Research map / Reinforcement learning / RL from feedback
ETS: Energy-Guided Test-Time Scaling for Training-Free RL Alignment
Xiuyu Li, Jinkai Zhang, Mingyang Yi, Yu Li, Longqiang Wang, Yue Wang, Ju Fan
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 |
|---|---|---|---|
| use ETS energy-guided sampling instead of Best-of-N | Qwen3-8B · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 66.31 -> 69.34 | Improved |
| use ETS instead of Best-of-N | Qwen3-1.7B · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 51.70 -> 54.76 | Improved |
| use ETS instead of Best-of-N | LLaDA-8B-Instruct · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 48.90 -> 51.01 | Improved |
| replace GRPO post-training with training-free ETS | Qwen3-8B · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 65.98 -> 69.34 | Improved |
| replace GRPO with training-free ETS | Qwen3-1.7B · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 48.08 -> 54.76 | Improved |
| replace VRPO with training-free ETS | LLaDA-8B-Instruct · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 46.40 -> 51.01 | Improved |
| use importance-sampling acceleration (ETS-IS) instead of full ETS | Qwen3-8B · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 69.34 -> 66.45 | Hurt |
| use ETS-IS instead of full ETS | LLaDA-8B-Instruct · Avg (MATH500, GSM8K, GPQA, HumanEval) | accuracy: 51.01 -> 49.36 | Hurt |
| use ETS instead of Power Sampling on MATH500 | Qwen3-8B · MATH500 | accuracy: 70.8 -> 72.4 | Improved |
| use ETS instead of Base decoding | Qwen3-8B · MATH500 | accuracy: 60.0 -> 72.4 | Improved |
| increase total samples M x K (ablation) | ETS on Qwen3-8B · HumanEval | accuracy | Improved |
| increase guidance steps I (ablation) | ETS on Qwen3-8B · HumanEval | accuracy | Improved |
Closest papers by meaning
- Compute Aligned Training: Optimizing for Test Time Inference (NeurIPS 2026 · similarity 0.90)
- Leftover Lunch: Advantage-based Offline Reinforcement Learning for Language Models (ICLR 2024 · similarity 0.89)
- TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- Is Best-of-N the Best of Them? Coverage, Scaling, and Optimality in Inference-Time Alignment (ICML 2025 · similarity 0.89)
- Theoretical Limits of Language Model Alignment (NeurIPS 2026 · similarity 0.89)
- Do Not Let Low-Probability Tokens Over-Dominate in RL for LLMs (ICLR 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.
