Chiyu Ma

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.

Haoming Meng, Kexin Huang, Shaohang Wei et al. · pku

Reinforcement learning with verifiable rewards (RLVR) has significantly improved reasoning in large language models (LLMs), yet the token-level mechanisms underlying these improvements remain unclear. We present a systematic empirical study of RLVR's distributional effects organized around three main analyses: (1) token-level characterization of distributional shifts between base and RL models, (2) the impact of token-level distributional shifts on sequence-level reasoning performance through cross-sampling interventions, and (3) fine-grained mechanics of these shifts at the token level. We find that RL fine-tuning induces highly sparse and targeted changes, with only a small fraction of token distributions exhibiting meaningful divergence between the base and RL policies. We further characterize the structure and evolution of these shifts through analyses of token entropy, positional concentration, and reallocation of probability mass. To assess the functional importance of these sparse changes, we conduct cross-sampling experiments that selectively swap token choices between the base and RL models with varying intervention budgets. We show that inserting only a small fraction of RL-sampled tokens into base generations progressively recovers RL performance gains, while injecting a similarly small number of base token choices into otherwise RL-generated sequences collapses performance to base levels, isolating a small set of token-level decisions directly responsible for RLVR's performance gains. Finally, we explore divergence-weighted variants of the advantage signal as a diagnostic intervention, finding that they can yield improvements over baselines. Together, our results shed light on the distributional changes induced by RLVR and provide a fine-grained, token-level lens for understanding RLVR fine-tuning as a targeted refinement process.

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.

Chiyu Ma, Brandon Zhao, Chaofan Chen et al.

We present ProtoConcepts, a method for interpretable image classification combining deep learning and case-based reasoning using prototypical parts. Existing work in prototype-based image classification uses a ``this looks like that'' reasoning process, which dissects a test image by finding prototypical parts and combining evidence from these prototypes to make a final classification. However, all of the existing prototypical part-based image classifiers provide only one-to-one comparisons, where a single training image patch serves as a prototype to compare with a part of our test image. With these single-image comparisons, it can often be difficult to identify the underlying concept being compared (e.g., ``is it comparing the color or the shape?''). Our proposed method modifies the architecture of prototype-based networks to instead learn prototypical concepts which are visualized using multiple image patches. Having multiple visualizations of the same prototype allows us to more easily identify the concept captured by that prototype (e.g., ``the test image and the related training patches are all the same shade of blue''), and allows our model to create richer, more interpretable visual explanations. Our experiments show that our ``this looks like those'' reasoning process can be applied as a modification to a wide range of existing prototypical image classification networks while achieving comparable accuracy on benchmark datasets.

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.

Xingjian Diao, Chunhui Zhang, Keyi Kong et al.

While large language models have demonstrated impressive reasoning abilities, their extension to the audio modality, particularly within large audio-language models (LALMs), remains underexplored. Addressing this gap requires a systematic approach that involves a capable base model, high-quality reasoning-oriented audio data, and effective training algorithms. In this work, we present a comprehensive solution for audio logical reasoning (ALR) tasks: we introduce SoundMind, a dataset of 6,446 audio-text annotated samples specifically curated to support complex reasoning. Building on this resource, we propose SoundMind-RL, a rule-based reinforcement learning (RL) algorithm designed to equip audio-language models with robust audio-text reasoning capabilities. By fine-tuning Qwen2.5-Omni-7B on the proposed SoundMind dataset using SoundMind-RL, we achieve strong and consistent improvements over state-of-the-art baselines on the SoundMind benchmark. This work highlights the benefit of combining high-quality, reasoning-focused datasets with specialized RL techniques, and contributes to advancing auditory intelligence in language models. The code and dataset introduced in this work are publicly available at https://github.com/xid32/SoundMind.

Wenhao You, Xingjian Diao, Chunhui Zhang et al.

While recent Multimodal Large Language Models exhibit impressive capabilities for general multimodal tasks, specialized domains like music necessitate tailored approaches. Music Audio-Visual Question Answering (Music AVQA) particularly underscores this, presenting unique challenges with its continuous, densely layered audio-visual content, intricate temporal dynamics, and the critical need for domain-specific knowledge. Through a systematic analysis of Music AVQA datasets and methods, this position paper identifies that specialized input processing, architectures incorporating dedicated spatial-temporal designs, and music-specific modeling strategies are critical for success in this domain. Our study provides valuable insights for researchers by highlighting effective design patterns empirically linked to strong performance, proposing concrete future directions for incorporating musical priors, and aiming to establish a robust foundation for advancing multimodal musical understanding. This work is intended to inspire broader attention and further research, supported by a continuously updated anonymous GitHub repository of relevant papers: https://github.com/xid32/Survey4MusicAVQA.

Chiyu Ma, Jon Donnelly, Wenjun Liu et al.

We present ProtoViT, a method for interpretable image classification combining deep learning and case-based reasoning. This method classifies an image by comparing it to a set of learned prototypes, providing explanations of the form ``this looks like that.'' In our model, a prototype consists of \textit{parts}, which can deform over irregular geometries to create a better comparison between images. Unlike existing models that rely on Convolutional Neural Network (CNN) backbones and spatially rigid prototypes, our model integrates Vision Transformer (ViT) backbones into prototype based models, while offering spatially deformed prototypes that not only accommodate geometric variations of objects but also provide coherent and clear prototypical feature representations with an adaptive number of prototypical parts. Our experiments show that our model can generally achieve higher performance than the existing prototype based models. Our comprehensive analyses ensure that the prototypes are consistent and the interpretations are faithful.

Alan Sun, Chiyu Ma, Kenneth Ge et al.

We present knowledge continuity, a novel definition inspired by Lipschitz continuity which aims to certify the robustness of neural networks across input domains (such as continuous and discrete domains in vision and language, respectively). Most existing approaches that seek to certify robustness, especially Lipschitz continuity, lie within the continuous domain with norm and distribution-dependent guarantees. In contrast, our proposed definition yields certification guarantees that depend only on the loss function and the intermediate learned metric spaces of the neural network. These bounds are independent of domain modality, norms, and distribution. We further demonstrate that the expressiveness of a model class is not at odds with its knowledge continuity. This implies that achieving robustness by maximizing knowledge continuity should not theoretically hinder inferential performance. Finally, to complement our theoretical results, we present several applications of knowledge continuity such as regularization, a certification algorithm, and show that knowledge continuity can be used to localize vulnerable components of a neural network.

Yaning Jia, Chunhui Zhang, Xingjian Diao et al.

Curriculum learning (CL) - ordering training data from easy to hard - has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction - forward or reverse - is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that (i) no curriculum strategy dominates universally - the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) task-aligned curricula focus on shaping the model's final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.

Yaning Jia, Shenyang Deng, Chiyu Ma et al.

Graph Neural Networks (GNNs) have achieved impressive success across a wide range of graph-based tasks, yet they remain highly vulnerable to small, imperceptible perturbations and adversarial attacks. Although numerous defense methods have been proposed to address these vulnerabilities, many rely on heuristic metrics, overfit to specific attack patterns, and suffer from high computational complexity. In this paper, we propose Kernel Complexity-Based Edge Sanitization (KCES), a training-free, model-agnostic defense framework. KCES leverages Graph Kernel Complexity (GKC), a novel metric derived from the graph's Gram matrix that characterizes GNN generalization via its test error bound. Building on GKC, we define a KC score for each edge, measuring the change in GKC when the edge is removed. Edges with high KC scores, typically introduced by adversarial perturbations, are pruned to mitigate their harmful effects, thereby enhancing GNNs' robustness. KCES can also be seamlessly integrated with existing defense strategies as a plug-and-play module without requiring training. Theoretical analysis and extensive experiments demonstrate that KCES consistently enhances GNN robustness, outperforms state-of-the-art baselines, and amplifies the effectiveness of existing defenses, offering a principled and efficient solution for securing GNNs.

Chiyu Ma, Enpei Zhang, Yilun Zhao et al.

LLM-as-Judge has emerged as a scalable alternative to human evaluation, enabling large language models (LLMs) to provide reward signals in trainings. While recent work has explored multi-agent extensions such as multi-agent debate and meta-judging to enhance evaluation quality, the question of how intrinsic biases manifest in these settings remains underexplored. In this study, we conduct a systematic analysis of four diverse bias types: position bias, verbosity bias, chain-of-thought bias, and bandwagon bias. We evaluate these biases across two widely adopted multi-agent LLM-as-Judge frameworks: Multi-Agent-Debate and LLM-as-Meta-Judge. Our results show that debate framework amplifies biases sharply after the initial debate, and this increased bias is sustained in subsequent rounds, while meta-judge approaches exhibit greater resistance. We further investigate the incorporation of PINE, a leading single-agent debiasing method, as a bias-free agent within these systems. The results reveal that this bias-free agent effectively reduces biases in debate settings but provides less benefit in meta-judge scenarios. Our work provides a comprehensive study of bias behavior in multi-agent LLM-as-Judge systems and highlights the need for targeted bias mitigation strategies in collaborative evaluation settings.

Lin Shi, Chiyu Ma, Wenhua Liang et al.

LLM-as-a-Judge has emerged as a promising alternative to human evaluators across various tasks, yet inherent biases - particularly position bias, the tendency to favor solutions based on their position within the prompt - compromise its reliability. This exploratory study evaluates position bias in LLM judges across pairwise and list-wise comparison settings, introducing three metrics: repetition stability, position consistency, and preference fairness. Our experiments, involving 15 LLM judges across MTBench and DevBench with 22 tasks and approximately 40 solution-generating models, result in over 150,000 evaluation instances. We identify Judge-Level, Candidate-Level, and Task-Level factors contributing to bias. The findings confirm that position bias is not due to random chance and varies significantly across judges and tasks. While position bias is weakly influenced by the length of prompt components, it is strongly affected by the quality gap between solutions. Our agreement and disagreement analysis among judges further provides insights into the distribution of judging difficulty across the dataset, and highlights the potential for dataset modifications.