Jinda Lu

Chiyu Ma, Shuo Yang, Kexin Huang et al.

We present Future-KL Influenced Policy Optimization (FIPO), a reinforcement learning algorithm designed to overcome reasoning bottlenecks in large language models. While GRPO style training scales effectively, it typically relies on outcome-based rewards (ORM) that distribute a global advantage uniformly across every token in a trajectory. We argue that this coarse-grained credit assignment imposes a performance ceiling by failing to distinguish critical logical pivots from trivial tokens. FIPO addresses this by incorporating discounted future-KL divergence into the policy update, creating a dense advantage formulation that re-weights tokens based on their influence on subsequent trajectory behavior. Empirically, FIPO enables models to break through the length stagnation seen in standard baselines. Evaluated on Qwen2.5-32B, FIPO extends the average chain-of-thought length from roughly 4,000 to over 10,000 tokens and increases AIME 2024 Pass@1 accuracy from 50.0% to a peak of 58.0% (converging at approximately 56.0\%). This outperforms both DeepSeek-R1-Zero-Math-32B (around 47.0%) and o1-mini (approximately 56.0%). Our results suggest that establishing dense advantage formulations is a vital path for evolving ORM-based algorithms to unlock the full reasoning potential of base models. We open-source our training system, built on the verl framework.

Jinda Lu, Junkang Wu, Jinghan Li et al.

Extending Reinforcement Learning with Verifiable Rewards (RLVR) to multimodal large language models (MLLMs) faces a fundamental challenge: their responses inherently interleave perception-related tokens, which ground visual content, with reasoning-related tokens, which construct reasoning chains. These token types instantiate distinct yet interdependent capacities -- visual grounding and symbolic reasoning -- making isolated optimization insufficient. Through token-level empirical analysis, we demonstrate that optimizing either perception- or reasoning-only tokens consistently underperforms full optimization, underscoring their inherent coupling. To address this, we propose a plug-and-play Token-Reweighting (ToR) strategy that explicitly models this interdependence by identifying critical tokens of both types and dynamically reweighting them during RLVR training. Applied on top of existing methods (e.g., GRPO and DAPO), ToR delivers consistent performance gains across multiple multi-modal reasoning benchmarks, achieving state-of-the-art performance with both accurate visual grounding and coherent reasoning.

Kexin Huang, Haoming Meng, Junkang Wu et al.

Reinforcement learning with verifiable rewards (RLVR) has substantially improved the reasoning capabilities of large language models. While existing analyses identify that RLVR-induced changes are sparse, they primarily focus on the \textbf{magnitude} of these updates, largely overlooking their \textbf{direction}. In this work, we argue that the direction of updates is a more critical lens for understanding RLVR's effects, which can be captured by the signed, token-level log probability difference $Δ\log p$ between the base and final RLVR models. Through statistical analysis and token-replacement interventions, we demonstrate that $Δ\log p$ more effectively identifies sparse, yet reasoning-critical updates than magnitude-based metrics (\eg divergence or entropy). Building on this insight, we propose two practical applications: (1) a \textit{test-time extrapolation} method that amplifies the policy along the learned $Δ\log p$ direction to improve reasoning accuracy without further training; (2) a \textit{training-time reweighting} method that focuses learning on low-probability (corresponding to higher $Δ\log p$) tokens, which improves reasoning performance across models and benchmarks. Our work establishes the direction of change as a key principle for analyzing and improving RLVR.

Jinghan Li, Junfeng Fang, Jinda Lu et al.

Reinforcement Learning with Verifiable Rewards (RLVR) and Group Relative Policy Optimization (GRPO) have significantly advanced the reasoning capabilities of large language models. Extending these methods to multimodal settings, however, faces a critical challenge: the instability of std-based normalization, which is easily distorted by extreme samples with nearly positive or negative rewards. Unlike pure-text LLMs, multimodal models are particularly sensitive to such distortions, as both perceptual and reasoning errors influence their responses. To address this, we characterize each sample by its difficulty, defined through perceptual complexity (measured via visual entropy) and reasoning uncertainty (captured by model confidence). Building on this characterization, we propose difficulty-aware group normalization (Durian), which re-groups samples by difficulty levels and shares the std within each group. Our approach preserves GRPO's intra-group distinctions while eliminating sensitivity to extreme cases, yielding significant performance gains across multiple multimodal reasoning benchmarks.

Jinda Lu, Junkang Wu, Jinghan Li et al.

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) for multimodal large language models (MLLMs) have mainly focused on improving final answer correctness and strengthening visual grounding. However, a critical bottleneck remains: although models can attend to relevant visual regions, they often fail to effectively incorporate visual evidence into subsequent reasoning, leading to reasoning chains that are weakly grounded in visual facts. To address this issue, we propose Trajectory-Guided Reinforcement Learning (TGRL), which guides the policy model to integrate visual evidence into fine-grained reasoning processes using expert reasoning trajectories from stronger models. We further introduce token-level reweighting and trajectory filtering to ensure stable and effective policy optimization. Extensive experiments on multiple multimodal reasoning benchmarks demonstrate that TGRL consistently improves reasoning performance and effectively bridges the gap between visual perception and logical reasoning.

Shuo Yang, Jinda Lu, Kexin Huang et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has become a promising paradigm for scaling reasoning capabilities of Large Language Models (LLMs). However, the sparsity of binary verifier rewards often leads to low efficiency and optimization instability. To stabilize training, existing methods typically impose token-level constraints relative to a reference policy. We identify that such constraints penalize deviations indiscriminately; this can flip verifier-determined direction when the policy attempts to outperform the reference, thereby suppressing gains. To resolve this, we propose One-Way Policy Optimization (OWPO), a method based on the principle of decoupling optimization direction from update magnitude. In OWPO, the verifier dictates the update direction, while the reference policy serves only to adjust the magnitude. Specifically, OWPO applies asymmetric reweighting: it performs Accelerated Alignment for inferior deviations (where the policy lags behind the reference) and Gain Locking for superior deviations (where the policy surpasses the reference). Furthermore, by incorporating iterative reference updates, OWPO creates a ``Ratchet Effect'' that continuously consolidates gains. Experimental results demonstrate that OWPO outperforms strong baselines, including DAPO, OPD, and MOPD, breaking the bottleneck of fixed priors to enable continuous self-evolution without reliance on external reference models.

Shuo Yang, Jinda Lu, Chiyu Ma et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has emerged as a central paradigm for scaling LLM reasoning, yet its optimization often suffers from training instability and suboptimal convergence. Through a systematic dissection of clipping-based GRPO-style objectives, we identify the rigid clipping decision induced by hard clipping as a key practical bottleneck in the studied RLVR setups. Specifically, our analysis suggests that informative signals can lie in the near-boundary region just beyond the clipping threshold, and are therefore discarded by the standard hard-clipping rule. Notably, once this bottleneck is precisely identified, even simple stochastic perturbations at the boundary can recover meaningful performance gains. Building on this finding, we propose Near-boundary Stochastic Rescue (NSR), a minimal, plug-and-play modification that stochastically retains these slightly out-of-bound tokens to recover lost signals. While NSR, via stochastic sampling, can be interpreted as inducing an implicit gradient decay in expectation, our ablations reveal that its stochastic, boundary-local rescue mechanism is consistently more effective than deterministic gradient decay. Validated by extensive experiments across model sizes from 7B to 30B and both dense and MoE architectures, as a plug-and-play solution, NSR substantially improves training stability and delivers consistent gains over strong baselines such as DAPO and GSPO.

Tao Yu, Yi-Fan Zhang, Chaoyou Fu et al.

Large language models (LLMs) can handle a wide variety of general tasks with simple prompts, without the need for task-specific training. Multimodal Large Language Models (MLLMs), built upon LLMs, have demonstrated impressive potential in tackling complex tasks involving visual, auditory, and textual data. However, critical issues related to truthfulness, safety, o1-like reasoning, and alignment with human preference remain insufficiently addressed. This gap has spurred the emergence of various alignment algorithms, each targeting different application scenarios and optimization goals. Recent studies have shown that alignment algorithms are a powerful approach to resolving the aforementioned challenges. In this paper, we aim to provide a comprehensive and systematic review of alignment algorithms for MLLMs. Specifically, we explore four key aspects: (1) the application scenarios covered by alignment algorithms, including general image understanding, multi-image, video, and audio, and extended multimodal applications; (2) the core factors in constructing alignment datasets, including data sources, model responses, and preference annotations; (3) the benchmarks used to evaluate alignment algorithms; and (4) a discussion of potential future directions for the development of alignment algorithms. This work seeks to help researchers organize current advancements in the field and inspire better alignment methods. The project page of this paper is available at https://github.com/BradyFU/Awesome-Multimodal-Large-Language-Models/tree/Alignment.

Jinda Lu, Shuo Wang, Yanbin Hao et al.

Recent adaptations can boost the low-shot capability of Contrastive Vision-Language Pre-training (CLIP) by effectively facilitating knowledge transfer. However, these adaptation methods are usually operated on the global view of an input image, and thus biased perception of partial local details of the image. To solve this problem, we propose a Visual Content Refinement (VCR) before the adaptation calculation during the test stage. Specifically, we first decompose the test image into different scales to shift the feature extractor's attention to the details of the image. Then, we select the image view with the max prediction margin in each scale to filter out the noisy image views, where the prediction margins are calculated from the pre-trained CLIP model. Finally, we merge the content of the aforementioned selected image views based on their scales to construct a new robust representation. Thus, the merged content can be directly used to help the adapter focus on both global and local parts without any extra training parameters. We apply our method to 3 popular low-shot benchmark tasks with 13 datasets and achieve a significant improvement over state-of-the-art methods. For example, compared to the baseline (Tip-Adapter) on the few-shot classification task, our method achieves about 2\% average improvement for both training-free and training-need settings.

Junxiang Qiu, Shuo Wang, Jinda Lu et al.

Diffusion Transformer (DiT) is a crucial method for content generation. However, it needs a lot of time to sample. Many studies have attempted to use caching to reduce the time consumption of sampling. Existing caching methods accelerate generation by reusing DiT features from the previous time step and skipping calculations in the next, but they tend to locate and cache low-error modules without focusing on reducing caching-induced errors, resulting in a sharp decline in generated content quality when increasing caching intensity. To solve this problem, we propose the \textbf{E}rror-\textbf{O}ptimized \textbf{C}ache (\textbf{EOC}). This method introduces three key improvements: \textbf{(1)} Prior knowledge extraction: Extract and process the caching differences; \textbf{(2)} A judgment method for cache optimization: Determine whether certain caching steps need to be optimized; \textbf{(3)} Cache optimization: reduce caching errors. Experiments show that this algorithm significantly reduces the error accumulation caused by caching, especially excessive caching. On the ImageNet dataset, without substantially increasing the computational load, this method improves the FID of the generated images when the rule-based model FORA has a caching level of \textbf{75}\%, \textbf{50}\%, and \textbf{25}\%, and the training-based model Learning-to-cache has a caching level of \textbf{22}\%. Specifically, the FID values change from 30.454 to 21.690 (\textbf{28.8}\%), from 6.857 to 5.821 (\textbf{15.1}\%), from 3.870 to 3.692 (\textbf{4.6}\%), and from 3.539 to 3.451 (\textbf{2.5}\%) respectively. Code is available at https://github.com/qiujx0520/EOC_MM2025.git.

Junxiang Qiu, Lin Liu, Shuo Wang et al.

Feature caching has emerged as an effective strategy to accelerate diffusion transformer (DiT) sampling through temporal feature reuse. It is a challenging problem since (1) Progressive error accumulation from cached blocks significantly degrades generation quality, particularly when over 50\% of blocks are cached; (2) Current error compensation approaches neglect dynamic perturbation patterns during the caching process, leading to suboptimal error correction. To solve these problems, we propose the Gradient-Optimized Cache (GOC) with two key innovations: (1) Cached Gradient Propagation: A gradient queue dynamically computes the gradient differences between cached and recomputed features. These gradients are weighted and propagated to subsequent steps, directly compensating for the approximation errors introduced by caching. (2) Inflection-Aware Optimization: Through statistical analysis of feature variation patterns, we identify critical inflection points where the denoising trajectory changes direction. By aligning gradient updates with these detected phases, we prevent conflicting gradient directions during error correction. Extensive evaluations on ImageNet demonstrate GOC's superior trade-off between efficiency and quality. With 50\% cached blocks, GOC achieves IS 216.28 (26.3\% higher) and FID 3.907 (43\% lower) compared to baseline DiT, while maintaining identical computational costs. These improvements persist across various cache ratios, demonstrating robust adaptability to different acceleration requirements. Code is available at https://github.com/qiujx0520/GOC_ICCV2025.git.

Zhe Xu, Zhicai Wang, Junkang Wu et al.

Large Vision-Language Models (LVLMs) often suffer from object hallucination, making erroneous judgments about the presence of objects in images. We propose this primar- ily stems from spurious correlations arising when models strongly associate highly co-occurring objects during train- ing, leading to hallucinated objects influenced by visual con- text. Current benchmarks mainly focus on hallucination de- tection but lack a formal characterization and quantitative evaluation of spurious correlations in LVLMs. To address this, we introduce causal analysis into the object recognition scenario of LVLMs, establishing a Structural Causal Model (SCM). Utilizing the language of causality, we formally de- fine spurious correlations arising from co-occurrence bias. To quantify the influence induced by these spurious correla- tions, we develop Causal-HalBench, a benchmark specifically constructed with counterfactual samples and integrated with comprehensive causal metrics designed to assess model ro- bustness against spurious correlations. Concurrently, we pro- pose an extensible pipeline for the construction of these coun- terfactual samples, leveraging the capabilities of proprietary LVLMs and Text-to-Image (T2I) models for their genera- tion. Our evaluations on mainstream LVLMs using Causal- HalBench demonstrate these models exhibit susceptibility to spurious correlations, albeit to varying extents.

Yuan Wang, Borui Liao, Huijuan Huang et al.

Recent advances in video reward models and post-training strategies have improved text-to-video (T2V) generation. While these models typically assess visual quality, motion quality, and text alignment, they often overlook key structural distortions, such as abnormal object appearances and interactions, which can degrade the overall quality of the generative video. To address this gap, we introduce REACT, a frame-level reward model designed specifically for structural distortions evaluation in generative videos. REACT assigns point-wise scores and attribution labels by reasoning over video frames, focusing on recognizing distortions. To support this, we construct a large-scale human preference dataset, annotated based on our proposed taxonomy of structural distortions, and generate additional data using a efficient Chain-of-Thought (CoT) synthesis pipeline. REACT is trained with a two-stage framework: ((1) supervised fine-tuning with masked loss for domain knowledge injection, followed by (2) reinforcement learning with Group Relative Policy Optimization (GRPO) and pairwise rewards to enhance reasoning capability and align output scores with human preferences. During inference, a dynamic sampling mechanism is introduced to focus on frames most likely to exhibit distortion. We also present REACT-Bench, a benchmark for generative video distortion evaluation. Experimental results demonstrate that REACT complements existing reward models in assessing structutal distortion, achieving both accurate quantitative evaluations and interpretable attribution analysis.

Lin Liu, Huixia Ben, Shuo Wang et al.

Controllable generative models have been widely used to improve the realism of synthetic visual content. However, such models must handle control conditions and content generation computational requirements, resulting in generally low generation efficiency. To address this issue, we propose a Hybrid-Grained Cache (HGC) approach that reduces computational overhead by adopting cache strategies with different granularities at different computational stages. Specifically, (1) we use a coarse-grained cache (block-level) based on feature reuse to dynamically bypass redundant computations in encoder-decoder blocks between each step of model reasoning. (2) We design a fine-grained cache (prompt-level) that acts within a module, where the fine-grained cache reuses cross-attention maps within consecutive reasoning steps and extends them to the corresponding module computations of adjacent steps. These caches of different granularities can be seamlessly integrated into each computational link of the controllable generation process. We verify the effectiveness of HGC on four benchmark datasets, especially its advantages in balancing generation efficiency and visual quality. For example, on the COCO-Stuff segmentation benchmark, our HGC significantly reduces the computational cost (MACs) by 63% (from 18.22T to 6.70T), while keeping the loss of semantic fidelity (quantized performance degradation) within 1.5%.

Xingyu Zhu, Shuo Wang, Jinda Lu et al.

Few-shot learning (FSL) based on manifold regularization aims to improve the recognition capacity of novel objects with limited training samples by mixing two samples from different categories with a blending factor. However, this mixing operation weakens the feature representation due to the linear interpolation and the overlooking of the importance of specific channels. To solve these issues, this paper proposes attentive feature regularization (AFR) which aims to improve the feature representativeness and discriminability. In our approach, we first calculate the relations between different categories of semantic labels to pick out the related features used for regularization. Then, we design two attention-based calculations at both the instance and channel levels. These calculations enable the regularization procedure to focus on two crucial aspects: the feature complementarity through adaptive interpolation in related categories and the emphasis on specific feature channels. Finally, we combine these regularization strategies to significantly improve the classifier performance. Empirical studies on several popular FSL benchmarks demonstrate the effectiveness of AFR, which improves the recognition accuracy of novel categories without the need to retrain any feature extractor, especially in the 1-shot setting. Furthermore, the proposed AFR can seamlessly integrate into other FSL methods to improve classification performance.

Chenlu Ding, Jiancan Wu, Yancheng Yuan et al.

The advent of Large Language Models (LLMs) has revolutionized natural language processing, enabling advanced understanding and reasoning capabilities across a variety of tasks. Fine-tuning these models for specific domains, particularly through Parameter-Efficient Fine-Tuning (PEFT) strategies like LoRA, has become a prevalent practice due to its efficiency. However, this raises significant privacy and security concerns, as models may inadvertently retain and disseminate sensitive or undesirable information. To address these issues, we introduce a novel instance-wise unlearning framework, LLMEraser, which systematically categorizes unlearning tasks and applies precise parameter adjustments using influence functions. Unlike traditional unlearning techniques that are often limited in scope and require extensive retraining, LLMEraser is designed to handle a broad spectrum of unlearning tasks without compromising model performance. Extensive experiments on benchmark datasets demonstrate that LLMEraser excels in efficiently managing various unlearning scenarios while maintaining the overall integrity and efficacy of the models.

Jinda Lu, Junkang Wu, Jinghan Li et al.

Direct Preference Optimization (DPO) has shown effectiveness in aligning multi-modal large language models (MLLM) with human preferences. However, existing methods exhibit an imbalanced responsiveness to the data of varying hardness, tending to overfit on the easy-to-distinguish data while underfitting on the hard-to-distinguish data. In this paper, we propose Data- and Model-aware DPO (DAMA) to dynamically adjust the optimization process from two key aspects: (1) a data-aware strategy that incorporates data hardness, and (2) a model-aware strategy that integrates real-time model responses. By combining the two strategies, DAMA enables the model to effectively adapt to data with varying levels of hardness. Extensive experiments on five benchmarks demonstrate that DAMA not only significantly enhances the trustworthiness, but also improves the effectiveness over general tasks. For instance, on the Object-HalBench, our DAMA-7B reduces response-level and mentioned-level hallucination by 90.0% and 95.3%, respectively, surpassing the performance of GPT-4V.

Jinda Lu, Jinghan Li, Yuan Gao et al.

Preference alignment through Direct Preference Optimization (DPO) has demonstrated significant effectiveness in aligning multimodal large language models (MLLMs) with human preferences. However, existing methods focus primarily on language preferences while neglecting the critical visual context. In this paper, we propose an Adaptive Vision-enhanced Preference optimization (AdaViP) that addresses these limitations through two key innovations: (1) vision-based preference pair construction, which integrates multiple visual foundation models to strategically remove key visual elements from the image, enhancing MLLMs' sensitivity to visual details; and (2) adaptive preference optimization that dynamically balances vision- and language-based preferences for more accurate alignment. Extensive evaluations across different benchmarks demonstrate our effectiveness. Notably, our AdaViP-7B achieves 93.7% and 96.4% reductions in response-level and mentioned-level hallucination respectively on the Object HalBench, significantly outperforming current state-of-the-art methods.