Behavioral Self-Rewarding Lifts Qwen3-14B on SWE-bench
The method requires no new environment interactions, so teams can layer it onto existing reward setups without extra rollout cost. Researchers and engineers training coding agents on sparse test-based rewards get a concrete fix for zero-variance groups, with the bounded-reward design showing how to keep self-judgement from overriding ground-truth correctness. It compares against standard GRPO-style binary reward training on Qwen3-14B.
A paper from Penn State and Microsoft Research addresses a weakness in training SWE agents: binary test rewards give no learning signal when every trajectory in a group passes or all fail. Their proposed Behavioral Self-Rewarding method has the policy compare its own trajectory pairs, converting these preferences into small bounded rewards added on top of test results. The bounded design ensures a self-judged preference can rank two passing trajectories but never lift a failing one above a passing one. On SWE-bench Verified, Qwen3-14B improves from 32.1% to 34.7% average accuracy at a 400-step budget, and on SWE-bench-Live from 6.33% to 8.11%.
// Behavioral Self-Rewarding for SWE agents //
It's standard practice to train SWE agents on binary test rewards.
This paper shows the problem with that. When every trajectory in a group passes, or every one fails, the reward has no contrast, and the group teaches nothing.
Researchers from Penn State and Microsoft Research propose Behavioral Self-Rewarding. The policy compares its own trajectories in pairs, and those preferences become small bounded rewards added on top of the test result.
The added rewards are bounded, so a self-judged preference can rank two passing trajectories but can never lift a failing one above a passing one.
No new environment interactions are needed. On SWE-bench Verified, Qwen3-14B improves from 32.1% to 34.7% average accuracy at a 400-step budget, and on SWE-bench-Live from 6.33% to 8.11%.
https://t.co/ekD2h7ya4y
Chat with Paper: https://t.co/6vUbYmIXRR
