AI 速览：作者构建两个故意训练出特定错位行为的模型生物，形式化事后行为抑制问题，并提出 PEARL 方法以追求即时抑制、训练后持久性与通用能力保持的平衡。
Undesired model behaviors can emerge or re-emerge at any stage of training. Once observed, how effectively can we suppress them? To study this question, we first construct two model organisms--language models deliberately trained to exhibit a specific misalignment--that differ in what triggers the behavior. The narrow-trigger organism misbehaves only when the prompt contains a specific cue: it cheats on multiple-choice questions by following a planted hint that points to a wrong answer, and behaves normally otherwise. The broad-trigger organism has no such cue: after fine-tuning on insecure code, it gives harmful responses across a wide range of open-ended prompts. We then formalize post-hoc behavioral suppression and evaluate methods along three criteria: immediate suppression of the target behavior, durability under further training, and preservation of general capability. We compare standard methods--SFT, GRPO, and gradient-ascent unlearning--along these dimensions. Building on prior work on alignment shallowness, we hypothesize that more durable suppression requires the ability to recover from a wider range of misalignment states. We introduce Prefix-Expanded Adversarial Reinforcement Learning (PEARL), a GRPO variant that augments rollouts with continuations from intermediate states sampled along cached organism trajectories. Across both organism settings and two model families (Qwen3-4B, GPT-OSS-20B), PEARL achieves strong immediate suppression while preserving task accuracy, and remains durable under further fine-tuning on either the organism's data or unrelated data. It obtains the greatest benefit when the undesirable behavior is broad and cannot be addressed through targeted SFT. Together, our framework and method offer a step toward more durable post-hoc removal of undesired behaviors.
Dialectal Arabic speech recognition remains constrained by the scarcity of dialect-resolved transcribed speech despite the abundance of web audio. We introduce GigaSpeech 3, a corpus of 39,000 hours and 10 million weakly transcribed speech segments covering Algerian, Egyptian, Iraqi, Saudi, and Syrian Arabic. Starting from 60 hours of manual transcription per dialect, we train labeling models using speech-recognition-biased self-supervised learning and apply them to dialect-matched web audio. Data selection and four rounds of noisy-student refinement for Saudi and Syrian Arabic yield the final corpus. We validate its utility through dialect-specific and multi-dialect speech recognition training and cross-domain evaluation. On GigaSpeechBench, the final dialect-specific models outperform their labeling models across all five dialects and achieve a macro-average word error rate of 23.86%, a 16.6% relative reduction from the strongest third-party system covering all five dialects, while outperforming the best third-party result on three dialects. A multi-dialect model using self-predicted dialect identities achieves a macro-average word error rate of 25.51% while also supporting Modern Standard Arabic. Across eight conditions from three external corpora, a GigaSpeech 3-trained Zipformer achieves the lowest word error rate among six evaluated models in every condition, while adapting the Qwen3 speech recognizer on GigaSpeech 3 improves performance across all eight conditions. We release the corpus and recipes as a foundation for dialect-aware Arabic speech processing.
Large reasoning models achieve strong performance on challenging tasks, but their long chain-of-thought often incurs substantial inference costs by continuing to deliberate after sufficient information has been obtained. Existing reasoning compression methods primarily operate on reasoning trajectories, paying less attention to the local decision of whether further reasoning is necessary given the current state. We propose LoCo, a local preference learning method for reasoning compression. LoCo detects candidate overthinking states from the model's own traces, freezes the corresponding reasoning prefixes, regenerates shorter verified continuations from the same states, and trains on the resulting state-conditioned preference pairs with DPO. This formulation teaches the model when to continue reasoning and when to conclude, rather than simply encouraging shorter trajectories overall. We construct training data entirely from self-generated mathematical traces and evaluate LoCo on five mathematical and three code generation benchmarks. LoCo reduces completion length by 41.7\%--45.4\% on mathematics and 23.4\%--36.3\% on code, while maintaining or improving accuracy and achieving the best AUC across all code benchmarks. LoCo consistently improves the accuracy-efficiency trade-off on mathematics and transfers effectively to code generation despite math-only training, demonstrating the cross-domain generality of local preference learning for reasoning compression. The code and model will be publicly released at https://github.com/MuyaoYuan/LoCo.
Small language models (<=9B parameters) can support local medical question answering (QA) while reducing privacy and latency concerns, but effective fine-tuning remains challenging. This paper compares two chain-of-thought (CoT) supervised fine-tuning methods: Uniform-across-Sequence (UaS), which applies equal loss weighting to the full CoT-and-answer sequence, and Segment-weighted (SwS), which separately weights CoT and answer losses. We evaluate six general- and medical-domain models on multiple-choice QA (MedQA and EMS-MCQA) and free-text QA (Med-O1). Results show that neither method is universally best: on multiple-choice QA (MedQA and EMS-MCQA), SwS generally performs better, with general-purpose models benefiting most; notably, Qwen3.5-4B improves by up to 44.8% over its zero-shot baseline on MedQA. In contrast, medical-domain models show smaller or even negative gains on MCQs. On free-text QA (Med-O1), UaS performs best for nearly all models and provides consistent improvements across both model types, with the 3B Llama-3.2 model approaching the performance of much larger 7B medical-domain models. We also find that excessively long CoT generation is associated with lower performance in some models on either multiple-choice or free-text QA tasks, suggesting that limiting the CoT length during training may benefit selected models. Overall, the results show that CoT loss weighting should be matched to the task format and the knowledge already encoded in the base model, rather than using a single fine-tuning strategy for all clinical QA tasks.
Zhaoyang Wang, Qianhui Wu, Boyu Gou, Demi Wang, Jiacheng Zhang, Peichun Hua, Qize Li, Shi Qiu 等 57 位作者
GUI web agents require realistic and interactive environments that remain stable during both evaluation and training. Live website evaluation suffers from changing content, access restrictions, and limited coverage of authenticated workflows, while existing synthetic environments for training often lack functional depth and visual fidelity. We propose WebHarbor, a suite of 153 executable and dockerized website environments that faithfully reconstruct the visual interfaces, content, and functional workflows of real-world web services under deterministic control. We build WebHarbor through task-driven human-agent collaboration, where environments evolve as new tasks expand their functional scope, coding agents implement the required changes, and human contributors audit visual fidelity, functional correctness, and task validity. The resulting suite comprises 3,518 tasks spanning multi-step and post-authentication workflows, where existing works often fail to cover. Beyond reliable evaluation, WebHarbor provides high-fidelity synthetic web environments for cost-efficient reinforcement learning (RL). We demonstrate that a Qwen 4B-level GUI agent trained with RL in WebHarbor generalizes to unseen tasks and performs competitively with leading open-source GUI web agents across four benchmarks. The code and data are publicly available at https://github.com/aiming-lab/WebHarbor.
Rapid advances in cowork agents are shifting human-AI collaboration toward real-world automation, delivering software, designs, and other artifacts. However, such productivity work demands both high-quality delivery on complex productivity tasks and continued monitoring and decision-making over long horizons, a combination that remains an open challenge. We introduce CoworkOS, a general-purpose harness for long-horizon work that separates task strategy from execution management: the model decides actions, delegation, and acceptance, while the runtime maintains state and feedback. Its three-layer architecture enables direct execution or delegation to Subagents, synchronously or asynchronously, under stable responsibilities. Through a shared application layer, humans and agents inspect and revise the same evolving documents, presentations, CAD models, and Blender scenes, so human-in-the-loop participation can enter both through conversation and through direct modification of the work itself. To sustain continuity across agents and sessions, CoworkOS further introduces hierarchical workspaces and persistent semantic memory, so working files, key decisions, and long-running Job logs remain shareable, reusable, and resumable after interruption. Across 12 benchmarks and 6 categories of real-world applications, CoworkOS proves highly competitive with frontier agent applications such as Codex and DeepSeek Harness. As a sustained-work case study, CoworkOS uses GLM-5.3 as its backbone to conduct RSI-driven automated post-training on Qwen3.6-35B-A3B. Over five rounds, the model's average score across eight benchmarks rose from 33.19 to 42.95 (+9.76). Looking ahead, we believe CoworkOS opens a meaningful path toward general cowork-agent intelligence.
Large language models (LLMs) can generate visual assets through text-based formats such as SVG and CAD, but these symbolic representations describe images using high-level primitives rather than explicit raster values. We ask whether LLMs can instead generate pixel images directly through their existing vocabulary, without a learned image tokenizer or decoder. We introduce PixelBench, a benchmark for text-to-pixel (T2P) and image-to-pixel (I2P) generation across multiple resolutions, together with a complementary pixel-understanding evaluation. Models generate an RGB color palette followed by a grid of indices specifying the palette color of each output pixel. In T2P, the strongest models produce recognizable images directly from text. In I2P, the strongest models outperform palette-matched downsampling and specialized pixelization algorithms at low resolutions. Pixel-understanding performance correlates strongly with generation performance. A resolution-scaling study shows that pixel-generation performance often improves initially before declining at high resolutions. Finally, we introduce Pixel Reconstruction Tuning (PRT), which augments visual instruction tuning with caption-palette-grid trajectories. We post-train Molmo2-4B from its pretraining checkpoint by mixing general instruction data with our curated PRT data, improving performance across 16 visual-centric benchmarks, indicating that explicit pixel reconstruction can provide useful auxiliary supervision.
Whether an aligned language model refuses depends in part on which words a prompt contains, regardless of whether those words carry the request's meaning. Harm words shift the hidden state at the generation boundary along a component shared across prompts and words, which crossed differencing extracts as a single direction. In eight models, writing along this direction raises and lowers refusal monotonically with dose, ahead of random and KL-matched directions. Used as a ruler over the full vocabulary, the direction gives every word a shift, from "feared" words at one end to "courage" words at the other; a word's shift is predictable from its own representation, and appending a feared word to a benign prompt is enough to make the model refuse it. Two held-out tests validate the ranking in all eight models: refusal after inserting an unseen word rises with its table shift, and replacing a prompt's most feared word by its most courageous meaning-preserving synonym significantly raises the answer rate on benign and harmful prompts alike, by an amount that follows the shift the replacement actually causes.
Sreehari Sankar, Debargha Ganguly, Vikash Singh, Mohsen Hariri, Weicong Chen, Yanyan Zhang, Shouren Wang, Wang Yang 等 20 位作者
Tool-using language-model agents often execute actions that their domain policies forbid. We present LeanGuard, a runtime that checks every proposed tool call against a policy written in Lean4 before the call reaches the environment. Policies combine state predicates with scoped past-time temporal operators over the execution history, evaluation is fail-closed, and the decision procedure is proved equivalent to a LeanLTL semantics. A denied call can be corrected in two ways. With feedback, the agent receives the violated rule and a hint; with proof-guided repair, Lean metaprogramming searches for a nearby call that preserves the user's intent and proves that the policy admits it, without the model. On Tau-Bench with five Qwen models (0.8B-27B), unguarded agents execute at least 33 calls that the Lean policy forbids, while LeanGuard executes none; with feedback, task success does not change significantly. On DTap, LeanGuard reduces direct prompt-injection success from 57.9% to 5.9%, and in telecom, where benign tasks need no writes, benign success is unchanged. Feedback reduces repeated blocked calls and raises task success for four of five models over a bare denial. On replayed denials, repair proves an admissible edit for 88 of 252 denials caused by argument values, in a median of 2,ms.
Rongyu Zhang, Mutian Bao, Jinjian Zhang, Yuetian Du, He Xu, Wei Zhou, Xiaojun Zhu, Ming Kong 等 11 位作者
When user queries are incomplete, large language models should ask targeted questions before committing to an answer, a capability we call proactive elicitation, yet evaluation on real user-assistant queries across seven LLMs (8B to frontier) shows that fewer than 11% of underspecified queries trigger a follow-up question. To benchmark and train this capability, we introduce VagueBench, where each seed query is annotated with structured intent and detail units then degraded into intent-vague and fully-vague variants via controlled removal of linguistic cues, yielding known hidden information sets for evaluating both when and what to ask. We propose DialogGRPO, an offline reinforcement learning method that organizes multi-turn elicitation as a dialogue-state directed acyclic graph, where candidate questions at each node receive dense turn-level rewards combining immediate information gain with lookahead value through group-relative comparison, enabling fine-grained credit assignment without runtime simulation. DialogGRPO elicits missing information on over 88% of vague queries with substantial gains in both recovery efficiency and answer coverage, without over-asking on complete queries, outperforms prompting and test-time-scaling baselines, and generalizes zero-shot to MultiWOZ 2.4 and AmbigQA.
Language agents read code, tool output, and dialogue in contexts that grow at every step. Dynamic sparse attention lets a small indexer choose the earlier tokens each query attends to, but the indexer still scores the whole prefix at every layer. Cross-layer reuse methods such as IndexCache rerun the indexer at only some layers, with one fixed pattern for every input. In an agent, the relevant evidence can change abruptly when a traceback or test result arrives, so fresh selections matter more at some steps than at others. We propose Adaptive Index Reuse (AIR), which keeps IndexCache's execution and learns, for each segment of an agent episode, how many layers refresh their selection. A small controller reads signals from completed computation and the latest tool event, chooses a pattern from a nested library, and learns from terminal task reward under an expected compute budget. On Qwen3-4B-Instruct-2507 with a 128K context, AIR improves success over the strongest static schedule at matched cost on SWE-bench Verified (27.0% to 32.4%), Terminal-Bench 2.0, and three $τ^2$-bench domains, and it lowers time per successful task on all five. Across 50 backbone, context, and benchmark settings, it matches or exceeds that schedule in 44.
AI 速览：作者提出 Assess→Decide→Respond 框架，让防御方在多轮对话的关键轮次先显式做出可解析的安全决策，再据此生成回复，以利用对齐的“浅层性”进行防御。
Safety alignment has enabled large language models to reliably refuse harmful requests directly, yet the same request distributed across a multi-turn conversation still succeeds far too often. This brittleness has a mechanistic root: Alignment is shallow, with safety behavior concentrated in a model's first few output tokens, and existing defenses inherit the weakness, supervising the surface form of the response without ever giving the model an explicit safety decision to learn. We take the opposite view of shallowness: if a short span determines the trajectory of an entire response, that span is the highest-leverage place to intervene, and the defender can occupy it first. We propose Assess $\to$ Decide $\to$ Respond, a framework in which the defender assesses the user's intent, commits to an explicit, parseable safety decision, and only then generates a reply conditioned on it, with the commitment trained by reinforcement learning at the payload turn. Because the decision is a token span, the reward can be a rule: a self-verified reward checks the committed label against the conversation's known nature, requiring no judge and no annotation, while the same format also supports an LLM-as-judge reward over a graded decision space with safe redirection. Across two model families and fourteen benchmarks spanning safety, utility, and helpfulness, our defenders achieve the strongest multi-turn safety in a controlled comparison, maintaining their lead in safety while preserving utility at a modest cost to helpfulness. A per-token KL analysis and interventional probes confirm why this works: training concentrates its effect on the brief deliberation that sets a response's trajectory, and the decision committed there is a causal control point. The very shallowness attackers exploit becomes the mechanism of defense.
AI 速览：PIVOT 将教师生成的纠正片段插入学生 rollout，并对这些片段施加 token 级 KL 监督，与 RLVR 目标联合训练，以缓解困难推理任务中稀疏奖励无学习信号的问题。
Reinforcement learning from verifiable rewards (RLVR) is limited by sparse outcome-level feedback on hard reasoning tasks: when all rollouts in a group fail, rewards provide no learning signal. Stronger teachers can provide additional guidance, but this raises a central question: how can such signals be incorporated into RLVR to support successful rollouts while remaining learnable for the student? We distinguish two complementary mechanisms: trajectory-level intervention, which improves access to reward-bearing trajectories, and token-level supervision, which provides dense targets to guide student learning. Our controlled experiments show that either mechanism alone can remain insufficient, motivating their combination. We therefore introduce PIVOT, which inserts teacher-generated corrections into student rollouts and applies token-level KL supervision to these segments alongside the RLVR objective. Beyond their combination, we systematically analyze where to intervene and how to supervise, yielding a recipe for effective guidance. Across hard logic games and mathematical reasoning benchmarks, PIVOT consistently outperforms existing methods, nearly doubling mean avg@8 on logic games from 19.5% for the best-performing baseline to 35.6%.
Agent safeguards must determine whether an agent's behavior violates the safety rules of the current task, environment, and user. Because these rules vary across contexts, a safeguard needs to select the relevant policy before judging the behavior. Existing policy-aware safeguards can apply a supplied policy, but generally assume that the correct policy is already known. They therefore do not learn how to select a policy from a library, even though an incorrect choice can lead to an incorrect safety judgment. We introduce RePolicy, a safeguard that jointly learns policy selection and safety assessment. Given an agent trajectory and a set of candidate policies, RePolicy selects the policy that best matches the situation, applies its requirements, and predicts whether the behavior is safe. We construct PolicyTraj-20K, a dataset of more than 20,000 agent trajectories annotated with applicable policies, safety labels, and explanations. We train RePolicy with supervised demonstrations followed by reinforcement learning, while randomizing policy order, identifiers, and distractors to reduce shortcut learning. Across six agent-safety benchmarks and 19 baselines, RePolicy-4B achieves leading overall safety-detection performance and improves policy-selection accuracy. These results show that safeguards can learn which safety requirements to use for assessing agent behavior.
In high-risk text interactions, a system may need to estimate whether the current speaker is likely to be a minor before applying age-conditioned safeguards. Many text-based dialogue systems lack strong identity signals such as account information, verified identity, facial data, or device data, and must rely on weak cues that emerge across multiple turns. We formulate implicit minor detection under weak identity information as a high-risk user-attribute inference task and propose a constrained framework comprising a large language model, a Harness, and a Maker-Checker protocol. The framework controls when inference is invoked, requires structured dialogue-linked evidence, and subjects validation-selected Harness revisions to version-level review. On a controlled Chinese benchmark comprising 8,328 multi-turn dialogues, our method improves average precision and minor-class F1 by 7.60 and 6.37 percentage points, respectively, over direct inference using the same LLM family.
Automating optimization modeling with LLMs is a promising path toward scalable decision intelligence, but existing approaches either rely on agentic pipelines built on closed-source LLMs with high inference latency or fine-tune smaller LLMs using costly process supervision that often overfits to a single solver API. We study how reinforcement learning from inexpensive outcome labels can improve optimization modeling and support reuse across heterogeneous solver environments. To this end, we propose Optimization Modeling from Solver Outcomes (OMSO), which uses reinforcement learning (RL) with a mathematical programming solver as a deterministic, interactive environment. Given a natural-language problem and a target solver, OMSO generates solver-specific code, executes it in a sandboxed harness, and converts execution outcomes into scalar rewards in a closed-loop generate-execute-feedback-update process. This RL formulation learns optimization modeling without process-level supervision and supports cross-solver generalization by switching the solver environment rather than reconstructing solver-specific datasets. Experiments on NL4OPT, MAMO, IndustryOR, and OptiBench across Gurobi, OR-Tools, and COPT show that OMSO matches or outperforms process-supervised SFT, supports zero-shot training-free transfer, direct RL training, and SFT-cold-started RL, and improves modeling reasoning and semantic correctness beyond solver API usage.
We introduce Nemotron 3.5 Super and Nemotron 3.5 Lightning, open and efficient Mixture-of-Experts Hybrid Mamba-Transformer models designed for reasoning, long-context understanding, and agentic workloads. Super is a 120B-parameter, 12B-active multimodal model, while Lightning is a smaller 30B-parameter, 3B-active text model; both designed for efficient deployment. Both models combine large-scale distillation, long-context training, and native speculative decoding, while Super further incorporates Indic continued pretraining and agentic software-engineering mid-training. We develop a comprehensive post-training pipeline spanning supervised fine-tuning and reinforcement learning (RLVR and RLHF), complemented in Super by specialized teacher training and multi- teacher on-policy distillation (MOPD). This transforms Super from a text-only base model into a multimodal agent capable of reasoning over text, images, video, and interactive tool environments. We further explore test-time scaling through parallel agent execution and self-verification. Across training and deployment, we jointly optimize capability and efficiency. Indic continued pretraining improves Indian-language benchmarks by up to 9.82 percentage points, while agentic mid-training strengthens software-engineering capabilities and robustness across agent harnesses. The post-trained Super model achieves 51.78% on SWE-bench Pro, surpassing Nemotron 3 Ultra despite using approximately 4.6x fewer active parameters. Post-training MTP boosting improves speculative- decoding acceptance length by up to 21.8% under probabilistic draft sampling, while NVFP4 quantization achieves 99.62% median accuracy recovery relative to BF16. Inference optimizations enable up to 4,856 output tokens/s/GPU on GB200 while sustaining at least 50 tokens/s/user.
Sihan Yang, Weichen Bai, Ruichuan An, Yiming Cheng, Zhiyuan Feng, Tianyi Jiang, Renye Yan, Ning Ding 等 12 位作者
As executable environments for language agents scale, generating and verifying their rollouts becomes increasingly expensive. Under finite rollout budgets, this raises an important question: can data management at the environment level make agentic reinforcement learning (RL) more efficient by organizing, evaluating, and scheduling task instances together with the environments that determine their tools, states, transitions, and verification? To address this question, we introduce EnvSched (Environment-aware Scheduling), an online scheduler that learns from rollout history. EnvSched assigns scheduling value through two complementary modeling choices: exploitation prioritizes tasks likely to provide informative reward differences, while coverage spreads each batch across underobserved environments and interaction patterns. EnvSched organizes environment--task pairs as nodes in a dual-relation graph, with within-environment edges capturing task semantics and shared tool, state, and verifier structure, and cross-environment edges capturing abstract execution and reward topology. By propagating historical rollout statistics along both edge types, it predicts within-task reward variance. These predictions guide a scheduler that allocates each batch to tasks with high predicted reward variance and underobserved environments and tasks. Environment quotas and a similarity penalty further limit concentration and redundancy, yielding informative and diverse batches without additional pilot rollouts. In RL on Qwen3-8B, EnvSched achieves an average score of 48.81 across BFCL-v3, tau-Bench, and ACEBench-Agent after 40 training steps, exceeding the highest score attained by random task selection after 120 steps. It also improves downstream performance across model sizes at the same training budget.
Henrique Schechter Vera, Sahil Dua, Jovan Andonov, Sara Smoot, Sebastian Russo, Ivan Llanos, Max Matveev, Jae Hyeon Yoo 等 73 位作者
We present EmbeddingGemma 2, an open multimodal embedding model which natively represents text, images, audio, video, and combinations thereof in a unified space. Our approach introduces improvements such as simplified per-layer embeddings and a multi-stage curriculum with modality and objective progression which ensure strong quality across unimodal, cross-modal, and mixed-modal evaluations. Notably, in comprehensive embedding benchmarks across all modalities, EmbeddingGemma 2 establishes a new Pareto frontier, outperforming both specialist models and models two to twelve times larger. Furthermore, our design offers flexibility for resource-constrained environments, supporting selective modality encoder loading, adjustable vision token budgets, truncatable embedding dimensions, and quantized weights. We present ablations analyzing key decisions and release EmbeddingGemma 2 to promote further research.
Reinforcement learning with verifiable rewards (RLVR) has shown strong empirical gains in improving the reasoning ability of large language models (LLMs) on tasks such as mathematics and programming. However, recent studies have shown that RLVR rollouts often collapse onto a narrow set of reasoning trajectories, limiting the explored reasoning space. Motivated by findings that thinking tokens at critical steps serve as critical structural elements in LLM reasoning pathways, we propose Guided-RLVR, a novel approach that augments RLVR with a lightweight controller. The controller strategically injects semantically meaningful steering tokens into the sampling distribution, encouraging the model to approach problems from multiple reasoning perspectives and thereby expanding the exploration space. Experiments on mathematical reasoning and code generation benchmarks show that Guided-RLVR consistently outperforms existing RLVR methods while producing more diverse reasoning trajectories.
The advancement of Large Language Models (LLMs) has catalyzed the development of autonomous agents capable of navigating complex environments. However, existing evaluations primarily adopt a deductive paradigm, where agents execute tasks based on explicitly provided rules and static goals, often within limited planning horizons. Crucially, this neglects the inductive necessity for agents to discover latent transition laws from experience autonomously, which is the cornerstone for enabling agentic foresight and sustaining strategic coherence. To bridge this gap, we introduce OdysseyArena, which re-centers agent evaluation on long-horizon, active, and inductive interactions. We formalize and instantiate four primitives, translating abstract transition dynamics into concrete interactive environments. Building upon this, we establish OdysseyArena-Lite for standardized benchmarking, providing a set of 120 tasks to measure an agent's inductive efficiency and long-horizon discovery. Pushing further, we introduce OdysseyArena-Challenge to stress-test agent stability across extreme interaction horizons (e.g., > 200 steps). Extensive experiments on 15+ leading LLMs reveal that even frontier models exhibit a deficiency in inductive scenarios, identifying a critical bottleneck in the pursuit of autonomous discovery in complex environments. Project homepage: https://os-copilot.github.io/OdysseyArena-Home/
Haoyu Sun, Wenxuan Wang, Mingyang Song, Jujie He, Weinan Zhang, Yang Liu, Yang Yang, Yu Cheng
Planning is central to LLM agents: before acting, an agent must decompose goals, select tools, reason over constraints, and decide when a task is infeasible. Yet existing agent evaluations often report only end-to-end success, making it difficult to determine whether failures stem from planning or execution. We introduce Agent Planning Benchmark (APB), a planning-specific diagnostic benchmark with 4,209 multimodal cases across 22 domains and five settings, covering holistic planning, feedback-conditioned step-wise planning, and robustness under extraneous tools, broken tools, and unsolvable tasks. Across 12 MLLMs, APB reveals systematic weaknesses in long-horizon planning, tool-noise robustness, calibrated refusal, and inference-time refinement. We further validate APB on 200 ToolSandbox tasks and 200 $τ^2$-bench tasks, where APB-guided refinement consistently improves plan correctness, plan grade, and downstream execution metrics across three representative models. APB thus serves as an upstream diagnostic complement to execution benchmarks. The APB benchmark and code are available in \href{https://github.com/Mikivishy/AgentPlanningBenchmark}{this URL}.
My Chiffon Nguyen, Tianyi Hu, Vinod Anbalagan, Henry Gagnier, Pouya Sadeghi, Farah Abdou, Tidor-Vlad Pricope, Diego MMS Oliveira 等 164 位作者
As LLMs expand their language coverage, evaluating how well they understand and represent diverse cultures becomes increasingly important. However, many benchmarks remain English-centric, with limited cultural coverage and a focus on factual knowledge over its application in real-world cultural contexts. We propose CulturalRiddles, a benchmark of original, culturally grounded riddles created entirely by humans through a participatory community effort, spanning 62 cultures (45 countries, 51 languages). These riddles are deeply rooted in their respective cultures and require abductive and compositional reasoning, alongside detailed factual and social knowledge, language skills, and an understanding of culturally specific associations. We benchmark 28 LLM configurations using both automatic and large-scale human evaluation. Even frontier models achieve at most 22% accuracy in some communities. We publicly release all artifacts alongside an open contribution process to expand the benchmark's linguistic and cultural coverage.
Automated fake news detection is vital on social media, where fake news is easily rewritten to fool detectors. However, existing approaches typically rely on superficial writing cues or single-style text rewrites, resulting in poor robustness under diverse cosmetic modifications. We construct MANTIS-Bench, a multi-axis presentation-attack benchmark that pairs original news with rewritten views across writing style, emotional expression, and dissemination framing. We further propose MANTIS, an instance-adaptive adversarial framework enabling dynamic worst-case training tailored to each news article with zero inference-time overhead. This reframes fake news defense from heuristic format smoothing into targeted, fact-invariant credibility assessment. Experiments on MANTIS-Bench show that MANTIS reduces the Union Attack Success Rate (UASR) by 3.55 percentage points relative to SheepDog on GossipCop and consistently outperforms prior single-axis defenses on Macro-F1 while maintaining competitive clean accuracy, demonstrating the effectiveness of multi-axis and instance-adaptive worst-case training.
Large language models (LLMs) often produce plausible but incorrect answers, motivating reinforcement learning methods that reward correct answers, penalize incorrect answers, and allow abstention. However, fixed outcome-based rewards do not distinguish incorrect responses with different confidence levels or account for differences in answering reliability across queries. Consequently, such rewards may inadequately penalize overconfident errors and fail to discourage unnecessary abstention. We propose CalibraRL, a reinforcement learning framework that incorporates response confidence and query-level answering evidence into reward design. For non-abstaining responses, CalibraRL combines correctness-based utility with a Brier-based penalty that encourages agreement between confidence and correctness. For abstentions, it estimates conditional answering reliability using a Beta posterior derived from the correctness of non-abstaining responses within each sampled group. The abstention reward increases with the posterior probability that the model's accuracy on non-abstaining responses is below a predefined threshold. CalibraRL reuses training rollouts without requiring additional annotations indicating whether the model knows the answer to each question. Experiments on three QA datasets demonstrate that CalibraRL improves answer precision and reduces hallucinations compared with strong baselines. Further analysis shows that CalibraRL reduces mean expected calibration error (ECE) by 30.8\% across two backbone models and three datasets, highlighting its potential to improve the reliability of retrieval-augmented generation (RAG) in web search.
Junxing Hu, Zicheng Zhang, Zhiqian Zhang, Ai Han, Yifeng Zhang, Pengzhang Liu, Ling Li
Reasoning models trained with Group Relative Policy Optimization (GRPO) now produce chains of thought thousands of tokens long, and recent studies reveal that such training is prone to policy collapse. Entropy regularization is the standard remedy, but it works by pushing probability mass away from the few tokens that carry it and toward the vast low-probability tail of the vocabulary, and it applies this force once for every token of the reasoning trace. This puts the bonus in direct conflict with the entropy decay that convergence requires, and makes its coefficient difficult to set in the long-trace regime. To resolve this tension, we propose Info-GRPO, an information-theoretic framework that cultivates correlation between the policy and a latent prior. Rather than raising entropy everywhere, Info-GRPO builds on the mutual information between a latent variable and the trajectory it induces: augmenting prompts with latent seeds during training lets the model explore diverse policies correlated with the prior, while the seed-conditioned entropy is guided toward convergence. Diversity is therefore carried across whole reasoning traces rather than by per-token randomness, and exploration no longer fights entropy reduction. Experiments show that Info-GRPO outperforms vanilla GRPO and entropy-regularized GRPO across diverse reasoning benchmarks. On AIME24, it improves $Avg@8$ over GRPO by 3.75%, 1.66%, 4.16%, and 4.17% on four backbones, and it reaches the same majority-vote accuracy with $8\times$ fewer samples. Code and models will be publicly released.
A benchmark is not merely a collection of test questions; it naturally induces an empirical measure over difficulty. In this work, we develop this measure-based view of benchmark evaluation. This perspective allows us to identify and formalize two central failure modes: measure mismatch, where the benchmark difficulty distribution differs from the intended difficulty distribution, and coverage gaps, where uncovered regions along the difficulty axis prevent benchmarks from detecting meaningful localized training progress. We propose two diagnostic metrics that capture these two failure modes and provide rigorous theoretical guarantees. Empirically, we analyze 25 popular benchmarks spanning 7 domains using carefully estimated item difficulties. Our results show that many benchmarks implicitly overweight certain difficulty regions, so their leaderboard conclusions should be interpreted with caution; others leave substantial portions of the difficulty axis uncovered, limiting their ability to monitor the model improvement process. Finally, we turn diagnosis into action by introducing BenchFill, a simple difficulty-targeted item rewriting algorithm that effectively reduces local coverage gaps.
Raymond Alavo University of Notre Dame, Xinyuan Zhang University of Notre Dame, Gemza Ademaj, Junhui Cai University of Notre Dame, Hyeokhyen Kwon Emory University, Robert Cotes Emory University, Gari D. Clifford Emory University and Georgia Institute of Technology, Ahmed Abbasi University of Notre Dame
We present Dash-M5H, an interactive dashboard for multi-modal, multi-model mental health assessment that helps clinicians and researchers jointly inspect multimodal behavioral data with multi-model signal outputs of recorded clinical interviews. Guided by signal detection and integrated sensemaking theories, Dash-M5H synchronizes transcript text, audio, and facial behavior (action units and gaze) to support overview-to-detail evidence tracing; and it integrates extracted signals (e.g., sentiment and facial activity) with a clinically grounded VLM prediction pipeline that produces DSM-5-aligned depression predictions. Dash-M5H (https://dash-m5h.io) is implemented in a lightweight, browser-based stack (Quarto + Observable JS + D3), supports local data import and time-synced clinical annotation with export. We demonstrate Dash-M5H through a depression screening scenario, evaluate its note-taking and screening capabilities through a user experiment, and release a live demo (https://youtu.be/w3qCJ02k6bw) and code (https://github.com/nd-hal/M5H-Dashboard-VLM) to facilitate reproducible evaluation.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Haochen Luo, Yifan LI, Ho Tin Ko City University of Hong Kong, An Binh Minh, Junjie Xu, Tang Pok Hin City University of Hong Kong, Wang Chak Wong, Gao Yuan 等 11 位作者
Large language model (LLM) agents are increasingly applied to financial decision-making tasks that require interaction with external tools, including market data retrieval, news analysis, and trade execution. However, existing trading systems rely on fragmented and task-specific APIs, resulting in inconsistent schemas, complex integration, and limited reproducibility. We present QFinZero, a unified trading environment for LLM-based financial agents. QFinZero standardizes three core capabilities: (i) multi-frequency market and derivatives data access, (ii) structured news and event retrieval, and (iii) stateful brokerage simulation with explicit order lifecycle management. All tools adopt consistent JSON schemas and time-aligned interfaces, enabling agents to acquire information and execute trades within a coherent framework. By abstracting financial interaction into composable, agent-invokable primitives, QFinZero reduces engineering overhead and supports reproducible evaluation through comprehensive logging and deterministic replay. We argue that such a standardized trading environment is essential for scalable research on LLM-based financial agents.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Fan Zhang, Mingzi Song Yokohama National University, Tokyo Institute of Technology, Rania Elbadry, Yankai Chen Cornell University, Shaobo Wang, Yixi Zhou, Xunwen Zheng, Yueru He Columbia University 等 22 位作者
Financial reporting systems increasingly leverage Large Language Models (LLMs) to extract and summarize corporate disclosures. However, most existing approaches assume a single-market setting and overlook structural differences across jurisdictions. Variations in accounting taxonomies, tagging infrastructures (e.g., XBRL vs. PDF), and aggregation conventions introduce substantial challenges for semantic alignment and reliable verification. Here, we aim to bridge this gap. We present FinReporting, an agentic workflow for localized cross-jurisdiction financial reporting. The system constructs a unified canonical ontology spanning the income statement, balance sheet, and cash flow statement, and decomposes reporting into auditable stages, including filing acquisition, extraction, canonical mapping, and anomaly logging. Rather than treating LLMs as free-form generators, FinReporting employs them as constrained verifiers operating under explicit decision rules with evidence grounding.Evaluated on annual filings from the USA, Japan, and China, FinReporting improves consistency and reliability under heterogeneous reporting regimes. We further release an interactive demo that enables cross-market inspection and supports structured export of localized financial statements. Our demo is available at https://huggingface.co/spaces/BoomQ/FinReporting-Demo. A video describing our system is available at https://www.youtube.com/watch?v=f65jdEL31Kk.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表