SWE-Verifier: a 4B model lifts coding agents ~1.7pp on SWE-Bench at 9–13% extra cost
radamihalcea · x · 2026-10-30
Researchers introduce SWE-Verifier, a lightweight 4B generative verifier that selectively guides coding agents roughly 70× its size at critical decision points.
How it works: instead of trajectory-level or all-step verification, it scores each proposed action and only triggers candidate sampling plus ranking when needed. It's trained from Qwen3.5-4B on preference data grouping contrasting actions with environment observations and rubric-guided teacher scores.
Results (SWE-Bench Verified, mini-swe-agent):
- MiniMax-M2.7: resolution up from 75.47% to 77.20% at 1.09× cost
- DeepSeek-V4-Flash: generalizes without training on its actions, 72.60% → 74.27% at 1.13× cost
- Fewer than 4% of steps are verified, matching or beating trajectory and all-step baselines at far lower cost
More from coding & agent
- VibeGame from Nanjing University builds complete 2D games from prompts with adversarial agent teams — nanjinguniv · 2026-10-30
- WebAnything Auto-Rebuilds Real Websites into Scalable Web Agent Environments — Tianci Xie · 2026-10-30
- WorkFactory Scales Professional Task Generation for Agent Training, Yields WorkPilot — IQuestLab · 2026-10-30
- Kimi launches 3D Make: connects to Blender and CAD tools to build models, scenes, and games — Kimi_Moonshot · 2026-10-30
- Kimi team shares its internal 3D workflow: parametric vs mesh modeling with K3.1 Pro — Kimi_Moonshot · 2026-10-30
- Animated short made with ComfyUI: ChatGPT poses, Qwen restyle, MiniMax H3 — bseoan · 2026-10-30