Test-Time Scaling via Error Localization
TTEL keeps correct reasoning prefixes and restarts generation at the localized error point.
The paper proposes an inference-time method that uses feedback to identify where a model’s trajectory first went wrong at token level. It then truncates from that point and branches, instead of throwing away the whole attempt. In the reported LiveCodeBench result with Qwen3-8B, TTEL reaches 71.0% pass@64 while using about half the generated tokens of independent sampling. The authors also report gains on AIME-2025 and HMMT-2025 across Qwen3-8B and Qwen3-4B-Thinking-2507. ArXiv · AI/CL/LG's note
The paper proposes an inference-time method that uses feedback to identify where a model’s trajectory first went wrong at token level. It then truncates from that point and branches, instead of throwing away the whole attempt. In the reported LiveCodeBench result with Qwen3-8B, TTEL reaches 71.0% pass@64 while using about half the generated tokens of independent sampling. The authors also report gains on AIME-2025 and HMMT-2025 across Qwen3-8B and Qwen3-4B-Thinking-2507. ArXiv · AI/CL/LG's note
score 6