Research map / Large language models / Reasoning and chain of thought
Beyond Majority Voting: Self-Reflective Test-Time Reinforcement Learning for LLM Reasoning
Sitong Wu, Haoru Tan, Xichen Zhang, Bin Xia, Shaofeng Zhang, Xiaojuan Qi, Bei Yu, Jiaya Jia
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes a self-reflective test-time RL variant that replaces majority-vote pseudo-labels for LLM reasoning, based only on the abstract's motivation.
Problem
Test-time reinforcement learning needs rewards without ground truth; existing TTRL methods use majority voting to create pseudo-labels, which often fails on complex reasoning tasks where the majority answer is not correct.
Approach
Moves beyond majority voting by having the model self-reflect on sampled trajectories to estimate rewards at test time; the abstract does not give details of the mechanism.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace majority-vote pseudo-labels with self-reflective pseudo-labels | TTRL · AIME24 | Pass@1 accuracy (%): 46.7 -> 55.8 (Qwen3-8B) | Improved |
| replace majority-vote pseudo-labels with self-reflective pseudo-labels | TTRL · average over 4 benchmarks, 5 models | Pass@1 accuracy (%): additional +7.9 to +11.2 over TTRL | Improved |
| replace majority-vote pseudo-labels with self-reflective pseudo-labels | TTRL · MATH-500 (easier, majority usually correct) | Pass@1 accuracy (%): +2.6 to +8.1 over TTRL | Improved |
| remove candidate trajectory pool construction (feed redundant trajectories, ignore frequency by prompt) (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy (%): 55.8 -> 54.3 when removed | Improved |
| remove trajectory summarization (use full-length trajectories) (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy (%): 55.8 -> 50.4 when removed | Improved |
| remove both pool construction and summarization (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy (%): 55.8 -> 48.8 when removed | Improved |
| use self-reflective pseudo-labels with N=8 samples versus TTRL with N=64 (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy: comparable performance with 8x fewer samples | Improved |
| increase sampled trajectories N beyond 32 (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy | No clear effect |
| increase sampled trajectories N from 8 to 64 for majority voting (ablation) | Maj@N pseudo-label · AIME24, Qwen3-8B | Maj@N accuracy: Maj@N flat at 43.3 while Pass@N 53.3 -> 80.0 | No clear effect |
| use ground-truth labels instead of self-reflective pseudo-labels (ablation) | SR-TTRL · AIME24, Qwen3-8B | Pass@1 accuracy (%): 55.8 -> 65.4 | Improved |
| add self-reflection step to pseudo-labeling | TTRL · AIME24, Qwen3-8B | time per step: 160 s -> 205 s | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- TTRL — TTRL: Test-Time Reinforcement Learning (NeurIPS 2025) · better on 11 of 12 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Surprising findings
- Majority voting pseudo-labels frequently break down on complex reasoning tasks, where correct solutions are not the most frequent answer (per abstract).
Limitations
- Full text unavailable; method details and results are not verified.
Experimental setup
Baselines: TTRL
Builds on: TTRL
Closest papers by meaning
- TTRL: Test-Time Reinforcement Learning (NeurIPS 2025 · similarity 0.94)
- Tool Verification for Test-Time Reinforcement Learning (NeurIPS 2026 · similarity 0.92)
- Optimizing Test-Time Compute via Meta Reinforcement Finetuning (ICML 2025 · similarity 0.90)
- SELF-HARMONY: LEARNING TO HARMONIZE SELF-SUPERVISION AND SELF-PLAY IN TEST-TIME REINFORCEMENT LEARNING (ICLR 2026 · similarity 0.90)
- Tricks or Traps? A Deep Dive into RL for LLM Reasoning (ICLR 2026 · similarity 0.89)
- Meta-TTRL: A Metacognitive Framework for Self-Improving Test-Time Reinforcement Learning for T2I Generation in Unified Multimodal Models (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.
