ScenarioBench is a policy-grounded, trace-aware benchmark for evaluating Text-to-SQL and retrieval-augmented generation in compliance contexts. Each YAML scenario includes a no-peek gold-standard package with the expected decision, a minimal witness trace, the governing clause set, and the canonical SQL, enabling end-to-end scoring of both what a system decides and why. Systems must justify outputs using clause IDs from the same policy canon, making explanations falsifiable and audit-ready. The evaluator reports decision accuracy, trace quality (completeness, correctness, order), retrieval effectiveness, SQL correctness via result-set equivalence, policy coverage, latency, and an explanation-hallucination rate. A normalized Scenario Difficulty Index (SDI) and a budgeted variant (SDI-R) aggregate results while accounting for retrieval difficulty and time. Compared with prior Text-to-SQL or KILT/RAG benchmarks, ScenarioBench ties each decision to clause-level evidence under strict grounding and no-peek rules, shifting gains toward justification quality under explicit time budgets.

Evaluating language models fairly is becoming harder as static benchmarks available on the internet risk contamination by training data. This makes it unclear whether models are truly reasoning or just recalling answers. In this paper, we introduce BeyondBench, an evaluation framework that avoids this problem by using algorithmic problem generation. Unlike traditional benchmarks that risk contamination from internet-scale training data, BeyondBench creates mathematically grounded problems on the fly, ensuring each test remains fresh and uncontaminated. Our framework covers 44 algorithmic tasks with a total of 117 variations, grouped into three difficulty levels: the Easy Suite (29 tasks) for basic arithmetic and statistics, the Medium Suite (5 tasks, 49 variations) for sequence patterns and reasoning, and the Hard Suite (10 tasks, 68 variations) tackling NP-complete and constraint satisfaction problems. Each task generates problems from a combinatorial space larger than 10^15 unique instances, with solutions verified deterministically by mathematical proofs. We evaluated 101 language models, including 85 open-source and 16 closed-source models, spanning sizes from 0.5B to 141B parameters and multiple quantization schemes. Our results show consistent reasoning deficiencies across model families, with performance degrading sharply as problem complexity increases from polynomial to exponential. In our Hard Suite evaluations, models such as Gemini-2.5-pro, Llama-3.3-70B, and Qwen2.5-72B achieved average accuracies of 56.38%, 26.91%, and 33.60%, respectively. Moreover, we observe that performance drops drastically without tool usage, with GPT-5, GPT-5-mini, and GPT-5-nano showing a decline of 16.81%, 28.05%, and 47.59% accuracy on the hard suite. Our leaderboard is publicly available at https://ctrl-gaurav.github.io/BeyondBench/

Online alignment (e.g., GRPO) is generally more performant than offline alignment (e.g., DPO) -- but why? Drawing on prospect theory from behavioral economics, we propose a human-centric explanation. We prove that online on-policy sampling better approximates the human-perceived distribution of what the model can produce, and PPO/GRPO-style clipping -- originally introduced to just stabilize training -- recovers a perceptual bias in how humans perceive probability. In this sense, PPO/GRPO act as perceptual losses already. Our theory further suggests that the online/offline dichotomy is itself incidental to maximizing human utility, since we can achieve the same effect by selectively training on any data in a manner that mimics human perception, rather than restricting ourselves to online on-policy data. Doing so would allow us to post-train more quickly, cheaply, and flexibly without sacrificing performance. To this end, we propose a design pattern that explicitly incorporates perceptual distortions of probability into objectives like DPO/KTO/GRPO, creating humanline variants of them. Surprisingly, we find that these humanline variants, even when trained with offline off-policy data, can match the performance of their online counterparts on both verifiable and unverifiable tasks.

The past few years have witnessed rapid progress in reinforcement learning (RL) for large language models (LLMs). This began with reinforcement learning from human feedback (RLHF) [Bai et al., 2022, Ouyang et al., 2022] that aligns pre-trained LLMs with human preferences, followed by reasoning-oriented RL that enables LLMs to produce long chains of thought [OpenAI, 2024, DeepSeek-AI, 2025, Kimi-Team, 2025b, Zhang et al., 2025b]. More recently, agentic RL [Kimi-Team, 2025a, Gao et al., 2025, Zhang et al., 2025a] aims to train LLMs for agentic capabilities such as tool use, long-horizon planning, and multi-step task execution in dynamic environments. Alongside these developments, off-policy RL has been attracting growing interest. In the "era of experience" [Silver and Sutton, 2025], LLM-powered agents need to be continually updated through interaction with the environment. Practical constraints in real-world deployment and the complexity of LLM-RL infrastructure often render on-policy training impractical [Noukhovitch et al., 2025]: rollout generation and model training can proceed at mismatched speeds, data might be collected from different policies, reward feedback might be irregular or delayed, and the environment may be too costly or unstable to query for fresh trajectories.

Large language models (LLMs) are increasingly used in high-stakes settings, where overconfident responses can mislead users. Reliable confidence estimation has been shown to enhance trust and task accuracy. Y et existing methods face practical barriers: logits are often hidden, multi-sampling is computationally expensive, and verbalized numerical uncertainty (e.g., giving a 0-100 score) deviates from natural communication. We revisit linguistic confidence (LC), where models express uncertainty through hedging language (e.g., probably, might), offering a lightweight and human-centered alternative. To advance this direction, we 1) release the first diverse, large-scale dataset of hedging expressions with human-annotated confidence scores, and 2) propose a lightweight mapper that converts hedges into confidence scores at near-zero cost. Building on these resources, we 3) conduct the first systematic study of LC across modern LLMs and QA benchmarks, revealing that while most LLMs underperform in expressing reliable LC, carefully designed prompting achieves competitive calibration and discriminability. Finally, we 4) introduce a fine-tuning framework that further improves LC reliability. Taken together, our work positions linguistic confidence as a scalable, efficient, and human-aligned approach to LLM uncertainty estimation, and calls for deeper exploration of this promising yet underexplored direction. The code and dataset are anonymously available at https://anonymous. Large language models (LLMs) are increasingly deployed in real-world applications, from education and healthcare to law and scientific discovery. While their capabilities make them powerful assistants, LLMs are also prone to hallucinations and factual errors, and human overreliance on their outputs can lead to serious consequences. For instance, a U.S. lawyer once submitted fabricated cases generated by ChatGPT, resulting in professional sanctions (ABC News, 2023). Recent social experiments demonstrate that people adjust their reliance on AI depending on how confident the model appears: reliable expressions of uncertainty can enhance trust, satisfaction, and task accuracy (Kim et al., 2024; Xu et al., 2025). These findings highlight the importance of associating reliable uncertainty estimates with LLM responses to support human decision-making. Ultimately, the conveyance of confidence plays a central role in shaping trust and guiding human-AI interaction. A growing body of work explores the extraction and representation of confidence in LLM outputs. These methods are simple and inexpensive but require access to model logits, which are typically unavailable in commercial LLM APIs. However, such scores rarely align with common user behavior or natural communication, as users do not typically phrase queries with explicit instructions like "Please output your confidence along with the answer."

A recently discovered class of entangled neurons, known as Wasserstein neurons, is disproportionately critical in large language models despite constituting only a very small fraction of the network: their targeted removal collapses the model, consistent with their unique role in differentiating similar inputs. Interestingly, in Wasserstein neurons immediately preceding smooth activation functions, such differentiation manifests in the negative pre-activation space, especially in early layers. Pairs of similar inputs are driven to highly distinct negative values, and these pairs involve syntactic tokens such as determiners and prepositions. We show that this negative region is functional rather than simply favorable for optimization. A minimal, sign-specific intervention that zeroes only the negative pre-activations of a small subset of entangled neurons significantly weakens overall model function and disrupts grammatical behavior, while both random and perplexity-matched controls leave grammatical performance largely unchanged. Part of speech analysis localizes the excess surprisal to syntactic scaffolding tokens, and layer-specific interventions reveal that small local degradations accumulate across depth. Over training checkpoints, the same ablation impairs grammatical behavior as Wasserstein neurons emerge and stabilize. Together, these results identify negative differentiation in a sparse subset of entangled neurons as a crucial mechanism that language models rely on for syntax.

Search-augmented LLMs often struggle with complex reasoning tasks due to ineffective multi-hop retrieval and limited reasoning ability. We propose AceSearcher, a cooperative self-play framework that trains a single large language model (LLM) to alternate between two roles: a decomposer that breaks down complex queries and a solver that integrates retrieved contexts for answer generation. AceSearcher couples supervised fine-tuning on a diverse mixture of search, reasoning, and decomposition tasks with reinforcement fine-tuning optimized for final answer accuracy, eliminating the need for intermediate annotations. Extensive experiments on three reasoning-intensive tasks across 10 datasets show that AceSearcher outperforms state-of-the-art baselines, achieving an average exact match improvement of 7.6%. Remarkably, on document-level finance reasoning tasks, AceSearcher-32B matches the performance of the DeepSeek-V3 model using less than 5% of its parameters. Even at smaller scales (1.5B and 8B), AceSearcher often surpasses existing search-augmented LLMs with up to 9x more parameters, highlighting its exceptional efficiency and effectiveness in tackling complex reasoning tasks. Our code will be published at https://github.com/ritaranx/AceSearcher and https://huggingface.co/AceSearcher.

While vision-language models (VLMs) have made significant progress in multimodal perception (e.g., open-vocabulary object detection) with simple language queries, state-of-the-art VLMs still show limited ability to perceive complex queries involving descriptive attributes and relational clauses. Our in-depth analysis shows that these limitations mainly stem from text encoders in VLMs. Such text encoders behave like bags-of-words and fail to separate target objects from their descriptive attributes and relations in complex queries, resulting in frequent false positives. To address this, we propose restructuring linguistic representations according to the hierarchical relations within sentences for language-based object detection. A key insight is the necessity of disentangling textual tokens into core components-objects, attributes, and relations ("talk in pieces")-and subsequently aggregating them into hierarchically structured sentence-level representations ("see in whole"). Building on this principle, we introduce the TaSe framework with three main contributions: (1) a hierarchical synthetic captioning dataset spanning three tiers from category names to descriptive sentences; (2) Talk in Pieces, the three-component disentanglement module guided by a novel disentanglement loss function, transforms text embeddings into subspace compositions; and (3) See in Whole, which learns to aggregate disentangled components into hierarchically structured embeddings with the guide of proposed hierarchical objectives. The proposed TaSe framework strengthens the inductive bias of hierarchical linguistic structures, resulting in fine-grained multimodal representations for language-based object detection. Experimental results under the OmniLabel benchmark show a 24% performance improvement, demonstrating the importance of linguistic compositionality.

As Large Language Models (LLMs) are increasingly proposed for high-stakes medical applications, there has emerged a critical need for reliable and accurate evaluation methodologies. Traditional accuracy metrics fail inadequately as they neither capture question characteristics nor offer topic-specific insights. To address this gap, we introduce \textsc{MedIRT}, a rigorous evaluation framework grounded in Item Response Theory (IRT), the gold standard in high-stakes educational testing. Unlike previous research relying on archival data, we prospectively gathered fresh responses from 80 diverse LLMs on a balanced, 1,100-question USMLE-aligned benchmark. Using one unidimensional two-parameter logistic IRT model per topic, we estimate LLM's latent model ability jointly with question difficulty and discrimination, yielding more stable and nuanced performance rankings than accuracy alone. Notably, we identify distinctive ``spiky'' ability profiles, where overall rankings can be misleading due to highly specialized model abilities. While \texttt{GPT-5} was the top performer in a majority of domains (8 of 11), it was outperformed in Social Science and Communication by \texttt{Claude-3-opus}, demonstrating that even an overall 23rd-ranked model can hold the top spot for specific competencies. Furthermore, we demonstrate IRT's utility in auditing benchmarks by identifying flawed questions. We synthesize these findings into a practical decision-support framework that integrates our multi-factor competency profiles with operational metrics. This work establishes a robust, psychometrically grounded methodology essential for the safe, effective, and trustworthy deployment of LLMs in healthcare.

GUI agents powered by vision-language models (VLMs) show promise in automating complex digital tasks. However, their effectiveness in real-world applications is often limited by scarce training data and the inherent complexity of these tasks, which frequently require long-tailed knowledge covering rare, unseen scenarios. We propose RAG-GUI , a lightweight VLM that leverages web tutorials at inference time. RAG-GUI is first warm-started via supervised finetuning (SFT) and further refined through self-guided rejection sampling finetuning (RSF). Designed to be model-agnostic, RAG-GUI functions as a generic plug-in that enhances any VLM-based agent. Evaluated across three distinct tasks, it consistently outperforms baseline agents and surpasses other inference baselines by 2.6% to 13.3% across two model sizes, demonstrating strong generalization and practical plug-and-play capabilities in real-world scenarios.