cs.CLNov 17, 2026

Tracking Representations in LLM Reasoning Traces for Iterated Games

Authors: Chaipat Tirapongprasert, Rajarshi Das, Parijat Dube

Organizations: Columbia University · MQube Cognition

Abstract

Large language model (LLM) agents achieve state-of-the-art performance across domains through chain-of-thought (CoT) reasoning. Despite their ubiquitous use, little is understood about the relationship between the externalized trace and the model's actual computation. We address this by investigating how LLM agents make decisions in iterative games with incomplete information. Using a series of mechanistic interpretability techniques, we find that the model forms and iteratively revises an internal representation of the game throughout the CoT. Once the model prematurely commits to an incorrect belief, it rarely recovers. This failure is manifested in both the literal trace and the layer activations. Ultimately, as multi-agent systems are more prevalent, our work opens the door to a more focused diagnosis of their strategic reasoning errors. For illustrative purposes, we offer an interactive visualizer of our methodology on selected sample traces at https://interp-self.github.io/gametrace/

Figures & tables

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Oct 29, 2026

Reasoning or Pattern Exploitation? Mechanistic Insights from RL-Trained Language Models

Large language models (LLMs) are increasingly described as acquiring "reasoning" skills after reinforcement learning with verifiable rewards. However, whether these gains reflect genuine reasoning or more superficial adaptations remains underexplored. In this work, we probe LLMs trained in a finite and exhaustively analyzable logical domain, namely Tic-Tac-Toe, and trace how internal representations evolve across reinforcement learning (RL) with Group Relative Policy Optimization (GRPO). Quantitatively, RL yields high accuracy and robustness across prompt variations, suggesting the emergence of game reasoning. Mechanistic interpretability, however, paints a different picture: features extracted with sparse autoencoders (SAEs) reveal that models primarily adapt to rely more heavily on information already explicit in the prompt, such as whose turn it is, game progression and board occupancy. In comparison, high-level concepts like board symmetries and strategic forks remain weakly represented, raising concerns that reasoning benchmarks overestimate model abilities. This suggests that RL-driven "reasoning" may conflate task-specific adaptations to input information with reasoning ability. We provide, for the first time: (i) a systematic interpretability pipeline tracing representation dynamics across RL training in small-scale LLMs; (ii) an extension of SAE-based feature discovery to hypothesis-driven testing in a finite logical domain; and (iii) an interpretability-based demonstration that RL post-training neither adds nor removes concept information but reorganizes which features dominate the representation, with causal interventions locating the decision downstream of it, in the late-layer state.
Apr 30, 2026

Why Do LLMs Struggle in Strategic Play? Broken Links Between Observations, Beliefs, and Actions

Large language models (LLMs) are increasingly tasked with strategic decision-making under incomplete information, such as in negotiation and policymaking. While LLMs can excel at many such tasks, they also fail in ways that are poorly understood. We shed light on these failures by uncovering two fundamental gaps in the internal mechanisms underlying the decision-making of LLMs in incomplete-information games, supported by experiments with open-weight models Llama 3.1, Qwen3, and gpt-oss. First, an observation-belief gap: LLMs encode internal beliefs about latent game states that are substantially more accurate than their own verbal reports, yet these beliefs are brittle. In particular, the belief accuracy degrades with multi-hop reasoning, exhibits primacy and recency biases, and drifts away from Bayesian coherence over extended interactions. Second, a belief-action gap: The implicit conversion of internal beliefs into actions is weaker than that of the beliefs externalized in the prompt, yet neither belief-conditioning consistently achieves higher game payoffs. These results show how analyzing LLMs' internal processes can expose systematic vulnerabilities that warrant caution before deploying LLMs in strategic domains without robust guardrails.
Aug 21, 2026

Level-k Distinguishable Mechanisms for Evaluating Bounded Rationality in LLMs

Strategic depth of reasoning is essential for human interaction of Large Language Models (LLMs) operating in boundedly rational environments. However, existing evaluations are primarily based on canonical games prevalent in pretraining corpora, making it difficult to disentangle true strategic reasoning from memorisation. To address this, we formalise a necessary level-K distinguishability condition for strategic depth inference and construct a suite of novel game structures that meet this standard. Using these games, we evaluate strategic depth in LLMs from both the Chain-of-Thought tokens and actual actions under recursive reasoning and an inductive trace of opponent game-play data. Across experimental trials spanning four LLMs, four game structures, and ten levels of iterated reasoning, we find that model models maintain accurate strategic depth under recursive reasoning, with strong internal consistency between stated reasoning and actions at every level. Errors arise from using the wrong number of iterated depth of reasoning steps, not from computing best responses incorrectly. However, inductive inference from opponent play degrades accuracy sharply and unevenly across games, and explicit strategic mentalizing in the chain of thought substantially improves overall performance.