论文检索

Deep Learning · Large Language Models

Jiaming Li, Longze Chen, Ze Gong, Yukun Chen, Lu Wang, Wanwei He, Zhihao Yang, Minzheng Wang, Lei Zhang, Haoran Ye 等

Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have empowered large language models (LLMs) to tackle challenging reasoning tasks such as mathematics and programming. Despite its promise, the RLVR paradigm poses significant challenges, as existing methods often suffer from sparse reward signals and unstable policy gradient updates, inherent to RL-based approaches. To address the challenges, we propose **PACS**, a novel RLVR framework that achieves im**P**licit **A**ctor **C**ritic coupling via a **S**upervised learning framework. By treating the outcome reward as a predictable label, we reformulate the RLVR problem into a supervised learning task over a score function parameterized by the policy model and optimized using cross-entropy loss. A detailed gradient analysis shows that this supervised formulation inherently recovers the classical policy gradient update while providing more stable and efficient training. Extensive experiments demonstrate that PACS significantly outperforms strong open-source models and RLVR baselines, yielding substantial average gains of **+8.26%** (4B) and **+9.57%** (8B) over base models offering a promising avenue for LLMs post-training with verifiable rewards.

Deep Learning · Large Language Models

Truong Nguyen, Tien-Phat Nguyen, Linh Van, Duy Nguyen, Khoa Doan, Trung Le

Direct Preference Optimization (DPO) is a widely used RL-free method for aligning language models from pairwise preferences, but it models preferences over full sequences even though generation is driven by per-token decisions. Existing token-level extensions typically decompose a sequence-level Bradley–Terry objective across timesteps, leaving per-prefix (state-wise) optimality implicit. We study how to recover **token-level** preference optimality using only standard sequence-level pairwise comparisons. We introduce **Token-level Bregman Preference Optimization (TBPO)**, which posits a token-level Bradley--Terry preference model over next-token actions conditioned on the prefix, and derive a Bregman-divergence density-ratio matching objective that generalizes the logistic/DPO loss while preserving the optimal policy induced by the token-level model and maintaining DPO-like simplicity. We introduce two instantiations: TBPO-Q, which explicitly learns a lightweight state baseline, and TBPO-A, which removes the baseline through advantage normalization. Across instruction following, helpfulness/harmlessness, and summarization, TBPO improves alignment quality and training stability and increases output diversity relative to strong sequence-level and token-level baselines.

Deep Learning · Large Language Models

Yiping Wang, Shao-Rong Su, Zhiyuan Zeng, Eva Xu, Liliang Ren, Xinyu Yang, Zeyi Huang, Xuehai He, Luyao Ma, Baolin Peng 等

Recent advances in large language models (LLMs) have enabled breakthroughs in mathematical discovery, exemplified by AlphaEvolve, a closed-source system that evolves programs to improve bounds on open problems. However, it relies on ensembles of frontier LLMs to achieve new bounds and is a pure inference system that models cannot internalize the evolving strategies. We introduce ThetaEvolve, an open-source framework that simplifies and extends AlphaEvolve to efficiently scale both in-context learning and Reinforcement Learning (RL) at test time, allowing models to continually learn from their experiences in improving open optimization problems. ThetaEvolve features a single LLM, a large program database for enhanced exploration, batch sampling for higher throughput, lazy penalties to discourage stagnant outputs, and optional reward shaping for stable training signals, etc. ThetaEvolve is the first evolving framework that enable a small open-source model, like DeepSeek-R1-0528-Qwen3-8B, to achieve new best-known bounds on open problems (circle packing and first auto-correlation inequality) mentioned in AlphaEvolve. Besides, across two models and four open tasks, we find that ThetaEvolve with RL at test-time consistently outperforms inference-only baselines, and the model indeed learns evolving capabilities, as the RL-trained checkpoints demonstrate faster progress and better final performance on both trained target task and other unseen tasks. We will release our code publicly.

Reinforcement Learning · Deep RL

wenlong deng, Yushu Li, Boying Gong, YI REN, Christos Thrampoulidis, Xiaoxiao Li

Tool-integrated (TI) reinforcement learning (RL) enables large language models (LLMs) to perform multi-step reasoning by interacting with external tools such as search engines and retrievers. Group Relative Policy Optimization (GRPO), exemplified by the recent Search-R1, offers fast convergence and a value-free formulation that makes it appealing for this setting, yet consistently suffers from training collapse. We identify Lazy Likelihood Displacement (LLD), a systematic reduction or stagnation in the likelihood of both correct and incorrect responses, as the core mechanism driving this failure. LLD emerges early and triggers a self-reinforcing LLD Death Spiral, where declining likelihood leads to low-confidence responses, inflating gradients, and ultimately causing collapse. We empirically characterize this process across models on a Search-R1-style, search-integrated question answering task, revealing a consistent three-phase trajectory: early stagnation, steady decay, and accelerated collapse. To address this, we propose a likelihood-preserving regularization LLDS that activates only when a response action’s likelihood decreases, and regularizes only the tokens responsible. This fine-grained structure mitigates LLD with minimal interference. Our method stabilizes training, prevents gradient explosion, and yields substantial performance improvements across seven benchmarks, including relative improvements of +45.2% on Qwen2.5-3B and +37.1% on Qwen2.5-7B over vanilla GRPO training. Our results establish LLD as a previously overlooked bottleneck in GRPO- based TIRL and provide a practical path toward stable, scalable training of tool-integrated RL.

Deep Learning · Large Language Models

Ke Wang, ZEHAO Yu, Luwei Wang, Yongchao Huang

Recent work shows that chain based sampling for power shaped trajectory distributions can deliver large test time gains from a fixed base LLM and can approach RL trained reasoners such as GRPO. Deployment is the bottleneck. Autoregressive Metropolis Hastings is inherently serial, limits GPU utilization, and exhibits extreme tail latency at high budgets, reaching p95 $=1318$s on MATH500 at $128\times$. We propose Adaptive Sequential Monte Carlo (ASMC), a parallel particle inference method that targets power shaped trajectory distributions while adapting particle populations to problem hardness. To make resampling practical for Transformers, we introduce cache coherent resampling, which realizes ancestry updates by reordering KV caches and other particle bound tensors, avoiding prefix recomputation. On MATH500 at the same budget, ASMC attains $80.6\%$ exact match accuracy with p95 $=73.7$s, improving the accuracy to tail latency trade off over both sequential MCMC and best of $n$. We further analyze particle degeneracy and find that collapse severity, measured by low $\mathrm{ESS}_{\min}/N$, strongly predicts failures, while sensitivity to the resampling scheme is limited.

Deep Learning · Large Language Models

Yingru Li, Jiacai Liu, Jiawei Xu, Yuxuan Tong, Ziniu Li, Baoxiang Wang

Policy gradient methods for Large Language Models (LLMs) optimize a policy $\pi_\theta$ via a surrogate objective computed from samples of a rollout policy $\pi_{\text{roll}}$. However, modern LLM-RL pipelines suffer from unavoidable implementation divergences—such as backend discrepancies, Mixture-of-Experts routing discontinuities, and distributed training staleness. These factors cause an off-policy mismatch ($\pi_{\text{roll}} \neq \pi_\theta$), leading to approximation errors between the surrogate and true objectives, often precipitating training collapse. We demonstrate that classical trust region bounds on this error scale as $O(T^2)$ with sequence length $T$, rendering them vacuous for long-horizon tasks. To address this, we derive two tighter bounds: a *Pinsker-Marginal* bound scaling as $O(T^{3/2})$ and a *Mixed* bound scaling as $O(T)$. Crucially, both bounds depend on $\mathcal{D}_{\text{KL}}^{\max}$—the maximum token-level KL divergence across the sequence. As this is a *sequence-level* quantity, it cannot be controlled by token-independent methods like PPO clipping. We propose Trust Region Masking (TRM), which masks entire sequences that violate the trust region. TRM theoretically provides the first non-vacuous monotonic improvement guarantees and empirically improves training stability for long-horizon LLM-RL.

Deep Learning · Large Language Models

Yingru Li, Jiawei Xu, Ziniu Li, Jiacai Liu, Wei Liu, Yuxuan Tong, Longtao Zheng, Zhenghai Xue, Yaxiang Zhang, Tianle Cai 等

Reinforcement Learning for Large Language Models (LLMs) often suffers from training collapse in long-horizon tasks due to exploding gradient variance. To mitigate this, baseline is commonly introduced for advantage computation; however, traditional value models remain difficult to optimize, and standard group-based baselines overlook sequence heterogeneity. Although classic optimal baseline theory can achieve global variance reduction, it neglects token heterogeneity and requires prohibitive gradient-based computation. In this work, we derive the Optimal Token Baseline (OTB) from first principles, proving that gradient updates should be weighted inversely to their cumulative gradient norm. To ensure efficiency, we propose the Logit-Gradient Proxy that approximates the gradient norm using only forward-pass probabilities. Our method achieves training stability and matches the performance of large group sizes ($N=32$) with only $N=4$, reducing token consumption by over 65\% across single-turn and tool-integrated reasoning tasks.

Deep Learning · Large Language Models

Junying Chen, Xinyuan Xie, Ziniu Li, Benyou Wang

Domain adaptation transforms general-purpose LLMs into specialized experts for specific domains or tasks. This process typically follows a two-stage recipe: first, Supervised Fine-Tuning (SFT) to inject domain knowledge or induce specific behaviors (e.g., reasoning patterns), followed by Reinforcement Learning (RL) for self-improvement. However, *does RL truly require a pre-SFT as cold-start phase?* We argue that pre-SFT is inherently problematic: (1) it indiscriminately reinforces knowledge and behaviors from references regardless of whether the LLM has already acquired them, leading to distribution contraction that constrains subsequent exploration; (2) it introduces substantial overhead in multi-stage training and data curation. While our pilot studies reveal that, without pre-SFT, RL struggles to acquire off-policy knowledge from scratch, we bridge this gap with **One-stage Policy Optimization (OnePO)**. OnePO is an SFT-free paradigm that enables LLMs to selectively internalize off-policy knowledge and behaviors directly during RL evolution. Crucially, we design an **Adaptive Objective Evolution** mechanism for rapid knowledge injection and a **Teacher Retirement** mechanism that prevents off-policy anchoring. Experiments demonstrate that OnePO successfully transforms the Qwen3-8B-Base model into a high-performance medical LLM in one RL stage, achieving competitive performance on HealthBench (67.2) and other benchmarks using only 20K samples. This highlights SFT-free RL can efficiently cultivate domain experts without the need for traditional multi-stage pipelines.

Deep Learning · Large Language Models

Ziniu Li, Congliang Chen, Tianyun Yang, Tian Ding, Ruoyu Sun, Ge Zhang, Wenhao Huang, Zhiquan Luo

Large Language Models (LLMs) can improve via reinforcement learning by generating trajectories to discover better solutions. This exploration process represents an investment of finite GPU compute to obtain learning signals. However, current methods typically allocate a small, uniform budget to every task, which is inefficient and ineffective: easy tasks consistently succeed while difficult tasks consistently fail. For policy optimization algorithms such as Group Relative Policy Optimization (GRPO), both edge cases produce zero gradients, resulting in wasted computation. We address this by reframing exploration budget allocation as a resource optimization problem. Viewing each task's exploration as an `"item'' with a distinct "learning value'' and "computational cost'', we establish a connection to the classical knapsack problem and derive an optimal assignment rule. When applied to GRPO, our method increases the ratio of effective gradients by 40\%. As a computational "free lunch'', it enables substantially larger budgets (e.g., 93) for challenging tasks—allocations that would be expensive under a uniform allocation framework. These efficiency gains translate to meaningful improvements on mathematical reasoning benchmarks, with average gains of 2--4 points and peak gains of 9 points. Notably, achieving comparable performance with traditional homogeneous allocation would require approximately $2\times$ the computational resources.

Reinforcement Learning · Deep RL

Shaojun Xu, Xiaoling Zhou, Yihan Lin, Yapeng Meng, Xinglong Ji, Luping Shi, Rong Zhao

Model-Based Reinforcement Learning (MBRL) leverages latent imagination for sample efficiency, yet remains constrained by **Historical Tethering**: imagination is typically initialized from observed states. This creates a learning asymmetry, where the world model’s manifold discovery outpaces the policy's sparse-reward optimization. We propose **Mind Dreamer (MD)**, a framework that operationalizes **Active Counterfactual Reasoning (ACR)** to transcend Markovian continuity. MD reformulates discovery as the minimization of a global Relay Manifold Expected Free Energy (R-EFE); by invoking a latent-space $do$-operator, MD utilizes an adversarial generator to synthesize non-continuous **latent jumps** to epistemic blind spots that are physically plausible yet cognitively challenging. To resolve the credit assignment paradox across these spatial ruptures, we derive the **Relay Value Function (RVF)** and **Relay Uncertainty Function (RUF)**. These potentials treat synthesized anchors as latent bridges, propagating pragmatic and epistemic value through a principled Bellman-style formulation. Notably, we prove that the Uncertainty propagation across discontinuities necessitates a quadratic discount $\gamma^2$, establishing a formal epistemic horizon. Theoretically, MD acts as an optimal importance sampler that expands the manifold's spectral gap, reducing the hitting time to critical bottleneck states. Empirically, MD achieves a **1.67$\times$ average speedup** over DreamerV3 on DeepMind Control Suite, reaching **8.8$\times$** in sparse-reward tasks.

Deep Learning · Generative Models and Autoencoders

Jaeyeon Kim, Seunggeun Kim, Taekyun Lee, David Pan, Hyeji Kim, Sham Kakade, Sitan Chen

A natural desideratum for generative models is \emph{self-correction}--detecting and revising low-quality tokens at inference. While Masked Diffusion Models (MDMs) have emerged as a promising approach for generative modeling in discrete spaces, their capacity for self-correction remains poorly understood. Prior attempts to incorporate self-correction into MDMs either require overhauling MDM architectures/training or rely on imprecise proxies for token quality, limiting their applicability. Motivated by this, we introduce PRISM--Plug-in Remasking for Inference-time Self-correction of Masked Diffusions--a lightweight, model-agnostic approach that applies to any pretrained MDM. Theoretically, PRISM defines a self-correction loss that provably learns per-token quality scores, without RL or a verifier. These quality scores are computed in the same forward pass with MDM and used to detect low-quality tokens. Empirically, PRISM advances MDM inference across domains and scales: Sudoku; unconditional text (170M); and code with LLaDA (8B).

Reinforcement Learning · Deep RL

Haitong Ma, Ofir Nabati, Aviv Rosenberg, Bo Dai, Oran Lang, Craig Boutilier, Na Li, Shie Mannor, Lior Shani, Guy Tennenholtz

Reinforcement learning (RL) struggles to scale to large, combinatorial action spaces common in many real-world problems. This paper introduces a novel framework for training discrete diffusion models as highly effective policies in these complex settings. Our key innovation is an efficient online training process that ensures stable and effective policy improvement. By leveraging policy mirror descent (PMD) to define an ideal, regularized target policy distribution, we frame the policy update as a distributional matching problem, training the expressive diffusion model to replicate this stable target. This decoupled approach stabilizes learning and significantly enhances training performance. Our method achieves state-of-the-art results and superior sample efficiency across a diverse set of challenging combinatorial benchmarks, including DNA sequence generation, RL with macro-actions, and multi-agent systems. Experiments demonstrate that our diffusion policies attain superior performance compared to other baselines. Crucially, our extensive empirical analysis reveals a key trade-off: FKL demonstrates superior sample efficiency and faster initial convergence, whereas RKL ensures greater training stability and higher asymptotic performance.

Reinforcement Learning · Deep RL

Youssef Mroueh

Group Relative Policy Optimization (GRPO) was introduced recently and used to train DeepSeek\textendash R1 for promoting reasoning in LLMs under verifiable (binary) rewards. We show that the mean{+}variance calibration of these rewards induces a contrastive loss in which the contrastive samples are synthetic data drawn from the previous policy. While GRPO was originally paired with clipping to keep updates near the old policy, we analyze variants that differ in reward normalization (mean-only vs.\ mean{+}variance) and in how they regularize updates using KL divergence: either penalizing divergence from the previous model (\emph{mirror}), penalizing divergence from a fixed reference model $\pi_{\mathrm{ref}}$, or combining both forms of regularization. For each, the optimal policy $\pi_n$ admits an explicit form in terms of the binary reward and the first and second order statistics of the reward under $\pi_{n-1}$, as well as the policies $\pi_{n-1}$ and $\pi_{\mathrm{ref}}$. Iterating results in a sequence $\{\pi_n\}$ whose \emph{probability of success (PoS)} obeys a simple recurrence that converges to a fixed point determined by the reference PoS and the regularization strength. We further show that this fixed point exceeds the reference, demonstrating that GRPO amplifies the policy's probability of success.

Deep Learning · Large Language Models

Luke Huang, Zhuoyang Zhang, Qinghao Hu, Shang Yang, Song Han

Reinforcement learning (RL) is widely used to improve large language models (LLMs) on reasoning tasks, and asynchronous RL training is attractive because it increases end-to-end throughput. However, for widely adopted critic-free policy-gradient methods such as REINFORCE and GRPO, high asynchrony makes the policy-gradient estimator **markedly higher variance**: stale off-policy rollouts induce heavy-tailed importance ratios, causing a small fraction of samples to dominate each update. This amplification makes gradients noisy and learning unstable relative to matched on-policy training. Across math and reasoning benchmarks, we find collapse is preceded by sharp drops in effective sample size (ESS) and unstable gradient norms. Motivated by this diagnosis, we propose **V**ariance **C**ontrolled **P**olicy **O**ptimization (VCPO), a drop-in stabilization method for REINFORCE/GRPO-style algorithms that (i) **rescales learning rate according to effective sample size** to dampen unreliable updates, and (ii) applies a **closed-form minimum-variance baseline** for the off-policy setting, avoiding an auxiliary value model and adding minimal overhead. Empirically, VCPO substantially improves robustness in highly asynchronous regimes across models sizes and tasks, reducing long-context, multi-turn training compute by 1.96×. Overall, our results demonstrate explicitly controlling policy-gradient variance is key to making asynchronous RL reliable at scale.

Deep Learning · Generative Models and Autoencoders

Harold Haodong Chen, Xinxiang Yin, Wenjie Shu, Hongfei (Faye) Zhang, Zixin Zhang, Chenfei Liao, Litao Guo, Qifeng Chen, YINGCONG CHEN

Text-to-image (T2I) generation has achieved remarkable progress, yet existing methods often lack the ability to dynamically reason and refine during generation--a hallmark of human creativity. Current reasoning-augmented paradigms mostly rely on explicit thought processes, where intermediate reasoning is decoded into discrete text at fixed steps with frequent image decoding and re-encoding, leading to inefficiencies, information loss, and cognitive mismatches. To bridge this gap, we introduce **LatentMorph**, a novel framework that seamlessly integrates implicit latent reasoning into the T2I generation process. At its core, LatentMorph introduces four lightweight components: (***i***) a **condenser** for summarizing intermediate generation states into compact visual memory, (***ii***) a **translator** for converting latent thoughts into actionable guidance, (***iii***) a **shaper** for dynamically steering next image token predictions, and (***iv***) an RL-trained **invoker** for adaptively determining when to invoke reasoning. By performing reasoning entirely in continuous latent spaces, LatentMorph avoids the bottlenecks of explicit reasoning and enables more adaptive self-refinement. Extensive experiments demonstrate that LatentMorph **(I)** enhances the base model Janus-Pro by 16% on GenEval and 25% on T2I-CompBench; **(II)** outperforms explicit paradigms (*e.g.*, TwiG) by 15% and 11% on abstract reasoning tasks like WISE and IPV-Txt, **(III)** while reducing inference time by 44% and token consumption by 51%; and **(IV)** exhibits 71% cognitive alignment with human intuition on reasoning invocation.

Deep Learning · Large Language Models

Jingchu Gai, Guanning Zeng, Huaqing ZHANG, Aditi Raghunathan

It is widely recognized that reinforcement learning (RL) fine-tuning of large language models often leads to \textit{diversity collapse}, where outputs lack variety. Prior work has proposed a range of heuristics to counteract this effect, but these methods are ad hoc: they frequently trade off correctness for diversity, their effectiveness varies across tasks, and in some cases they even contradict one another. In this work, we place these observations on a rigorous foundation. We first provide a formal proof of why RL fine-tuning exhibits diversity collapse via a selection and reinforcement bias. Next, we make a key observation that any reward modification to address diversity collapse only needs to be applied on the correct trajectories. Building directly on this analysis, we introduce a principled method---\textit{differential smoothing}---that provably improves both correctness and diversity, outperforming vanilla RL as well as widely used entropy-based heuristics. Our theory precisely characterizes when existing heuristics help and why they fail, while showing that differential smoothing is universally superior. Extensive experiments with models from 1B to 7B parameters, across domains including CountDown and real-world mathematical reasoning, demonstrate consistent gains. Differential smoothing improves both Pass@1 and Pass@k, with up to 6.7\% improvements on AIME24 dataset.

Deep Learning · Large Language Models

Zeping Min, Weihang Xu, Zhengzhong You, Wotao Yin, Xinshang Wang

Natural language to optimization (NL2Opt) requires translating unstructured text into executable mathematical models. Beyond simple syntax errors, this task suffers from silent modeling failures, where incorrect formulations execute successfully but yield invalid results. We propose Draft-and-Audit RL (DA-RL), a framework that learns optimization modeling as a two-step iterative workflow. Unlike inference-time scaffolds that rely on intermediate solver feedback to guide repairs, DA-RL optimizes a shared-parameter policy using terminal-only verification: the model is rewarded solely based on the execution of the final audited program. This constraint forces the model to internalize rubric-guided revision as a learned capability and encourages the emergence of cross-turn synergy, where the policy learns to generate drafts that are structurally amenable to self-correction.

Deep Learning · Large Language Models

Qiang Zhang, Boli Chen, Fanrui Zhang, Ruixue Ding, Shihang Wang, Qiuchen Wang, Yinfeng Huang, Haonan Zhang, Rongxiang Zhu, Xin Li 等

Reinforcement learning (RL) has advanced LLM agents on verifiable tasks but remains challenging for open-ended tasks with vast solution spaces (e.g., complex travel planning). Lacking objective ground truth, current RL algorithms rely on reward models assigning scalar scores to individual responses. We contend such pointwise scoring induces discrimination collapse: reward model fails to distinguish subtle advantages among trajectories, compressing intra-group rewards into a narrow range. This drowns effective reward signals in reward model noise, causing optimization stagnation. To address this, we propose ArenaRL, a reinforcement learning paradigm shifting from pointwise scalar scoring to intra-group relative ranking. ArenaRL introduces a process-aware pairwise evaluation with multi-level rubrics for fine-grained relative scoring. Meanwhile, we construct an intra-group adversarial arena and devise a tournament-based ranking scheme to obtain stable advantage signals. ArenaRL achieves high-precision advantage estimation with only $O(N)$ computational complexity, striking a favourable balance between efficiency and accuracy. Furthermore, to address the lack of full-cycle benchmarks for open-ended agents, we introduce two high-quality benchmarks: Open-Travel and Open-DeepResearch, encompassing full training and multi-dimensional evaluation pipelines. Extensive experiments across three open-ended tasks validate the effectiveness of ArenaRL.

Reinforcement Learning · Deep RL

Huy Le, Onur Celik, Denis Blessing, Tai Hoang, Claas Voelcker, Axel Brunnbauer, Felix Richter, Michael Volpp, Gerhard Neumann

Reinforcement learning with massively parallel simulations has become an emerging trend; however, most existing approaches still rely on simple Gaussian policy parameterizations. Diffusion models provide a more expressive policy class and have shown strong performance on challenging control problems, yet most diffusion-based RL methods are designed for offline or off-policy training. In this work, we ask whether diffusion policies can be trained effectively in the massively parallel, on-policy regime. To this end, we introduce Trust-region Diffusion Policies (TruDi), which enables diffusion policies for on-policy RL with massively parallel simulations. This setting is particularly challenging because the data distribution changes quickly across updates, making stable training with complex policies difficult. TruDi addresses this by integrating a trust-region optimization rule to enforce a KL-divergence constraint over the entire diffusion trajectory, rather than only at the final denoising step. Empirically, we evaluate TruDi on a diverse set of 4 massively parallel RL benchmarks comprising a total of 73 tasks. Across these tasks, TruDi consistently outperforms or is on-par with strong baselines on standard tasks and achieves clear gains on more challenging humanoid control tasks, establishing a strong new baseline for massively parallel on-policy RL.

Deep Learning · Large Language Models

Zhe Zheng, Wenqi Zhang, Xiaohe Zhou, Guiyang Hou, Xinyu Mei, Tianyang Mao, Anni PIAO, Shengpei Jiang, Yongliang Shen, Yueting Zhuang 等

Multimodal large language models (MLLMs) have demonstrated remarkable reasoning capabilities over internalized knowledge. However, current research overlooks contextual reasoning, the ability to reason based on the relevant information present in the context. To investigate this issue, we construct the Visual Contextual Reasoning Benchmark (ContextReasonV-Bench), and our analysis reveals two predominant failure modes: \textit{context neglect}, where models rely on pre-trained knowledge instead of contextual information, and \textit{superficial pattern matching}, where models exploit shallow correlations rather than genuine patterns. To address these limitations, we propose a two-stage approach that progressively establishes and reinforces contextual pattern acquisition. The first stage establishes an "analyze-then-solve" reasoning paradigm through supervised fine-tuning (SFT). We then employ a context-aware reinforcement learning (RL) framework that integrates context-aware reward modeling with hierarchical advantage estimation to encourage the model to identify genuine contextual patterns. This approach yields Context-Aware Reasoner (CAR), a model that achieves 38.14\% accuracy on ContextReasonV-Bench, improving the base model by 22.09\%. Notably, CAR exhibits strong generalization to tasks not seen during training, confirming that our approach enhances genuine contextual reasoning capability.