a case-level fusion function. In patient-level three-fold cross-validation of 859 examinations from 646 patients, the proposed method integrated heterogeneous outputs within a single framework, performed comparably to the best-performing reference strategy evaluated, and outperformed a general-purpose LLM without task-specific training. These results suggest that a supervised LLM can combine multiple upstream models without model-specific output conversion or feature engineering that requires medical knowledge.
Multi-hop QA benchmarks often reward Large Language Models (LLMs) for spurious correctness, where models reach correct answers through invalid intermediate reasoning. We propose SAFE, an LLM-as-verifier framework for evidence-grounded multi-hop QA. Rather than judging only the final answer after generation, SAFE verifies reasoning during generation by checking intermediate steps against the provided passages and previous reasoning trajectory. To make this process checkable, SAFE decomposes reasoning into atomic, evidence-grounded units represented with Knowledge Graph (KG) triples. At train-time, SAFE verifies benchmark supervision under KG-grounded constraints and constructs reliable verifier training data. At inference-time, an external verifier checks each generated step, identifies invalid reasoning, and provides correction feedback before errors propagate. Across three multi-hop QA benchmarks, SAFE improves accuracy by 8.8 pp on average. These results show that evidence-grounded multi-hop QA benefits from shifting LLM-based evaluation from post-hoc answer judgment to stepwise reasoning verification.
Cyber threat intelligence (CTI) reports describe adversary behavior in natural language, but converting them into structured MITRE ATT&CK mappings remains difficult: existing methods decompose reports to sentence-level granularity for extraction, while modern LLMs can work with full reports in a single pass but with noticeable performance degradation. We present ARBITRA, which orchestrates multiple role-specialized LLM experts that propose candidate ATT&CK entities from the full report, which are then verified through a per-candidate quantitative bipolar argumentation framework (QBAF) grounded in document evidence and rationales. We additionally construct reasoning-specialized corpora for fine-tuning so experts can act as effective participants during this verification process. Across ATT&CK extraction benchmarks, ARBITRA reaches a stronger precision-oriented operating point than sentence-level pipelines, recent LLM-based extraction systems, and frontier models including GPT-5.5. Under adversarial perturbations simulating conflicting TTP statements, cross-report contamination, and semantic poisoning, ARBITRA limits mean precision degradation to roughly half that of frontier-model baselines for a more robust extraction.
Multimodal large language models can recognize visual content yet struggle to reason reliably over it across multiple steps. In counting, a model may see all objects but lose track of which ones it has counted. We argue that such failures, commonly attributed to perception, reveal a reference gap: the failure to bind language to the relevant visual entities, regions, or locations. Motivated by the theory of visual routines, we test whether external references such as numbering can narrow this gap. We introduce ReGap-Bench, a paired benchmark of five visual reasoning tasks in which the scene, question, and answer are fixed and only the reference varies. Results show that task-matched references improve accuracy by up to 25\%, while mismatched references provide little benefit or even hurt performance. These findings identify the reference gap as a bottleneck distinct from visual perception and show that constructing task-appropriate visual references can serve as a training-free visual skill for improving multimodal reasoning.
Large language models (LLMs) have shown strong performance on software engineering (SE) tasks, driving the need for high-quality datasets for reliable training and evaluation. SWE-bench-style datasets (SWE datasets) are widely used to evaluate LLM-based coding agents on real-world software issues, where task clarity and test coverage directly affect whether instances are solvable and verifiable. However, quality issues persist in existing SWE datasets, and manual annotation is costly and labor-intensive, motivating LLM-based annotation. Existing approaches typically rely on static, overly general rubrics, limiting their ability to handle diverse and nuanced scenarios and often yielding high false positive rates. In contrast, human annotators draw on accumulated experience to make context-aware judgments. Inspired by this, we propose ExLabel, an LLM-based, self-evolving, experience-guided assistant for task clarity and test coverage assessment in SWE datasets. ExLabel maintains an evolving experience bank that distills cross-instance knowledge into stable yet adaptive judgment criteria and reuses relevant experience to guide the assessment of new instances. For systematic evaluation, we construct SWE-TQ, a new benchmark spanning eight programming languages, by cleaning human annotations from SWE-bench and Multi-SWE-bench. Experiments with three backbone LLMs show that ExLabel achieves the best F1 across all settings, with relative improvements of 44.2%/49.9% and 24.0%/18.4% over the best baseline for task clarity and test coverage assessment on the Python/Multilingual subsets, respectively. Moreover, a coding model fine-tuned on SWE-rebench data filtered by ExLabel resolves 66.2% of issues on SWE-bench Verified, outperforming the model trained on data filtered by the best baseline by 14.5%.
Yi Wen, Xiangjue Dong, Bingxuan Li, Hangxiao Zhu, Leo Zhicheng Liu, Yu Zhang, James Caverlee
Deep research agents (DRAs) conduct long-horizon reasoning, search, and literature synthesis to provide long-form answers to complex, domain-specific questions. These agents are typically trained and evaluated on fully specified user queries. However, real-world user queries can often be underspecified. We introduce UNDERSPEC-DR, a framework for systematically evaluating how underspecified queries affect agents' final answer quality and intermediate search behavior. Across three DRAs we evaluated, underspecification substantially reduces how well DRA's answers fulfill users' information needs, even though they remain grounded in cited evidence. Our analysis of intermediate search behavior suggests a possible explanation: DRAs conduct broader searches that draw less on the original user query but rely more on their learned prior during training, which can be misaligned with users' intent. Our findings reveal the vulnerability of current DRAs to underspecified user queries and inform future efforts to build DRAs that proactively recognize when additional user input is needed to reliably complete deep research tasks.
When annotators disagree on a label, the disagreement itself carries signal - and the number of annotators needed to capture it depends on the evaluation metric. We fine-tune NLI models on label distributions subsampled from ChaosNLI, a dataset providing 100 independent annotator judgments per item, and identify metric-dependent saturation. In our 3-class NLI setting, entropy correlation - whether the model identifies which items elicit disagreement - requires N ~ 20-50 annotators to converge, while distributional match (KL divergence) saturates by N ~ 10 (87-95% of improvement across five model seeds); the specific thresholds are estimates for this setting, and it is the phenomenon, not the constants, that we expect to generalize. This finding rests on a prior observation: soft labels carry item-specific signal that label smoothing cannot replicate. Across five smoothing intensities, entropy correlation clusters at r ~ 0.45-0.49, while soft labels reach r = 0.643 (p < 0.001); per-item analysis traces this gap to smoothing's inability to distinguish ambiguous items from clear ones. The soft label advantage replicates across two architectures (DeBERTa, RoBERTa) and a non-NLI-pretrained baseline; differential saturation additionally appears in an exploratory cross-domain evaluation on content safety. These results suggest that annotation budgets should be informed by the target evaluation metric rather than set uniformly.
LLM agentic workflows increasingly employ different models for specialized roles. Users may prefer different trade-offs between task quality and deployment cost, sometimes also considering latency. Finding good assignments incurs a separate search cost from evaluating multiple candidates. We introduce Cost-Coupled Gittins (CC-Gittins), a multi-objective extension of single-objective cost-aware Gittins search for Pareto recommendation. It exploits the coupling between evaluation and deployment costs: assignments that cost more to deploy also cost more to evaluate. The search alternates between seeking higher quality and lower deployment cost, reusing each evaluation to inform both searches. After each batch, CC-Gittins recommends assignments that are nondominated under uncertainty-adjusted estimates of quality and deployment cost. We evaluate CC-Gittins on eight quality--cost matrices from workflows with two to five roles and up to 1,681 candidate assignments. We also evaluate the method with latency as a third objective. Across the eight benchmarks, CC-Gittins recovers a substantial fraction of the true Pareto frontier at roughly 10% of the exhaustive-evaluation cost, and recovers most of the frontier with few false positives by 30%.
Reasoning-trained language models can select, zero-shot, a small set of relevant labels from thousands to hundreds of thousands of candidates. We ask how they do it mechanistically, and whether the mechanism can be distilled. We treat each decision as a token-level event scored by the model's own decision margin: the token that picks a coarse region of the label space, the tokens that pick a label within it, and the token where the output departs from a near-miss named earlier in the reasoning. Attribution, exact mean-ablation, knock-in into another example's context and a null calibration give individual attention heads causal standing. On clinical coding of MIMIC-IV discharge summaries, with all 5,651 candidate codes in context, a small, global, phase-structured set of heads is necessary and sufficient on essentially every summary; distinct head families read the candidate region and the earlier near-miss, and blocking what they read removes their effect; for mentions elicitable in the reasoning, a disjoint mid-layer set makes the region elicitable several tokens before the code. The structure recurs in a second reasoner. We introduce MISTILL, which, unlike chain-of-thought distillation, also supervises the student's pooled attention at these decision events. Read on heads found after training, it nearly doubles the causal recovery of the contrastive decision in a cross-family student and adds a small, seed-stable gain in one that already carries most of it, with no detected task difference at matched precision; under matched selection it exceeds a frozen random-head pool on both recovery and task score. A pre-registered evidence-flip test detected no improvement in evidence-dependent decisions through the tested routes.
Chatbots are increasingly consulted about public affairs, and their judgments of policies are useful only if they are robust. We measure this robustness in the setting of an ordinary user, through consumer chat interfaces. For 30 days, five chatbot configurations (Mistral, Gemini Flash, Claude Sonnet, and Claude Opus at two effort settings) were asked twice a day to score the long-term effect on the United States of 40 actions of the last four presidential administrations and of 15 hypothetical reforms, each with a confidence and a short comment. The judgments are robust: all configurations agree on the sign and the ranking of the administrations' mean scores, and two different configurations asked on the same day give opposite signs to the same action in about 4% of cases. Responses from different days differ more than same-day responses, but the difference grows slowly over the month, except where the service itself changed. We identify three such changes, which shifted the judgments while leaving little or no trace in the interface. Treating the comments as a retrieval task, we find them specific to each item and consistent across runs and chatbots, with lexical and embedding similarities agreeing. We release the dataset of 16,500 judgments.
Toni J. B. Liu, Jiajun Bao, Gurbir Arora, Mitchell Ostrow, Yizhou Liu, Raphaël Sarfati, Nicolas Boullé, Christopher J. Earls
What does a language model predict when it has few clues?The answer lies in its unembedding geometry: a single direction of the unembedding matrix encodes the unigram distribution of the training corpus, which serves as the Bayesian prior the model falls back on when uncertain. This structure, termed the direction of ignorance, appears across all model families examined (Llama, Qwen, Gemma, and Pythia), ranging from 0.4B to 405B parameters. Projecting the final prediction state onto this direction yields a per-token prior loading factor $λ$, which declines steadily as the context becomes more informative. Formally, the same projection decomposes the prediction state into two orthogonal vectors interpretable as the two components of a tempered Bayesian update: a unigram prior raised to the exponent $λ$ and a context-driven likelihood. This geometric probabilistic connection makes $λ$ meaningfully comparable across model sizes and families. Larger models generally exhibit lower prior reliance, and reasoning models exhibit lower reliance than their base counterparts. Finally, we show that $λ$ serves as a causal, inference-time control: reducing it can promote context-driven continuations and improve Llama-3.2-3B-Instruct on GSM8K, whereas amplifying it or applying matched perturbations along random directions degrades performance.
A long-lived assistant that brings up memories without being asked must decide at every turn whether anything in its store is worth an interruption. We show that this decision is a multiple-testing problem whose difficulty grows with the store size N. On streaming replays of LongMemEval, a similarity threshold tuned to 5% false alarms at N=8 reaches 84-100% at N=2048 across embedding models and cue types, and LoCoMo and LLM judges show it too. We prove that a gate which thresholds the best of N independent irrelevant scores at false-alarm rate alpha operates at a per-item false-positive rate of about -ln(1-alpha)/N, which pins its hit rate to the far left of the single-comparison ROC curve and, under regular variation, makes it regularly varying in N. A finite-pool form of this bound predicts the hit rate of a label-tuned threshold on held-out questions to within 6.1 points. At matched false alarms the deployed score recalls 14-26 points more than raw cosine. A cue-swap conformal gate, which ranks the current cue among other cues on the same store, has a finite-sample false-alarm guarantee under exchangeability. With a background drawn from the user's unfiltered history it keeps false alarms at 3-6% on LongMemEval at the hit rate of a label-tuned threshold; on LoCoMo it overshoots or turns conservative depending on where its reference cues come from. In a chronological replay with the user's own earlier messages as references, it speaks on 6-8% of session-opening turns, against 57% for the deployed gate at N=2048, at about 50 ms per decision on one CPU thread. A cross-encoder reranker pays the same price of scale, and on a prospective-memory benchmark, similarity gates recall at most 9% of constraints at controlled false alarms. One human and two LLM annotators judge most false alarms unhelpful for a response. Code and data: https://github.com/RiverChenStillCompiling/price-of-scale
Research on medical item difficulty has largely focused on prediction rather than on identifying interpretable patterns associated with empirical difficulty. This study used empirical difficulty labels to guide large language models (LLMs) in inducing difficulty rubrics and examined how rubric guidance affected LLM difficulty judgments. Using training items and their empirical difficulty estimates from the United States Medical Licensing Examination (USMLE) and the Physician Assistant National Certifying Examination (PANCE), Claude Opus 5 induced separate rubrics for absolute and pairwise difficulty judgment. None of the induced rubrics treated surface features, such as stem length, option count, or images, as major sources of difficulty. Several patterns emerged across both examinations, including cue convergence, misleading cues, and option competition, although their specific forms differed. Retrieval of knowledge that could not be derived from the stem was particularly prominent in the PANCE rubrics, whereas the USMLE rubrics placed greater emphasis on integrating multiple clinical cues. We compared commercial and open-weight LLMs on held-out items under unguided and rubric-guided conditions. Rubric guidance changed how models used the difficulty scale but did not consistently improve performance. In absolute judgments, guidance encouraged broader use of the scale and generally improved coarse-grained agreement, without consistently improving rank ordering. In pairwise comparisons, its effects varied across models and examinations, and judgments became more sensitive to presentation order. These findings suggest that the value of label-guided rubric induction may lie more in generating interpretable, testable hypotheses from empirical difficulty data than in reliably improving difficulty judgments.
The decoding order of masked diffusion language models (dLLMs) can substantially affect generation quality, yet common decoders typically apply the same reveal rule throughout generation. Confidence provides a strong default, but it measures how decisively a position can be predicted now, not how useful revealing that position will be for subsequent decoding. We propose Cord, a lightweight prompt-conditioned policy that retains confidence decoding as a stable reference and learns when and how to depart from it through geometric residuals over generation direction, proximity to revealed context, and two-sided infilling. The residuals adapt to both the prompt and the evolving partial response. Rather than sampling a new policy action at every decoding step, Cord samples a single low-dimensional coefficient matrix per rollout, which induces an entire state-dependent reveal trajectory. This yields a contextual-bandit formulation trained directly from sequence-level rewards, without step-level credit assignment or backbone updates. Across three dLLMs and five reasoning, code, and constraint-satisfaction benchmarks, Cord improves average task success from 32.0% under confidence decoding to 48.2%, and further improves over the best per-domain fixed residual. Behavioral audits show that the learned residuals induce substantial, task-dependent changes in reveal locality and two-sided infilling. Cord adds only $0.53$M trainable parameters, requires no additional backbone forward passes, and incurs only 5.39% inference latency overhead.
Yuan Pu, Yudong Lu, Ting Wang, Zhicong Hong, Qinuo Li, Xuansheng Wu, Na Zou, Xia Hu 等 9 位作者
Energy-based models have shown promising properties for domain-specific tasks but remain challenging in modeling general-purpose scenarios. To bridge this gap, we design energy-steered models (ESMs) for foundational language modeling. ESMs integrate energy steering and token generation within a single architecture to leverage the energy-guided search to refine token generation. Specifically, ESMs first optimize a latent state over the energy landscape, then leverages the refined latent state for token generation. By performing such energy-guided pre-training on large-scale corpora, ESMs learn generally effective energy landscape where favorable tokens correspond to lower energy values. To explore the scaling behavior of ESMs, we pre-train ESMs on three large-scale corpora, then evaluate across 26 downstream tasks spanning a broad range of domains. Our results demonstrate that ESMs have promising scaling behaviors with increasing model size and training data, and outperform GPTs on most downstream tasks. These findings highlight ESMs as a promising architecture for large-scale pre-training and general purpose problem solving. We fully release the model weights and training infrastructure to enable the community to completely reproduce and extend our results. GitHub: https://github.com/datamllab/openebfm HuggingFace: https://huggingface.co/collections/guan-wang/openebfm
LLM-based search agents are widely used for information-seeking tasks, but their reliance on external tool returns introduces a critical security risk: web content retrieved during execution is untrusted, exposing agents to prompt injection and goal hijacking. Prior work on search-agent safety primarily focuses on static web-content injection, but modern agents issue follow-up queries and cross-check competing sources, so a single injected page is often diluted or rejected. We show that the channel delivering search and page observations is a fragile security boundary: beyond exposing the agent to a single poisoned page, a mediated search interface can repeatedly steer how the agent gathers evidence and forms its final answer. Under a constrained tool-intermediary threat model, appending only one controlled result per query can substantially increase attack success when the evidence is coordinated across the agent's trajectory. We study this setting with a strategy-driven long-horizon attack system and introduce Authority-Chain Hijack (ACH), an expert-refined strategy that turns isolated search-result and page-content manipulations into a coherent evidence chain across seemingly corroborating sources. ACH achieves the highest Overall ASR among all baselines, reaching 55.9% / 83.3% ASR / MaxN ASR on the full SafeSearch test split. We further introduce Trace-Guided Strategy Evolution (TGSE), which automatically improves attacker strategies from execution traces, replacing manual redesign with trace-driven refinement; its strongest single setting reaches 71.4% / 95.0% in held-out evaluation. Code is available at https://github.com/MBOSSLESS/Breadcrumbing-Search-Agents.
Sam Agnoli, Brigid Reilly, Lily Barna, Emőke-Ágnes Horvát
Google's AI Overviews (AIOs) are the first result users encounter when searching the web, shifting the work of selecting and evaluating sources from the user to Gemini. This may be especially consequential for health information, given that Google's traditional algorithm was designed to prioritize quality sources, and inaccurate information carries health risks. We evaluated 3,611 benchmark health questions, comparing each AIO against that question's top five traditional results on source selection and quality. We also evaluated the AIOs' statement grounding in their cited sources and how accurate AIOs were against physician-written rubrics. Overviews appeared for 96.4% of questions but drew on substantially different evidence than traditional search: only 44.4% of the top five results were cited by the corresponding AIO, which instead overrepresented commercial health publishers (e.g., WebMD, Healthline) at the expense of academic and government sources (e.g., Cleveland Clinic, National Health Service). AIO-cited pages scored lower than traditional results on four of six well-established credibility criteria. The grounding of AIO statements to their sources was poor: after correcting for classifier error, an estimated 56.7% of cited statements were supported by their cited page, and a further 20.8% of all statements were factual claims with no citation at all. When AIOs were graded against physicians' rubrics, they earned 38.2% of all available points (e.g., accuracy 54.2%, completeness 21.6%), and among the subset of AIOs whose physician rubric anticipated a specific inaccurate or potentially harmful claim, 22.2% of AIOs contained the harmful content. Notably, the accuracy of AIOs was unrelated to the grounding of their statements. We recommend AIO literacy interventions for users and holding AIOs at least to Google's own health-ranking standards.
Large Language Models excel at reasoning and generation but their autoregressive, token-by-token inference makes them unnecessarily slow, expensive, and poorly calibrated for the many routine decisions (routing, gating, scoring, classification) that modern AI systems require at sub-second latency. A growing class of models addresses this gap: System 1 decision models, architectures that map an input directly to a typed, structured decision in a single forward pass. This guide provides a comprehensive, practitioner-oriented treatment of the field. We propose a five-family taxonomy covering encoder-based classifiers, reward models, non-autoregressive decision engines, contrastive decision models, and extraction models, and we provide in-depth coverage of each family's architecture, training methodology, and deployment trade-offs. A central theme is calibration: we develop the theory of proper scoring rules, analyze Reinforcement Learning for Calibrated Decisions (RLCD), and examine recently discovered phenomena including IIA violations, ordinal-scale bias, and few-shot recalibration under distribution shift. On the integration side, we describe how System 1 models serve as decision infrastructure within agentic workflows and System 1 + System 2 cascades that route between fast and slow models based on calibrated confidence. The guide covers 18 application domains, maps the rapidly growing open-source ecosystem (over 70 model reproductions in weeks), provides runnable code throughout, and includes a practitioner's checklist, an open-source model zoo, and 22 self-assessment questions with detailed answers. Spanning 190 references, it is intended as a unified entry point for researchers and engineers working at the intersection of fast decision-making, calibration, and production AI systems.
Shu Wei, Jingjing Wu, Qunyi Xie, Lingshu Zhang, Xiangcheng Du, Yannian Fu, Mengjie Wang, Zuyi Wang 等 12 位作者
Unlimited OCR ushers the era of long-horizon parsing, substantially mitigating the growth of working memory as document lengthens. However, as document length scales to hundreds of pages, the visual context grows substantially, sharply increasing the computational cost of each decoding step and limiting the scalability of ultra-long document parsing. Inspired by human book transcription via localized focus without re-scanning preceding pages, we introduce Unlimited OCR 2, a document parsing model that looks only at what it transcribes. Building on Unlimited OCR, we upgrade all the decoder attention modules to Sparse Reference Sliding Window Attention (SR-SWA), bounding active reference tokens to a fixed capacity and thus sustaining constant computational costs across arbitrarily long documents. By combining SR-SWA with Unlimited-Horizon Decoding Paradigm, Unlimited OCR 2 enables document parsing over virtually unlimited lengths without prohibitive memory overhead. At the hundred-page scale, Unlimited OCR 2 achieves 7.0$\times$ and 5.6$\times$ per-step speedups over Full Attention and R-SWA, respectively. Furthermore, it delivers state-of-the-art end-to-end performance on both MPDocBench and MarathonDocBench, our newly proposed benchmark targeting hundred-page document parsing. Ultimately, the capacity for uninterrupted, one-shot document parsing signifies a fundamental paradigm shift into the OCR 3.0 era.
Developing seamless, high-performance, native intelligent full-duplex Spoken Language Models (SLMs) remains a critical challenge and long-standing goal for the speech and NLP community. Despite notable progress, recent endeavors are fundamentally constrained by severe modality interference, which causes substantial knowledge degradation and compromises semantic integrity -- ultimately making full-duplex SLMs feel unnatural and unintelligent. In this paper, through an exhaustive fine-grained analysis of model optimization dynamics, we uncover the root cause of such performance degradation, revealing that modality interference arises from inherent gradient conflicts between acoustic and semantic modeling when the two modalities are forced to share a deep parameter space. Guided by this key insight, we introduce Lychee-FD, a native end-to-end full-duplex framework designed to mitigate modality interference. Importantly, we propose a hierarchical parameter separation strategy that decouples conflicting modalities in deep layers while preserving cross-modality coherence via a dedicated semantic alignment channel. Extensive experiments on multiple full-duplex benchmarks demonstrate that our method significantly advances the state of the art, yielding substantial improvements in both speech intelligence (+7.4% on Spoken QA) and full-duplex interaction fluidity (+28.5% on FullDuplexBench 1.5) without compromising inference efficiency. To the best of our knowledge, this work is the first to achieve two key advances: 1) uncovering and elucidating the root cause of modality interference in full-duplex SLMs, and 2) designing an elegant hierarchical model together with a practical solution for seamless, high-performance, native intelligent full-duplex SLMs.
Recent research on Multimodal Sentiment Analysis (MSA) has focused on learning from language, visual, and acoustic modalities with incomplete data to infer human sentiment. Most studies typically compensate for missing information by reconstructing modality features or designing complicated fusion mechanisms. However, these methods still suffer from spurious generation and noisy guidance due to the lack of high-level semantic grounding in partially observed multimodal evidence. To address these issues, we propose SemMSA, a latent semantic-aided framework that constructs rich sentiment-relevant semantics with LLMs, fully integrating with all modalities via anchor-free spectral alignment. It mainly consists of Cross-modal Semantic Refinement (CSR) and Cross-modal Spectral Alignment (CSA). Specifically, CSR first adaptively extracts visual and acoustic representations by corresponding adapters to form a unified multimodal prefix with language in the frozen LLM embedding space. It then iteratively produces continuous discriminative semantic states through a token-efficient latent refinement process without decoding explicit text. Next, CSA simultaneously aligns the refined semantics with all modalities by enhancing the dominant spectral component of their kernel Gram matrix. This captures global nonlinear dependencies among all representations without relying on a predefined anchor modality. In addition, an instance-level spectral separation constraint preserves cross-sample discriminability and mitigates representation collapse. Extensive experiments on SIMS, MOSI, and MOSEI benchmarks demonstrate that SemMSA achieves state-of-the-art performance.
Generative recommendation (GR) represents items as semantic identifiers (SIDs) and autoregressively generates the next item's SID from interaction history. However, most methods assign the same SID length to all items despite highly unequal interaction frequencies. This uniform design repeatedly incurs the cost of fine-grained codes for popular items, while shortening codes to reduce this cost also compresses the semantic representations of long-tail items. Inspired by the compact forms of frequent expressions in natural language, we propose JevRec to encode items as variable-length words and generate the next item from interaction context. Linguistic encoding retains frequency-dependent semantic prefixes from shared codebooks, producing shorter SIDs for popular items and richer descriptions for long-tail items. Yet predicting a sequence-level length and generating token-level content impose different optimization requirements that may conflict under a shared output head. Our Jev-like decision modeling component therefore predicts a length distribution with a dedicated head and generates content conditioned on length through a separate head while sharing the interaction-history representation. Experiments on three public datasets show that JevRec outperforms all compared baselines across the evaluated recommendation metrics. Further analyses show lower training and inference time than TIGER and reduced gradient conflict between length and content prediction.
Reinforcement learning with verifiable rewards (RLVR) has become a powerful tool for improving multimodal reasoning. However, standard outcome supervision suffers from credit assignment ambiguity, as a model may receive positive rewards by exploiting language priors or dataset shortcuts rather than grounding its reasoning in the requisite visual evidence. This makes high-quality negative responses critical, since effective negatives should remain plausible under the current policy while exposing the shortcut that should be suppressed. In this paper, we propose EVAN, a multimodal reinforcement learning framework utilizing EVidence-Ablated Negatives to train LMMs against their own language-prior failures. EVAN first identifies highly visually dependent examples using a scaled image information gain metric and preserves data diversity with a mixed-ranking strategy. It then constructs evidence-ablated images by masking the most discriminative image patches, eliciting plausible but visually ungrounded responses from the current policy. During policy optimization, these negative responses are incorporated as bounded auxiliary signals under the clean image, penalizing shortcut learning while maintaining RL stability. Experiments across challenging multimodal reasoning benchmarks demonstrate that EVAN mitigates modality imbalance and improves over strong RLVR baselines. Our code is available at https://github.com/DawnLIU35/EVAN.git.
Large language model (LLM) agents have demonstrated strong capabilities in automating complex information-seeking and transactional tasks across the open web. However, their real-world reliability remains limited by access barriers, such as dead links, authentication walls, rate limits, and anti-scraping defenses. These barriers block the primary execution path, causing agents to abort prematurely, loop indefinitely, or quietly compromise the user's original constraints. Existing evaluations either implicitly assume that target resources are freely accessible or focus on transient perturbations that simple retries can resolve, failing to capture scenarios where agents must actively reroute around impassable bottlenecks. To bridge this gap, this paper introduces AgentHeal, a comprehensive benchmark evaluating goal-preserving resource recovery, defined as the ability of web agents to detect non-transient access barriers, pivot to legitimate alternative sources, and fulfill the user's exact original objective. Spanning nine barrier families across a counterfactual sandbox and real-website replay, our experiments on leading LLM backbones show that average task success drops sharply from 89.4% to 43.7% under access barriers. Further analysis reveals that agents frequently fall for deceptive look-alikes, fail to escalate to users when necessary, violate access policies, or waste computation on futile retries, providing actionable diagnostics for developing resilient web agents.
We present XLaw-JP-27B, an open-weight model specialized for Japanese law that runs on a single 48\,GB GPU under the Apache~2.0 licence. It adapts Qwen3.8-27B with hybrid retrieval over all current statutes and 31k case summaries, a light QLoRA fine-tune on its own verified answers to past bar-exam questions (self-distillation), a contract-specific LoRA adapter, and four-sample voting. We evaluate it on eight Japanese legal tasks---the multiple-choice section of the bar examination in 2024, 2025 and 2026, statement-level bar questions (JBE-QA), statute questions (LawQA-JP), the patent attorney exam, and contract review and drafting built from government model contracts---against open models of 12--33B under identical prompts and retrieval. XLaw-JP-27B ranks first on six tasks, ties on one and is third on one (mean rank 1.3 against 2.8--3.1 for the baselines). It is the only system that passes the bar exam in all three years (129, 120 and 130 of 175, meeting every subject minimum; no test-year material was used in training), while open Japanese 32--33B models fail at least one year. A controlled study of each component shows that retrieval, voting and self-distillation help, and that thinking mode, retrieval-grounded synthetic tuning, self-verification, per-statement decomposition and more self-distillation do not. The largest remaining gaps to frontier models are in contract review and patent law. The model is available as a free web service at https://x-law.jp/ with the weights available at https://huggingface.co/X-LawJapan/xlaw-jp-27b under Apache~2.0.
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 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.
Large language models (LLMs) are increasingly described as acquiring "reasoning" skills after reinforcement learning with verifiable rewards. However, whether these gains reflect genuine reasoning or more superficial adaptations remains underexplored. In this work, we probe LLMs trained in a finite and exhaustively analyzable logical domain, namely Tic-Tac-Toe, and trace how internal representations evolve across reinforcement learning (RL) with Group Relative Policy Optimization (GRPO). Quantitatively, RL yields high accuracy and robustness across prompt variations, suggesting the emergence of game reasoning. Mechanistic interpretability, however, paints a different picture: features extracted with sparse autoencoders (SAEs) reveal that models primarily adapt to rely more heavily on information already explicit in the prompt, such as whose turn it is, game progression and board occupancy. In comparison, high-level concepts like board symmetries and strategic forks remain weakly represented, raising concerns that reasoning benchmarks overestimate model abilities. This suggests that RL-driven "reasoning" may conflate task-specific adaptations to input information with reasoning ability. We provide, for the first time: (i) a systematic interpretability pipeline tracing representation dynamics across RL training in small-scale LLMs; (ii) an extension of SAE-based feature discovery to hypothesis-driven testing in a finite logical domain; and (iii) an interpretability-based demonstration that RL post-training neither adds nor removes concept information but reorganizes which features dominate the representation, with causal interventions locating the decision downstream of it, in the late-layer state.
Few-shot voice cloning makes it cheap to produce daily Japanese financial commentary with several synthetic speakers, but kanji words with more than one reading, such as the word for "market" (shijo vs. ichiba) or "overall trend" (taisei vs. oze), are still misread. A common quality gate transcribes each synthesized sentence with automatic speech recognition (ASR), converts both the input and the transcript to kana, and regenerates on mismatch. We show that this gate is unreliable in both directions: when the audio contains a misreading, the ASR language model frequently writes the expected kanji back, hiding the error, and when the audio is correct, ordinary recognition errors trigger needless regeneration. We propose a reading-level check that scores the expected reading and its confusable alternatives with a connectionist temporal classification (CTC) hiragana recognizer and flags a word when an alternative explains the audio better. On 540 utterances from three voice-cloned production voices covering 30 financial terms, with misreadings injected under controlled conditions, the transcript-based gate detected 78.9% of misreadings at a 44.4% false-alarm rate, whereas the proposed check detected 91.7% at 5.0% (AUC 0.990). On natural kanji input, it caught a front-end error that the transcript-based gate missed for one voice, and reduced regeneration requests from 38.9% to 6.1% of sentences.