MDP-GRPO: Stabilized Group Relative Policy Optimization for Multi-Constraint Instruction Following (2026.acl-long)
Copied to clipboard
| Challenge: | Large language models (LLMs) can follow many natural-language instructions, yet they remain brittle when a request bundles multiple explicit constraints, such as asking the LLM to respond in a particular structure with an exact ending phrase. |
| Approach: | They propose a method which stabilizes learning through multi-temperature sampling to increase reward dispersion, dual-anchor advantages to restore gradients in homogeneous groups, prospect-theoretic shaping to bound updates and penalize violations based on Kahneman Tversky’s theory and asymmetric KL regularization. |
| Outcome: | The proposed method outperforms standard GRPO on FollowBench, IFEval, and a curated multi-constraint dataset, improving strict constraint satisfaction by up to 5.0% on Llama-3.2-3B. |
Similar Papers
BiasGRPO: Stabilizing Bias Mitigation in High-Variance Reward Landscapes via Group-Relative Policy Optimization (2026.findings-acl)
Copied to clipboard
| Challenge: | Recent preference-based fine-tuning methods have limited exploration in offline training . previous methods have been limited by the lack of exploration inherent in offline learning . |
| Approach: | They propose a method that normalizes rewards across a group of completed tasks to mitigate social bias in Large Language Models. |
| Outcome: | The proposed approach outperforms DPO and PPO in multiple benchmarks . it can overcome limitations of previous preference-based methods . |
DRA-GRPO: Your GRPO Needs to Know Diverse Reasoning Paths for Mathematical Reasoning (2026.findings-acl)
Copied to clipboard
Xiwen Chen, Wenhui Zhu, Peijie Qiu, Xuanzhao Dong, Hao Wang, Haiyu Wu, Huayu Li, Aris Sotiras, Yalin Wang, Abolfazl Razi
| Challenge: | Existing methods for group-relative policy optimization rely on scalar correctness rewards that are often non-injective with respect to semantic content. |
| Approach: | They propose a framework that calibrates the reward signal using the semantic density of sampled groups. |
| Outcome: | The proposed framework outperforms strong baselines on five math benchmarks with 7,000 samples and 55 cost. |
Graph-GRPO: Stabilizing Multi-Agent Topology Learning via Group Relative Policy Optimization (2026.findings-acl)
Copied to clipboard
Yueyang Cang, Xiaoteng Zhang, Erlu Zhao, Zehua Ji, Yuhang Liu, Yuchen He, Zhiyuan Ning, Chen Yijun, Wenge Que, Li Shi
| Challenge: | Recent approaches to optimize communication topology rely on single-sample policy gradients with absolute rewards. |
| Approach: | They propose a topology optimization framework that integrates Group Relative Policy Optimization. |
| Outcome: | The proposed topology optimization framework outperforms state-of-the-art methods on reasoning and code generation benchmarks. |
Rewarding the Unlikely: Lifting GRPO Beyond Distribution Sharpening (2025.emnlp-main)
Copied to clipboard
| Challenge: | Reinforcement learning is emerging as a primary driver for improving language model reasoning capabilities. |
| Approach: | They propose a method for explicitly up-weighting rare but correct solutions to overcome rank bias in group relative policy optimization (GRPO) . |
| Outcome: | The proposed method mitigates rank bias and improves pass@N across a large range of N in both synthetic and real theorem proving settings. |
GRPO-LEAD: A Difficulty-Aware Reinforcement Learning Approach for Concise Mathematical Reasoning in Language Models (2025.emnlp-main)
Copied to clipboard
| Challenge: | Existing methods for group-relative policy optimization face challenges in reward sparsity, verbosity and inadequate focus on problem difficulty. |
| Approach: | They propose a method to improve group relative policy optimization with length-regularized rewards and explicit penalties for incorrect solutions. |
| Outcome: | The proposed method achieves state-of-the-art performance for 14B-scale models . it improves reasoning accuracy, conciseness, and efficiency . |
Too Correct to Learn: Reinforcement Learning on Saturated Reasoning Data (2026.acl-short)
Copied to clipboard
| Challenge: | Strong base models saturate benchmarks, resulting in weaker performance, a paradox . a new approach to Reinforcement Learning (RL) is needed to improve performance . |
| Approach: | They propose a method that uses constrained uniform top-k sampling to flatten the local optimization landscape by sampling uniformly from constrained high-confidence candidates. |
| Outcome: | Experiments show that the proposed approach prevents policy degeneration and boosts out-of-domain generalization. |
Auto-Weighted Group Relative Preference Optimization for Multi-Objective Text Generation Tasks (2025.emnlp-industry)
Copied to clipboard
| Challenge: | Failing to balance the objectives in advance can lead to overfitting or insufficient learning of each reward function. |
| Approach: | They propose a method that adjusts reward weights according to learning progress . they evaluate AW-GRPO on advertising text generation problem . |
| Outcome: | The proposed method outperforms GRPO on advertising text generation tasks . it overfits BLEURT and jReadability at the expense of BLeurT performance . |
On the Hidden Objective Biases of Group-based Reinforcement Learning (2026.acl-short)
Copied to clipboard
| Challenge: | Recent studies have reported unexpected behaviors during training, including lengthrelated biases, formatting tokens, and reward hacking in multi-objective settings. |
| Approach: | They propose to analyze group-based reinforcement learning methods within a unified surrogate formulation. |
| Outcome: | The proposed methods exhibit structural mismatches between reward optimization and the underlying training objective. |
Token-Level Policy Optimization: Linking Group-Level Rewards to Token-Level Aggregation via sequence-level likelihood (2026.acl-long)
Copied to clipboard
| Challenge: | Group Relative Policy Optimization (GRPO) has significantly advanced the reasoning ability of large language models (LLMs). |
| Approach: | They propose a token-level framework that leverages sequence-level likelihood to link group-level rewards with individual tokens via token- level aggregation and introduces a KL-Divergence mask constraint that targets tokens with positive advantages and decreasing entropy to mitigate abrupt policy updates. |
| Outcome: | Experiments show that TEPO achieves state-of-the-art performance on mathematical reasoning benchmarks and reduces convergence time by 50% compared with GRPO/DAPO. |
MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization (2026.acl-long)
Copied to clipboard
Yang Zhao, Hepeng Wang, Xiao Ding, Yangou Ouyang, Bibo Cai, Kai Xiong, Jinglong Gao, Zhouhao Sun, Li Du, Bing Qin, Ting Liu
| Challenge: | Group-Relative Policy Optimization (GRPO) has emerged as an efficient paradigm for aligning Large Language Models (LLMs), but its efficacy is confined to domains with verifiable ground truths. |
| Approach: | They propose a meta-cognitive orchestration layer that treats reward scalarization as a dynamic latent policy, leveraging the model’s terminal hidden states as 'a semantic bottleneck' . Across seven benchmarks, MAESTRO consistently outperforms single-reward and static multi-objective baselines while preserving the efficiency advantages of GRPO. |
| Outcome: | The proposed model outperforms single-reward and static multi-objective baselines while preserving efficiency advantages. |