AI Paper Atlas

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

ICML 2026 · Large language models › Reasoning and chain of thought · approach: RL with verifiable rewards for reasoning

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 changedWhereMeasuredEffect
replace majority-vote pseudo-labels with self-reflective pseudo-labelsTTRL · AIME24Pass@1 accuracy (%): 46.7 -> 55.8 (Qwen3-8B)Improved
replace majority-vote pseudo-labels with self-reflective pseudo-labelsTTRL · average over 4 benchmarks, 5 modelsPass@1 accuracy (%): additional +7.9 to +11.2 over TTRLImproved
replace majority-vote pseudo-labels with self-reflective pseudo-labelsTTRL · MATH-500 (easier, majority usually correct)Pass@1 accuracy (%): +2.6 to +8.1 over TTRLImproved
remove candidate trajectory pool construction (feed redundant trajectories, ignore frequency by prompt) (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracy (%): 55.8 -> 54.3 when removedImproved
remove trajectory summarization (use full-length trajectories) (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracy (%): 55.8 -> 50.4 when removedImproved
remove both pool construction and summarization (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracy (%): 55.8 -> 48.8 when removedImproved
use self-reflective pseudo-labels with N=8 samples versus TTRL with N=64 (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracy: comparable performance with 8x fewer samplesImproved
increase sampled trajectories N beyond 32 (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracyNo clear effect
increase sampled trajectories N from 8 to 64 for majority voting (ablation)Maj@N pseudo-label · AIME24, Qwen3-8BMaj@N accuracy: Maj@N flat at 43.3 while Pass@N 53.3 -> 80.0No clear effect
use ground-truth labels instead of self-reflective pseudo-labels (ablation)SR-TTRL · AIME24, Qwen3-8BPass@1 accuracy (%): 55.8 -> 65.4Improved
add self-reflection step to pseudo-labelingTTRL · AIME24, Qwen3-8Btime per step: 160 s -> 205 sHurt

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.

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

Limitations

Experimental setup

Baselines: TTRL

Builds on: TTRL

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy