Large Language Models (LLMs) can enhance their reasoning by interacting with external tools, a paradigm known as Tool-Integrated Reasoning (TIR). However, extending TIR to multi-turn settings using Reinforcement Learning (RL) often exhibits training instability and degraded performance. We attribute the instability to harmful negative samples resulting from distributional drift and compounding errors induced by using external tool outputs during multi-turn rollout. To address this issue, we introduce SimpleTIR, a simple method that stabilizes multi-turn TIR training via filtering out trajectories with "void turns", i.e., turns that yield neither a code block nor a final answer. Specifically, we remove those trajectories from the policy update to block harmful gradients, while retaining them in advantage estimation to keep the estimate unbiased. Extensive experiments show that SimpleTIR effectively mitigates gradient norm explosion and stabilizes multi-turn RL training from base models. It achieves state-of-the-art performance on challenging math reasoning benchmarks, including an AIME24 score of 50.5 starting from the Qwen2.5-7B base model. SimpleTIR also promotes more diverse reasoning behaviors such as self-correction and cross-validation, outperforming prior methods trained from stronger instruction-tuned models.

High-quality kernel is critical for scalable AI systems, and enabling LLMs to generate such code would advance AI development. However, training LLMs for this task requires sufficient data, a robust environment, and the process is often vulnerable to and . In these cases, models may hack training rewards or prioritize trivial correctness over meaningful speedup. In this paper, we systematically study reinforcement learning (RL) for kernel generation. We first design , a robust distributed GPU environment that supports reward hacking check, data collection from multi-turn interactions and long-term RL training. Building on KernelGYM, we investigate effective multi-turn RL methods and identify a biased policy gradient issue caused by self-inclusion in GRPO. To solve this, we propose Turn-level Reinforce-Leave-One-Out () to provide unbiased advantage estimation for multi-turn RL. To alleviate lazy optimization, we incorporate mismatch correction for training stability and introduce Profiling-based Rewards () and Profiling-based Rejection Sampling () to overcome the issue. The trained model, Dr. Kernel-14B, reach performance competitive with Claude-4.5-Sonnet in Kernelbench. Finally, we study sequential test-time scaling for Dr. Kernel-14B, which even GPT-5 and Claude-4.5-Sonnet in the Kernelbench level-2 subset.

Modern language models are trained almost exclusively on token sequences produced by a fixed tokenizer, an external lossless compressor often over UTF‑8 byte sequences, thereby coupling the model to that compressor. This work introduces proxy compression, an alternative training scheme that preserves the efficiency benefits of compressed inputs while providing an end-to-end, raw-byte interface at inference time. During training, one language model is jointly trained on raw byte sequences and compressed views generated by external compressors; through the process, the model learns to internally align compressed sequences and raw bytes. This alignment enables strong transfer between the two formats, even when training predominantly on compressed inputs which are discarded at inference. Extensive experiments on code language modeling demonstrate that proxy compression substantially improves training efficiency and significantly outperforms pure byte-level baselines given fixed compute budgets. As model scale increases, these gains become more pronounced, and proxy-trained models eventually match or rival tokenizer approaches, all while operating solely on raw bytes and retaining the inherent robustness of byte-level modeling.

Recent advances in coding agents suggest rapid progress toward autonomous software development, yet existing benchmarks primarily evaluate short-horizon behaviors such as localized code generation, scaffolded completion, or repository repair, leaving it unclear whether agents can sustain coherent reasoning, planning, and execution over the extended horizons demanded by real-world repository construction. To address this gap, we introduce NL2Repo-Bench, a benchmark explicitly designed to evaluate the long-horizon repository generation from scratch: given only a single natural-language requirements document and an empty workspace, agents must autonomously design the architecture, manage dependencies, and produce a fully installable Python library. Experiments across state-of-the-art open- and closed-source models reveal that long-horizon repository generation remains largely unsolved, with even the strongest agents achieving merely 40\% average test pass rates and rarely completing an entire repository correctly. Further analysis identifies systematic long-horizon failure modes, including premature termination, loss of global coherence, fragile cross-file dependencies, and inadequate planning over hundreds of interaction steps. These results position NL2Repo-Bench as a rigorous, execution-based testbed for evaluating sustained agentic competence and highlight long-horizon reasoning as a key bottleneck for autonomous coding agents. Our data and code are available at https://anonymous.4open.science/r/nl2repobench-foricml-F4ED/.

Vision-Language-Action (VLA) models are emerging as a promising paradigm for end-to-end autonomous driving, valued for their potential to leverage world knowledge and reason about complex driving scenes. However, existing methods suffer from two critical limitations: a persistent misalignment between language instructions and action outputs, and the inherent inefficiency of typical auto-regressive action generation. In this paper, we introduce LinkVLA, a novel architecture that directly addresses these challenges to enhance both alignment and efficiency. First, we establish a structural link by unifying language and action tokens into a shared discrete codebook, processed within a single multi-modal model. This structurally enforces cross-modal consistency from the ground up. Second, to create a deep semantic link, we introduce an auxiliary action understanding objective that trains the model to generate descriptive captions from trajectories, fostering a bidirectional language–action mapping. Finally, we replace the slow, step-by-step generation with a two-step coarse-to-fine generation method (C2F) that efficiently decodes the action sequence, saving 86% inference time. Experiments on closed-loop driving benchmarks show consistent gains in instruction following accuracy and driving performance, alongside reduced inference latency.

Low-resource languages challenge multilingual LLMs due to limited high-quality training data, leading to weaker performance on complex reasoning and knowledge tasks. To address this, we propose improving training data quality through data synthesis, moving beyond simple resource scaling. First, we introduce SynTrans, which translates high-quality, knowledge-rich English data into low-resource languages during pre-training to inject world knowledge, though at the cost of semantic fluency. To overcome low-quality data issues while maintaining fluency, we also propose SynRank. SynRank leverages synthetic data as positive samples to train a classifier that ranks and filters noisy real-world data, enabling the extraction of high-quality subsets without expensive human cleaning. Experiments show SynRank matches handcrafted rule-based filtering by human experts and significantly improves knowledge-intensive task performance at the same filtering rate. Remarkably, higher filtering rates even improve performance with less data, demonstrating the efficiency and effectiveness of our method, surpassing expert filtering. Lastly, we introduce DA-QwenScore, a training-free metric that evaluates corpus quality by normalizing model loss with diversity measures, further enhancing evaluation efficiency. Our insights into knowledge injection could advance low-resource multilingual LLM development.

Multi-frame story illustration requires long-horizon coherence beyond single-image text-to-image generation, including narrative decomposition and persistent character identity, layout, and affect across frames. We propose Story-to-Executable Descriptions (S2ED), a training-free, model-agnostic, prompt-layer framework that converts a full story into a sequence of explicit, editable executable descriptions for more consistent rendering. S2ED coordinates three agents to segment the narrative, ground canonical character attributes, and enrich spatial and affective cues, enabling interpretable prompt-carried state propagation and local edits to repair drift without retraining the generator. Experiments on Flintstones and Shakoo Maku show that S2ED improves sequence-level consistency and character fidelity over strong prompting, large-model planning, and a reference training-based method, under both automatic metrics and human judgments. We also deploy S2ED in an end-to-end story-to-storybook system for children's illustrated stories, with a supplementary video.

Code performance optimization is paramount in real-world software engineering and critical for production-level systems. While Large Language Models (LLMs) have demonstrated impressive capabilities in code generation and bug fixing, their proficiency in enhancing code performance at the repository level remains largely unexplored. To address this gap, we introduce SWE-Perf, the first benchmark specifically designed to systematically evaluate LLMs on code performance optimization tasks within authentic repository contexts. SWE-Perf comprises 140 carefully curated instances, each derived from performance-improving pull requests from popular GitHub repositories. Each benchmark instance includes the relevant codebase, target functions, performance-related tests, expert-authored patches, and executable environments. Through a comprehensive evaluation of representative methods that span file-level and repo-level approaches (e.g., Agentless and OpenHands), we reveal a substantial capability gap between existing LLMs and expert-level optimization performance, highlighting critical research opportunities in this emerging field.

Assessing mental disorder causes and severity from social media history or clinical interviews is essential for early intervention and monitoring population mental health. However, existing methods face several challenges, especially under long-context: the inherently noisy and lengthy nature of online content obscures crucial clinical signals, while the computational intensity of Transformer-based models limits practical deployment. To address these issues, this paper proposes MindSpan, a novel framework that enhances automated mental health profiling by knowledge-augmented compression with output-controlled generation. Specifically, MindSpan first compresses the input long-context by actively retrieving and retaining the top symptom-relevant snippets, effectively filtering out irrelevant information. Then it applies a profiling-specific generation template that constrains the Large Language Models (LLMs) to produce structured, consistent, and clinically interpretable severity ratings. Experiments on two public benchmarks demonstrate that MindSpan achieves better performance. This work provides a practical and efficient solution for scalable, automated mental health assessment

Reinforcement Learning (RL) for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, a baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes (N=32) with only N=4, reducing token consumption by over 65

Multimodal emotion recognition in conversations (MERC) is important for applications such as emotional companion systems, intelligent customer service, and mental health interactions. However, most methods based on multimodal large language models (MLLMs) are evaluated in an offline setting, where the full conversation context is available at prediction time. This does not match the causal constraints in real-time deployment. In this paper, we systematically measure the performance gap between offline and real-time settings on the MELD dataset. We further design a lightweight Causal Emotion Memory (CEM) module that uses a gated recurrent mechanism to compress historical emotional states. We also incorporate emotion trajectory prompting, which explicitly encodes the history of emotion changes to complement the implicit memory modeling. Experiments show that our method improves real-time weighted F1 by up to 3.76

Large Language Models (LLMs) have demonstrated impressive capabilities and are increasingly being explored for applications in the legal field. Despite these advancements, LLMs still face challenges in providing reliable legal assistance to legal professionals and laypeople due to the lack of legal knowledge. Under this limitation, recent research has begun to explore the integration of legal knowledge into LLMs to enhance LLMs' comprehension, analysis, and decision-making abilities. We first explore the different types of legal knowledge that are or can be used by LLMs. Then, we comprehensively review four types of current methods for fusing this knowledge. Next, we discuss the challenges faced by legal LLMs and propose future research directions, highlighting the importance of collaboration between academia and legal practitioners to advance applications in the legal field.

Real-world enterprise data intelligence workflows encompass data engineering that turns raw sources into analytical-ready tables and data analysis that convert those tables into decision-oriented insights. We introduce DAComp, a benchmark of 236 tasks that mirrors these complex workflows. Data engineering (DE) tasks require repository-level engineering on industrial schemas, including designing and building multi-stage SQL pipelines from scratch and evolving existing systems under evolving requirements. Data analysis (DA) tasks pose open-ended business problems that demand strategic planning, exploratory analysis through iterative coding, interpretation of intermediate results, and the synthesis of actionable recommendations. Engineering tasks are scored through execution-based, multi-metric evaluation. Open-ended tasks are assessed by a reliable, experimentally validated LLM-judge, which is guided by hierarchical, meticulously crafted rubrics. Our experiments reveal that even state-of-the-art agents falter on DAComp. Performance on DE tasks is particularly low, with success rates under 10\%, exposing a critical bottleneck in holistic pipeline orchestration, not merely code generation. Scores on DA tasks also average below 40\%, highlighting profound deficiencies in open-ended reasoning and demonstrating that engineering and analysis are distinct capabilities. By clearly diagnosing these limitations, DAComp provides a rigorous and realistic testbed to drive the development of truly capable autonomous data agents for enterprise settings. Our data and code are available at \url{https://anonymous.4open.science/r/DAComp-397A}.

Despite growing interest in graph-based models for knowledge recombination prediction using academic knowledge graphs, existing approaches suffer from significant limitations: they fail to learn informative and robust knowledge entity representations by neglecting high-order information, inadequately account for real-world dynamics, and crucially, cannot provide readable rationales for their predictions. We address these challenges with H2GLM, which reformulates traditional graph learning as heterogeneous hypergraph learning to capture high order information, incorporating a variational autoencoder (VAE) mechanism to enhance informativeness and robustness. Our approach then integrates large language models (LLMs) with the learned graph contextual information through a step-wise methodology, enabling evidence-supported decisions with clear, readable rationales. Experimental results highlight that H2GLM outperforms previous strong graph-based and LLM-based baselines by 4% to 8% accuracy, 3% to 9% in AUC and 5% to 8% in F1 on extensive academic knowledge graphs containing over 1,000,000 nodes, with a small amount of training data. Visualizations and case studies further illustrate our method's substantial utility over existing approaches in real-world scenarios. Further explainability and efficiency analyses underscore the practical value of our method

We present BaziQA-Benchmark, a standardized benchmark for evaluating symbolic and temporally compositional reasoning in large language models. The benchmark is derived from 200 professionally curated, multiple-choice problems from the Global Fortune-teller Competition (2021–2025), where each instance requires structured inference over a fixed symbolic chart and interacting temporal conditions. Unlike anecdotal or prompt-driven evaluations, BaziQA-Benchmark enables objective scoring and controlled comparison across years, domains, and model families. We evaluate contemporary language models under a multi-turn setting and analyze performance variation across temporal difficulty, reasoning domains, and inference protocols.To further probe reasoning behavior, we introduce a lightweight Structured Reasoning Protocol that constrains inference order without adding domain knowledge. Results show that models consistently outperform chance but remain far from saturation, exhibiting pronounced sensitivity to temporal composition and reasoning order, as well as systematic failures on precise temporal localization and multi-condition symbolic judgments.

Reinforcement Learning (RL) for training Large Language Models is notoriously unstable. While recent studies attribute this to "training inference mismatch stemming" from inconsistent hybrid engines, standard remedies, such as Importance Sampling, might fail during extended training runs. In this work, we analyze this instability through the lens of optimization, demonstrating that gradient noise and training-inference mismatch escalate in tandem as training progresses. Meanwhile, we find that the mismatch can be effectively suppressed by shrinking the update size. Taken together, we deduce that the mismatch is not merely a static numerical discrepancy, but a dynamic failure coupled with the model's optimization. Based on this insight, we propose a simple yet effective solution: a specialized Learning Rate (LR) scheduler. Instead of pre-defined decay schedule in traditional LR scheduler, our method dynamically triggers LR decay based on response length, which we identify as a reliable early-warning signal for impending instability. Empirical evidence suggests that by reducing the learning rate as gradient noise rises, we can consistently stabilize RL training and keep the training-inference mismatch at a safe level.

Real-world enterprise text-to-SQL workflows often involve complex cloud or local data across various database systems, multiple SQL queries in various dialects, and diverse operations from data transformation to analytics.We introduce Spider 2.0, an evaluation framework comprising $632$ real-world text-to-SQL workflow problems derived from enterprise-level database use cases. The databases in Spider 2.0 are sourced from real data applications, often containing over 1,000 columns and stored in local or cloud database systems such as BigQuery and Snowflake.We show that solving problems in Spider 2.0 frequently requires understanding and searching through database metadata, dialect documentation, and even project-level codebases. This challenge calls for models to interact with complex SQL workflow environments, process extremely long contexts, perform intricate reasoning, and generate multiple SQL queries with diverse operations, often exceeding $100$ lines, which goes far beyond traditional text-to-SQL challenges.Our evaluations indicate that based on o1-preview, our code agent framework successfully solves only 21.3\% of the tasks, compared with 91.2\% on Spider 1.0 and 73.0\% on BIRD.Our results on Spider 2.0 show that while language models have demonstrated remarkable performance in code generation --- especially in prior text-to-SQL benchmarks --- they require significant improvement in order to achieve adequate performance for real-world enterprise usage.Progress on Spider 2.0 represents crucial steps towards developing intelligent, autonomous, code agents for real-world enterprise settings.Our code, baseline models, and data are available at [spider2-sql.github.io](spider2-sql.github.io) .

Language models (LMs) like GPT and Claude have shown impressive abilities in a range of natural language processing (NLP) tasks. Among these tasks, code understanding and generation have quickly become one of the most popular applications of LMs, given its nature of executable logic forms. However, there is a practical understanding of how programming knowledge can be combined with natural language to automate software development. Moreover, recent studies also empirically demonstrate that code can be a better form for complex reasoning and agentic task automation, but they do not indicate their significance. In this tutorial, we deem such superior capabilities brought by code modeling as Code Intelligence, and aim to provide a coherent overview of recent advances in this topic. We will start by first providing preliminaries of training foundation models on code and their common practices. We will then focus on downstream tasks in the domain of code and their evaluations. Then, we will cover how code can contribute to advancements in general tasks, and the opportunities of future research on Code Intelligence.

Chatbot Arena is a popular platform for evaluating LLMs by pairwise battles, where users vote for their preferred response from two randomly sampled anonymous models. While Chatbot Arena is widely regarded as a reliable LLM ranking leader-board, we show that crowdsourced voting can be rigged to improve (or decrease) the ranking of a target model mt. We first introduce a straightforward target-only rigging strategy that focuses on new battles involving m(t), identifying it via watermarking or a binary classifier, and exclusively voting formt wins. However, this strategy is practically inefficient because there are over 190 models on Chatbot Arena and on average only about 1% of new battles will involve mt. To overcome this, we propose omnipresent rigging strategies, exploiting the Elo rating mechanism of Chatbot Arena that any new vote on a battle can influence the ranking of the target model mt, even if mt is not directly involved in the battle. We conduct experiments on around 1.7 million historical votes from the Chatbot Arena Notebook, showing that omnipresent rigging strategies can improve model rankings by rigging only hundreds of new votes. While we have evaluated several defense mechanisms, our findings highlight the importance of continued efforts to prevent vote rigging. Code is publicly available to reproduce all experiments.

Language model pretraining involves training on extensive corpora, where data quality plays a pivotal role. In this work, we aim to directly estimate the contribution of data during pretraining and select pretraining data in an efficient manner. Specifically, we draw inspiration from recent findings showing that compression efficiency (i.e., normalized loss) of diverse models on certain text correlates strongly with their downstream performance, when the text domain aligns with the downstream benchmarks (Huang et al., 2024). Building on this observation, we hypothesize that data on which model losses are predictive of downstream abilities also contribute effectively to learning, which shares similar intuition with Thrush et al. (2024). To leverage this insight, we introduce predictive data selection (PreSelect), a lightweight and efficient data selection method that requires training and deploying only a fastText-based scorer. Through comprehensive experiments with 1B and 3B parameter models, we demonstrate that models trained on 30B tokens selected with PreSelect surpass the performance of the vanilla baseline trained on 300B tokens, achieving a 10x reduction in compute requirements. Furthermore, PreSelect significantly outperforms other competitive data selection baselines, such as DCLM and FineWeb-Edu on a scale of 3B models trained on 100B tokens. We open-source our trained data selection scorer along with the curated datasets at https://github.com/hkust-nlp/PreSelect.