Training on 69,000 DeepSeek-R1 Failed Attempts Boosts 7B Math Scores
imjustnewatai · x · 2026-08-06
A recent paper suggests that a strong model's failed trajectories can be more valuable for training than its successes. Researchers collected 69,000 failed math reasoning attempts from DeepSeek-R1 and trained smaller models to locate the mistakes, repair the reasoning, and solve the problems.
Experiments showed that a 7B model trained this way achieved significant math improvements:
- Average scores across 6 math benchmarks rose from 37.0 to 42.4.
- On 3 out-of-domain tests, scores doubled from 20.9 to 40.0.
The core insight is that a strong model's wrong answer typically contains long stretches of correct reasoning followed by a single critical error. Teaching a model to locate and recover from that mistake provides more learning signal than copying a polished answer. This implies that the massive amounts of failed coding runs and agent trajectories stored in AI labs hold immense untapped training value.
More from Research
- Latent Labs Launches Latent-Y, First Lab-Validated Drug Design Agent — chaitjo · 2026-08-06
- From GPT-3 to Agentic Sandboxes: A History of AI Evaluation — natolambert · 2026-08-06
- DeepSeek-V4 Writeup Reveals Use of Agentic Trajectories in Mid-Training — cwolferesearch · 2026-08-06
- Solving Tool Overload in AI Agents: Skill-Aware Decomposition — bendee983 · 2026-08-06
- Snorkel AI proposes milestone-based evaluation to pinpoint long-horizon agent bottlenecks — ajratner · 2026-08-06
- FPTalks 2026 Focuses on Low-Precision LLM Pretraining — nmwsharp · 2026-08-06