StopLoss aborts likely-failing coding-agent runs early
Original title: StopLoss: Teaching Agents to Predict Failure and Stop Early
On SWE-bench Verified it cut cost per retained solve by 4.5 percent while keeping 99 percent of successful runs.
Coding agents often keep working on runs that will fail verification, and frontier judges are expensive to watch them. StopLoss is a small external monitor trained on completed trajectories and verifier labels to score recent tool steps and abort likely failures without changing the executor. Trained on GPT-family SWE-rebench data and tested with GPT and Claude executors, it beat six general-purpose judges on cost per retained solve in all four settings.
- A threshold fit on successful runs turns the score into an abort.
- Savings reach up to 31 percent on a SWE-bench Pro subset.
- A Qwen3-8B backbone reaches 48 percent savings on that subset.
- Outcome supervision lifts first-check failure prediction from near chance.
From the original article
1 min read at huggingface.co
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. We introduce StopLoss, a small external monitor that learns failure recognition from completed trajectories and their verifier outcomes. 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.
By Jiarong Liang, Xueguang Ma, Dongfu Jiang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.