Aishik Nagar and Cheston Tan: Reasoning or Pattern Exploitation?

Aishik Nagar and Cheston Tan: Reasoning or Pattern Exploitation?

Authors Aishik Nagar and Cheston Tan examine whether reinforcement learning with verifiable rewards produces reasoning or more superficial adaptations. They train small language models in Tic-Tac-Toe, a finite logical domain they can analyze exhaustively. The models achieve high accuracy and robustness across prompt variations, which the authors say suggests the emergence of game reasoning. But their mechanistic interpretability analysis points elsewhere: sparse autoencoders show that models rely more heavily on information already explicit in the prompt, while concepts such as board symmetries and strategic forks remain weakly represented. The authors argue that this gap raises concerns that reasoning benchmarks overestimate model abilities. They also report that reinforcement learning reorganizes which features dominate representations rather than adding or removing concept information, and that causal interventions locate the decision downstream, in the late-layer state. The study offers a way to examine what changes inside models during training, but its evidence comes from a small, exhaustively analyzable game. How well do these findings about Tic-Tac-Toe apply to reasoning benchmarks in other domains?

Publication: af011bd1-8786-4aca-8eec-0c50befeb390

attribution — pass: The reading correctly credits the paper and its claims to authors Aishik Nagar and Cheston Tan. It keeps findings and concerns attributed to them, for example 'the authors say' and 'the authors argue', rather than presenting them as established fact.

Authors:Aishik Nagar, Cheston Tan

summary — pass: The summary faithfully captures the central tension: strong accuracy and robustness that 'suggest' game reasoning, versus SAE evidence of reliance on prompt-explicit information and weak strategic concepts. It also preserves the authors' benchmark concern, keeps the limits of Tic-Tac-Toe in view, and poses a non-leading question. Terms such as sparse autoencoders, mechanistic interpretability and late-layer state should be explained through term-library tooltips. The phrase 'downstream, in the late-layer state' drops the source's 'of it', meaning downstream of the representation. Restoring that would sharpen the sentence, but the omission does not block understanding.

This suggests that RL-driven "reasoning" may conflate task-specific adaptations to input information with reasoning ability.

date — pass: originalPublished 2026-10-29 matches the arXiv v1 submission date. The feed date of Oct 31 is a later announcement and is correctly not used.

[Submitted on 29 Oct 2026]

support — pass: Every claim is backed by verbatim abstract text, including the description of the models as small, which the abstract calls small-scale LLMs. The draft adds no unsupported benefits, costs or forecasts. The generalization question and the stated limitations follow directly from the finite domain and do not assert a conclusion.

RL post-training neither adds nor removes concept information but reorganizes which features dominate the representation

media — pass: The selected media is the verified direct link to the original arXiv abstract page, which is the most relevant way to access the source.