Xiaozhe Li

Haodong Duan, Xinyu Fang, Junming Yang et al. · pku

We present VLMEvalKit: an open-source toolkit for evaluating large multi-modality models based on PyTorch. The toolkit aims to provide a user-friendly and comprehensive framework for researchers and developers to evaluate existing multi-modality models and publish reproducible evaluation results. In VLMEvalKit, we implement over 200+ different large multi-modality models, including both proprietary APIs and open-source models, as well as more than 80 different multi-modal benchmarks. By implementing a single interface, new models can be easily added to the toolkit, while the toolkit automatically handles the remaining workloads, including data preparation, distributed inference, prediction post-processing, and metric calculation. Although the toolkit is currently mainly used for evaluating large vision-language models, its design is compatible with future updates that incorporate additional modalities, such as audio and video. Based on the evaluation results obtained with the toolkit, we host OpenVLM Leaderboard, a comprehensive leaderboard to track the progress of multi-modality learning research. The toolkit is released on https://github.com/open-compass/VLMEvalKit and is actively maintained.

Xiaozhe Li, Tianyi Lyu, Yizhao Yang et al.

Large Language Models (LLMs) struggle with long-horizon tasks due to the "context bottleneck" and the "lost-in-the-middle" phenomenon, where accumulated noise from verbose environments degrades reasoning over multi-turn interactions. To address this issue, we introduce a symbiotic framework that decouples context management from task execution. Our architecture pairs a lightweight, specialized policy model, ContextCurator, with a powerful frozen foundation model, TaskExecutor. Trained via reinforcement learning, ContextCurator actively reduces information entropy in the working memory. It aggressively prunes environmental noise while preserving reasoning anchors, that is, sparse data points that are critical for future deductions. On WebArena, our framework improves the success rate of Gemini-3.0-flash from 36.4% to 41.2% while reducing token consumption by 8.8% (from 47.4K to 43.3K). On DeepSearch, it achieves a 57.1% success rate, compared with 53.9%, while reducing token consumption by a factor of 8. Remarkably, a 7B ContextCurator matches the context management performance of GPT-4o, providing a scalable and computationally efficient paradigm for autonomous long-horizon agents.

Xiaozhe Li, Jixuan Chen, Xinyu Fang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning and tool use. However, the fundamental cognitive faculties essential for problem solving, including perception, reasoning, and memory, remain the stable core of intelligence. Unlike memorizing specific patterns, humans succeed in novel environments by applying these intrinsic faculties to adapt and optimize. Yet, whether LLMs possess this essential capacity, namely the ability to continuously refine solutions in response to dynamic environmental feedback, remains underexplored. To address this challenge, we introduce OPT-BENCH, a benchmark for evaluating self-improvement capabilities in large-scale search spaces. By combining 20 machine learning tasks with 10 classic NP-hard problems, OPT-BENCH provides a rigorous setting to assess whether agents can adapt through intrinsic self-reflection rather than rote tool application. We further propose OPT-Agent, a framework that emulates human-like cognitive adaptation. It operates through a general perception, memory, and reasoning loop, iteratively refining solutions based on environmental feedback. Through extensive experiments on 19 LLMs from 7 model families, including reasoning models, general models, and open-source models ranging from 3B to 235B parameters, we demonstrate that stronger models are more effective at leveraging feedback signals for self-improvement. However, this upper-bound adaptability remains fundamentally constrained by the models' base capacity, and even the most advanced LLMs still fall short of human expert performance.

Yichuan Ma, Linyang Li, Yongkang chen et al.

As large language models (LLMs) increasingly tackle complex reasoning tasks, test-time scaling has become critical for enhancing capabilities. However, in agentic scenarios with frequent tool calls, the traditional generation-length-based definition breaks down: tool latency decouples inference time from generation length. We propose Timely Machine, redefining test-time as wall-clock time, where models dynamically adjust strategies based on time budgets. We introduce Timely-Eval, a benchmark spanning high-frequency tool calls, low-frequency tool calls, and time-constrained reasoning. By varying tool latency, we find smaller models excel with fast feedback through more interactions, while larger models dominate high-latency settings via superior interaction quality. Moreover, existing models fail to adapt reasoning to time budgets. We propose Timely-RL to address this gap. After cold-start supervised fine-tuning, we use reinforcement learning to enhance temporal planning. Timely-RL improves time budget awareness and consistently boosts performance across Timely-Eval. We hope our work offers a new perspective on test-time scaling for the agentic era.

Peiji Li, Linyang Li, Handa Sun et al.

Large language models have demonstrated strong reasoning capabilities in complex tasks through tool integration, which is typically framed as a Markov Decision Process and optimized with trajectory-level RL algorithms such as GRPO. However, a common class of reasoning tasks, iterative optimization, presents distinct challenges: the agent interacts with the same underlying environment state across turns, and the value of a trajectory is determined by the best turn-level reward rather than cumulative returns. Existing GRPO-based methods cannot perform fine-grained, turn-level optimization in such settings, while black-box optimization methods discard prior knowledge and reasoning capabilities. To address this gap, we propose Turn-Level GRPO (TL-GRPO), a lightweight RL algorithm that performs turn-level group sampling for fine-grained optimization. We evaluate TL-GRPO on analog circuit sizing (ACS), a challenging scientific optimization task requiring multiple simulations and domain expertise. Results show that TL-GRPO outperforms standard GRPO and Bayesian optimization methods across various specifications. Furthermore, our 30B model trained with TL-GRPO achieves state-of-the-art performance on ACS tasks under same simulation budget, demonstrating both strong generalization and practical utility.

Xiaozhe Li, Tianyi Lyu, Yang Li et al.

Reinforcement learning can train LLM agents from sparse task rewards, but long-horizon credit assignment remains challenging: a single success-or-failure signal must be distributed across many actions. Existing methods rely on trajectory-level rewards or proxy signals, without fully leveraging per-step environmental feedback. Multi-turn agent settings are underexplored, where feedback can include error messages, page changes, observations, or reference trajectories. We systematically study five feedback sources and two insertion granularities and introduce SERL, a selective environment-reweighted learning framework. SERL uses the task reward to determine update direction, while environment feedback adjusts placement and magnitude, focusing on critical actions. On ALFWorld and WebShop, SERL achieves 90.0% and 80.1% success, outperforming strong RL and distillation baselines. Analysis shows that grounded, action-relevant feedback at meaningful points consistently outperforms indiscriminate use of longer or richer context.

Xiaozhe Li, Yang Li, Xinyu Fang et al.

On-policy reinforcement learning methods like GRPO suffer from mode collapse: they exhibit reduced solution diversity, concentrating probability mass on a single solution once discovered and ceasing exploration of alternative strategies. We show this stems from reverse KL minimization's mode-seeking behavior, which reinforces the first high-reward trajectory found rather than maintaining a distribution over multiple diverse solutions. We propose DMPO (Distribution-Matching Policy Optimization), which prevents mode collapse through principled approximation of forward KL minimization. DMPO constructs a group level target distribution over sampled trajectories proportional to their rewards, then aligns the policy distribution to this target. This provides mode-covering behavior without requiring sampling from the intractable global target distribution, enabling sustained exploration throughout training. We validate DMPO on NP-hard combinatorial optimization, where exponentially many feasible solutions exist but only a few approach optimality, an ideal testbed for evaluating exploration. DMPO achieves 43.9% Quality Ratio on text-based NP-Bench (vs. GRPO's 40.1%) and 43.1% on vision-based NP-Bench (vs. 38.4%), demonstrating 9% and 12% relative improvements respectively. These gains generalize to mathematical reasoning (+2.0%) and out-of-domain tasks (+2.3%), showing that diversity-preserving training enhances general reasoning capabilities across modalities. Our work establishes distribution matching as a practical, principled approach to preventing mode collapse in on-policy RL, with consistent quality improvements demonstrating sustained exploration across diverse reasoning tasks.

Xiaozhe Li, Jixuan Chen, Xinyu Fang et al. · pku

Large Language Models (LLMs) have shown remarkable capabilities in solving diverse tasks. However, their proficiency in iteratively optimizing complex solutions through learning from previous feedback remains insufficiently explored. To bridge this gap, we present OPT-BENCH, a comprehensive benchmark designed to evaluate LLM agents on large-scale search space optimization problems. OPT-BENCH includes 20 real-world machine learning tasks sourced from Kaggle and 10 classical NP problems, offering a diverse and challenging environment for assessing LLM agents on iterative reasoning and solution refinement. To enable rigorous evaluation, we introduce OPT-Agent, an end-to-end optimization framework that emulates human reasoning when tackling complex problems by generating, validating, and iteratively improving solutions through leveraging historical feedback. Through extensive experiments on 9 state-of-the-art LLMs from 6 model families, we analyze the effects of optimization iterations, temperature settings, and model architectures on solution quality and convergence. Our results demonstrate that incorporating historical context significantly enhances optimization performance across both ML and NP tasks. All datasets, code, and evaluation tools are open-sourced to promote further research in advancing LLM-driven optimization and iterative reasoning. Project page: \href{https://github.com/OliverLeeXZ/OPT-BENCH}{https://github.com/OliverLeeXZ/OPT-BENCH}.

Chunyi Li, Xiaozhe Li, Zicheng Zhang et al.

With the emergence of Multimodal Large Language Models (MLLMs), hundreds of benchmarks have been developed to ensure the reliability of MLLMs in downstream tasks. However, the evaluation mechanism itself may not be reliable. For developers of MLLMs, questions remain about which benchmark to use and whether the test results meet their requirements. Therefore, we propose a critical principle of Information Density, which examines how much insight a benchmark can provide for the development of MLLMs. We characterize it from four key dimensions: (1) Fallacy, (2) Difficulty, (3) Redundancy, (4) Diversity. Through a comprehensive analysis of more than 10,000 samples, we measured the information density of 19 MLLM benchmarks. Experiments show that using the latest benchmarks in testing can provide more insight compared to previous ones, but there is still room for improvement in their information density. We hope this principle can promote the development and application of future MLLM benchmarks. Project page: https://github.com/lcysyzxdxc/bench4bench

Xiaozhe Li, Tianyi Lyu, Siyi Yang et al.

Understanding human intent is a high-level cognitive challenge for Large Language Models (LLMs), requiring sophisticated reasoning over noisy, conflicting, and non-linear discourse. While LLMs excel at following individual instructions, their ability to distill Collective Intent - the process of extracting consensus, resolving contradictions, and inferring latent trends from multi-source public discussions - remains largely unexplored. To bridge this gap, we introduce COIN-BENCH, a dynamic, real-world, live-updating benchmark specifically designed to evaluate LLMs on collective intent understanding within the consumer domain. Unlike traditional benchmarks that focus on transactional outcomes, COIN-BENCH operationalizes intent as a hierarchical cognitive structure, ranging from explicit scenarios to deep causal reasoning. We implement a robust evaluation pipeline that combines a rule-based method with an LLM-as-the-Judge approach. This framework incorporates COIN-TREE for hierarchical cognitive structuring and retrieval-augmented verification (COIN-RAG) to ensure expert-level precision in analyzing raw, collective human discussions. An extensive evaluation of 20 state-of-the-art LLMs across four dimensions - depth, breadth, informativeness, and correctness - reveals that while current models can handle surface-level aggregation, they still struggle with the analytical depth required for complex intent synthesis. COIN-BENCH establishes a new standard for advancing LLMs from passive instruction followers to expert-level analytical agents capable of deciphering the collective voice of the real world. See our project page on COIN-BENCH.

Xiaozhe Li, Xinyu Fang, Shengyuan Ding et al.

Large Language Models (LLMs) have achieved remarkable success on reasoning benchmarks through Reinforcement Learning with Verifiable Rewards (RLVR), excelling at tasks such as math, coding, logic, and puzzles. However, existing benchmarks evaluate only correctness, while overlooking optimality, namely the ability to find the best solutions under constraints. We propose OPT-BENCH, the first comprehensive framework for training and evaluating LLMs on NP-hard optimization problems through quality-aware RLVR. OPT-BENCH provides three key components: a scalable training infrastructure with instance generators, quality verifiers, and optimal baselines across 10 tasks; a rigorous benchmark with 1,000 instances evaluating both feasibility, measured by Success Rate, and quality, measured by Quality Ratio; and quality-aware rewards that enable continuous improvement beyond binary correctness. Training on Qwen2.5-7B-Instruct-1M with 15K examples achieves 93.1% SR and 46.6% QR, significantly outperforming GPT-4o, which achieves 29.6% SR and 14.6% QR. Beyond optimization, training on OPT-BENCH transfers to diverse tasks, including mathematics (+2.2%), logic (+1.2%), knowledge (+4.1%), and instruction following (+6.1%). Our analysis reveals that quality-aware rewards improve solutions by 28.8% over binary rewards, and that task diversity drives generalization more than data quantity, offering insights into RLVR scaling for complex reasoning.

Xiaozhe Li, Kai WU, Siyi Yang et al.

Recent advancements in text-to-video (T2V) generation have leveraged diffusion models to enhance visual coherence in videos synthesized from textual descriptions. However, existing research primarily focuses on object motion, often overlooking cinematic language, which is crucial for conveying emotion and narrative pacing in cinematography. To address this, we propose a threefold approach to improve cinematic control in T2V models. First, we introduce a meticulously annotated cinematic language dataset with twenty subcategories, covering shot framing, shot angles, and camera movements, enabling models to learn diverse cinematic styles. Second, we present CameraDiff, which employs LoRA for precise and stable cinematic control, ensuring flexible shot generation. Third, we propose CameraCLIP, designed to evaluate cinematic alignment and guide multi-shot composition. Building on CameraCLIP, we introduce CLIPLoRA, a CLIP-guided dynamic LoRA composition method that adaptively fuses multiple pre-trained cinematic LoRAs, enabling smooth transitions and seamless style blending. Experimental results demonstrate that CameraDiff ensures stable and precise cinematic control, CameraCLIP achieves an R@1 score of 0.83, and CLIPLoRA significantly enhances multi-shot composition within a single video, bridging the gap between automated video generation and professional cinematography.\textsuperscript{1}

Xiaozhe Li, Xinyu Fang, Shengyuan Ding et al. · pku

Large Language Models (LLMs) have shown strong reasoning capabilities, with models like OpenAI's O-series and DeepSeek R1 excelling at tasks such as mathematics, coding, logic, and puzzles through Reinforcement Learning with Verifiable Rewards (RLVR). However, their ability to solve more complex optimization problems - particularly NP-hard tasks - remains underexplored. To bridge this gap, we propose NP-ENGINE, the first comprehensive framework for training and evaluating LLMs on NP-hard problems. NP-ENGINE covers 10 tasks across five domains, each equipped with (i) a controllable instance generator, (ii) a rule-based verifier, and (iii) a heuristic solver that provides approximate optimal solutions as ground truth. This generator-verifier-heuristic pipeline enables scalable and verifiable RLVR training under hierarchical difficulties. We also introduce NP-BENCH, a benchmark derived from NP-ENGINE-DATA, specifically designed to evaluate LLMs' ability to tackle NP-hard level reasoning problems, focusing not only on feasibility but also on solution quality. Additionally, we present QWEN2.5-7B-NP, a model trained via zero-RLVR with curriculum learning on Qwen2.5-7B-Instruct, which significantly outperforms GPT-4o on NP-BENCH and achieves SOTA performance with the same model size. Beyond in-domain tasks, we demonstrate that RLVR training on NP-ENGINE-DATA enables strong out-of-domain (OOD) generalization to reasoning tasks (logic, puzzles, math, and knowledge), as well as non-reasoning tasks such as instruction following. We also observe a scaling trend: increasing task diversity improves OOD generalization. These findings suggest that task-rich RLVR training is a promising direction for advancing LLM's reasoning ability, revealing new insights into the scaling laws of RLVR.

Xiaozhe Li, TianYi Lyu, Siyi Yang et al.

Understanding human intent is a complex, high-level task for large language models (LLMs), requiring analytical reasoning, contextual interpretation, dynamic information aggregation, and decision-making under uncertainty. Real-world public discussions, such as consumer product discussions, are rarely linear or involve a single user. Instead, they are characterized by interwoven and often conflicting perspectives, divergent concerns, goals, emotional tendencies, as well as implicit assumptions and background knowledge about usage scenarios. To accurately understand such explicit public intent, an LLM must go beyond parsing individual sentences; it must integrate multi-source signals, reason over inconsistencies, and adapt to evolving discourse, similar to how experts in fields like politics, economics, or finance approach complex, uncertain environments. Despite the importance of this capability, no large-scale benchmark currently exists for evaluating LLMs on real-world human intent understanding, primarily due to the challenges of collecting real-world public discussion data and constructing a robust evaluation pipeline. To bridge this gap, we introduce \bench, the first dynamic, live evaluation benchmark specifically designed for intent understanding, particularly in the consumer domain. \bench is the largest and most diverse benchmark of its kind, supporting real-time updates while preventing data contamination through an automated curation pipeline.