MIT and Sakana AI's SIFT boosts self-improving coding agents at one-tenth DGM's cost
dair_ai · x · 2026-09-20
A paper from MIT and Sakana AI shows self-improving coding agents work, at a fraction of the cost:
- Method: Self-Improvement via Fast Tree-search (SIFT) attacks the runtime bottleneck—benchmark evaluation—by using an LLM judge to rank candidate self-modifications before full evaluation.
- Results: 35.1% on Polyglot with o3-mini after 30 expansions, vs DGM's 30.7% after 80 tree-search nodes.
- Cost: SIFT runs in under 50 CPU hours and 5 wall-clock hours; the Qwen3-30B config completes full search at 224 CPU hours and $34 API spend, a tenth of the DGM baseline.
- Judge quality decides the run: on TerminalBench, gpt-5.4-high as pairwise judge finds a 36.7% agent vs a 29.2% starting point; gpt-5 finds 34.5%.
More from coding & agent
- SnapBridge plugin one-clicks Chrome screenshots into Google's Antigravity IDE — AI_Andrew · 2026-09-20
- SnapBridge goes open source: a CDP-powered macOS daemon that one-clicks screenshots into Antigravity — AI_Andrew · 2026-09-20
- Claude-Code-Game-Studios turns one Claude Code session into a studio of 49 AI agents — tom_doerr · 2026-09-20
- Generating song lyrics with Qwen3.8 + YuE2 music pipeline, full prompt workflow shared — cocktailpeanut · 2026-09-20
- Appshots tip: click this to see exactly what the model sees — nickbaumann_ · 2026-09-20
- Agent-built 3D websites bloat the web: 5MB loads, 1GB RAM, 50% CPU at idle — dbreunig · 2026-09-20