Searcharxiv⌕ Search

arXiv subjects

Sirui Han

Publications and source records attributed to Sirui Han.

At least 55 records · Page 3Linked to original sources
arXiv2026

Bit-by-Bit: Progressive QAT Strategy with Outlier Channel Splitting for Stable Low-Bit LLMs

Training LLMs at ultra-low precision remains a formidable challenge. Direct low-bit QAT often suffers from convergence instability and substantial training costs, exacerbated by quantization noise from heavy-tailed outlier channels and error accumulation across layers. To address these issues, we present Bit-by-Bit, a progressive QAT framework with outlier channel splitting. Our approach integrates three key components: (1) block-wise progressive training that reduces precision stage by stage, ensuring stable initialization for low-bit optimization; (2) nested structure of integer quantization grids to enable a "train once, deploy any precision" paradigm, allowing a single model to support multiple bit-widths without retraining; (3) rounding-aware outlier channel splitting, which mitigates quantization error while acting as an identity transform that preserves the quantized outputs. Furthermore, we follow microscaling groups with E4M3 scales, capturing dynamic activation ranges in alignment with OCP/NVIDIA standards. To address the lack of efficient 2-bit kernels, we developed custom operators for both W2A2 and W2A16 configurations, achieving up to 11$\times$ speedup over BF16. Under W2A2 settings, Bit-by-Bit significantly outperforms baselines like BitDistiller and EfficientQAT on both Llama2/3, achieving a loss of only 2.25 WikiText2 PPL compared to full-precision models.

cs.LG↗
arXiv2026

Follow-Your-Motion: Video Motion Transfer via Efficient Spatial-Temporal Decoupled Finetuning

Recently, breakthroughs in the video diffusion transformer have shown remarkable capabilities in diverse motion generations. As for the motion-transfer task, current methods mainly use two-stage Low-Rank Adaptations (LoRAs) finetuning to obtain better performance. However, existing adaptation-based motion transfer still suffers from motion inconsistency and tuning inefficiency when applied to large video diffusion transformers. Naive two-stage LoRA tuning struggles to maintain motion consistency between generated and input videos due to the inherent spatial-temporal coupling in the 3D attention operator. Additionally, they require time-consuming fine-tuning processes in both stages. To tackle these issues, we propose Follow-Your-Motion, an efficient two-stage video motion transfer framework that finetunes a powerful video diffusion transformer to synthesize complex motion. Specifically, we propose a spatial-temporal decoupled LoRA to decouple the attention architecture for spatial appearance and temporal motion processing. During the second training stage, we design the sparse motion sampling and adaptive RoPE to accelerate the tuning speed. To address the lack of a benchmark for this field, we introduce MotionBench, a comprehensive benchmark comprising diverse motion, including creative camera motion, single object motion, multiple object motion, and complex human motion. We show extensive evaluations on MotionBench to verify the superiority of Follow-Your-Motion.

cs.CV↗
arXiv2026

LABSHIELD: A Multimodal Benchmark for Safety-Critical Reasoning and Planning in Scientific Laboratories

Artificial intelligence is increasingly catalyzing scientific automation, with multimodal large language model (MLLM) agents evolving from lab assistants into self-driving lab operators. This transition imposes stringent safety requirements on laboratory environments, where fragile glassware, hazardous substances, and high-precision laboratory equipment render planning errors or misinterpreted risks potentially irreversible. However, the safety awareness and decision-making reliability of embodied agents in such high-stakes settings remain insufficiently defined and evaluated. To bridge this gap, we introduce LABSHIELD, a realistic multi-view benchmark designed to assess MLLMs in hazard identification and safety-critical reasoning. Grounded in U.S. Occupational Safety and Health Administration (OSHA) standards and the Globally Harmonized System (GHS), LABSHIELD establishes a rigorous safety taxonomy spanning 164 operational tasks with diverse manipulation complexities and risk profiles. We evaluate 20 proprietary models, 9 open-source models, and 3 embodied models under a dual-track evaluation framework. Our results reveal a systematic gap between general-domain MCQ accuracy and Semi-open QA safety performance, with models exhibiting an average drop of 32.0% in professional laboratory scenarios, particularly in hazard interpretation and safety-aware planning. These findings underscore the urgent necessity for safety-centric reasoning frameworks to ensure reliable autonomous scientific experimentation in embodied laboratory contexts. The full dataset will be released soon.

cs.AI↗
arXiv2026

Not Just the Destination, But the Journey: Reasoning Traces Causally Shape Generalization Behaviors

Chain-of-Thought (CoT) is often viewed as a window into LLM decision-making, yet recent work suggests it may function merely as post-hoc rationalization. This raises a critical alignment question: Does the reasoning trace causally shape model generalization independent of the final answer? To isolate reasoning's causal effect, we design a controlled experiment holding final harmful answers constant while varying reasoning paths. We construct datasets with \textit{Evil} reasoning embracing malice, \textit{Misleading} reasoning rationalizing harm, and \textit{Submissive} reasoning yielding to pressure. We train models (0.6B--14B parameters) under multiple paradigms, including question-thinking-answer (QTA), question-thinking (QT), and thinking-only (T-only), and evaluate them in both think and no-think modes. We find that: (1) CoT training could amplify harmful generalization more than standard fine-tuning; (2) distinct reasoning types induce distinct behavioral patterns aligned with their semantics, despite identical final answers; (3) training on reasoning without answer supervision (QT or T-only) is sufficient to alter behavior, proving reasoning carries an independent signal; and (4) these effects persist even when generating answers without reasoning, indicating deep internalization. Our findings demonstrate that reasoning content is causally potent, challenging alignment strategies that supervise only outputs.

cs.CL↗
arXiv2026

ThinkPatterns-21k: A Systematic Study on the Impact of Thinking Patterns in LLMs

Large language models (LLMs) have demonstrated enhanced performance through the \textit{Thinking then Responding} paradigm, where models generate internal thoughts before final responses (aka, System 2 thinking). However, existing research lacks a systematic understanding of the mechanisms underlying how thinking patterns affect performance across model sizes. In this work, we conduct a comprehensive analysis of the impact of various thinking types on model performance and introduce ThinkPatterns-21k, a curated dataset comprising 21k instruction-response pairs (QA) collected from existing instruction-following datasets with five thinking types. For each pair, we augment it with five distinct internal thinking patterns: one unstructured thinking (monologue) and four structured variants (decomposition, self-ask, self-debate and self-critic), while maintaining the same instruction and response. Through extensive evaluation across different model sizes (3B-32B parameters), we have two key findings: (1) smaller models (<30B parameters) can benefit from most of structured thinking patterns, while larger models (32B) with structured thinking like decomposition would degrade performance and (2) unstructured monologue demonstrates broad effectiveness across different model sizes. Finally, we released all of our datasets, checkpoints, training logs of diverse thinking patterns to reproducibility, aiming to facilitate further research in this direction.

cs.CL↗
arXiv2026

DC-W2S: Dual-Consensus Weak-to-Strong Training for Reliable Process Reward Modeling in Biological Reasoning

In scientific reasoning tasks, the veracity of the reasoning process is as critical as the final outcome. While Process Reward Models (PRMs) offer a solution to the coarse-grained supervision problems inherent in Outcome Reward Models (ORMs), their deployment is hindered by the prohibitive cost of obtaining expert-verified step-wise labels. This paper addresses the challenge of training reliable PRMs using abundant but noisy "weak" supervision. We argue that existing Weak-to-Strong Generalization (W2SG) theories lack prescriptive guidelines for selecting high-quality training signals from noisy data. To bridge this gap, we introduce the Dual-Consensus Weak-to-Strong (DC-W2S) framework. By intersecting Self-Consensus (SC) metrics among weak supervisors with Neighborhood-Consensus (NC) metrics in the embedding space, we stratify supervision signals into distinct reliability regimes. We then employ a curriculum of instance-level balanced sampling and label-level reliability-aware masking to guide the training process. We demonstrate that DC-W2S enables the training of robust PRMs for complex reasoning without exhaustive expert annotation, proving that strategic data curation is more effective than indiscriminate training on large-scale noisy datasets.

cs.CL↗
arXiv2026

Unlocking Data Value in Finance: A Study on Distillation and Difficulty-Aware Training

Large Language Models (LLMs) have demonstrated strong general capabilities, yet their deployment in finance remains challenging due to dense domain-specific terminology, stringent numerical reasoning requirements, and low tolerance for factual errors. We conduct a controlled empirical study showing that in specialized vertical domains, performance is largely determined by the quality and difficulty/verifiability profile of post-training data. We introduce \textbf{ODA-Fin-SFT-318k}, constructed via multi-stage distillation and verification to produce high-quality Chain-of-Thought supervision, and \textbf{ODA-Fin-RL-12k}, curated for hard-but-verifiable tasks that balance reward precision and task diversity. Using standard SFT and RL pipelines, we show that high-quality CoT distillation establishes a robust foundation during SFT, while difficulty- and verifiability-aware sampling improves RL generalization. Evaluated on nine benchmarks spanning general financial tasks, sentiment analysis, and numerical reasoning, our ODA-Fin-RL-8B consistently surpasses open-source state-of-the-art (SOTA) financial LLMs of comparable size. We release our ODA-Fin-SFT-318k and ODA-Fin-RL-12k datasets, along with trained models to advance data-centric financial AI research.

cs.LG↗
arXiv2026

What, Whether and How? Unveiling Process Reward Models for Thinking with Images Reasoning

The rapid advancement of Large Vision Language Models (LVLMs) has demonstrated excellent abilities in various visual tasks. Building upon these developments, the thinking with images paradigm has emerged, enabling models to dynamically edit and re-encode visual information at each reasoning step, mirroring human visual processing. However, this paradigm introduces significant challenges as diverse errors may occur during reasoning processes. This necessitates Process Reward Models (PRMs) for distinguishing positive and negative reasoning steps, yet existing benchmarks for PRMs are predominantly text-centric and lack comprehensive assessment under this paradigm. To address these gaps, this work introduces the first comprehensive benchmark specifically designed for evaluating PRMs under the thinking with images paradigm. Our main contributions are: (1) Through extensive analysis of reasoning trajectories and guided search experiments with PRMs, we define 7 fine-grained error types and demonstrate both the necessity for specialized PRMs and the potential for improvement. (2) We construct a comprehensive benchmark comprising 1,206 manually annotated thinking with images reasoning trajectories spanning 4 categories and 16 subcategories for fine-grained evaluation of PRMs. (3) Our experimental analysis reveals that current LVLMs fall short as effective PRMs, exhibiting limited capabilities in visual reasoning process evaluation with significant performance disparities across error types, positive evaluation bias, and sensitivity to reasoning step positions. These findings demonstrate the effectiveness of our benchmark and establish crucial foundations for advancing PRMs in LVLMs.

cs.CV↗
arXiv2026

Reimagining Legal Fact Verification with GenAI: Toward Effective Human-AI Collaboration

Fact verification is a critical yet underexplored component of non-litigation legal practice. While existing research has examined automation in legal workflow and human-AI collaboration in high-stakes domains, little is known about how GenAI can support fact verification, a task that demands prudent judgment and strict accountability. To address this, we conducted semi-structured interviews with 18 lawyers to understand their current verification practices, attitudes toward GenAI adoption, and expectations for future systems. We found that while lawyers use GenAI for low-risk tasks like drafting and language optimization, concerns over accuracy, confidentiality, and liability are currently limiting its adoption for fact verification. These concerns translate into core design requirements for AI systems that are trustworthy and accountable. Based on these, we contribute design insights for human-AI collaboration in legal fact verification, emphasizing the development of auditable systems that balance efficiency with professional judgment and uphold ethical and legal accountability in high-stakes practice.

cs.HC↗
arXiv2026

Controllable Video Generation: A Survey

With the rapid development of AI-generated content (AIGC), video generation has emerged as one of its most dynamic and impactful subfields. In particular, the advancement of video generation foundation models has led to growing demand for controllable video generation methods that can more accurately reflect user intent. Most existing foundation models are designed for text-to-video generation, where text prompts alone are often insufficient to express complex, multi-modal, and fine-grained user requirements. This limitation makes it challenging for users to generate videos with precise control using current models. To address this issue, recent research has explored the integration of additional non-textual conditions, such as camera motion, depth maps, and human pose, to extend pretrained video generation models and enable more controllable video synthesis. These approaches aim to enhance the flexibility and practical applicability of AIGC-driven video generation systems. In this survey, we provide a systematic review of controllable video generation, covering both theoretical foundations and recent advances in the field. We begin by introducing the key concepts and commonly used open-source video generation models. We then focus on control mechanisms in video diffusion models, analyzing how different types of conditions can be incorporated into the denoising process to guide generation. Finally, we categorize existing methods based on the types of control signals they leverage, including single-condition generation, multi-condition generation, and universal controllable generation. For a complete list of the literature on controllable video generation reviewed, please visit our curated repository at https://github.com/mayuelala/Awesome-Controllable-Video-Generation.

cs.GR↗
arXiv2026

Perception, Understanding and Reasoning, A Multimodal Benchmark for Video Fake News Detection

The advent of multi-modal large language models (MLLMs) has greatly advanced research on video fake news detection (VFND) tasks. Existing benchmarks typically focus on the detection accuracy, while failing to provide fine-grained assessments for the entire detection process. To address these limitations, we introduce {POVFNDB (Process-oriented Video Fake News Detection Benchmark)}, a process-oriented benchmark comprising 10 tasks designed to systematically evaluate MLLMs' perception, understanding, and reasoning capabilities in VFND. This benchmark contains \textit{36,240} human-annotated question-answer (QA) in structured or open-ended formats, spanning 15 distinct evaluation dimensions that characterize different aspects of the video fake news detection process. Using POVFNDB, we conduct comprehensive evaluations on both proprietary and open-source MLLMs. Moreover, we establish a strong benchmark baseline by fine-tuning Qwen2.5VL-7B-Instruct on process-oriented chain-of-thought data constructed with our proposed POVFND-CoT framework, achieving state-of-the-art performance on VFND.

cs.CV↗
arXiv2026

LRAS: Advanced Legal Reasoning with Agentic Search

While Large Reasoning Models (LRMs) have demonstrated exceptional logical capabilities in mathematical domains, their application to the legal field remains hindered by the strict requirements for procedural rigor and adherence to legal logic. Existing legal LLMs, which rely on "closed-loop reasoning" derived solely from internal parametric knowledge, frequently suffer from lack of self-awareness regarding their knowledge boundaries, leading to confident yet incorrect conclusions. To address this challenge, we present Legal Reasoning with Agentic Search (LRAS), the first framework designed to transition legal LLMs from static and parametric "closed-loop thinking" to dynamic and interactive "Active Inquiry". By integrating Introspective Imitation Learning and Difficulty-aware Reinforcement Learning, LRAS enables LRMs to identify knowledge boundaries and handle legal reasoning complexity. Empirical results demonstrate that LRAS outperforms state-of-the-art baselines by 8.2-32\%, with the most substantial gains observed in tasks requiring deep reasoning with reliable knowledge. We will release our data and models for further exploration soon.

cs.AI↗
arXiv2026

MMFCTUB: Multi-Modal Financial Credit Table Understanding Benchmark

The advent of multi-modal language models (MLLMs) has spurred research into their application across various table understanding tasks. However, their performance in credit table understanding (CTU) for financial credit review remains largely unexplored due to the following barriers: low data consistency, high annotation costs stemming from domain-specific knowledge and complex calculations, and evaluation paradigm gaps between benchmark and real-world scenarios. To address these challenges, we introduce MMFCTUB (Multi-Modal Financial Credit Table Understanding Benchmark), a practical benchmark, encompassing more than 7,600 high quality CTU samples across 5 table types. MMFCTUB employ a minimally supervised pipeline that adheres to inter-table constraints and maintains data distributions consistency. The benchmark leverages capacity-driven questions and mask-and-recovery strategy to evaluate models' cross-table structure perception, domain knowledge utilization, and numerical calculation capabilities. Utilizing MMFCTUB, we conduct comprehensive evaluations of both proprietary and open-source MLLMs, revealing their strengths and limitations in CTU tasks. MMFCTUB serves as a valuable resource for the research community, facilitating rigorous evaluation of MLLMs in the domain of CTU.

cs.CE↗
arXiv2026

Reinforcement Learning of Large Language Models for Interpretable Credit Card Fraud Detection

E-commerce platforms and payment solution providers face increasingly sophisticated fraud schemes, ranging from identity theft and account takeovers to complex money laundering operations that exploit the speed and anonymity of digital transactions. However, despite their theoretical promise, the application of Large Language Models (LLMs) to fraud detection in real-world financial contexts remains largely unexploited, and their practical effectiveness in handling domain-specific e-commerce transaction data has yet to be empirically validated. To bridge this gap between conventional machine learning limitations and the untapped potential of LLMs in fraud detection, this paper proposes a novel approach that employs Reinforcement Learning (RL) to post-train lightweight language models specifically for fraud detection tasks using only raw transaction data. We utilize the Group Sequence Policy Optimization (GSPO) algorithm combined with a rule-based reward system to fine-tune language models of various sizes on a real-life transaction dataset provided by a Chinese global payment solution company. Through this reinforcement learning framework, the language models are encouraged to explore diverse trust and risk signals embedded within the textual transaction data, including patterns in customer information, shipping details, product descriptions, and order history. Our experimental results demonstrate the effectiveness of this approach, with post-trained language models achieving substantial F1-score improvements on held-out test data. Our findings demonstrate that the observed performance improvements are primarily attributable to the exploration mechanism inherent in reinforcement learning, which allows models to discover novel fraud indicators beyond those captured by traditional engineered features.

cs.AI↗
arXiv2026

Wow, wo, val! A Comprehensive Embodied World Model Evaluation Turing Test

As world models gain momentum in Embodied AI, an increasing number of works explore using video foundation models as predictive world models for downstream embodied tasks like 3D prediction or interactive generation. However, before exploring these downstream tasks, video foundation models still have two critical questions unanswered: (1) whether their generative generalization is sufficient to maintain perceptual fidelity in the eyes of human observers, and (2) whether they are robust enough to serve as a universal prior for real-world embodied agents. To provide a standardized framework for answering these questions, we introduce the Embodied Turing Test benchmark: WoW-World-Eval (Wow,wo,val). Building upon 609 robot manipulation data, Wow-wo-val examines five core abilities, including perception, planning, prediction, generalization, and execution. We propose a comprehensive evaluation protocol with 22 metrics to assess the models' generation ability, which achieves a high Pearson Correlation between the overall score and human preference (>0.93) and establishes a reliable foundation for the Human Turing Test. On Wow-wo-val, models achieve only 17.27 on long-horizon planning and at best 68.02 on physical consistency, indicating limited spatiotemporal consistency and physical reasoning. For the Inverse Dynamic Model Turing Test, we first use an IDM to evaluate the video foundation models' execution accuracy in the real world. However, most models collapse to $\approx$ 0% success, while WoW maintains a 40.74% success rate. These findings point to a noticeable gap between the generated videos and the real world, highlighting the urgency and necessity of benchmarking World Model in Embodied AI.

cs.RO↗
arXiv2025

GSRender: Deduplicated Occupancy Prediction via Weakly Supervised 3D Gaussian Splatting

Weakly-supervised 3D occupancy perception is crucial for vision-based autonomous driving in outdoor environments. Previous methods based on NeRF often face a challenge in balancing the number of samples used. Too many samples can decrease efficiency, while too few can compromise accuracy, leading to variations in the mean Intersection over Union (mIoU) by 5-10 points. Furthermore, even with surrounding-view image inputs, only a single image is rendered from each viewpoint at any given moment. This limitation leads to duplicated predictions, which significantly impacts the practicality of the approach. However, this issue has largely been overlooked in existing research. To address this, we propose GSRender, which uses 3D Gaussian Splatting for weakly-supervised occupancy estimation, simplifying the sampling process. Additionally, we introduce the Ray Compensation module, which reduces duplicated predictions by compensating for features from adjacent frames. Finally, we redesign the dynamic loss to remove the influence of dynamic objects from adjacent frames. Extensive experiments show that our approach achieves SOTA results in RayIoU (+6.0), while also narrowing the gap with 3D- supervised methods. This work lays a solid foundation for weakly-supervised occupancy perception. The code is available at https://github.com/Jasper-sudo-Sun/GSRender.

cs.CV↗
arXiv2025

InterMT: Multi-Turn Interleaved Preference Alignment with Human Feedback

As multimodal large models (MLLMs) continue to advance across challenging tasks, a key question emerges: What essential capabilities are still missing? A critical aspect of human learning is continuous interaction with the environment -- not limited to language, but also involving multimodal understanding and generation. To move closer to human-level intelligence, models must similarly support multi-turn, multimodal interaction. In particular, they should comprehend interleaved multimodal contexts and respond coherently in ongoing exchanges. In this work, we present an initial exploration through the InterMT -- the first preference dataset for multi-turn multimodal interaction, grounded in real human feedback. In this exploration, we particularly emphasize the importance of human oversight, introducing expert annotations to guide the process, motivated by the fact that current MLLMs lack such complex interactive capabilities. InterMT captures human preferences at both global and local levels into nine sub-dimensions, consists of 15.6k prompts, 52.6k multi-turn dialogue instances, and 32.4k human-labeled preference pairs. To compensate for the lack of capability for multi-modal understanding and generation, we introduce an agentic workflow that leverages tool-augmented MLLMs to construct multi-turn QA instances. To further this goal, we introduce InterMT-Bench to assess the ability of MLLMs in assisting judges with multi-turn, multimodal tasks. We demonstrate the utility of \InterMT through applications such as judge moderation and further reveal the multi-turn scaling law of judge model. We hope the open-source of our data can help facilitate further research on aligning current MLLMs to the next step. Our project website can be found at https://pku-intermt.github.io .

cs.AI↗
arXiv2025

MedInsightBench: Evaluating Medical Analytics Agents Through Multi-Step Insight Discovery in Multimodal Medical Data

In medical data analysis, extracting deep insights from complex, multi-modal datasets is essential for improving patient care, increasing diagnostic accuracy, and optimizing healthcare operations. However, there is currently a lack of high-quality datasets specifically designed to evaluate the ability of large multi-modal models (LMMs) to discover medical insights. In this paper, we introduce MedInsightBench, the first benchmark that comprises 332 carefully curated medical cases, each annotated with thoughtfully designed insights. This benchmark is intended to evaluate the ability of LMMs and agent frameworks to analyze multi-modal medical image data, including posing relevant questions, interpreting complex findings, and synthesizing actionable insights and recommendations. Our analysis indicates that existing LMMs exhibit limited performance on MedInsightBench, which is primarily attributed to their challenges in extracting multi-step, deep insights and the absence of medical expertise. Therefore, we propose MedInsightAgent, an automated agent framework for medical data analysis, composed of three modules: Visual Root Finder, Analytical Insight Agent, and Follow-up Question Composer. Experiments on MedInsightBench highlight pervasive challenges and demonstrate that MedInsightAgent can improve the performance of general LMMs in medical data insight discovery.

cs.AI↗