The Internal Anatomy of Strategic Choice in Large Language Models
Vinícius Ferraz, Leon Houf, Enrico Ferrea
cs.AI · cs.GT
Quick verdict
LLMs encode 2×2-game incentives but differ in using them; rigorous, candid study, though causal steering holds in Qwen only and multiplicity is uncorrected.
Summary
The authors record residual-stream activations from four open-weight LLMs (Qwen2.5-72B, Qwen2.5-72B-Instruct, Llama-3.1-70B-Instruct, GPT-OSS-120B) playing all 144 strict ordinal 2×2 games one-shot, tracking a prespecified payoff incentive from prompt through activations to choice. Behaviorally, dense models mirror the human decline in canonical play with game complexity (ρ≈−0.40 to −0.42 vs −0.44), fit quantal level-k near depth 1, and respond to the incentive gap (λ≈1.9–2.0), while GPT-OSS sits apart at fitted depth 2.07. Internally, own incentive and realized choice decode in every model (AUC 0.767–0.916; opponent incentive at chance); the matched base–instruct Qwen pair chooses identically at baseline (96.4% agreement) yet differs in whether incentive reaches choice (behavioral bridge +0.047 vs −0.012). Causal steering along the letter-orthogonalized incentive direction shifts graded policy in double-dominance games for both Qwen models (9/9 games) but is heterogeneous for Llama and untested for GPT-OSS. The paper claims a three-way dissociation — availability, recruitment, susceptibility — and argues post-training reshapes the route from incentive to decision. Limitations: no multiplicity correction, only three permutation controls, 8-bit quantized weights, one cue wording per disposition, and GPT-OSS recorded only post-reasoning. Code and derived data are publicly released.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 3
● 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
- In all four models the payoff incentive and the eventual choice are linearly decodable from activations, while the opponent's incentive is not — a self-regarding strategic representation only.
- Availability, recruitment and susceptibility dissociate: e.g., Qwen2.5 base decodes the incentive strongly (AUC 0.847) yet shows no behavioral bridge (−0.012), and can still be steered (+0.060 in double-dominance games).
- The base and instruction-tuned Qwen2.5 checkpoints choose almost identically at baseline (96.4% agreement) but differ in whether the internal incentive reaches the choice, implying post-training rewires the path from representation to decision while preserving behavior.
- Dense models mirror the human decline of canonical play with game complexity, but after controlling for the incentive gap only Llama retains the decline, so the raw similarity is largely incentive-confounded.
- Fixed decision cues are internally encoded (LDA 0.97–1.00) but their behavioral uptake is constrained by game structure, with inequity aversion producing the largest displacement (mean 0.643 over 17 identifiable games).
Strengths
Census, not sample, of the game space
All 144 strict ordinal 2×2 games, 8 conditions and 4 counterbalanced forms (letter assignment × option order) yield 18,432 trials with 99.4% usable choices (18,326); behavioral conclusions cannot hinge on game selection, and J/P counterbalancing defeats answer-letter artifacts.
Carefully controlled dissociation framework
Availability/recruitment/susceptibility is operationalized with game-held-out probes (PCA, training-fold standardization), a decision–incentive angle tested against a per-layer joint game-permutation null recalibrated to activation geometry (Suppl. Fig. 4 shows why a fixed 90° reference fails), and letter-orthogonalized steering directions; the layer-79 vs layer-65 ablation cleanly separates readout overlap from upstream incentive representation.
Clean base–instruct contrast isolating post-training
Unusual candor about negative and inconclusive evidence
Dual human calibration with confound correction
Weaknesses
No multiplicity correction across many layerwise tests
Causal specificity rests on three permutation directions
Causal story generalizes to only 2 of 4 models
All dense-model internal results under 8-bit quantization
One wording per disposition cue
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper asks how four open-weight LLMs (Qwen2.5-72B, Qwen2.5-72B-Instruct, Llama-3.1-70B-Instruct, GPT-OSS-120B) convert payoff incentives into choices in one-shot play of the complete catalogue of 144 strict ordinal 2×2 games. Its central claim is a three-way dissociation: **availability** (incentive and choice decode from activations in every model), **recruitment** (whether the internal incentive predicts choice beyond the objective gap — model-dependent), and **susceptibility** (whether pushing the internal signal moves the policy — selectively yes for the Qwen pair). The headline result is the matched base–instruct Qwen2.5 contrast: near-identical baseline behavior (96.4% agreement, 555/576; identical policies in 124/144 games) yet divergent recruitment (bridge +0.047 [.011,.085] vs −0.012 [−.040,.019]) and answer-axis expression (λlens 1.43 vs 0.17) — post-training rewiring the route while preserving the choice.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 5 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
What do availability, recruitment and susceptibility mean in this paper?
Availability is whether the payoff incentive can be decoded from the model's activations (yes, in all four models, AUC 0.767–0.916). Recruitment is whether that internal signal actually predicts the model's choice beyond the objective payoff gap (varies: strong in Qwen2.5-Instruct, absent in Qwen2.5 base). Susceptibility is whether artificially amplifying the internal signal changes the choice (yes, selectively in double-dominance games, for both Qwen models). The paper's core point is that these three properties can come apart.
Which models and games were used in the study?
Qwen2.5-72B (base), Qwen2.5-72B-Instruct, Llama-3.1-70B-Instruct and GPT-OSS-120B each played every one of the 144 strict ordinal 2×2 games (the complete catalogue of Bruns 2015) in one-shot play with no history, under baseline, five decision cues (risk aversion, loss aversion, inequity aversion, selfish maximisation, maximin), a neutral control, and as player 2 — 18,432 trials total, 99.4% yielding a usable choice.
Did post-training change how the model decides, according to this paper?
Yes, internally though not behaviorally: the base and instruction-tuned Qwen2.5 models made the same choice in 96.4% of matched baseline presentations and had similar incentive decodability, but differed in whether the internal incentive predicted the choice (bridge +0.047 vs −0.012) and in answer-axis expression (λlens 1.43 vs 0.17). The authors conclude post-training reshaped the internal route from incentive to decision while leaving baseline behavior largely intact.
Is the code and data for this paper available?
Yes. The paper states a full reproduction package (code, derived data and instructions) is publicly available at github.com/vferraz/strategic-choice-anatomy, with activation data in a separate deposit that has access requirements. The authors honestly note that immutable checkpoint/tokenizer hashes and a locked software-environment specification are not yet archived and will appear in the public release.
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.
- The provided text is a sampled excerpt with [...sampling gap...] sections (e.g., parts of the Figure 3–4 captions, part of Section 2.2, and the Supplementary 'Causal interventions' apparatus description); conclusions touching those regions were cross-checked against visible Results text, Table 1 and Suppl. Analyses B.4.
- The excerpt ends mid-reference list (within the Moore et al. entry); no content beyond that point was visible, so late references and any closing matter were not assessed.
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.