arXiv Machine Learning

DataFlex-RL: An Evaluation Platform for RLVR Data Policies

arXiv AI
6d ago

Are Dense Rewards Worth Their Tokens in GRPO?

The paper evaluates whether dense, token‑level rewards improve training of language models compared to the standard group‑relative policy optimization (GRPO) algorithm. Using a matched‑compute protocol that equalizes generated tokens and GPU‑hours across interventions, the authors find that dense rewards—such as Monte‑Carlo step‑wise scoring—do not outperform GRPO, and in some cases perform worse even when matched for compute. Instead, a simple strategy of training on prompts predicted to be near the 50% success threshold yields consistent accuracy gains across multiple models and settings.

By Vikash Singh, Weicong Chen, Debargha Ganguly, Yanyan Zhang, Mohsen Hariri, Sreehari Sankar, Wang Yang, Chaoda Song, Alexander Nemecek, Osama Zafar, Yuting Shao, Kai Ye, Biyao Zhang, Xinpeng Li, Shouren Wang, Erman Ayday, Jing Ma, Sanmukh Rao Kuppannagari, Vipin Chaudhary
arXiv AI
Oct 31

SCARL: Sequence Confidence-Aligned Reinforcement Learning

SCARL (Sequence Confidence-Aligned Reinforcement Learning) addresses a misalignment between a reinforcement learning policy’s confidence and its reward signal in language-model reasoning. By reshaping each advantage with an alignment gap that reflects the difference between a response’s group-relative confidence and the outcome it certifies, SCARL drives responses with equal advantage toward equal confidence and mitigates entropy collapse. Experiments on Qwen3‑4B‑Base show a 2.4‑point lift in Pass@32 over DAPO and consistent improvements across multiple estimators and backbones while preserving policy entropy.

By Zhenghong Ke, Jingbo Zhou, Hua Wu
arXiv AI
Nov 6

Measuring the Marginal Value of Verifier-Based RL Beyond Supervised Fine-Tuning

The paper evaluates the true benefit of verifier‑based reinforcement learning (RLVR) over supervised fine‑tuning (SFT) using a deterministic benchmark of 1,877 synthetic GPU and LLM training failure diagnosis tasks. In controlled experiments, SFT alone achieved high performance (0.785 on the hard split), while RLVR added no detectable improvement in‑distribution and only a modest gain under distribution shift (+0.33 points). The study concludes that most performance gains come from SFT, with verifier‑based RL contributing only small additional benefits.

By Amit Sagtani, Saloni Garg
Hugging Face Trending Papers
Jul 30

LoRA Scaffolded Policy Optimization (LSPO): A Sampling-Time Low-Rank Scaffold for Recovering Reinforcement-Learning Gradient on Zero-Reward Cliff Prompts

Reinforcement learning from verifiable rewards (RLVR) for mathematical reasoning suffers from a structural blind spot: on "cliff" prompts-those on which every sampled rollout in a group fails-the group-normalized advantage is identically zero, so GRPO produces no gradient on precisely the prompts at the frontier of the model's capability. We introduce LoRA Scaffolded Policy Optimization (LSPO), a sampling-time mechanism that recovers this lost gradient.

arXiv Machine Learning
Jul 31

LoRA Scaffolded Policy Optimization (LSPO): A Sampling-Time Low-Rank Scaffold for Recovering Reinforcement-Learning Gradient on Zero-Reward Cliff Prompts

arXiv:2607. 27787v1 Announce Type: new Abstract: Reinforcement learning from verifiable rewards (RLVR) for mathematical reasoning suffers from a structural blind spot: on "cliff" prompts-those on which every sampled rollout in a group fails-the group-normalized advantage is identically zero, so GRPO produces no gradient on precisely the prompts at the frontier of the model's capability.

By Ken Ding
arXiv AI
Oct 31

DUET: Optimize Token-Budget Allocation for Reinforcement Learning with Verifiable Rewards

DUET is a token‑budget allocation method for reinforcement learning with verifiable rewards that jointly optimizes which prompts receive rollouts and how long each rollout runs. By using a lightweight surrogate to assign rollouts and a marker‑gated abort rule to stop them, DUET improves reasoning quality and reduces training time compared to full‑budget GRPO and other baseline methods. Experiments on Qwen3‑1.7B and other LLMs show that DUET outperforms baselines even with only 50% of the token budget, achieving up to 2.51× speedup while maintaining or improving performance.

By Haoyu Hu, Xuandong Zhao, Xuhai "Orson'' Xu, Nori Jacoby
arXiv Machine Learning
Oct 12

Cross-Behavior Policy Optimization: Unified Learning from Offline Demonstrations and On- and Off-Policy Rollouts via a Shared Regression Target

The paper introduces Cross-Behavior Policy Optimization (CBPO), a method that unifies offline demonstrations and policy-generated rollouts into a single regression loss using a fixed-reference target. CBPO eliminates the need for separate supervised fine‑tuning and reinforcement learning objectives, allowing positive and negative offline completions, reference responses, and current and historical rollouts to share a common reward‑labeled format and replay buffer. Experiments on two model scales and four mathematical reasoning benchmarks show that CBPO, especially the CBPO‑Unified variant, outperforms baseline methods in avg@32 and pass@32 metrics, with historical rollouts providing additional gains.

By Du Cheng, Junyu Lu, Jingfan Zhang, Yuzhuo Zhang, Haoxuan Jia, Shichao Weng
arXiv Machine Learning
Aug 11

CODS: Iterative Bellman-Residual Data Selection for Reusable Offline Reinforcement Learning

arXiv:2608. 07719v1 Announce Type: new Abstract: Offline reinforcement learning repeatedly trains policies from a fixed transition pool, making redundant data costly across seeds and hyperparameters, while naive subsampling can remove rare transitions needed for long-horizon credit assignment.

By Ibne Farabi Shihab, Sanjeda Akter, Abu Sa-Adat Mohamed Moon-Im Al Ahsan, Md Najmus Swaqeeb, Anuj Sharma
arXiv Machine Learning
Jun 26

Reinforcement Learning without Ground-Truth Solutions can Improve LLMs

arXiv:2606. 27369v1 Announce Type: new Abstract: Reinforcement learning with verifiable rewards (RLVR) for training LLMs typically rely on ground-truth answers to assign rewards, limiting their applicability to tasks where the ground-truth solution is unknown.

By Yingyu Lin, Qiyue Gao, Nikki Lijing Kuang, Xunpeng Huang, Kun Zhou, Tongtong Liang, Zhewei Yao, Yi-An Ma, Yuxiong He