Google DeepMind’s TTEL localizes errors to cut test-time search costs in reasoning

burny_tech · x · 2026-07-25

• Google DeepMind researchers present Test-Time Scaling via Error Localization (TTEL), a method for improving reasoning and programming performance at inference time.
• TTEL uses feedback to localize where an error happens in a generation, then prunes the search trajectory and branches from the valid prefix instead of retrying from scratch.
• The paper reports better Pareto tradeoffs on sequential reasoning tasks and says TTEL can reach 71.0% pass@64 on LiveCodeBench with Qwen3-8B, while using about half as many tokens as independent sampling.
• It also says TTEL outperforms competing test-time baselines on AIME-2025 and HMMT-2025 for the tested models.

Original post →

More from Research

• MCP fault injection study: agent task scores drop 60% under realistic tool failures, scale doesn't help — dair_ai · 2026-10-26
• Study: AI reviewing AI misses most reward hacking; best models miss 25%+ of hacks — ronbodkin · 2026-10-26
• FLUX.3 Klein step-distilled to 4-8 steps with CFG baked in — Fit-Structure-6489 · 2026-10-26
• NLI alone rejects 5/6 new agent memories; a 0.45 similarity gate fixes it all — Alternative_Pin9598 · 2026-10-26
• AI folks, stop saying "biology" when you mean drug development — KevinKaichuang · 2026-10-26
• ostris pre-training "Primal VAE": 1-channel latents match modern 8x VAEs via patching — ostrisai · 2026-10-26

Research channel →