Doing What They Say, Not What They Reason: Locating the Faithfulness Gap in LLM Agents
Title: Aligning Action with Explanation: Pinpointing the Faithfulness Discrepancy in LLM Agents
Abstract: Do large language model agents actually execute the plans they articulate? This issue of process fidelity is critical for employing LLMs in social simulations, but it remains difficult to quantify due to the absence of a benchmark for correct behavior. To investigate this, we utilize a controlled Texas Hold’em poker simulator that provides verifiable reference actions for every decision. By breaking down the faithfulness gap into two distinct stages—reasoning-to-conclusion and conclusion-to-action—we observe that these two steps exhibit opposing behaviors.
Source: arXiv Generated at: 2026-06-02 00:00:00 UTC
