REVIEW 8 cited by
On Designing Effective RL Reward at Training Time for LLM Reasoning
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
Reward models have been increasingly critical for improving the reasoning capability of LLMs. Existing research has shown that a well-trained reward model can substantially improve model performances at inference time via search. However, the potential of reward models during RL training time still remains largely under-explored. It is currently unclear whether these reward models can provide additional training signals to enhance the reasoning capabilities of LLMs in RL training that uses sparse success rewards, which verify the correctness of solutions. In this work, we evaluate popular reward models for RL training, including the Outcome-supervised Reward Model (ORM) and the Process-supervised Reward Model (PRM), and train a collection of LLMs for math problems using RL by combining these learned rewards with success rewards. Surprisingly, even though these learned reward models have strong inference-time performances, they may NOT help or even hurt RL training, producing worse performances than LLMs trained with the success reward only. Our analysis reveals that an LLM can receive high rewards from some of these reward models by repeating correct but unnecessary reasoning steps, leading to a severe reward hacking issue. Therefore, we introduce two novel reward refinement techniques, including Clipping and Delta. The key idea is to ensure the accumulative reward of any reasoning trajectory is upper-bounded to keep a learned reward model effective without being exploited. We evaluate our techniques with multiple reward models over a set of 1.5B and 7B LLMs on MATH and GSM8K benchmarks and demonstrate that with a carefully designed reward function, RL training without any additional supervised tuning can improve all the evaluated LLMs, including the state-of-the-art 7B LLM Qwen2.5-Math-7B-Instruct on MATH and GSM8K benchmarks.
Forward citations
Cited by 8 Pith papers
-
On the Emergence of Implicit Curriculum in RLVR Learning Dynamics
RLVR training on transformers for compositional tasks follows an implicit curriculum from easy to hard problems, with difficulty spectrum smoothness determining steady relay progress or grokking phase transitions.
-
TraversRL: Traversable Pedestrian Pathway Generation With Reinforcement Learning
An iterative reinforcement-learning model that traces pedestrian networks step-by-step from aerial images beats a segmentation baseline on overlap and connectivity.
-
Reward-Aware Population Scaling of Evolutionary Strategies in LLM Fine-Tuning
Binary-reward ES collapses at N=2 because of z-score advantage normalization, not because two perturbation pairs are intrinsically too few.
-
TimeRFT: Stimulating Generalizable Time Series Forecasting for TSFMs via Reinforcement Finetuning
TimeRFT fine-tunes time-series foundation models with step-wise reward signals and difficulty-filtered data, beating supervised fine-tuning on eight benchmarks across data regimes.
-
Encouraging Good Processes Without the Need for Good Answers: Reinforcement Learning for LLM Agent Planning
Decoupling LLM-agent planning from summarization and rewarding tool-call completeness rather than final-answer correctness improves planning by 8-12% and end-to-end answers by 5-6% over end-to-end RL baselines.
-
GM-PRM: A Generative Multimodal Process Reward Model for Multimodal Mathematical Reasoning
A generative multimodal process reward model that produces step-level critiques and corrections improves average math accuracy for six multimodal LLMs by 2.9 to 5.9 points under a refinement-based Best-of-N strategy.
-
EMO-R3: Reflective Reinforcement Learning for Emotional Reasoning in Multimodal Large Language Models
EMO-R3, which combines a three-step emotional reasoning prompt with a reward for the model agreeing with its own image–emotion judgments, raises visual emotion-recognition accuracy by about one point over plain GRPO.
-
Enhancing Large Language Models through Structured Reasoning
Structured reasoning tags plus a max-flow reward let a 1.5B model match the math accuracy of models trained for far longer, but the gains are within statistical noise.
Discussion (0). Sign in to comment.