@HuggingPapers: FlowBalance: verifier-grounded self-improvement for reasoning models Improves math reasoning by +2.12 avg over GRPO on …
Summary
FlowBalance introduces a verifier-grounded self-improvement technique that improves math reasoning performance by an average of 2.12 over GRPO on the Qwen3-8B model, offering faster training, enhanced stability, and greater solution diversity.
View Cached Full Text
Cached at: 09/09/26, 03:52 PM
FlowBalance: verifier-grounded self-improvement for reasoning models
Improves math reasoning by +2.12 avg over GRPO on Qwen3-8B, with faster training, better stability, and higher solution diversity. https://t.co/Ec5knYSXPz
Similar Articles
FlowBalance: Verifier-Grounded Self-Improvement from On-Policy Reasoning Experience
FlowBalance is a verifier-grounded self-improvement method that calibrates on-policy reasoning experience to enhance AI model performance and stability in mathematical reasoning tasks.
Flow Reasoning Models: Scaling Reasoning Through Iterative Self-Refinement
Flow Reasoning Models (FRMs) introduce a training and test-time-scaling framework for discrete flow models on structured reasoning tasks. By using self-verification and self-conditioning, FRMs achieve nearly 100% solve rates on Sudoku and Zebra puzzles with far fewer passes than previous baselines.
Efficient Reasoning with Balanced Thinking
This paper introduces ReBalance, a training-free, plug-and-play method that dynamically balances overthinking and underthinking in large reasoning models, improving efficiency and accuracy across multiple benchmarks.
Improving mathematical reasoning with process supervision
OpenAI demonstrates that process supervision—rewarding intermediate reasoning steps rather than just final answers—improves mathematical reasoning while reducing alignment costs. This approach produces more interpretable, human-aligned reasoning without sacrificing model performance.
Questioning the Questions: Sustaining Self-Evolution in Reasoning Models
The paper identifies why self-evolving reasoning models suffer performance collapse — invalid questions and mathematically-equivalent question repetition — and proposes R-Quest, which uses validity and novelty feedback to sustain self-evolution, outperforming R-Zero by 17.32 points across 12 benchmarks over ten rounds.