Zicheng Lin

Yiming Ren, Yiran Xu, Zicheng Lin et al.

We identify a new dimension for enhancing rollout diversity in Group Relative Policy Optimization (GRPO) for LLMs. While GRPO relies on diverse rollouts, prevailing strategies primarily increase diversity by injecting more token-level randomness, which may introduce step-wise noise and lead to incoherent trajectories. We uncover that smaller models within the same model family inherently exhibit higher policy-level diversity, indicated by their superior pass@k relative to larger counterparts as sample counts increase. Unlike token-level noise, this diversity is temporally correlated, preserves logical consistency, and provides structured exploration signals for gradient estimation. We thus propose S2L-PO (Small-to-Large Policy Optimization), a framework that leverages fixed small models as natural explorers to train larger models. To balance exploration and exploitation, we design a progressive annealing strategy that transitions from offline small-model rollouts to the large learner's own sampling. This shift elegantly avoids mid-training performance drops caused by the small model's capacity limits, achieving faster convergence and unlocking a higher performance ceiling. S2L-PO improves accuracy on diverse mathematical reasoning benchmarks (e.g., +8.8% on AIME 24 using a 1.7B explorer to guide the 8B model) while reducing rollout compute.

Ruilin Luo, Liyuan Wang, Binghuai Lin et al.

Large Language Models (LLMs) have emerged as powerful tools for Text-to-SQL tasks, exhibiting remarkable reasoning capabilities. Different from tasks such as math word problems and commonsense reasoning, SQL solutions have a relatively fixed pattern. This facilitates the investigation of whether LLMs can benefit from categorical thinking, mirroring how humans acquire knowledge through inductive reasoning based on comparable examples. In this study, we propose that employing query group partitioning allows LLMs to focus on learning the thought processes specific to a single problem type, consequently enhancing their reasoning abilities across diverse difficulty levels and problem categories. Our experiments reveal that multiple advanced LLMs, when equipped with PTD-SQL, can either surpass or match previous state-of-the-art (SOTA) methods on the Spider and BIRD datasets. Intriguingly, models with varying initial performances have exhibited significant improvements, mainly at the boundary of their capabilities after targeted drilling, suggesting a parallel with human progress. Code is available at https://github.com/lrlbbzl/PTD-SQL.

Tianming Liang, Qirui Du, Jian-Fang Hu et al.

Segmentation based on language has been a popular topic in computer vision. While recent advances in multimodal large language models (MLLMs) have endowed segmentation systems with reasoning capabilities, these efforts remain confined by the frozen internal knowledge of MLLMs, which limits their potential for real-world scenarios that involve up-to-date information or domain-specific concepts. In this work, we propose \textbf{Seg-ReSearch}, a novel segmentation paradigm that overcomes the knowledge bottleneck of existing approaches. By enabling interleaved reasoning and external search, Seg-ReSearch empowers segmentation systems to handle dynamic, open-world queries that extend beyond the frozen knowledge of MLLMs. To effectively train this capability, we introduce a hierarchical reward design that harmonizes initial guidance with progressive incentives, mitigating the dilemma between sparse outcome signals and rigid step-wise supervision. For evaluation, we construct OK-VOS, a challenging benchmark that explicitly requires outside knowledge for video object segmentation. Experiments on OK-VOS and two existing reasoning segmentation benchmarks demonstrate that our Seg-ReSearch improves state-of-the-art approaches by a substantial margin. Code and data will be released at https://github.com/iSEE-Laboratory/Seg-ReSearch.

Zhengchao Huang, Bin Xia, Zicheng Lin et al.

The rapid advancement of deepfake technologies has sparked widespread public concern, particularly as face forgery poses a serious threat to public information security. However, the unknown and diverse forgery techniques, varied facial features and complex environmental factors pose significant challenges for face forgery analysis. Existing datasets lack descriptive annotations of these aspects, making it difficult for models to distinguish between real and forged faces using only visual information amid various confounding factors. In addition, existing methods fail to yield user-friendly and explainable results, hindering the understanding of the model's decision-making process. To address these challenges, we introduce a novel Open-World Face Forgery Analysis VQA (OW-FFA-VQA) task and its corresponding benchmark. To tackle this task, we first establish a dataset featuring a diverse collection of real and forged face images with essential descriptions and reliable forgery reasoning. Based on this dataset, we introduce FFAA: Face Forgery Analysis Assistant, consisting of a fine-tuned Multimodal Large Language Model (MLLM) and Multi-answer Intelligent Decision System (MIDS). By integrating hypothetical prompts with MIDS, the impact of fuzzy classification boundaries is effectively mitigated, enhancing model robustness. Extensive experiments demonstrate that our method not only provides user-friendly and explainable results but also significantly boosts accuracy and robustness compared to previous methods.

Zicheng Lin, Tian Liang, Jiahao Xu et al.

Mathematical reasoning tasks pose significant challenges for large language models (LLMs) because they require precise logical deduction and sequence analysis. In this work, we introduce the concept of critical tokens -- elements within reasoning trajectories that significantly influence incorrect outcomes. We present a novel framework for identifying these tokens through rollout sampling and demonstrate their substantial divergence from traditional error tokens. Through extensive experiments on datasets such as GSM8K and MATH500, we show that identifying and replacing critical tokens significantly improves model accuracy. We propose an efficient methodology for pinpointing these tokens in large-scale datasets using contrastive estimation and extend this framework to enhance model training processes with direct preference optimization (DPO). Experimental results on GSM8K and MATH500 benchmarks with the widely used models Llama-3 (8B and 70B) and Deepseek-math (7B) demonstrate the effectiveness of the proposed approach, cDPO. Our results underscore the potential of leveraging critical tokens to reduce errors in reasoning tasks, advancing the development of AI systems capable of robust logical deduction. Our code, annotated datasets, and trained models are available at https://github.com/chenzhiling9954/Critical-Tokens-Matter to support and encourage future research in this promising field.

Yifan Chen, Fei Yin, Qingyan Bai et al.

We introduce MMCL-Bench, a benchmark for multimodal context learning: learning task-local rules, procedures, and empirical patterns from visual or mixed-modality teaching context and applying them to new visual instances. Unlike text-only context learning or standard multimodal question answering, this setting requires models to recover and localize relevant evidence from images, screenshots, manuals, videos, and frame sequences before they can reason over the learned context. MMCL-Bench contains 102 tasks spanning three categories: rule system application, procedural task execution, and empirical discovery and induction. We evaluate frontier multimodal models with strict rubric-based scoring and find that current systems remain far from robust multimodal context learning, with even the strongest model solving fewer than one-third of tasks under strict evaluation. Diagnostic ablations and error analysis show that failures arise throughout the context-to-answer pipeline, including context anchoring, visual evidence extraction, context reasoning, and response construction. MMCL-Bench thus highlights multimodal context learning as an important unsolved capability bottleneck for current multimodal models.

Zicheng Lin, Zhibin Gou, Tian Liang et al. · tsinghua

The ability of Large Language Models (LLMs) to critique and refine their reasoning is crucial for their application in evaluation, feedback provision, and self-improvement. This paper introduces CriticBench, a comprehensive benchmark designed to assess LLMs' abilities to critique and rectify their reasoning across a variety of tasks. CriticBench encompasses five reasoning domains: mathematical, commonsense, symbolic, coding, and algorithmic. It compiles 15 datasets and incorporates responses from three LLM families. Utilizing CriticBench, we evaluate and dissect the performance of 17 LLMs in generation, critique, and correction reasoning, i.e., GQC reasoning. Our findings reveal: (1) a linear relationship in GQC capabilities, with critique-focused training markedly enhancing performance; (2) a task-dependent variation in correction effectiveness, with logic-oriented tasks being more amenable to correction; (3) GQC knowledge inconsistencies that decrease as model size increases; and (4) an intriguing inter-model critiquing dynamic, where stronger models are better at critiquing weaker ones, while weaker models can surprisingly surpass stronger ones in their self-critique. We hope these insights into the nuanced critique-correct reasoning of LLMs will foster further research in LLM critique and self-improvement.

Ruilin Luo, Zhuofan Zheng, Yifan Wang et al.

Process Reward Models (PRMs) have shown promise in enhancing the mathematical reasoning capabilities of Large Language Models (LLMs) through Test-Time Scaling (TTS). However, their integration into multimodal reasoning remains largely unexplored. In this work, we take the first step toward unlocking the potential of PRMs in multimodal mathematical reasoning. We identify three key challenges: (1) the scarcity of high-quality reasoning data constrains the capabilities of foundation Multimodal Large Language Models (MLLMs), which imposes further limitations on the upper bounds of TTS and reinforcement learning (RL); (2) a lack of automated methods for process labeling within multimodal contexts persists; (3) the employment of process rewards in unimodal RL faces issues like reward hacking, which may extend to multimodal scenarios. To address these issues, we introduce URSA, a three-stage Unfolding multimodal Process-Supervision Aided training framework. We first construct MMathCoT-1M, a high-quality large-scale multimodal Chain-of-Thought (CoT) reasoning dataset, to build a stronger math reasoning foundation MLLM, URSA-8B. Subsequently, we go through an automatic process to synthesize process supervision data, which emphasizes both logical correctness and perceptual consistency. We introduce DualMath-1.1M to facilitate the training of URSA-8B-RM. Finally, we propose Process-Supervised Group-Relative-Policy-Optimization (PS-GRPO), pioneering a multimodal PRM-aided online RL method that outperforms vanilla GRPO. With PS-GRPO application, URSA-8B-PS-GRPO outperforms Gemma3-12B and GPT-4o by 8.4% and 2.7% on average across 6 benchmarks. Code, data and checkpoint can be found at https://github.com/URSA-MATH.

Ruilin Luo, Tianle Gu, Haoling Li et al.

Temporal Knowledge Graph Completion (TKGC) is a complex task involving the prediction of missing event links at future timestamps by leveraging established temporal structural knowledge. This paper aims to provide a comprehensive perspective on harnessing the advantages of Large Language Models (LLMs) for reasoning in temporal knowledge graphs, presenting an easily transferable pipeline. In terms of graph modality, we underscore the LLMs' prowess in discerning the structural information of pivotal nodes within the historical chain. As for the generation mode of the LLMs utilized for inference, we conduct an exhaustive exploration into the variances induced by a range of inherent factors in LLMs, with particular attention to the challenges in comprehending reverse logic. We adopt a parameter-efficient fine-tuning strategy to harmonize the LLMs with the task requirements, facilitating the learning of the key knowledge highlighted earlier. Comprehensive experiments are undertaken on several widely recognized datasets, revealing that our framework exceeds or parallels existing methods across numerous popular metrics. Additionally, we execute a substantial range of ablation experiments and draw comparisons with several advanced commercial LLMs, to investigate the crucial factors influencing LLMs' performance in structured temporal knowledge inference tasks.

Yiming Ren, Zhiqiang Lin, Yu Li et al.

Controllable captioning is essential for precise multimodal alignment and instruction following, yet existing models often lack fine-grained control and reliable evaluation protocols. To address this gap, we present the AnyCap Project, an integrated solution spanning model, dataset, and evaluation. We introduce AnyCapModel (ACM), a lightweight plug-and-play framework that enhances the controllability of existing foundation models for omni-modal captioning without retraining the base model. ACM reuses the original captions from base models while incorporating user instructions and modality features to generate improved captions. To remedy the data scarcity in controllable multimodal captioning, we build AnyCapDataset (ACD), covering three modalities, 28 user-instruction types, and 300\,k high-quality data entries. We further propose AnyCapEval, a new benchmark that provides more reliable evaluation metrics for controllable captioning by decoupling content accuracy and stylistic fidelity. ACM markedly improves caption quality across a diverse set of base models on AnyCapEval. Notably, ACM-8B raises GPT-4oś content scores by 45\% and style scores by 12\%, and it also achieves substantial gains on widely used benchmarks such as MIA-Bench and VidCapBench.

Shuai Bai, Yuxuan Cai, Ruizhe Chen et al.

We introduce Qwen3-VL, the most capable vision-language model in the Qwen series to date, achieving superior performance across a broad range of multimodal benchmarks. It natively supports interleaved contexts of up to 256K tokens, seamlessly integrating text, images, and video. The model family includes both dense (2B/4B/8B/32B) and mixture-of-experts (30B-A3B/235B-A22B) variants to accommodate diverse latency-quality trade-offs. Qwen3-VL delivers three core pillars: (i) markedly stronger pure-text understanding, surpassing comparable text-only backbones in several cases; (ii) robust long-context comprehension with a native 256K-token window for both text and interleaved multimodal inputs, enabling faithful retention, retrieval, and cross-referencing across long documents and videos; and (iii) advanced multimodal reasoning across single-image, multi-image, and video tasks, demonstrating leading performance on comprehensive evaluations such as MMMU and visual-math benchmarks (e.g., MathVista and MathVision). Architecturally, we introduce three key upgrades: (i) an enhanced interleaved-MRoPE for stronger spatial-temporal modeling across images and video; (ii) DeepStack integration, which effectively leverages multi-level ViT features to tighten vision-language alignment; and (iii) text-based time alignment for video, evolving from T-RoPE to explicit textual timestamp alignment for more precise temporal grounding. Under comparable token budgets and latency constraints, Qwen3-VL achieves superior performance in both dense and Mixture-of-Experts (MoE) architectures. We envision Qwen3-VL serving as a foundational engine for image-grounded reasoning, agentic decision-making, and multimodal code intelligence in real-world workflows.

Zicheng Lin, Weichao Pan

Road damage detection is a critical task for ensuring traffic safety and maintaining infrastructure integrity. While deep learning-based detection methods are now widely adopted, they still face two core challenges: first, the inadequate multi-scale feature extraction capabilities of existing networks for diverse targets like cracks and potholes, leading to high miss rates for small-scale damage; and second, the substantial parameter counts and computational demands of mainstream models, which hinder their deployment for efficient, real-time detection in practical applications. To address these issues, this paper proposes a high-precision and lightweight model, YOLO - Road Orthogonal Compact (YOLO-ROC). We designed a Bidirectional Multi-scale Spatial Pyramid Pooling Fast (BMS-SPPF) module to enhance multi-scale feature extraction and implemented a hierarchical channel compression strategy to reduce computational complexity. The BMS-SPPF module leverages a bidirectional spatial-channel attention mechanism to improve the detection of small targets. Concurrently, the channel compression strategy reduces the parameter count from 3.01M to 0.89M and GFLOPs from 8.1 to 2.6. Experiments on the RDD2022_China_Drone dataset demonstrate that YOLO-ROC achieves a mAP50 of 67.6%, surpassing the baseline YOLOv8n by 2.11%. Notably, the mAP50 for the small-target D40 category improved by 16.8%, and the final model size is only 2.0 MB. Furthermore, the model exhibits excellent generalization performance on the RDD2022_China_Motorbike dataset.

Zicheng Lin, Xiaoqiang Li, Yichao Wang et al.

Particle Image Velocimetry (PIV) is fundamental to fluid dynamics, yet deep learning applications face significant hurdles. A critical gap exists: the lack of comprehensive evaluation of how diverse optical flow models perform specifically on PIV data, largely due to limitations in available datasets and the absence of a standardized benchmark. This prevents fair comparison and hinders progress. To address this, our primary contribution is a novel, large-scale synthetic PIV benchmark dataset generated from diverse CFD simulations (JHTDB and Blasius). It features unprecedented variety in particle densities, flow velocities, and continuous motion, enabling, for the first time, a standardized and rigorous evaluation of various optical flow and PIV algorithms. Complementing this, we propose Multi Cost Volume PIV (MCFormer), a new deep network architecture leveraging multi-frame temporal information and multiple cost volumes, specifically designed for PIV's sparse nature. Our comprehensive benchmark evaluation, the first of its kind, reveals significant performance variations among adapted optical flow models and demonstrates that MCFormer significantly outperforms existing methods, achieving the lowest overall normalized endpoint error (NEPE). This work provides both a foundational benchmark resource essential for future PIV research and a state-of-the-art method tailored for PIV challenges. We make our benchmark dataset and code publicly available to foster future research in this area.