Research map / Reinforcement learning / RL from feedback
No Prompt Left Behind: Exploiting Zero-Variance Prompts in LLM Reinforcement Learning via Entropy-Guided Advantage Shaping
Thanh-Long V. Le, Myeongho Jeon, Kim Vu, Viet Dac Lai, Eunho Yang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add RL-ZVP entropy-guided advantage for zero-variance prompts | GRPO · AIME25 (Qwen3-8B-Base, DAPO-Math-17k) | Acc@8: 20.00 -> 24.58 | Improved |
| add RL-ZVP advantage for zero-variance prompts at equal rollout budget vs dynamic-sampling filtering | GRPO-DS / GRESO · six math benchmarks, large scale | Acc@8: +19.20 Acc@8 avg over r baselines | Improved |
| use RL-ZVP instead of filtering baselines given 3-5x more rollouts (-g settings) | GRPO-DS-g / GRESO-g · six math benchmarks, large scale | Acc@8: +5.51 Acc@8 avg | Improved |
| add reward/penalty for all-incorrect (negative) zero-variance prompts (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 47.75 -> 49.90 | Improved |
| add reward for all-correct (positive) zero-variance prompts (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 47.25 -> 49.90 | Improved |
| add token-entropy scaling of advantage magnitude (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 46.88 -> 49.90 | Improved |
| set alpha to 0.10 instead of 0.05 or 0.30 (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 46.78 (0.05), 45.81 (0.30) -> 49.90 (0.10) | Improved |
| increase alpha to 1.0 and 10.0 (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 49.90 -> 44.32 (1.0), 39.96 (10.0) | Hurt |
| use symmetric (sign-flipped entropy) advantage for negative prompts (ablation) | RL-ZVP · Minerva/AMC23/MATH500 average | Acc@8: 49.90 -> 46.64 | Hurt |
| train with RL-ZVP on NuminaMath-CoT subset | GRPO · Qwen3-1.7B-Base NuminaMath-CoT, six benchmarks | Acc@8/Pass@8: Pass@8 AMC23 -1.39, AIME24 -0.27; other gains up to +4.27 | Mixed |
| apply RL-ZVP to Llama-3.2-3B-Instruct | GRPO · Llama-3.2-3B-Instruct trained on MATH | Acc@8: +1.20 Minerva, +2.50 AMC23, +2.50 AIME24; 0.00 AIME25 | Improved |
| apply RL-ZVP trained on MATH to out-of-domain tasks | GRPO · LiveCodeBench/MMLU-Pro/GPQA-Diamond/BBH average | Acc@8: 27.84 -> 30.19 | Improved |
Closest papers by meaning
- The Surprising Effectiveness of Negative Reinforcement in LLM Reasoning (NeurIPS 2025 · similarity 0.90)
- Query-Dependent Prompt Evaluation and Optimization with Offline Inverse RL (ICLR 2024 · similarity 0.89)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.89)
- Prompted Policy Search: Reinforcement Learning through Linguistic and Numerical Reasoning in LLMs (NeurIPS 2025 · similarity 0.89)
- Absolute Zero: Reinforced Self-play Reasoning with Zero Data (NeurIPS 2025 · similarity 0.89)
- LLM-as-a-Tutor: Policy-Aware Prompt Adaptation for Non-Verifiable RL (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
