Behavioral Self-Rewarding lifts Qwen3-14B to 34.7% on SWE-bench Verified
dair_ai · x · 2026-10-15
Researchers from Penn State and Microsoft Research propose Behavioral Self-Rewarding (BSR) to fix a key weakness of binary test rewards for SWE agents: when every trajectory in a group passes or fails, the reward has no contrast and teaches nothing. BSR has the policy compare its own trajectories pairwise, turning those preferences into bounded auxiliary rewards added on top of the test result — bounded so self-judged preferences can rank two passing trajectories but never lift a failing one above a passing one, with no new environment interactions needed. Results: Qwen3-14B improves from 32.1% to 34.7% on SWE-bench Verified (400-step budget) and from 6.33% to 8.11% on SWE-bench-Live.
More from coding & agent
- Stop letting Claude or Codex pick your stack: write it in AGENTS.md and gate new dependencies — weswinder · 2026-10-15
- Dev closes all PRs: let agents patch open source via Nix overlays instead — kieranklaassen · 2026-10-15
- Prime Sandboxes Adds Filesystem Checkpoints to Fork Agent Rollouts Mid-Trajectory — samsja19 · 2026-10-15
- Every to host behind-the-scenes session on reporting Codex with Codex — every · 2026-10-15
- x402 transactions on Base jump 65% in three quarters, Chainalysis says — kleffew94 · 2026-10-15
- Pydantic AI author boosts PiDrei, a Python port of Pi using Rust-based Monty — samuelcolvin · 2026-10-15