Lilian Weng reports that a harness evolved on Terminal-Bench-2 transferred successfully to SWE-bench-verified without further evolution, suggesting it encodes general engineering experience rather than benchmark-specific optimization.
Lilian Weng ·
Supporting evidence
Harness Engineering for Self-Improvement
Original excerpt
On Terminal-Bench-2, AHE achieved better than human-designed harness (OpenCode, Terminus-2, Codex) except for Hard tier and a few other self-evolve baselines (ACE, TF-GRPO). The same frozen harness, without further evolving, transfers to SWE-bench-verified, indicating that the evolved harness is able to encode engineering experience into harness components rather than doing benchmark-specific optimization.
About this interpretation
The object-specific interpretation and Chinese translation were checked independently against the source. This is an AI semantic review, not playback verification. Reviewed Oct 5, 2026 · qwen3.8-max-0902
Report an issue