Do Chess Explanations Reflect Model Decisions? Behavioral and Token-Level Tests of LLM Reasoning Faithfulness
cs.CL · cs.AI
Quick verdict
Claims chess explanations are fluent but weakly faithful; masking/recoverability/token-scoring tests on 200 puzzles find modest, decoder-dependent signal.
Summary
This paper asks whether fluent LLM explanations of chess moves carry decision-relevant information beyond plausible language. On 200 Lichess endgame puzzles (ratings 500-2000), explanations from GPT-4.1-mini and five other hosted model families are tested with three instruments: move recoverability (decoders infer the intended move from masked explanations), decoder-side controls (FEN-only, explanation-only, strict/maximal masking, random and unmasked explanations), and token-level log-probability scoring of legal moves under open-weight models (Mistral-7B, Qwen2.5-32B, Llama-3.1-70B). Strictly masked explanations recover the generator's own move at 0.379-0.398 top-1 but the correct reference move at only 0.056-0.091; gains over the FEN-only baseline are small, decoder-dependent, with paired bootstrap intervals including zero. In the decoder-recovery control, unmasked explanations lift generated-move top-1 to 0.68-0.74, and random plausible explanations from other puzzles actively lower reference-move top-3 rank by 0.16-0.18. Motif-rich explanations improve recoverability without improving correctness. Limitations: control and scoring analyses use only 50 puzzles (layerwise 20, counterfactual 25), headline magnitudes come from GPT-4.1-mini self-decoding, exact move accuracy stays below 10% in nearly every condition, and no code or data release is stated.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Strictly masked chess explanations recover the model's own generated move (0.379-0.398 top-1, GPT-4.1-mini self-decoding) far better than the correct reference move (0.056-0.091).
- Under strict masking, explanations add only small, decoder-dependent gains over the board state alone, with paired bootstrap intervals including zero.
- Random but plausible explanations from other puzzles actively reduce the correct move's top-3 ranking (-0.16 to -0.18), showing irrelevant reasoning text is not ignored.
- Motif-rich explanations improve move recoverability without improving move correctness (pawn endings: recovery 0.259 to 0.548 but exact only 0.111 to 0.140; rook endings: exact falls 0.143 to 0.092).
- Linguistic plausibility, action consistency, and solution correctness are distinct properties of LLM explanations.
Strengths
Well-factored control design separating leakage from signal
Six decoder conditions on the same 50-puzzle subset (FEN-only, explanation-only, strict mask, maximal mask, random, unmasked; Table 2) cleanly isolate lexical leakage: unmasked explanations recover generated moves at 0.680-0.737 top-1 while masked conditions sit near FEN-only, and a masking audit (Table 15: 56.7% residual promotion vocabulary under strict vs 0% under maximal) quantifies the trade-off explicitly.
The generated-vs-reference target separation is a sharp, novel behavioral finding
Explanations recover the model's own move (0.379-0.398 top-1, Table 1) far better than the correct Lichess solution (0.056-0.091), and the ordering survives cross-model decoders (Tables 16-17), directly demonstrating that action consistency and correctness dissociate.
The random-explanation result is consequential and practically relevant
Candid, well-calibrated framing of limits
Broad triangulation within a compact design
Weaknesses
Generator monoculture and self-decoder magnitudes
Underpowered control experiments
Floor-level correctness undermines correctness-coupled analyses
No code, data, or artifact release stated
Subgroup analyses with tiny cells and no stated correction discipline
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper studies explanation faithfulness in a controlled domain: given an LLM's chess move and its free-text explanation, does the explanation contain decision-relevant information? Three properties are explicitly separated — linguistic plausibility, action consistency, and solution correctness. The headline claims: (1) strictly masked explanations let decoders recover the generator's own move at 0.379–0.398 top-1 (GPT-4.1-mini self-decoding, Table 1) but the correct Lichess solution at only 0.056–0.091; (2) masked explanations add only small, decoder-dependent gains over a FEN-only baseline (Table 2; paired bootstrap CIs include zero); (3) random plausible explanations from other puzzles actively reduce the correct move's token-level rank (Table 3: reference top-3 −0.16 brief [−0.30, −0.04], −0.18 calc [−0.30, −0.06]); (4) motif mentions raise recoverability without raising correctness (pawn endings: recovery 0.259→0.548, exact 0.111→0.140; rook endings: exact falls 0.143→0.092; Table 19). Claims are consistently hedged as behavioral evidence rather than causal access to internal computation — the framing is careful throughout.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
Are LLM chess explanations faithful to their moves?
According to this paper, only weakly. When explicit move mentions are masked, a decoder can recover the model's own chosen move at 0.379-0.398 top-1 but the actually-correct solution at just 0.056-0.091; beyond that, masked explanations add little over the board position alone, and the size of the gains depends heavily on which model reads them.
What is move recoverability in LLM evaluation?
It is a behavioral test: give a model a chess position plus an explanation whose explicit move strings (UCI/SAN) have been masked, and ask it to output the intended move in UCI notation. High recoverability means the explanation contains action-specific information. The paper contrasts recovering the model's generated move versus the actually-correct reference move to separate action consistency from correctness.
Can irrelevant chain-of-thought text hurt LLM outputs?
Yes - this paper shows that inserting a random but plausible chess explanation taken from a different puzzle lowers the correct move's top-3 ranking by 0.16-0.18. Language models do not simply ignore ungrounded reasoning text; it actively steers token-level preferences, which is a caution for contexts where explanation-like text is injected into prompts.
Is code or data available for this chess faithfulness study?
The paper text states no public code or data release. The underlying inputs are public (the Lichess puzzle database, Stockfish evaluations, and open-weight models Mistral-7B, Qwen2.5-32B, and Llama-3.1-70B), but the masking lexicons, prompt templates, and specificity-score weights are described only in prose, so exact replication would require reimplementation.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- No code or data availability statement appears anywhere in the provided text; the reproducibility assessment is based on methods prose alone.
- Figures appear in the extraction only as captions and interleaved axis labels; plotted values were assessed from the tables and in-text numbers, not from the figure images themselves.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.