SWE-Verifier, described in an October 21 paper, trains a Qwen3.5-4B model on preferences among replayed coding-agent actions.

The big change

  • What changed: It scores a proposed action; a low score triggers alternatives, ranked before execution.
  • Why it matters: Developers can target costly comparisons to suspect actions.
  • What to watch: Reported cost excludes local verifier inference.

Results

On 500 SWE-bench Verified tasks in mini-swe-agent, MiniMax-M2.7 moved from 75.47% to 77.20% resolution at 1.09 times modeled baseline cost. DeepSeek-V4-Flash Preview moved from 72.60% to 74.27% at 1.13 times. The Two-Stage Rubric variant compared alternatives at fewer than 4% of steps.

MiniMax thresholds used 202 tasks. The table leaves variation undefined and does not establish significance. BIG CHANGE did not rerun it. The linked repository returned HTTP 404 on October 25; artifacts and setup remain unverified.

Earlier coverage examines distinct harness attribution.