In the In‑Context Learning (ICL) paradigm, large language model (LLM) agents can improve their decisions by leveraging interaction histories. Yet it remains unclear whether this improvement stems from genuine recursive reasoning or merely from extrapolating statistical patterns in the data. To disentangle these mechanisms, we place LLM agents in multi‑agent incomplete‑information games that require recursive belief reasoning.
We construct a public‑goods game and deliberately perturb the statistical structure of historical feedback. Decision quality is measured against a history‑independent Rational Expectations Equilibrium (REE) benchmark.
Our experiments show that when the statistical regularities in the history are disrupted, the advantage of longer contexts virtually disappears, causing decision quality to drop to the no‑context baseline. This degradation is sharply amplified in settings with stronger strategic interdependence.
These findings suggest that, in highly interdependent strategic environments, ICL behavior aligns more with statistical extrapolation than with true recursive reasoning. Our work extends mechanistic studies of ICL to strategic multi‑agent settings, introduces REE as a diagnostic tool for distinguishing reasoning from extrapolation, and provides a reusable framework for probing the limits of LLM reasoning in recursive belief tasks.
Review: By carefully designing a public‑goods game, the paper demonstrates that LLM agents rely heavily on statistical patterns rather than deep reasoning in multi‑agent decisions, offering valuable insights for future efforts to enhance genuine reasoning capabilities.