Papers with RLVR
ScaleBox: Enabling High-Fidelity and Scalable Code Verification for Large Language Models (2026.acl-demo)
Copied to clipboard
Jiasheng Zheng, Xin Zheng, Boxi Cao, Pengbo Wang, Zhengzhao Ma, Qiming Zhu, Jiazhen Jiang, Yaojie Lu, Hongyu Lin, Xianpei Han, Le Sun
| Challenge: | Existing code sandboxes fail to provide accurate verification and efficiency under high-concurrency workloads. |
| Approach: | They propose a high-fidelity code verification system that provides sandbox feedback for RL training and evaluation. |
| Outcome: | The proposed system outperforms heuristic-matching baselines on LiveCodeBench and training stability on high-concurrency workloads. |
Online Difficulty Filtering for Reasoning Oriented Reinforcement Learning (2026.eacl-long)
Copied to clipboard
| Challenge: | Recent advances in reinforcement learning with verifiable rewards (RLVR) show that large language models enhance their reasoning abilities when trained with veriable signals. |
| Approach: | They propose a method for a problem-aware filtering system that maximizes learning efficiency by selecting tasks of intermediate difficulty. |
| Outcome: | The proposed model improves when trained with verifiable rewards, but training efficiency is bottleneck . the proposed model achieves +12% gains in less than half the training steps of standard GRPO . |
Plasticity vs. Rigidity: The Impact of Low-Rank Adapters on Reasoning on a Micro-Budget (2026.eacl-srw)
Copied to clipboard
| Challenge: | Recent advances in mathematical reasoning typically rely on massive scale . yet, can strong reasoning capabilities be induced in small language models under extreme constraints? |
| Approach: | They train small language models with a single GPU for under 24 hours . they find that adapters unlock significant plasticity in standard instruction-tuned models . |
| Outcome: | The proposed model training on a single GPU (48GB) achieves 40% Pass@1 on AIME 24 (an 11.1% improvement over baseline) the model training results show that the adapter capacity and initialization are critical factors. |
OpenRLHF: A Ray-based Easy-to-use, Scalable and High-performance RLHF Framework (2025.emnlp-demos)
Copied to clipboard
Jian Hu, Xibin Wu, Wei Shen, Jason Klein Liu, Weixun Wang, Songlin Jiang, Haoran Wang, Hao Chen, Bin Chen, Wenkai Fang, null Xianyu, Yu Cao, Haotian Xu, Yiming Liu
| Challenge: | Existing RLHF frameworks face inference bottlenecks and complexity barriers restricting their accessibility for newcomers. |
| Approach: | They propose an open-source RLHF framework that can be used to train large language models. |
| Outcome: | The proposed framework achieves superior training efficiency with speedups ranging from 1.22 to 1.68 across different model sizes compared to state-of-the-art frameworks, while requiring significantly fewer lines of code for implementation. |
Act as you think: Reinforcing Consistent Reasoning in Medical Visual Question Answering (2026.acl-long)
Copied to clipboard
Songtao Jiang, Yuan Wang, Ruizhe Chen, Yan Zhang, Ruilin Luo, Bohan Lei, Yeying Jin, Sibo Song, ZhiBo Yang, Jimeng Sun, Jian Wu, Zuozhu Liu
| Challenge: | Recent advances have improved the accuracy of medical visual question answering (Med-VQA) however, the high stakes nature of the medical domain has precipitated a shift towards interpretability and transparency of reasoning processes. |
| Approach: | They propose a reinforcement learning from verifiable rewards framework that rewards internal consistency and logical coherence. |
| Outcome: | The proposed framework rewards internal consistency and logical coherence, and is highly versatile, the authors show. |
BV-Blend: Uncertainty-Weighted Historical Baselines for Stable Critic-Free RL with Verifiable Rewards (2026.findings-acl)
Copied to clipboard
| Challenge: | Critic-free reinforcement learning with verifiable rewards (RLVR) is a practical paradigm for aligning Large Language Models. |
| Approach: | They propose a framework that stabilizes advantage estimation by combining prompt-local on-policy statistics with semantic-cluster-conditioned historical moments. |
| Outcome: | Experiments show that RLVR improves training stability and performance compared to critic-based methods . compared with other approaches, RL VR improves in cold-start regimes with binary verifiers . |
One Missing Piece for Open-Source Reasoning Models: A Dataset to Mitigate Cold-Starting Short CoT LLMs in RL (2025.acl-industry)
Copied to clipboard
Hyungjoo Chae, Dongjin Kang, Jihyuk Kim, Beong-woo Kwak, Sunghyun Park, Haeju Park, Jinyoung Yeo, Moontae Lee, Kyungjae Lee
| Challenge: | Existing large reasoning models are limited by their closed nature and high API costs and safety issues. |
| Approach: | They propose to build a long CoT dataset with existing short CoT LLMs that are not trained for inference-time scaling. |
| Outcome: | The proposed model achieves quality comparable to—or slightly below—R1 and is able to think longer and provide control over the thought budget to better manage the overthinking problem. |
PaperSearchQA: Learning to Search and Reason over Scientific Papers with RLVR (2026.eacl-long)
Copied to clipboard
James Burgess, Jan N. Hansen, Duo Peng, Yuhui Zhang, Alejandro Lozano, Min Woo Sun, Emma Lundberg, Serena Yeung-Levy
| Challenge: | Recent methods supervise only the final answer accuracy using reinforcement learning with verifiable rewards (RLVR). |
| Approach: | They propose to train search agents to search and reason over scientific papers and a factoid QA dataset with 60k biomedical paper abstracts. |
| Outcome: | The proposed model outperforms non-RL retrieval baselines and is scalable and extendable to other scientific domains. |
From Data-Centric to Sample-Centric: Enhancing LLM Reasoning via Progressive Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has recently advanced the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose a method that incorporates partial solution prefixes from expert demonstrations to guide the policy. |
| Outcome: | The proposed methods outperform strong baselines, yielding faster convergence and a higher performance ceiling. |
Think Outside the Policy: In-Context Steered Policy Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing Reinforcement Learning from Verifiable Rewards (RLVR) methods exhibit limited exploration due to reliance on on-policy rollouts which are limited to the current policy’s distribution, resulting in narrow trajectory diversity. |
| Approach: | They propose a framework that leverages the in-context learning capability of Large Reasoning Models to provide expert guidance using existing datasets. |
| Outcome: | The proposed framework improves RLVR performance and training stability on mathematical reasoning benchmarks. |
Rethinking Sample Polarity in Reinforcement Learning with Verifiable Rewards (2026.acl-long)
Copied to clipboard
| Challenge: | Large reasoning models are typically trained using reinforcement learning with verifiable reward (RLVR) positive and negative self-generated rollouts are used to update the model's policy . positive samples sharpen existing correct reasoning patterns, while negative samples encourage exploration of new reasoning paths. |
| Approach: | They propose a method that allocates advantage signals to key tokens across different polarities. |
| Outcome: | The proposed method improves the ability of large reasoning models to learn from their own generated rollouts. |
From Coarse to Fine: Benchmarking and Reward Modeling for Writing-Centric Generation Tasks (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing evaluation benchmarks for writing reward models are coarse-grained. |
| Approach: | They propose a benchmark and a fine-grained training framework to evaluate writing reward models. |
| Outcome: | The proposed model improves on various writing benchmarks and exhibits strong generalization. |
CODERL+: Improving Code Generation via Reinforcement with Execution Semantics Alignment (2026.acl-long)
Copied to clipboard
Xue Jiang, Yihong Dong, Mengyang Liu, Deng Hongyi, Tian Wang, Yongding Tao, Zhi Jin, Wenpin Jiao, Ge Li
| Challenge: | Large Language Models excel at code generation by learning from vast code corpora, but a fundamental semantic gap remains between training on textual patterns and the goal of functional correctness . reinforcement learning with verifiable rewards (RLVR) approaches are inefficient for establishing a well-aligned connection between the textual representation of code and its execution semantics. |
| Approach: | They propose a novel approach that integrates execution semantics alignment into the RLVR training pipeline for code generation. |
| Outcome: | The proposed model outperforms baseline training and RLVR and shows strong applicability across RL and LLMs. |
Crossing the Reward Bridge: Expanding Reinforcement Learning with Verifiable Rewards Across Diverse Domains (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has been effective on structured tasks, but its reliance on simple, rule-based verifiers creates a bottleneck. |
| Approach: | They propose a framework that uses a generative verifier to provide soft, probabilistic rewards. |
| Outcome: | The proposed framework outperforms existing models up to 10x their size and can be scalable and effective. |
G2RPO-A: Guided Group Relative Policy Optimization with Adaptive Guidance (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in reasoning-centric large language models (LLMs) have significantly expanded the performance boundaries of LLMs, showcasing the immense potential of reasoning-enhanced models. |
| Approach: | They propose an adaptive algorithm that injects ground-truth reasoning steps into roll-out trajectories to compensate for SLMs’ inherent weaknesses. |
| Outcome: | Experiments on mathematical reasoning and code-generation benchmarks confirm that G2RPO-A substantially outperforms vanilla GRPO. |
Writing-RL: Advancing Long-form Writing via Adaptive Curriculum Reinforcement Learning (2026.acl-long)
Copied to clipboard
Xuanyu Lei, Chenliang Li, Yuning Wu, Kaiming Liu, Weizhou Shen, Peng Li, Ming Yan, Fei Huang, Ya-Qin Zhang, Yang Liu
| Challenge: | Recent advances in Large Language Models (LLMs) have enabled strong performance in long-form writing, but current training paradigms remain limited. |
| Approach: | They propose an Adaptive Curriculum Reinforcement Learning framework to advance long-form writing capabilities beyond SFT. |
| Outcome: | Experiments on 7B-scale writer models show that Writing-RL improves long-form writing performance over strong SFT baselines. |
Visually-Guided Policy Optimization for Multimodal Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing RLVRs lack visual faithfulness due to text-dominated reasoning . a novel framework to reinforce visual focus during policy optimization is proposed . |
| Approach: | They propose a framework to reinforce visual focus during policy optimization using visual attention compensation mechanism. |
| Outcome: | The proposed framework exhibits better visual activation and superior performance in multimodal reasoning and visual-dependent tasks. |
Step Potential Advantage Estimation: Harnessing Intermediate Confidence and Correctness for Efficient Mathematical Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to RLVR provide sparse supervision since reward arrives only after the full generation is complete. |
| Approach: | They propose a step-level reward system that extracts confidence and correctness and combines them into a Step Potential signal that explicitly estimates reasoning state at each step. |
| Outcome: | The proposed method outperforms existing methods on multiple benchmarks and improves accuracy while reducing response length. |
Identification of Multiple Logical Interpretations in Counter-Arguments (2025.emnlp-main)
Copied to clipboard
Wenzhi Wang, Paul Reisert, Shoichi Naito, Naoya Inoue, Machi Shimmei, Surawat Pothong, Jungmin Choi, Kentaro Inui
| Challenge: | Counter-arguments (CAs) are a good way to improve learners' critical thinking skills . however, it is difficult to provide every learner tailored feedback due to limited human resources and heavy workloads. |
| Approach: | They propose to annotate a dataset of 134 CAs annotated with 13 logical predicate questions and train a model with Reinforcement Learning with Verifiable Rewards to identify multiple logical interpretations. |
| Outcome: | The proposed model performs on par with larger proprietary models. |
AGGC: Adaptive Group Gradient Clipping for Stabilizing Large Language Model Training (2026.findings-acl)
Copied to clipboard
| Challenge: | Adaptive group-wise gradient clipping (AGGC) is a new approach to stabilize training of Large Language Models. |
| Approach: | They propose a method to stabilize gradient clipping by partitioning parameters into groups based on functional types and a time-dependent scheduling mechanism to balance exploration and convergence. |
| Outcome: | The proposed algorithm outperforms standard LoRA and achieves 72.93% accuracy . it can be integrated into existing pipelines with negligible overhead. |
Do Not Step Into the Same River Twice: Learning to Reason from Trial and Error (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to RLVR train LMs based on their own on-policy responses and are constrained by the initial capability of LM. |
| Approach: | They propose an approach that hints LMs with their self-made mistakes without external guidance. |
| Outcome: | The proposed approach outperforms the normal group relative policy optimization and requires no external guidance. |
Reinforcing Agentic Search Via Reward Density Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) is a promising approach for enhancing agentic search, but its performance is often hindered by reward sparsity . |
| Approach: | They propose a new research problem to improve the reward obtained per unit of exploration cost by using a system that decomposes long-horizon tasks into intermediate objectives and assigns process-level rewards to provide denser learning signals. |
| Outcome: | The proposed framework outperforms strong baselines on several agentic search benchmarks and achieves comparable performance to that of advanced proprietary models. |
HarmRLVR: Weaponizing Verifiable Rewards for Harmful LLM Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have gained significant attention due to their objective and verifiably verifier reward signals. |
| Approach: | They propose to exploit RLVR for alignment reversibility by using GRPO to reverse alignment with merely 64 harmful prompts without responses. |
| Outcome: | The proposed method outperforms fine-tuning and RLHF in reasoning and code generation tasks while maintaining general capabilities. |
UR2 : Unify RAG and Reasoning through Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing attempts to unify large language models are limited to open-domain QA with fixed retrieval settings. |
| Approach: | They propose a general reinforcement learning framework that dynamically coordinates retrieval and reasoning. |
| Outcome: | The proposed framework outperforms existing paradigms on open-domain QA, MMLU-Pro, medical, and mathematical reasoning tasks. |
Adapt to Thrive! Adaptive Power-Mean Policy Optimization for Improved LLM Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for reinforcement learning with verifiable rewards (RLVR) rely on static objective functions and rigid clipping strategies that misalign with the model’s evolving reasoning capabilities. |
| Approach: | They propose to incorporate Power-Mean Policy Optimization (PMPO) and Feedback-Adaptive Clipping (FAC) to overcome limitations of static mechanisms. |
| Outcome: | Extensive experiments on nine reasoning tasks show the proposed paradigm outperforms state-of-the-art methods. |
Balancing Classification and Calibration Performance in Decision-Making LLMs via Calibration Aware Reinforcement Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are increasingly deployed in decision-making tasks where accuracy and reliable confidence estimates are essential. |
| Approach: | They propose a calibration-aware reinforcement learning formulation that directly adjusts decision-token probabilities. |
| Outcome: | The proposed model preserves RLVR’s accuracy level while mitigating overconfidence, reducing ECE scores up to 9 points. |
Prune as You Generate: Online Rollout Pruning for Faster and Better RLVR (2026.acl-long)
Copied to clipboard
Haobo Xu, Sirui Chen, Ruizhong Qiu, Yuchen Yan, Chen Luo, Monica Xiao Cheng, Jingrui He, Hanghang Tong
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) has improved reasoning capabilities of Large Language Models (LLMs). |
| Approach: | They propose an online pruning method that prunes rollouts while steering correct ones to enhance learning signals. |
| Outcome: | The proposed method improves average accuracy by +2.30 to +2.99 across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models. |
A Few Bad Apples Spoil the Bunch: Preventing Global Entropy Collapse Driven by a Small Set of Tokens in LLM Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards and Reinforced Learning from internal feedback fail to benefit from test-time compute due to entropy collapse and the resulting loss of reasoning diversity. |
| Approach: | They propose a strategy that assigns each generated token a redistribution score and applies selective KL regularization to only the top 5% of tokens under this score. |
| Outcome: | The proposed model improves on both RLVR and RLIF models on math reasoning benchmarks, showing that targeted entropy control at a vanishingly small subset of tokens is sufficient to sustain reasoning diversity and effective test-time scaling. |
Less Noise, More Voice: Reinforcement Learning for Reasoning via Instruction Purification (2026.findings-acl)
Copied to clipboard
| Challenge: | Experimental results show that LENS outperforms GRPO in delivering higher performance and faster convergence. |
| Approach: | They propose a framework that purifies prompts by identifying and removing interference tokens and then transfers successful rollouts to supervise policy optimization on original noisy prompts. |
| Outcome: | The proposed framework outperforms GRPO in the real-world, with a 3.88% gain and speedup. |
PRIME: A Process-Outcome Alignment Benchmark for Verifiable Reasoning in Mathematics and Engineering (2026.acl-long)
Copied to clipboard
Xiangfeng Wang, Hangyu Guo, Yanlin Lai, Mitt Huang, Liang Zhao, Chengyuan Yao, Yinmin Zhang, Qi Han, null Xiaoxiaoren, Chun Yuan, Tong Xu, Zheng Ge, Xiangyu Zhang, Daxin Jiang
| Challenge: | Current outcome-centric verification paradigms neglect potential errors in the derivation process. |
| Approach: | They propose a process-aware RLVR training paradigm utilizing verifiers selected via **PRIME**. |
| Outcome: | The proposed approach outperforms the baseline verification paradigm on AIME24, AIME25, and Beyond-AIME models. |
Beyond Length Scaling: Synergizing Breadth and Depth for Generative Reward Models (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in Generative Reward Models have demonstrated that scaling the length of Chain-of-Thought reasoning enhances reliability of evaluation. |
| Approach: | They propose a framework that reconfigures raw rationales into structured Breadth-CoT and Depth-Co T through a modular synthesis pipeline. |
| Outcome: | The proposed framework surpasses open-source RMs by an average of 8.2%. |
How to Allocate, How to Learn? Dynamic Rollout Allocation and Advantage Modulation for Policy Optimization (2026.findings-acl)
Copied to clipboard
Yangyi Fang, Jiaye Lin, Xiaoliang Fu, Cong Qin, Haolin Shi, Chaowen Hu, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing methods for reinforcement learning with verifiable rewards are limited by the complexity of the problem and the complexity. |
| Approach: | They propose a theoretically-grounded dual-pronged optimization framework for reinforcement learning with verifiable rewards that compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
| Outcome: | The proposed framework compensates for gradient attenuation of high-confidence correct actions while utilizing entropy changes as computable indicators to stabilize excessive update magnitudes. |
Warm Up Before You Train: Unlocking General Reasoning in Resource-Constrained Settings (2025.emnlp-main)
Copied to clipboard
| Challenge: | Reasoning-capable large language models (LLMs) have driven a major shift in artificial intelligence . these models generate long CoTs, capturing reasoning behaviors such as self-reflection, self-correction, and hypothesis testing. |
| Approach: | They propose a sample-efficient, two-stage training strategy to build reasoning LLMs . they "warm up" a model by distilling Long CoTs from a toy domain to acquire general reasoning skills . |
| Outcome: | The proposed training strategy outperforms existing models on a range of tasks. |
When and What to Ask: AskBench and Rubric-Guided RLVR for LLM Clarification (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models respond even when prompts omit critical details or include misleading information, leading to hallucinations or reinforced misconceptions. |
| Approach: | They propose an interactive benchmark that converts standard QA pairs into multi-turn interactions with explicit checkpoints. |
| Outcome: | The proposed benchmark improves accuracy, rubric adherence, and interaction efficiency with strong generalization to unseen domains. |
Understanding and Preventing Entropy Collapse in RLVR with On-Policy Entropy Flow Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing RLVR algorithms suffer from entropy collapse, leading to premature determinism and unstable optimization. |
| Approach: | They propose an adaptive entropy flow balancing mechanism that rescales entropic-increasing and enotro-decreazing updates according to their contributions to enthroy change. |
| Outcome: | The proposed method outperforms existing RLVR algorithms on six reasoning benchmarks. |
Revisiting Entropy Regularization: Adaptive Coefficient Unlocks Its Potential for LLM Reinforcement Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reasoning ability is a defining capability of Large Language Models (LLMs), but RLVR training suffers from policy entropy collapse, hindering exploration and limiting reasoning performance. |
| Approach: | They propose a framework that dynamically balances exploration and exploitation via three components: difficulty-aware coefficient allocation, initial-anchored target entropy, and dynamic global coefficient adjustment. |
| Outcome: | The proposed framework outperforms baselines on multiple mathematical reasoning benchmarks. |
HTMR: Hybrid Token Masking Reinforcement Learning with Verifiable Rewards for Event Argument Extraction with Multi-Perspective Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work formulates EAE with large language models as a structured conditional generation task and applies Reinforcement Learning with Verifiable Rewards (RLVR) to optimize sequence-level event structures. |
| Approach: | They propose a method that selectively updates policy gradients on high-entropy forking tokens and event-critical tokens that define event structure. |
| Outcome: | The proposed method outperforms full-token and high-entropy only methods and transfers effectively as a plug-and-play approach to other tasks such as named entity recognition and relation classification. |
Unlocking Exploration in RLVR: Uncertainty-aware Advantage Shaping for Deeper Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) has shown significant promise for enhancing the reasoning capabilities of large language models (LLMs). |
| Approach: | They propose a model-free method that refines credit assignment by leveraging the model's internal uncertainty signals. |
| Outcome: | Extensive experiments on five mathematical reasoning benchmarks show that the proposed method outperforms strong RLVR baselines on multiple model scales, including 1.5B and 7B. |
Rethinking Multiple-Choice Questions for RLVR: Unlocking Potential via Distractor Design (2026.findings-acl)
Copied to clipboard
Xu Guo, Qiming Ge, Jian Tong, Kedi Chen, Jin Zhang, Xiaogui Yang, Xuan Gao, Haijun Lv, Zhihui Lu, Yicheng Zou, Qipeng Guo
| Challenge: | Existing approaches to RLVR use multiple-choice questions as verifiable rewards . however, not all tasks provide reliable verification . |
| Approach: | They propose a framework that actively constructs high-quality distractors to block elimination shortcuts and promote deep reasoning. |
| Outcome: | The proposed method significantly improves reasoning capabilities of Large Language Models. |
EvoCoT: Overcoming the Exploration Bottleneck in Reinforcement Learning for LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing approaches to reinforcement learning with verifiable reward (RLVR) are limited by difficulty or lack of exploration. |
| Approach: | They propose a self-evolving curriculum learning framework based on chain-of-thought reasoning optimization that constrains exploration space by self-generating and verifying CoT trajectories. |
| Outcome: | The proposed framework enables LLMs to solve previously unsolved problems without external supervision and is compatible with various RL fine-tuning methods. |
Table-R1: Inference-Time Scaling for Table Reasoning Tasks (2025.emnlp-main)
Copied to clipboard
| Challenge: | In this study, we explore inference-time scaling on table reasoning tasks. |
| Approach: | They propose a large-scale dataset of reasoning traces and a reinforcement learning with verifiable rewards approach to enable inference-time scaling on table reasoning tasks. |
| Outcome: | The proposed model matches or exceeds GPT-4.1 and DeepSeek-R1 models on diverse table reasoning tasks. |
From Off-Policy to On-Policy: Enhancing GUI Agents via Bi-level Expert-to-Policy Assimilation (2026.acl-long)
Copied to clipboard
| Challenge: | Vision-language models are increasingly deployed as computer-use agents that operate desktops and browsers. |
| Approach: | They propose a method that turns static expert traces into policy-aligned guidance . they propose RLVR with a per-task, dynamically updated cache to decompose planning and execution . |
| Outcome: | The proposed model improves UITARS1.5-7B success from 22.87% to 32.13% on OSWorld-Verified and raises a held-out split from 5.74% to 10.30% on MMBench-GUI and Online-Mind2Web. |
I²B-LPO: Latent Policy Optimization via Iterative Information Bottleneck (2026.acl-long)
Copied to clipboard
Huilin Deng, Hongchen Luo, Yue Zhu, Long Li, Zhuoyue Chen, Xinghao Zhao, Ming LI, Chuyang Zhao, Jihai Zhang, MengChang Wang, Yang Cao, Yu Kang
| Challenge: | Existing methods for large language model reasoning suffer from exploration collapse due to the semantic homogeneity of random rollouts. |
| Approach: | They propose to use latent policy optimization via iterative information bottleneck to optimize reasoning trajectories by diversifying reasoning . |
| Outcome: | Empirical results show that the proposed method achieves state-of-the-art performance with margins of up to 5.3% in accuracy and 7.4% in diversity metrics. |
Powering Verifiable Learning via Automated Evolutionary Data Synthesis (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to building generalizable verifiable data are task-specific and lack a principled, universal evaluator of verifikatability. |
| Approach: | They propose a task-agnostic, strategy-guided, executably-checkable data synthesis framework that synthesizes problems, diverse candidate solutions and verification artifacts from a single source. |
| Outcome: | The proposed framework synthesizes problems, candidates, and verification artifacts from human-annotated and strategy-induced checks and iteratively discovers strategies. |
GeoRA: Geometry-Aware Low-Rank Adaptation for RLVR (2026.acl-long)
Copied to clipboard
| Challenge: | Existing parameter-efficient methods for RLVR face limitations . low-rank adaptation methods do not account for the distinct optimization dynamics . |
| Approach: | They propose a low-rank adaptation method tailored for RLVR that exploits the anisotropic structure of RL update subspace and extracts its principal directions via Singular Value Decomposition (SVD). |
| Outcome: | Experiments on large reasoning models show that GeoRA outperforms strong low-rank baselines across RLVR settings while showing stronger generalization and less forgetting on out-of-domain tasks. |
SEARL: Joint Optimization of Policy and Tool Graph Memory for Self-Evolving Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have demonstrated significant potential in single-turn reasoning tasks. |
| Approach: | They propose a tool-memory based self-evolving agentic framework that integrates planning with execution. |
| Outcome: | The proposed framework is able to extract explicit knowledge from historical data and leverage inter-trajectory correlations to densify reward signals. |
PARIF: Pushing the Pareto Frontier of Instruction Following and Reasoning with Curriculum Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing alignment methods struggle to balance general reasoning with instruction-following (IF) this is hindered by dependency on teacher models, reward hacking, and reasoning-answer inconsistencies. |
| Approach: | They propose a two-stage curriculum learning framework based on Reinforcement Learning from Verifiable Rewards to enhance both IF and general reasoning capabilities. |
| Outcome: | The proposed framework outperforms leading models on six representative IF tasks while achieving a 21.25% relative average improvement over the original model. |
Towards Stable and Effective Reinforcement Learning for Mixture-of-Experts (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) training with Mixture-of-Experts policies remains fragile and prone to reward collapse. |
| Approach: | They propose a router shift-based policy optimization method that computes a per-token router-shift ratio conditioned on the previously activated experts and applies stop-gradient and a lower-bound floor. |
| Outcome: | The proposed method achieves better performance and greater stability than previous methods. |
Verifying the Subjective: Structured Multilingual Rewards for Low-Resource Alignment (2026.findings-acl)
Copied to clipboard
| Challenge: | Structured Multilingual Reward Modeling Framework extends Reinforcement Learning with Verifiable Rewards (RLVR) to subjective and open-ended tasks. |
| Approach: | They propose a framework that extends Reinforcement Learning with Verifiable Rewards to subjective and open-ended tasks. |
| Outcome: | The proposed framework improves reasoning capability and response quality on 7 tasks across 50 low-resource languages. |
Low-probability Tokens Sustain Exploration in Reinforcement Learning with Verifiable Reward (2026.findings-acl)
Copied to clipboard
Guanhua Huang, Tingqiang Xu, Mingze Wang, Qi Yi, Xue Gong, Siheng Li, Ruibin Xiong, Kejiao Li, Yuhao Jiang, Bo Zhou
| Challenge: | Recent studies show that RLVR training is slow and results plateau as policy entropy collapses . low-probability regularization (Lp-Reg) reduces the number of low-quality exploratory tokens induced by RL training . |
| Approach: | They propose a method to reduce RLVR over-penalization by eliminating low-probability exploratory tokens . they propose 'Low-provability Regularization' to reduce the gradual elimination of low-quality exploratory entropy tokens. |
| Outcome: | The proposed method eliminates low-probability exploratory tokens and prevents suppression of potentially valuable low-property candidates. |
Revisiting Entropy in Reinforcement Learning for Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Renren Jin, Pengzhi Gao, Yuqi Ren, Zhuowen Han, Tongxuan Zhang, Wuwei Huang, Wei Liu, Jian Luan, Deyi Xiong
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has emerged as a paradigm for enhancing the reasoning capabilities of large language models. |
| Approach: | They propose a positive-advantage reweighting approach that regulates model entropy by adjusting the loss weights assigned to tokens with positive advantages during RLVR training. |
| Outcome: | The proposed approach regulates model entropy by adjusting loss weights assigned to tokens with positive advantages during RLVR training while maintaining competitive performance. |
AutoRubric: Rubric-Based Generative Rewards for Faithful Multimodal Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Multimodal large language models (MLLMs) have advanced from perception tasks to complex multi-step reasoning. |
| Approach: | They propose a framework that integrates reinforcement learning with verifiable rewards with process-level supervision through automatically collected rubric-based generative rewards. |
| Outcome: | The proposed framework achieves state-of-the-art performance on six multimodal reasoning benchmarks and significantly improves reasoning faithfulness in dedicated evaluations. |
Exploring Chain-of-Thought Reasoning for Steerable Pluralistic Alignment (2025.emnlp-main)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are typically trained to reflect a relatively uniform set of values, which limits their applicability to tasks that require understanding of nuanced human perspectives. |
| Approach: | They propose to use Chain-of-Thought reasoning techniques to build steerable pluralistic models by fine-tuning on human-authored CoT and synthetic explanations. |
| Outcome: | The proposed methods outperform others and demonstrate strong sample efficiency. |
Incentivizing In-depth Reasoning over Long Contexts with Process Advantage Shaping (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) has proven effective in enhancing LLMs’ short-context reasoning but falters in long-contemporal scenarios requiring precise grounding and multi-hop reasoning. |
| Approach: | They propose a framework that constructs high-difficulty, multi-hop long-context QA pairs with inherent reasoning chains to overcome this bottleneck. |
| Outcome: | The proposed framework outperforms RLVR baselines and matches frontier LLMs while using far fewer parameters. |
CURE: Critique-Driven Unified Reinforcement Learning for Test-Time Self-Improvement (2026.acl-long)
Copied to clipboard
| Challenge: | Existing critique-guided methods fail to equip models with the autonomous improvement capabilities required for test-time scaling. |
| Approach: | They propose a framework that jointly optimizes a single policy for standard solving, critiquing, and guided re-exploration. |
| Outcome: | The proposed framework maintains competitive single-turn performance and unlocks effective inference-time scaling. |
Generative Floor Plan Design with LLMs via Reinforcement Learning with Verifiable Rewards (2026.findings-acl)
Copied to clipboard
Luis Lara, Aristides Milios, ZhiHao Luo, Aditya Sharma, Ge Ya Luo, Christopher Beckham, Florian Golemo, Christopher Pal
| Challenge: | Existing generative models focus on respecting the requested connectivity between rooms, but do not support generating floor plans that respect numerical constraints. |
| Approach: | They propose a text-based approach that fine-tunes a large language model on real plans and applies reinforcement learning with verifiable rewards to improve adherence to topological and numerical constraints. |
| Outcome: | The proposed model outperforms existing methods on Realism, Compatibility, Diversity metrics. |
Forge: Quality-Aware Reinforcement Learning for NP-Hard Optimization in LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks focus on correctness, overlooking optimality . large language models excel at math, coding, logic and puzzles . |
| Approach: | They propose a framework for training and evaluating Large Language Models on NP-hard optimization problems through quality-aware RLVR. |
| Outcome: | The proposed framework outperforms existing benchmarks on math, coding, logic and puzzles. |
HEALing Entropy Collapse: Enhancing Exploration in Few-Shot RLVR via Hybrid-Domain Entropy Dynamics Alignment (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for training reasoning-oriented large language models assume high-resource settings with abundant data. |
| Approach: | They propose a framework that integrates high-value general-domain data to promote more diverse exploration. |
| Outcome: | The proposed framework matches or surpasses RLVR trained with 32 target-domain samples using 32 target domain samples. |
QuantumQA: Enhancing Scientific Reasoning via Physics-Consistent Dataset and Verification-Aware Reinforcement Learning (2026.acl-long)
Copied to clipboard
Songxin Qu, Tai-Ping Sun, Yun-Jie Wang, Huan-Yu Liu, Cheng Xue, Xiao-Fan Xu, Han Fang, Yang Yang, Yu-Chun Wu, Guo-Ping Guo, Zhao-Yun Chen
| Challenge: | Large language models lack reliability in scientific domains that require strict adherence to physical constraints. |
| Approach: | They propose a large-scale dataset constructed via a task-adaptive strategy and a hybrid verification protocol that combines deterministic solvers with semantic auditing to guarantee scientific rigor. |
| Outcome: | The proposed model outperforms baselines and general-purpose preference models and is competitive with proprietary models. |
VANE: Guiding High-Value Exploration in RLVR via Outcome-Process Novelty Shaping (2026.findings-acl)
Copied to clipboard
| Challenge: | Extensive experiments on large-scale mathematical reasoning and out-of-distribution tasks demonstrate the effectiveness and generalization of the proposed method. |
| Approach: | They propose a method that quantifies novelty across the outcome space and semantic process space by using reward or solution divergence. |
| Outcome: | Experiments on Qwen2.5-Math-7B demonstrate the proposed method is general and efficient. |
Rethinking Entropy Interventions in RLVR: An Entropy Change Perspective (2026.acl-long)
Copied to clipboard
Zhezheng Hao, Hong Wang, Haoyang Liu, Jian Luo, Jiarui Yu, Hande Dong, Qiang Lin, Can Wang, Jiawei Chen
| Challenge: | Large Language Models (LLMs) have remarkable reasoning capabilities in complex tasks such as mathematics and coding. |
| Approach: | They propose an entropy-modulation method that adaptively reweighs tokens based on theoretically-estimated entropic variations. |
| Outcome: | The proposed method outperforms state-of-the-art methods in six mathematical reasoning and three coding benchmarks. |
RubricHub: A Comprehensive and Highly Discriminative Rubric Dataset via Automated Coarse-to-Fine Generation (2026.acl-long)
Copied to clipboard
Sunzhu Li, Jiale Zhao, Huimin Ren, Zhenlin Wei, Yang Zhou, Jingwen Yang, Shunyu Liu, Kaike Zhang, Chen Wei
| Challenge: | Existing methods for generating open-ended rubrics suffer from scalability bottlenecks and coarse criteria resulting in a supervision ceiling effect. |
| Approach: | They propose a framework for automated Coarse-to-Fine Rubric Generation . their framework uses principle-guided synthesis, multi-model aggregation, difficulty evolution . |
| Outcome: | The proposed framework produces comprehensive and highly discriminative criteria capable of capturing the subtle nuances. |
Disentangling Reasoning Logic to Resolve Explicit Knowledge Conflicts (2026.acl-long)
Copied to clipboard
| Challenge: | Existing approaches to resolve explicit knowledge conflicts are based on semantic decoding and auxiliary embedding. |
| Approach: | They propose a framework that adjudicates conflicts by structuring the underlying logic. |
| Outcome: | Experiments show that the proposed framework improves on existing models. |
WIST: Web-Grounded Iterative Self-Play Tree for Domain-Targeted Reasoning Improvement (2026.acl-long)
Copied to clipboard
| Challenge: | Existing methods for self-improvement of large language models with verifiable rewards (RLVR) can drift over iterations, while corpus-grounded approaches rely on curated data environments. |
| Approach: | They propose a Web-grounded Iterative Self-play Tree framework for domain-targeted reasoning improvement that learns directly from the open-web without requiring any pre-arranged domain corpus. |
| Outcome: | The proposed framework outperforms both purely endogenous self-evolution and corpus-grounded self-play baselines and is domain-steerable. |
Backdoors in RLVR: Jailbreak Backdoors in LLMs From Verifiable Reward (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) is an emerging paradigm that significantly boosts a Large Language Model’s reasoning abilities on complex logical tasks. |
| Approach: | They propose a trigger mechanism that incentivizes the model to generate harmful responses for positive rewards while penalizing refusals. |
| Outcome: | The proposed attack exploits the RLVR training loop by assigning positive rewards for harmful responses and negative rewards for refusals. |
Good Reasoning Makes Good Demonstrations: Implicit Reasoning Quality Supervision via In-Context Reinforcement Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in large language models but treats all correct solutions equally, potentially reinforcing flawed traces that arrive at correct answers by chance. |
| Approach: | They propose a method that reweights rewards by a factor approximately proportional to Evidence Gain and assigns higher weights to high-quality traces without requiring costly computation. |
| Outcome: | Experiments on mathematical reasoning benchmarks show that Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in large language models but treats all correct solutions equally. |
DARL: Encouraging Diverse Answers for General Reasoning without Verifiers (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent efforts such as RLPR have extended RLVR to general domains, enabling training on broader datasets and achieving improvements over RL PR. |
| Approach: | They propose a framework that encourages the generation of diverse answers within a controlled deviation range from the reference while preserving alignment with it. |
| Outcome: | Extensive experiments on 13 benchmarks show that DARL surpasses RLPR in both reasoning accuracy and output diversity. |
Mitigating Lost in Multi-turn Conversation via Curriculum RL with Verifiable Accuracy and Abstention Rewards (2026.acl-long)
Copied to clipboard
Ming Li, Pei Chen, Zhenhao Zhang, Tao Yang, Xinyang Zhang, Han Li, Tianyu Cao, Ming Zeng, Zhuofeng Wu, Meng Jiang, Huasheng Li, Lihong Li, Bing Yin
| Challenge: | Large Language Models exhibit strong capabilities in single-turn instruction following but suffer from Lost-in-Conversation (LiC) when instructions are revealed progressively in multi-turn settings, models get "Lost in Conversation" |
| Approach: | They propose a framework that encourages models to generate correct answers and judge solvability in multi-turn conversations. |
| Outcome: | The proposed framework improves models' ability to balance problem-solving with abstention . it reduces premature answering behaviors that cause lost-in-conversation (LiC) |
VerIF: Verification Engineering for Reinforcement Learning in Instruction Following (2025.emnlp-main)
Copied to clipboard
| Challenge: | Best practices for RL in instruction following remain underexplored. |
| Approach: | They propose a verification method that combines rule-based code verification with LLM-based verification from a large reasoning model. |
| Outcome: | The proposed method achieves state-of-the-art performance among models of comparable size and generalizes well to unseen constraints. |
R1-RE: Cross-Domain Relation Extraction with RLVR (2026.acl-long)
Copied to clipboard
| Challenge: | Relation extraction (RE) is a core task in natural language processing. |
| Approach: | They propose a supervised learning task for relation extraction (RE) based on annotation guidelines. |
| Outcome: | The proposed model achieves an average OOD accuracy of 70%, on par with leading proprietary models such as GPT-4o. |
From Individual to Common: An Early Exploration of Consensus in Non-verifiable Data for Balanced Preference Optimization (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated remarkable effectiveness in boosting the objective performance of Large Language Models (LLMs). |
| Approach: | They propose a dataset where response pairs differ only by subtle nuances and a model with a non-verifiable dataset. |
| Outcome: | The proposed model outperforms models trained on data with explicit quality gaps while maintaining objective capabilities. |
Verifier-Free RL for LLMs via Intrinsic Gradient-Norm Reward (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable capabilities in reasoning tasks and general tasks. |
| Approach: | They propose a "Verifier-free Intrinsic Gradient-Norm Reward" that uses only the policy model itself. |
| Outcome: | The proposed reward outperforms the state-of-the-art RLIF baseline INTUITOR on math benchmarks and shows cross-domain transfer to code benchmarks when trained only on math data. |
Beyond Outcome Verification: Verifiable Process Reward Models for Structured Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work on reinforcement learning with verifiable rewards (RLVR) has shown that large language models can be substantially improved using outcome-level verification signals. |
| Approach: | They propose a framework where intermediate reasoning steps are checked by deterministic, rule-based verifiers. |
| Outcome: | The proposed framework achieves 20% higher F1 than state-of-the-art models and 6.5% higher than verifiable outcome rewards, with substantial gains in evidence grounding and logical coherence. |
TTVS: Boosting Self-Exploring Reinforcement Learning via Test-time Variational Synthesis (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing test-time methods are limited in specialized or novel domains where supervision is prohibitively expensive or unavailable. |
| Approach: | They propose a framework that augments training stream from unlabeled test queries. |
| Outcome: | Extensive experiments show TTVS outperforms state-of-the-art RL-based techniques on unlabeled test-time data. |
Beyond High-Entropy Exploration: Correctness-Aware Low-Entropy Segment-Based Advantage Shaping for Reasoning LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent work studies RLVR through token entropy, arguing that high-entropies drive exploration and should receive stronger updates. |
| Approach: | They propose a correctness-aware reinforcement framework that performs fine-grained advantage modulation over low-entropy segments. |
| Outcome: | The proposed framework improves accuracy over strong RL baselines across three backbones and six math benchmarks while maintaining high-entropy exploration. |
The Impact of Language Mixing on Bilingual LLM Reasoning (2025.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies show multilingual speakers intentionally switch languages during reasoning . enforcing monolingual decoding reduces accuracy by 5.6 percentage points . |
| Approach: | They find that multilingual speakers intentionally switch languages during reasoning . enforcing monolingual decoding reduces accuracy by 5.6 percentage points . authors suggest that language mixing is not merely a byproduct of multilingual training . |
| Outcome: | The proposed model can be used to predict whether a language switch would benefit or harm reasoning. |
Graph Reasoning Paradigm: Structured and Symbolic Reasoning with Topology-Aware Reinforcement Learning for Large Language Models (2026.acl-long)
Copied to clipboard
Runxuan Liu, Xianhao Ou, Xinyan Ma, Jiyuan Wang, Jiafeng Liang, Jiaqi Li, Tao He, Zheng Chu, Rongchuan Mu, Zekun Wang, Baoxin Wang, Dayong Wu, Ming Liu, Shijin Wang, Guoping Hu, Bing Qin
| Challenge: | Existing methods for long chain-of-thought (LCoT) are coarse-grained, reward hacking, and poor generalization. |
| Approach: | They propose a Long Chain-of-Thought (LCoT) model that integrates reinforcement learning with verifiable rewards with a process-aware verification approach. |
| Outcome: | The proposed model improves reasoning and code generation tasks while reducing the cost of training and performance bottlenecks. |
Turning Failures into Value: Negative Experience Replay for RLVR via Confidence Gating and Boundary Failure Sampling (2026.acl-long)
Copied to clipboard
| Challenge: | Existing experience replay methods for RLVR ignore sample inefficiency . expensive reasoning trajectories are discarded immediately after a single gradient update . |
| Approach: | They propose a method to replay failure trajectories to improve model refinement . they propose 'nexGRPO' which employs mid-confidence gating to filter invalid noise and saturated errors. |
| Outcome: | The proposed model outperforms strong baaselines and achieves improved out-of-distribution generalization. |
Beyond Reasoning Gains: Mitigating General-Capability Forgetting in Large Reasoning Models (2026.findings-acl)
Copied to clipboard
Hoang Phan, Xianjun Yang, Yuanshun Yao, Jingyu Zhang, Shengjie Bi, Xiaocheng Tang, Madian Khabsa, Lijuan Liu, Deren Lei
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) has delivered impressive gains in mathematical and multimodal reasoning . however, the recipe introduces a significant risk of capability regression, where models forget foundational skills after prolonged training without employing regularization strategies. |
| Approach: | They propose a replay strategy with dynamic objective reweighting for general knowledge preservation using short-horizon signals of convergence and instability. |
| Outcome: | The proposed method preserves general capabilities and improves reasoning . it can be applied to existing RLVR pipelines without training additional models or tuning . |
AG-GRPO: Answer-Guided GRPO for Masked Diffusion Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent work on large language models (LLMs) has emphasized not only final-answer accuracy but also reliability of reasoning on challenging tasks. |
| Approach: | They propose an answer-guided group-relative policy optimization for masked diffusion language models which generates text through iterative mangled token restoration. |
| Outcome: | The proposed approach improves over pretrained dLLMs and prior RL methods across mathematics, puzzle-solving, and code-generation benchmarks. |
Beyond Majority Voting: Towards Fine-grained and More Reliable Reward Signal for Test-Time Reinforcement Learning (2026.acl-long)
Copied to clipboard
| Challenge: | RLVR is a paradigm for improving reasoning ability of large language models . but voting results often induce confirmation bias and suffer from sparse rewards . |
| Approach: | They propose a framework integrating model confidence and dynamic subgroup partitioning to address these issues. |
| Outcome: | The proposed framework outperforms recent baselines on multiple models and benchmarks. |
BoundRL: Efficient Token-level Structured Text Segmentation through Reinforced Boundary Generation (2026.findings-acl)
Copied to clipboard
Haoyuan Li, Zhengyuan Shen, Sullam Jeoung, Yueyan Chen, Jiayu Li, Qi Zhu, Shuai Wang, Vassilis N. Ioannidis, Huzefa Rangwala
| Challenge: | Structured texts often contain elements beyond plain language, such as code snippets, which conventional sentence-level segmentation methods cannot handle effectively. |
| Approach: | They propose a token-level approach that performs efficient token-based text segmentation and label prediction for long structured texts. |
| Outcome: | The proposed approach outperforms existing models on short-shot prompts and SFT and standard RLVR models on complex LLM prompts. |
Better LLM Reasoning via Dual-Play (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have made remarkable progress through Reinforcement Learning with Verifiable Rewards (RLVR) however, external supervision remains a bottleneck for tasks and domains for which supervised data are scarce or non-existent. |
| Approach: | They propose a novel dual-play framework that adversarially trains two models initialized from the same base model. |
| Outcome: | The proposed framework improves the math reasoning performance of large language models. |
StepHint: Multi-level Stepwise Hints Enhance Reinforcement Learning to Reason (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) approaches face two challenges: the near-miss reward problem and exploration stagnation. |
| Approach: | They propose an algorithm that partitions valid reasoning chains into reasoning steps using multi-level stepwise hints. |
| Outcome: | The proposed method outperforms competing RLVR enhancement methods across six mathematical benchmarks and two out-of-domain benchmarks. |
AttnPO: Attention-Guided Process Supervision for Efficient Reasoning (2026.acl-long)
Copied to clipboard
Shuaiyi Nie, null Dingsiyu, Wenyuan Zhang, Linhao Yu, Tianmeng Yang, Yao Chen, Weichong Yin, Yu Sun, Hua Wu, Tingwen Liu
| Challenge: | Existing trajectory-level length penalties fail to effectively shorten reasoning length and degrade accuracy, as they treat all reasoning steps uniformly and lack fine-grained signals to distinguish redundancy from necessity. |
| Approach: | They propose a low-overhead process-supervised RL framework that leverages the model’s intrinsic attention signals for step-level credit assignment. |
| Outcome: | The proposed framework reduces reasoning length while improving performance across 9 benchmarks. |
Orchestrating Tokens and Sequences: Dynamic Hybrid Policy Optimization for RLVR (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing RLVR algorithms focus on different granularities and have complementary strengths and limitations. |
| Approach: | They propose a framework for reinforcement learning with verifiable rewards that bridges RLVR and GSPO . group-level importance ratios are used to update a policy, which preserves fine-grained credit assignment . |
| Outcome: | The proposed framework outperforms existing methods on seven reasoning benchmarks. |
Knowledge-to-Verification: Exploring RLVR for LLMs in Knowledge-Intensive Domains (2026.acl-long)
Copied to clipboard
Zhonghang Yuan, Zhefan Wang, Fang Hu, Zihong Chen, Jinzhe Li, Gang Li, Jie Ying, Huanjun Kong, Songyang Zhang, Nanqing Dong
| Challenge: | Recent large language models (LLMs) have demonstrated remarkable progress in reasoning, but their applications on knowledge-intensive domains have not been explored due to the scarcity of high-quality verifiable data. |
| Approach: | They propose a framework that extends reinforcement learning with verifiable rewards (RLVR) to knowledge-intensive domains through automated verififiability data synthesis while enabling verification of the LLM's reasoning process. |
| Outcome: | Extensive experiments show that the proposed framework enhances the reasoning of large language models in knowledge-intensive domains without significantly compromising the model’s general capabilities. |
Semantic-Space Exploration and Exploitation in RLVR for LLM Reasoning (2026.findings-acl)
Copied to clipboard
Fanding Huang, Guanbo Huang, Xiao Fan, Yi He, Xiao Liang, Xiao Chen, Qinting Jiang, Faisal Nadeem Khan, Jingyan Jiang, Zhi Wang
| Challenge: | Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have substantially improved the reasoning abilities of Large Language Models (LLMs). |
| Approach: | They propose a method that balances exploration and exploitation in the hidden-state space of response trajectories. |
| Outcome: | The proposed model yields consistent improvements across models, algorithms and reasoning benchmarks. |
MASPO: Unifying Gradient Utilization, Probability Mass, and Signal Reliability for Robust and Sample-Efficient LLM Reasoning (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Binbin Zheng, Chaowen Hu, Zekai Shao, Cong Qin, Lu Pan, Ke Zeng, Xunliang Cai
| Challenge: | Existing RLVR algorithms rely on rigid, uniform, and symmetric trust region mechanisms . current algorithms lack robustness, asymmetric signal reliability and inefficient gradient utilization . |
| Approach: | They propose a framework to harmonize three dimensions of RLVR algorithms, a paper argues . a binary cutoff is used to discard valuable reinforcement signals, they argue . |
| Outcome: | The proposed framework outperforms baselines in evaluating a robust RLVR solution. |
Teaching Language Models to Forecast Research Success Through Comparative Idea Evaluation (2026.findings-acl)
Copied to clipboard
| Challenge: | Language models are accelerating scientific research by automating hypothesis generation and implementation. |
| Approach: | They ask whether LMs can forecast the empirical success of research ideas before experiments . they frame evaluation as a reasoning task via Reinforcement Learning with Verifiable Rewards . |
| Outcome: | The proposed model outperforms off-the-shelf models in 77.1% of the evaluations . the model outpersforms GPT-5 in the evaluation of 11,488 idea pairs . |
From log 𝜋 to 𝜋: Taming Divergence in Soft Clipping via Bilateral Decoupled Decay of Probability Gradient Weight (2026.acl-long)
Copied to clipboard
Xiaoliang Fu, Jiaye Lin, Yangyi Fang, Chaowen Hu, Cong Qin, Zekai Shao, Binbin Zheng, Lu Pan, Ke Zeng
| Challenge: | Standard algorithms for Large Language Models (LLMs) enforce stability via "hard clipping" but relying on log-probability gradient yields divergent weights as probabilities vanish, destabilizing LLM training. |
| Approach: | They propose a decoupled gradient policy optimization that uses a decay mechanism to decouple the probability of a boundary token. |
| Outcome: | The proposed algorithm outperforms baselines on various mathematical benchmarks. |
RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization (2026.acl-long)
Copied to clipboard
Yihong Dong, Xue Jiang, Yongding Tao, Huanyu Liu, Kechi Zhang, Lili Mou, Rongyu Cao, Yingwei MA, Jue Chen, Binhua Li, Zhi Jin, Fei Huang, Yongbin Li, Ge Li
| Challenge: | Reinforcement Learning with Verifiable Reward (RLVR) has significantly advanced the complex reasoning abilities of Large Language Models (LLMs). |
| Approach: | They propose a hybrid-policy optimization approach that synergizes internal exploitation with external data to achieve stronger reasoning capabilities. |
| Outcome: | The proposed approach achieves state-of-the-art performance on six math reasoning benchmarks and superior performance on out-of distribution reasoning tasks. |
Entropy-Aware Reshaping of Reinforcement Signals for Multi-Answer Reasoning (2026.findings-acl)
Copied to clipboard
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) is a standard post-training paradigm for large language models. |
| Approach: | They propose a framework that reshapes how learning signals are normalized and aggregated. |
| Outcome: | Experiments on MCTACO and MMLU-Multi show that the proposed framework improves accuracy, training stability and cross-dataset transfer performance. |
LearnAlign: Data Selection for LLM Reinforcement Learning with Improved Gradient Alignment (2026.findings-acl)
Copied to clipboard
Shipeng Li, Zhiqin Yang, Shikun Li, Xiaobo Xia, Hengyu Liu, Xinghua Zhang, Gaode Chen, Dong Fang, Ying Tai, Zhe Peng
| Challenge: | Reinforcement learning with verifiable rewards (RLVR) is a key technique for enhancing LLMs’ reasoning abilities, yet its data inefficiency remains a major bottleneck. |
| Approach: | They propose a gradient-alignment-based method which intelligently selects the learnable and representative training reasoning data for RLVR post-training. |
| Outcome: | Experiments on five reasoning benchmarks show that the proposed method significantly reduces training data requirements while improving performance. |
AIPO: Adaptive Information Guided Token-Level Reinforcement Learning for Large Language Model Reasoning (2026.acl-long)
Copied to clipboard
| Challenge: | Existing RLVR methods focus on all generated tokens rather than on which tokens contribute to reasoning. |
| Approach: | They propose to use a Random–Fourier approximation of the Hilbert–Schmidt Independence Criterion to focus updates on decisive tokens discovered on the fly to improve the efficiency of mutual-information estimation. |
| Outcome: | The proposed approach yields +20% accuracy over strong RLVR baselines while updating merely 10% of tokens, demonstrating superior efficiency and effectiveness. |
Can Small LLMs Learn a Robust Theory of Mind via RLVR? Investigating Generalization through the False-Belief Task (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent advances in large language models (LLMs) have demonstrated emergent capabilities in complex reasoning, largely spurred by rule-based Reinforcement Learning (RL) techniques applied during post-training. |
| Approach: | They evaluate whether small-scale LLMs can acquire a robust and generalizable Theory of Mind (ToM) capability through RL with verifiable rewards. |
| Outcome: | The proposed model performs well on in-distribution tasks but fails to transfer to unseen ToM tasks with different characteristics. |
Beyond Single-Shot: Multi-step Tool Retrieval via Query Planning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large language models (LLMs) are evolving from text generation into integration within agentic workflows . tools such as APIs, databases, and software tools are expanding rapidly . |
| Approach: | They propose a lightweight framework that models retrieval as iterative query planning . instead of single-shot matching, ToolQP decomposes instructions into sub-tasks . |
| Outcome: | The proposed framework achieves state-of-the-art performance and robustness across retrievers. |
Breaking the Impasse: Dual-Scale Evolutionary Policy Training for Social Language Agents (2026.acl-long)
Copied to clipboard
| Challenge: | Reinforcement Learning with Verifiable Rewards (RLVR) is effective for closed-ended tasks, but it is not applicable to open-ended social language games. |
| Approach: | They propose a method that uses a time-scaled evolutionary perception mechanism to detect impasse by quantifying dual-scale value baseline divergence alongside match entropy. |
| Outcome: | Experiments on multiple social language games show that the proposed method outperforms baselines and avoids policy degeneration. |
Does RLVR Extend Reasoning Boundaries? Investigating Capability Expansion in Vision-Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | Recent studies suggest that RLVR amplifies behaviors inherent to the pre-training distribution rather than inducing new capabilities. |
| Approach: | They propose a framework for RLVR that extends the spatial reasoning boundary . they use a mapping framework where the difficulty is precisely regulated by path length and number of turns . |
| Outcome: | The proposed framework extends the spatial reasoning boundary on two real-world navigation benchmarks. |
SynthRL: Scaling Visual Reasoning with Verifiable Data Synthesis (2026.findings-acl)
Copied to clipboard
| Challenge: | SynthRL synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. |
| Approach: | They propose a scalable and guaranteed pipeline for automatic data scaling in reasoning-oriented RL training. |
| Outcome: | The proposed pipeline synthesizes over 3.3K additional verifiable, challenging questions from approximately 8K seed samples. |
Data Efficient RLVR via Off-Policy Influence Guidance (2026.acl-long)
Copied to clipboard
Erle Zhu, Dazhi Jiang, Yuan Wang, Xujun Li, Jiale Cheng, Yuxian Gu, Yilin Niu, Aohan Zeng, Jie Tang, Minlie Huang, Hongning Wang
| Challenge: | Existing data selection methods for RLVR are heuristic-based, lacking theoretical guarantees and generalizability. |
| Approach: | They propose an off-policy influence estimation method that approximates data influence using offline trajectories. |
| Outcome: | The proposed method reduces the computational cost of policy rollouts and improves storage and computation efficiency. |
Reinforced Efficient Reasoning via Semantically Diverse Exploration (2026.acl-long)
Copied to clipboard
Ziqi Zhao, Zhaochun Ren, Jiahong Zou, Liu Yang, Zhiwei Xu, Xuri Ge, Zhumin Chen, Xinyu Ma, Daiting Shi, Shuaiqiang Wang, Dawei Yin, Xin Xin
| Challenge: | Existing methods for reinforcement learning with verifiable rewards suffer from limited exploration diversity and inefficient reasoning. |
| Approach: | They propose a method that rewards concise and correct reasoning while penalizing unnecessarily long reasoning chains. |
| Outcome: | Extensive experiments on Qwen and Llama models validate the effectiveness and efficiency of ROSE. |