English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

TTPO: Test-Time Policy Optimization for Label-Free LLM Reasoning

Forum topic · 小凯 · 2026-08-29

Summary

TTPO (Test-Time Policy Optimization) is a new method for improving large language model mathematical reasoning without ground-truth labels. Post-training techniques like reinforcement learning (RL) and on-policy self-distillation (OPSD) depend on labeled data, which prevents test-time training. Using majority-vote pseudo-labels is fragile because a single wrong vote can corrupt the teacher model. The authors observe an asymmetric failure mode: rollouts that disagree with the pseudo-label are usually wrong regardless of whether the vote is correct. TTPO exploits this with an asymmetric objective—distilling agreeing rollouts via OPSD while penalizing disagreeing rollouts with Grouped RL—plus token-level selection that down-weights converged positions for distillation and penalizes only confident errors in RL. Without any labels, TTPO matches label-supervised OPSD on five competition-level benchmarks, improves Qwen3-1.7B from 38.0% to 45.2% in test-time training, gains +25.2% to +36.4% in non-thinking mode, and shows strong cross-task generalization. The paper (arXiv:2508.11369) is by Aozhe Wang, Zhengxi Lu, and Jianze Wang.

Paper Overview

  • Field: NLP (mathematical reasoning for LLMs)
  • Authors: Aozhe Wang, Zhengxi Lu, Jianze Wang
  • arXiv: 2508.11369
  • Abstract (English)

    Recent prominent post-training methods, such as Reinforcement Learning (RL) and On-Policy Self-Distillation (OPSD), have driven rapid progress in mathematical reasoning for large language models, yet their reliance on ground-truth labels precludes test-time training (TTT). Replacing ground truth with majority-vote pseudo-labels is a natural alternative, yet it is fragile: an incorrect vote corrupts the teacher and misleads every token. The authors observe that this failure mode is asymmetric: rollouts that disagree with the pseudo-label are typically wrong regardless of whether the vote itself is correct.

    Building on this observation, they propose Test-Time Policy Optimization (TTPO), an asymmetric objective that distills agreeing rollouts via OPSD and penalizes disagreeing rollouts with Grouped RL.

    Key Points

  • Asymmetric objective: distill rollouts consistent with the majority-vote pseudo-label; apply Grouped RL penalties only to inconsistent rollouts, which are typically wrong even when the vote is wrong.
  • Token-level selection: distillation down-weights already-converged positions, while RL penalizes only confident errors—keeping both updates well-defined despite frequent pseudo-label mistakes.
  • Tighter self-supervision: majority-vote routing yields progressively tighter supervision as the model improves.
  • Results:
  • Matches label-supervised OPSD on five competition-level benchmarks without any labels.
  • Improves Qwen3-1.7B from 38.0% to 45.2% in test-time training.
  • Gains of +25.2% to +36.4% in non-thinking mode.
  • Strong cross-task generalization.
*Auto-collected on 2026-08-29.*

Tags

#tttpo#test-time-training#llm-reasoning#reinforcement-learning#self-distillation#pseudo-labels#arxiv#nlp

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178634191