StopLoss: Teaching Agents to Predict Failure and Stop Early

Key points

• Coding agents keep working on runs that will fail verification, and asking a frontier model to watch for this is expensive in its own right.1
• The name is inspired by the stop-loss order in finance: StopLoss ends a run early when the run is likely to fail, which saves resources.1
• Trained on GPT-family SWE-rebench trajectories and evaluated on SWE-bench Verified and a SWE-bench Pro subset with GPT and Claude executors, StopLoss is compared against six general-purpose judges reading identical windows.1
• This translates into savings: relative to completing every task, StopLoss lowers cost per retained solve by 4.5\% on Verified while preserving 99\% of successful runs, and by up to 31\% on the costlier Pro subset, reaching 48\% with a Qwen3-8B backbone, and in all four settings its cost per retained solve is below that of every judge policy.1

Sources (2)

• [1]StopLoss: Teaching Agents to Predict Failure and Stop Early
  Hugging Face Daily Papers · Oct 18, 12:00 AM

  We introduce StopLoss, a small external monitor that learns failure recognition from completed trajectories and their verifier outcomes.

  Coding agents keep working on runs that will fail verification, and asking a frontier model to watch for this is expensive in its own right.

• [2]StopLoss: Teaching Agents to Predict Failure and Stop Early
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 18, 08:39 PM · same content

Extractive summary: sentences quoted from the sources.