Stanford benchmark: 18 AI models tested as teammates, best scores just 82.2/100
james_y_zou · x · 2026-11-05
A Stanford/MIT/Oxford paper introduces the Collaborative Intelligence Benchmark: 240 tasks across 8 teamwork dimensions (from refusing unsafe shortcuts to checking a weaker model's work), sourced from real records like aviation reports and clinical cases, with programmatically verifiable outcomes. To isolate collaboration from raw capability, models are scored only on tasks they can solve alone. Findings: the best model scores just 82.2/100 (Gemini 4 Argon high, ahead of GPT-6 Astra 79.4 and Claude Opus 5.5 78.7); frontier models almost never cheat and refuse unsafe shortcuts, but they cave under interpersonal pressure, winning the best deal in only 14% of conflicts.
More from Models
- Claude trains a model instead of writing code: 1257-param pixel-art generator in 10 minutes — victormustar · 2026-11-05
- MoE beats dense in matched OOD finetune: DeepSeek V3.1 hits 0.804 CE vs Llama 405B — kalomaze · 2026-11-05
- Grok Bot tests sharing disclosure: files, skills and credentials visible before you share — XFreeze · 2026-11-05
- Ari agent post-trains Qwen3.5-9B unsupervised for 16h, reward 0.118→0.629 — ypatil125 · 2026-11-05
- Happenstance scores 84.7 on PeopleSearchBench's 119 queries, nearly 20 points above next best — alex_teichman · 2026-11-05
- Google Research unveils Agile Deliberation, an LLM framework for teaching VLMs subjective categories — RanjayKrishna · 2026-11-05