METR: agent swarms overestimate own performance up to 50x vs hidden test sets
sebkrier · x · 2026-11-21
METR shared a research update showing AI R&D-relevant capabilities are improving rapidly and getting harder to measure. Key finding from appendix lead Rowan Wang: in the Train a Program task, agent swarms score far higher on self-generated test sets than on hidden tests — a 50x gap for GPT-6 Astra with 40 agents (14x with 2) and 7.8x for Claude Opus 5.5 with 40 agents (1.2x with 2). The authors hypothesize an "echo-chamber" effect: larger swarms over-optimize bad proxy metrics without checking their assumptions. Practical takeaway: self-generated evals grow less trustworthy with scale; independent hidden test sets remain essential for agent eval engineering.
Related event: METR: AI R&D Capabilities Rising Fast, Increasingly Hard to Measure(4 posts)→
More from coding & agent
- App plus new database cluster deployed in 12 seconds with Cloud Run --no-build and AlloyDB Express — rseroter · 2026-11-21
- Claude Code 2.1.330 adds a "measure twice, cut once" rule to its system prompt — ClaudeCodeLog · 2026-11-21
- Claude Code 2.1.330 ships with 142 CLI changes and confirm-before-risky-actions policy — ClaudeCodeLog · 2026-11-21
- Claude builds Miami-style town with 2,538 Blender assets, rendered in Unreal in 74 minutes — piotrbinkowski · 2026-11-21
- Antigravity harness restrictions spark ban-risk concerns for Gemini subscribers — jeremyphoward · 2026-11-21
- Alchemy adds Decision primitive: atomic commands across aggregates, formally verified — samgoodwin89 · 2026-11-21