While transformer-based Large Language Models (LLMs) theoretically support massive context windows, they suffer from severe performance degradation when processing long numerical sequences. We attribute this failure to the attention dispersion in the Softmax mechanism, which prevents the model from concentrating attention. To overcome this, we propose Separate Sequence (SepSeq), a training-free, plug-and-play framework to mitigate dispersion by strategically inserting separator tokens. Mechanistically, we demonstrate that separator tokens act as an attention anchor, recalibrating attention to focus on local segments while preserving global context. Extensive evaluations on 9 widely-adopted LLMs confirm the effectiveness of our approach: SepSeq yields an average relative accuracy improvement of 35.6% across diverse domains while reducing 16.4% inference token consumption.

The emergence of groundbreaking large language models capable of performing complex reasoning tasks holds significant promise for addressing various scientific challenges, including those arising in complex clinical scenarios. To enable their safe and effective deployment in real-world healthcare settings, it is urgently necessary to benchmark the diagnostic capabilities of current models systematically. Given the limitations of existing medical benchmarks in evaluating advanced diagnostic reasoning, we present DiagnosisArena, a comprehensive and challenging benchmark designed to rigorously assess professional-level diagnostic competence. DiagnosisArena consists of 1,113 pairs of segmented patient cases and corresponding diagnoses, spanning 28 medical specialties, deriving from clinical case reports published in 10 top-tier medical journals. The benchmark is developed through a meticulous construction pipeline, involving multiple rounds of screening and review by both AI systems and human experts, with thorough checks conducted to prevent data leakage. Our study reveals that even the most advanced reasoning models, o3-mini, o1, and DeepSeek-R1, achieve only 45.82%, 31.09%, and 17.79% accuracy, respectively. This finding highlights a significant generalization bottleneck in current large language models when faced with clinical diagnostic reasoning challenges. Through DiagnosisArena, we aim to drive further advancements in AI’s diagnostic reasoning capabilities, enabling more effective solutions for real-world clinical diagnostic challenges. We openly share the benchmark and evaluation tools for further research and development.

Recent advances in AI-assisted programming have empowered agents to execute complex workflows via command-line interfaces, however, existing benchmarks are limited by short task horizons, data contamination from GitHub scraping, and a lack of fine-grained evaluation metrics, fail to rigorously evaluate the long-horizon planning and execution capabilities essential for realistic software engineering. To address these gaps, we introduce LongCLI-Bench, a comprehensive benchmark designed to evaluate agentic capabilities across long-horizon, realistic, sequential engineering tasks. We curated 20 high-quality, long-horizon tasks from over 1,000 computer science assignments and real-world workflows, covering four engineering categories: from scratch, feature addition, bug fixing, and refactoring. LongCLI-Bench employs a dual-set testing protocol, which measures requirement fulfillment (fail(→)pass) and regression avoidance (pass(→)pass), and incorporates step-level scoring to pinpoint execution failures. Extensive experiments reveal that even state-of-the-art agents achieve pass rates below 20% in LongCLI-Bench. Step-level analysis further indicates that the majority of tasks stall at less than 30% completion, highlighting that critical failures often occur in the early stages. Although self-correction offers marginal gains, human-agent collaboration through plan injection and interactive guidance yields significantly higher improvements. These results highlight that future research must emphasize the development of synergistic human-agent workflows alongside advances in agents’ planning and execution capabilities to overcome key challenges in long-horizon task performance.

Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive benchmark derived from daily AI usage, evaluating 6 core agentic capabilities across 32 real-world scenarios, comprising 138 tasks with specific queries, deliverables, and rubrics. These scenarios require an average of 90 tool calls, 1 million tokens, and hours of execution time to resolve. To enable automated evaluation, we employ a user simulation agent to provide iterative feedback, and a Docker sandbox to conduct visual and functional rubric-based assessment. Experiments reveal that closed-source models significantly outperform open-source models (48.4% vs 32.1%). Further analysis reveals significant disparities across models in resource efficiency, feedback-driven self-correction, and specific tool-use preferences.

High-quality scientific data is critical for advancing LLMs, yet academic literature remains largely underutilized. This work addresses the fundamental question: How can we systematically unlock scientific data’s value for pre-training? First, we construct a large-scale raw scientific corpus but identify a critical Learnability Gap, revealing that direct pre-training yields negligible gains. To bridge this, we develop a multi-stage pipeline featuring content cleaning and pedagogical augmentation, resulting in SciPedia, a 900B-token corpus. Finally, we establish a controlled verification framework: we develop SciPedia-Eval benchmark and conduct 600B tokens of continued pre-training (CPT) starting from transparent base models (3B/7B) trained from scratch. Compared to a CPT baseline trained with general-purpose data, our approach with SciPedia data boosts average performance by +2.12 (3B) and +2.95 (7B), reaching +5.60 and +8.40 on in-domain tasks. This setup further allows us to derive empirical guidelines for data composition and model configurations.

Direct Preference Optimization (DPO) has become a widely used training method for the instruction fine-tuning of large language models (LLMs). In this work, we explore an under-investigated aspect of DPO – its dependency on the reference model or policy. Such reference policies, typically instantiated as the model to be further fine-tuned, are important since they can impose an upper limit on DPO’s effectiveness. Therefore, we address three related research questions in this work. First, we explore the optimal strength of the KL divergence constraint in DPO, which penalizes deviations from the reference policy, and find that DPO is sensitive to this strength. Next, we examine the necessity of the KL-constraint from the reference policies in DPO by providing both theoretical and empirical comparisons between DPO and related learning objectives, demonstrating DPO’s superiority in this controlled setting. Additionally, we investigate whether DPO benefits from stronger reference policies, finding that a stronger reference policy can lead to improved performance, but only when it is similar to the model being fine-tuned. Our findings highlight the confounding role of reference policies in DPO and offer insights for best practices, while also identifying open research questions for future studies.

Large Language Models (LLMs) with web search capabilities show significant potential for deep research, yet current methods—brittle prompt engineering or RAG-based reinforcement learning in controlled environments—fail to capture real-world complexities. In this paper, we introduce DeepResearcher, the first comprehensive framework for end-to-end training of LLM-based deep research agents through scaling reinforcement learning (RL) in real-world environments with authentic web search interactions. Unlike RAG approaches reliant on fixed corpora, DeepResearcher trains agents to navigate the noisy, dynamic open web. We implement a specialized multi-agent architecture where browsing agents extract relevant information from various webpage structures and overcoming significant technical challenges. Extensive experiments on open-domain research tasks demonstrate that DeepResearcher achieves substantial improvements of up to 28.9 points over prompt engineering-based baselines and up to 7.2 points over RAG-based RL agents. Our qualitative analysis reveals emergent cognitive behaviors from end-to-end RL training, such as planning, cross-validation, self-reflection for research redirection, and maintain honesty when unable to find definitive answers. Our results highlight that end-to-end training in real-world web environments is fundamental for developing robust research capabilities aligned with real-world applications. The source codefor DeepResearcher is released at: https://github.com/GAIR-NLP/DeepResearcher.

We introduce DavIR, a model-based data selection method for post-training Large Language Models. DavIR generalizes Reducible Holdout Loss to core-set selection problem of causal language modeling, and quantifies the learnability of a given datum with respect to a pre-trained LLM based on relative reduction in loss during fine-tuning, a metric we show to be closely related to the implicit reward model described in Direct Preference Optimization (DPO). We show that 6% of Alpaca dataset selected with DavIR can steer both the LLaMA and Gemma model family to produce superior performance compared to the same models trained on the full 52K dataset. We also show that Alpaca dataset compressed with DavIR can be combined with GSM8K dataset to effectively balance open-domain freeform QA and mathematical reasoning capabilities. Finally, we apply the DavIR objective to DPO and develop a normalized DavIR-DPO objective which improves alignment performance of Zephyr-7B-SFT model by 8% (relative) on AlpacaEval, compared against training on vanilla DPO objective.

As Large Language Models (LLMs) increasingly participate in human-AI interactions, evaluating their Theory of Mind (ToM) capabilities - particularly their ability to track dynamic mental states - becomes crucial. While existing benchmarks assess basic ToM abilities, they predominantly focus on static snapshots of mental states, overlooking the temporal evolution that characterizes real-world social interactions. We present DynToM, a novel benchmark specifically designed to evaluate LLMs’ ability to understand and track the temporal progression of mental states across interconnected scenarios. Through a systematic four-step framework, we generate 1,100 social contexts encompassing 5,500 scenarios and 78,100 questions, each validated for realism and quality. Our comprehensive evaluation of ten state-of-the-art LLMs reveals that their average performance underperforms humans by 44.7%, with performance degrading significantly when tracking and reasoning about the shift of mental states. This performance gap highlights fundamental limitations in current LLMs’ ability to model the dynamic nature of human mental states.