Reinforcement learning (RL) has shown strong performance in LLM post-training, but real-world deployment often involves noisy or incomplete supervision. In such settings, complex and unreliable supervision signals can destabilize training and harm generalization. While existing approaches such as worst-case optimization (e.g., RFQI, CQL) and mean-based methods (e.g., PPO, GRPO) can improve stability, they often overlook generalization and may produce overly conservative policies, leading to uneven performance across diverse real scenarios. To this end, we introduce DVPO (Distributional Value Modeling with Risk-aware Policy Optimization), a new RL framework that combines conditional risk theory with distributional value modeling to better balance robustness and generalization. DVPO learns token-level value distributions to provide fine-grained supervision, and applies an asymmetric risk regularization to shape the distribution tails: it contracts the lower tail to dampen noisy negative deviations, while expanding the upper tail to preserve exploratory diversity. Across extensive experiments and analysis in multi-turn dialogue, math reasoning, and scientific QA, DVPO consistently outperforms PPO, GRPO, and robust Bellman-based PPO under noisy supervision, showing its potential for LLM post-training in the real-world.

Machine learning for early prediction in medicine has recently shown breakthrough performance, however, the focus on improving prediction accuracy has led to a neglect of faithful explanations that are required to gain the trust of medical practitioners. The goal of this paper is to teach LLMs to follow medical consensus guidelines step-by-step in their reasoning and prediction process. Since consensus guidelines are ubiquitous in medicine, instantiations of verbalized medical inference rules to electronic health records provide data for fine-tuning LLMs to learn consensus rules and possible exceptions thereof for many medical areas. Consensus rules also enable an automatic evaluation of the model's inference process regarding its derivation correctness (evaluating correct and faithful deduction of a conclusion from given premises) and value correctness (comparing predicted values against real-world measurements). We exemplify our work using the complex Sepsis-3 consensus definition. Our experiments show that small fine-tuned models outperform one-shot learning of considerably larger LLMs that are prompted with the explicit definition and models that are trained on medical texts including consensus definitions. Since fine-tuning on verbalized rule instantiations of a specific medical area yields nearly perfect derivation correctness for rules (and exceptions) on unseen patient data in that area, the bottleneck for early prediction is not out-of-distribution generalization, but the orthogonal problem of generalization into the future by forecasting sparsely and irregularly sampled clinical variables. We show that the latter results can be improved by integrating the output representations of a time series forecasting model with the LLM in a multimodal setup.

Due to their architecture and vast pre-training data, large language models (LLMs) demonstrate strong text classification performance. However, LLM output - here, the category assigned to a text - depends heavily on the wording of the prompt. While literature on prompt engineering is expanding, few studies focus on classification tasks, and even fewer address domains like psychology, where constructs have precise, theory-driven definitions that may not be well represented in pre-training data. We present an empirical framework for optimizing LLM performance for identifying constructs in texts via prompt engineering. We experimentally evaluate five prompting strategies --codebook-guided empirical prompt selection, automatic prompt engineering, persona prompting, chain-of-thought reasoning, and explanatory prompting - with zero-shot and few-shot classification. We find that persona, chain-of-thought, and explanations do not fully address performance loss accompanying a badly worded prompt. Instead, the most influential features of a prompt are the construct definition, task framing, and, to a lesser extent, the examples provided. Across three constructs and two models, the classifications most aligned with expert judgments resulted from a few-shot prompt combining codebook-guided empirical prompt selection with automatic prompt engineering. Based on our findings, we recommend that researchers generate and evaluate as many prompt variants as feasible, whether human-crafted, automatically generated, or ideally both, and select prompts and examples based on empirical performance in a training dataset, validating the final approach in a holdout set. This procedure offers a practical, systematic, and theory-driven method for optimizing LLM prompts in settings where alignment with expert judgment is critical.

When using an LLM through an API provider, users expect the served model to remain consistent over time, a property crucial for the reliability of downstream applications and the reproducibility of research. Existing audit methods are too costly to apply at regular time intervals to the wide range of available LLM APIs. This means that model updates are left largely unmonitored in practice. In this work, we show that while LLM log probabilities (logprobs) are usually non-deterministic, they can still be used as the basis for cost-effective continuous monitoring of LLM APIs. We apply a simple statistical test based on the average value of each token logprob, requesting only a single token of output. This is enough to detect changes as small as one step of fine-tuning, making this approach more sensitive than existing methods while being 1,000x cheaper. We introduce the TinyChange benchmark as a way to measure the sensitivity of audit methods in the context of small, realistic model changes. LLM API providers typically offer version-pinned endpoints, signaling to users that a given endpoint will serve a consistent model. Users of APIs tend to rely on this consistency: developers want to avoid unexpected regressions in their applications; researchers seek reproducibility in their experiments; regulators perform initial compliance assessments, and assume that the API will keep serving the same model afterward (Y an & Zhang, 2022).

Reinforcement Learning (RL) has proven highly effective for autoregressive language models, but adapting these methods to diffusion large language models (dLLMs) presents fundamental challenges. The core difficulty lies in likelihood approximation: while autoregressive models naturally provide token-level conditional probabilities essential for token-level RL objectives (e.g., GRPO), dLLMs generate sequences through iterative non-autoregressive denoising steps that lack this factorization. To address this fundamental mismatch, we propose ELBO-based Sequence-level Policy Optimization (ESPO), a principled RL framework that treats entire sequence generation as a single action and uses the ELBO as a tractable sequence-level likelihood proxy. Our method incorporates per-token normalization of importance ratios and robust KL-divergence estimation to ensure stable large-scale training. Extensive experiments on mathematical reasoning, coding, and planning tasks demonstrate that ESPO significantly outperforms token-level baselines, achieving dramatic improvements of 20-40 points on the Countdown task, while maintaining consistent gains on math and coding benchmarks. Our approach establishes sequence-level optimization as a principled and empirically effective paradigm for RL in dLLMs. Our code is available at https://github.com/ML-GSAI/ESPO. Large language models (LLMs) (OpenAI, 2023) have become a cornerstone of modern natural language processing, achieving remarkable progress across math (Guo et al., 2025), coding (Hui et al., 2024), and planning tasks (Y ao et al., 2023). While autoregressive (AR) modeling has long dominated this field, recent advances in diffusion large language models (dLLMs) have demonstrated strong potential as an alternative formulation (Ou et al., 2024; Shi et al., 2024; Sahoo et al., 2024; Nie et al., 2025; Y e et al., 2025). With the advent of powerful pretrained dLLMs, the next frontier lies in post-training (Ouyang et al., 2022) to further enhance their capabilities. Among various post-training paradigms, reinforcement learning (RL) has emerged as a powerful approach that enables test-time scaling (Snell et al., 2025) through verifiable rewards (Guo et al., 2025). It has yielded substantial gains on reasoning tasks in recent AR models (OpenAI, 2024), such as math (Cobbe et al., 2021b), coding (Chen et al., 2021), and reasoning (Liu et al., 2023b).

Multimodal large language models (MLLMs) that think with images can interactively use tools to reason about visual inputs, but current approaches often rely on a narrow set of tools with limited real-world necessity and scalability. In this work, we first reveal a critical and previously overlooked weakness: even state-of-the-art MLLMs are surprisingly brittle, showing significant performance degradation on images with simple orientation changes or natural corruptions, underscoring the need for more robust tool-based reasoning. To address this, we propose CodeVision, a flexible and scalable code-as-tool framework where the model generates code as a universal interface to invoke any image operation, moving beyond fixed tool registries. We train our model using a two-stage methodology, beginning with Supervised Fine-Tuning (SFT) on a high-quality dataset curated for complex, multi-turn tool composition and error recovery, followed by Reinforcement Learning (RL) with a novel and dense process reward function to encourage strategic and efficient tool use. To facilitate this research, we construct new SFT and RL datasets and introduce a challenging new benchmark suite designed to rigorously evaluate robustness to orientation changes and multi-tool reasoning. Experiments on Qwen2.5-VL and Qwen3-VL series show that our approach significantly improves model performance and fosters emergent capabilities such as flexible tool composition, efficient chained execution, and robust error recovery from runtime feedback. Code is available at https://github.com/ByteDance-BandAI/CodeVision.

Accurate and reliable search on online healthcare platforms is critical for user safety and service efficacy. Traditional methods, however, often fail to comprehend complex and nuanced user queries, limiting their effectiveness. Large language models (LLMs) present a promising solution, offering powerful semantic understanding to bridge this gap. Despite their potential, deploying LLMs in this high-stakes domain is fraught with challenges, including factual hallucinations, specialized knowledge gaps, and high operational costs. To overcome these barriers, we introduce \textbf{AR-Med}, a novel framework for \textbf{A}utomated \textbf{R}elevance assessment for \textbf{Med}ical search that has been successfully deployed at scale on the Online Medical Delivery Platforms. AR-Med grounds LLM reasoning in verified medical knowledge through a retrieval-augmented approach, ensuring high accuracy and reliability. To enable efficient online service, we design a practical knowledge distillation scheme that compresses large teacher models into compact yet powerful student models. We also introduce LocalQSMed, a multi-expert annotated benchmark developed to guide model iteration and ensure strong alignment between offline and online performance. Extensive experiments show AR-Med achieves an offline accuracy of over 93\%, a 24\% absolute improvement over the original online system, and delivers significant gains in online relevance and user satisfaction. Our work presents a practical and scalable blueprint for developing trustworthy, LLM-powered systems in real-world healthcare applications.

Large Language Models (LLMs) have emerged as powerful tools for diverse applications. However, their uniform token processing paradigm introduces critical vulnerabilities in instruction handling, particularly when exposed to adversarial scenarios. In this work, we identify and propose a novel class of vulnerabilities, termed Tool-Completion Attack (TCA), which exploits function-calling mechanisms to subvert model behavior. To evaluate LLM robustness against such threats, we introduce the Tool-Completion benchmark, a comprehensive security assessment framework, which reveals that even state-of-the-art models remain susceptible to TCA, with surprisingly high attack success rates. To address these vulnerabilities, we introduce Context-Aware Hierarchical Learning (CAHL), a sophisticated mechanism that dynamically balances semantic comprehension with role-specific instruction constraints. CAHL leverages the contextual correlations between different instruction segments to establish a robust, context-aware instruction hierarchy. Extensive experiments demonstrate that CAHL significantly enhances LLM robustness against both conventional attacks and the proposed TCA, exhibiting strong generalization capabilities in zero-shot evaluations while still preserving model performance on generic tasks. Our code is available at https://github.com/S2AILab/CAHL.

We present AITutor-EvalKit, an application that uses language technology to evaluate the pedagogical quality of AI tutors, provides software for demonstration and evaluation, as well as model inspection and data visualization. This tool is aimed at education stakeholders as well as *ACL community at large, as it supports learning and can also be used to collect user feedback and annotations.

Large language models (LLMs) can reliably distinguish grammatical from ungrammatical sentences, but how grammatical knowledge is represented within the models remains an open question. We investigate whether different syntactic phenomena recruit shared or distinct components in LLMs. Using a functional localization approach inspired by cognitive neuroscience, we identify the LLM units most responsive to 67 English syntactic phenomena in seven open-weight models. These units are consistently recruited across sentences containing the phenomena and causally support the models' syntactic performance. Critically, different types of syntactic agreement (e.g., subject-verb, anaphor, determiner-noun) recruit overlapping sets of units, suggesting that agreement constitutes a meaningful functional category for LLMs. This pattern holds in English, Russian, and Chinese; and further, in a cross-lingual analysis of 57 diverse languages, structurally more similar languages share more units for subject-verb agreement. Taken together, these findings reveal that syntactic agreement-a critical marker of syntactic dependencies-constitutes a meaningful category within LLMs' representational spaces.