BCSS Preprint: Symbolic Models Reveal Humans and AI Make Surprisingly Similar Reasoning Errors
LakeBrenden · x · 2026-11-10
Brenden Lake's group (first author Solim LeGris) releases the preprint Synthesizing symbolic models of human and machine reasoning at scale, introducing Behavior-Conditioned Symbolic Synthesis (BCSS) — automated search for interpretable, deterministic symbolic programs that verifiably reproduce behavior, applied at scale to human and machine errors in the ARC domain.
Key findings
- Humans and AI share error signatures across hundreds of reasoning problems: oversimplifying, adopting rules that only partially explain observations, and misinferring dependencies between features
- Shared signatures are modulated by available resources: reasoning effort predicts how much machines simplify their explanations
- Results point to satisficing-like mechanisms in both people and machines; BCSS offers a general way to interpret and compare rule-based reasoning of humans and black-box AI
More from AGI Musings
- A 1996 LA Times column predicted the Internet would fail—and replace 'web' with 'agents' today — omooretweets · 2026-11-10
- Nat Lambert welcomes more labs competing with Anthropic on open AI research tools — natolambert · 2026-11-10
- 2007 video resurfaces: Larry Page anticipated the Bitter Lesson 20 years early — juanbenet · 2026-11-10
- Peter Diamandis: Data center backlash is the best thing to happen to AI — PeterDiamandis · 2026-11-10
- Mythos 5.2 hits 41% on Remote Labor Index, up from Opus 4.8's 8% — peterwildeford · 2026-11-10
- Candidate uses phone AI to answer interview questions live in viral video — TansuYegen · 2026-11-10