AI agents that fix bugs are usually graded like a true-false quiz: did the test pass, yes or no. A new method lets the agent grade the quality of its own debugging along the way, and that extra signal makes it better at the job.
Researchers describe Behavioral Self-Rewarding (BSR) in a paper posted to arXiv on October 12, 2026. The technique has a software engineering agent compare its own past attempts at the same task, even when a test-based verifier scored them identically as pass or fail. The agent's comparative judgments turn into small bonus rewards layered on top of the verifier's binary signal, without requiring any new runs in the coding environment. Tested on Qwen3-14B and Qwen3-32B models across the SWE-bench Verified and SWE-bench-Live benchmarks, the approach produced consistent gains over standard pass/fail training.
The real finding here is that pass/fail rewards throw away information. Two agents can both fail a test, but one flailed through ten irrelevant edits while the other nearly nailed it. Letting the model itself spot that difference, instead of paying for a separate reward model or more environment interactions, is a cheap way to recover that signal.
Still, this is self-grading: the agent is both the student and the one handing out the extra credit. The paper reports gains on two SWE-bench variants with two model sizes, not a broad sweep, so how far this generalizes beyond those setups, and whether it introduces its own biases over time, is still an open question.