Research map / Reinforcement learning / RL from feedback
ECHO: Entropy-Confidence Hybrid Optimization for Test-Time Reinforcement Learning
Chu Zhao, Enneng Yang, Yuting Liu, Jianzhe Zhao, Guibing Guo
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 EC-Tree entropy-confidence tree rollout (ablation) | ECHO · AIME2025 (Qwen2.5-7B-Base, AIME2024 training) | pass@16: 29.7 -> 54.3 | Improved |
| add EC-Tree entropy-confidence tree rollout (ablation) | ECHO · AMC (Qwen2.5-7B-Base, AIME2024 training) | pass@16: 54.2 -> 71.1 | Improved |
| add CA-Clip confidence-adaptive clipping (ablation) | ECHO · AIME2025 (Qwen2.5-7B-Base, AIME2024 training) | pass@16: 30.1 -> 54.3 | Improved |
| add E-SC Adv advantage shaping (ablation) | ECHO · AIME2025 (Qwen2.5-7B-Base, AIME2024 training) | pass@16: 53.5 -> 54.3 | Improved |
| add E-SC Adv advantage shaping (ablation) | ECHO · AMC (Qwen2.5-7B-Base, AIME2024 training) | pass@16: 78.7 -> 71.1 | Hurt |
| use ECHO instead of INTUITOR | test-time RL · Qwen2.5-7B avg (AIME2024 training) | pass@16 avg: 51.6 -> 55.3 | Improved |
| use ECHO instead of INTUITOR | test-time RL · Qwen3-8B avg (AIME2024 training) | pass@16 avg: 73.9 -> 76.1 | Improved |
| use ECHO instead of MM-UPT | unsupervised multimodal RL · Qwen3-VL-8B avg (GeoQA training) | pass@1 avg: 58.1 -> 60.5 | Improved |
| raise pruning threshold from 0.4 to 1.2 (ablation) | ECHO · AIME2024 | Pass@16: 30.0 -> 16.7 | Hurt |
| lower pruning threshold from 0.4 to 0.2 (ablation) | ECHO · AIME2024 | Pass@16: 30.0 -> 16.7 | Hurt |
| increase tail window from 2 to 8 (ablation) | ECHO · AMC | Pass@16: 60.2 -> 75.9 | Improved |
| train on ReClor instead of math data (Qwen2.5-7B-Base) | ECHO · MATH-500 | pass@16: 83.0 -> 81.9 | Hurt |
Closest papers by meaning
- STARE: Surprisal-Guided Token-Level Advantage Reweighting for Policy Entropy Stability (NeurIPS 2026 · similarity 0.87)
- Efficient Test-Time Adaptation For Robot Policies (NeurIPS 2026 · similarity 0.87)
- TGRL: Temperature-Grouped Reinforcement Learning for Efficient Exploration in LLMs (NeurIPS 2026 · similarity 0.87)
- Test-time Adapted Reinforcement Learning with Action Entropy Regularization (ICML 2025 · similarity 0.87)
- Optimizing Test-Time Compute via Meta Reinforcement Finetuning (ICML 2025 · similarity 0.87)
- DIME: Diffusion-Based Maximum Entropy Reinforcement Learning (ICML 2025 · similarity 0.87)
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.
