A paired guessing game exposes compounding communication errors
The study tests six self-paired model systems, not the reliability of every multi-agent collaboration.
The log2 N-Questions study runs 408 games in which one agent sees candidate Wikipedia lead paragraphs and another, from the same provider, sees the secret target. The questioner has exactly log2 N binary questions. It is a controlled information-asymmetry task.
The reported decline follows a geometric power of horizon length: more rounds have to succeed. Individual errors need not become more likely at later rounds for the whole game to fail. Model-specific errors and discrimination failures matter; the study’s pooled correlations must retain their stated subsets.
The original article extended a leading-five-model correlation to every model and treated this game as proof that agents are unready for any teamwork. The experiment does not establish that conclusion.
For an actual collaborative workflow, isolate the information boundary, define what a receiver can verify and measure error recovery over the intended horizon. This study motivates testing communication reliability separately from spending more reasoning tokens.
Editorial correction (2026-10-13 UTC): Narrowed the pooled correlation and collaboration inference to the tested game.