Stanford's Collaborative Intelligence Benchmark: Best Model Drops 20-47 Points When Teammates Misbehave
omarsar0 · x · 2026-11-07
Stanford researchers propose the Collaborative Intelligence Benchmark, measuring AI agents' ability to act as teammates: 240 tasks across 8 teamwork skills, sourced from real records like aviation accident reports, clinical cases, coding tasks, and negotiations.
Key findings:
- The best of 18 models scores 82.2/100, yet every model loses 20–47 points once a teammate creates a problem.
- Frontier models almost always refuse unsafe shortcuts and report misconduct, but holding a position under pressure is much harder.
- In negotiations where a teammate pushes, adds guilt, and threatens to walk away, models reach the best deal for the team only 14% of the time.
Essential reading for anyone building multi-agent systems or agent teams.
Related event: Stanford Releases Collaborative Intelligence Benchmark for AI Teammates(2 posts)→
More from coding & agent
- Google team confirms Hacktron AI submitted high-impact vulns, expands partnership — ithilgore · 2026-11-07
- Stardock CEO: orchestrators miss the point—teams need exhibits and shared agent data — draginol · 2026-11-07
- Researcher uses Claude to auto-fix LaTeX figure offsets: "ridiculously painless" — alfcnz · 2026-11-07
- vmfunc shows off Reflet, a fully native disassembler for Mach-O/AArch64 — amplifiedamp · 2026-11-07
- Agentic robot assembles autonomously for 2.5 hours, one gadget every 7 minutes — ihorbeaver · 2026-11-07
- Agentic robot control demo: 2.5 hours autonomous assembly at one gadget per 7 minutes — ihorbeaver · 2026-11-07