Attention Mean Fields Predict Average Representation Dynamics and Reveal Context-Specific Computation
Micah Adler, John W. Byers, Mark Crovella
cs.LG · cs.CL
Quick verdict
Introduces attention mean-field kernels that predict average representation dynamics in five LLMs and expose ICL; methods and evidence are largely convincing.
Summary
The paper develops an empirical mean-field theory of attention in trained transformers. A head's attention averaged over a corpus defines a type-to-type kernel; iterated with the model's frozen MLPs, this open-loop rollout predicts how the top-1000 token-type centroids evolve layer by layer. Across GPT-2, Gemma-2-2B, Llama-3.1/3.2 and Qwen-3-14B, predicted and measured centroids align closely (cross-layer centered-cosine means 0.907/0.830/0.816 for GPT-2/Gemma-2/Qwen-3-14B), beating context-free, co-occurrence and structure-ablated kernels. On Pythia-160m/1.4B/12B, early training is indistinguishable from the corpus mean field — substituting the kernel for attention at frequent query types barely changes loss — until divergence at the induction transition. The context-conditional mean field acts as a null model whose deviation tracks in-context learning: it spikes at induction copy boundaries, separates prompt classes, emerges with induction during training, and decomposes additively into attention-covariance and value-contextualization terms. Ranking heads by deviation recovers documented induction/name-mover heads plus positional heads; task-conditioned deviations serve as injectable function vectors with transfer accuracy comparable to causal-mediation selection at a fraction of the cost. Stated limitations include worse Llama rollouts from MLP curvature (Jensen gap), restriction to the top-1000 types (64–68% of positions), corpus-conditioned results, numerical instability in some Qwen-3 ablations, and no mentioned code release.
Review complete
Assessment at a glance
5
risk flags identified
● high 0
● medium 2
● low 3
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
- A corpus-averaged attention kernel, iterated open-loop with the model's frozen MLPs, accurately predicts layerwise type-centroid evolution (cross-layer cosine 0.907/0.830/0.816 on GPT-2/Gemma-2/Qwen-3-14B), beating context-free, co-occurrence and structure-ablated alternatives.
- Early in Pythia training (through ~step 128), the model is behaviorally indistinguishable from its corpus mean field — mean-field substitution at frequent query types leaves loss unchanged — and the kernel reorganizes rapidly around the induction transition.
- Context-conditional mean-field deviation is a task-agnostic measure of context-specific computation that rises with ICL and splits exactly into attention-covariance and value-contextualization terms (Eq. 15).
- Ranking GPT-2 heads by corpus-level deviation recovers documented induction and name-mover heads while revealing positional and content-tracking heads (Table 4).
- Task-conditioned deviation vectors act as function vectors: mean-field-based head selection achieves zero-shot transfer accuracy comparable to causal-mediation (AIE) selection on Gemma-2-2B and GPT-J-6B using a single forward pass.
Strengths
Broad, held-out multi-model validation
The open-loop rollout (Algorithm 1) is validated on five models from GPT-2 Small to Qwen-3-14B (Table 1: cross-layer cosine 0.907/0.830/0.816 for GPT-2/Gemma-2/Qwen-3-14B), with a document-interleaved train/test split so the rollout never sees the scored centroids, plus cross-corpus checks on FineWeb, WikiText and Pre-1919 books (App. F).
Unusually careful bias and error accounting
§4.3 decomposes rollout error and attributes the dominant share to the MLP-applied-once-at-the-centroid step rather than the mean-field attention (teacher-forced attention prediction: 0.78–0.87 cosine, Fig. 5); Appendix B quantifies the context-conditional kernel's ~5–7% typical under-prediction (Table 6) and fixes it with an IPF calibration (Algorithm 2), verifying all key results unchanged (head-ranking Spearman 0.9967).
Ablations include a genuine negative control
Convergent, multi-scale evidence linking deviation to ICL
Novel structural observation about circuits
Weaknesses
Cross-architecture accuracy claim does not extend to Llama
Coverage limited to the top-1000 types
Protocol asymmetry in the function-vector comparison
Some ablation numbers are unusable
No stated code or artifact release
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
The paper builds an empirical mean-field theory of attention in trained transformers. Its central object is a per-head, type-to-type attention kernel estimated in a single corpus pass; under a corpus-conditional reading it predicts the layerwise evolution of token-type centroids, and under a context-conditional reading it provides a per-input null model whose residual — the *mean-field deviation* — indexes context-specific computation. Four headline claims: (1) accurate open-loop centroid prediction across five models; (2) early Pythia training indistinguishable from its mean field, with divergence at the induction transition; (3) deviation as a task-agnostic ICL meter with an exact additive decomposition into covariance and contextualization terms (Eq. 15); (4) task-conditioned deviations usable as function vectors competitive with causal-mediation (AIE) selection. The scope is deliberately type-level (top-1000 frequent types) and the authors are explicit that no infinite-width or infinite-depth limit is taken — this is finite-sample averaging over contexts, distinct from the statistical-physics mean-field literature they review in §2.
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 is the 'mean field' of attention in this paper?
It is the average attention a given attention head places from one token type (the query, e.g. 'October') to another (the key), measured over a large text corpus. Treated as a fixed type-to-type table per head, it can be iterated together with the model's frozen MLP weights to predict how the average representation of each frequent token type evolves layer by layer — without running the model on any text.
What does 'mean-field deviation' measure and why does it matter?
For a specific input context, it measures how far a head's actual output departs from the corpus-average prediction for that query type. Large deviation signals computation that depends on the particular context. The paper shows deviation rises with in-context learning: it spikes at the copy boundary of repeated-text induction prompts, separates prompt classes, appears during training exactly when induction heads form, and decomposes into an attention-routing part and a value-contextualization part.
Which models were analyzed?
Centroid-prediction results cover GPT-2 Small, Gemma-2-2B, Llama-3.1-8B, Llama-3.2-3B and Qwen-3-14B (kernels estimated on OpenWebText, with checks on FineWeb, WikiText and pre-1919 books). Training dynamics use Pythia-160m, 1.4B and 12B checkpoints. The function-vector comparison against causal mediation uses Gemma-2-2B and GPT-J-6B.
Does replacing attention with its mean field break the model?
It depends on training stage and prompt type. Early in Pythia training, substituting the corpus kernel for attention on frequent query types leaves the loss essentially unchanged. In fully trained models, substitution is catastrophic exactly on prompts needing in-context computation — about +9.2 nats on induction prompts for GPT-2 (Table 2) — while being roughly harmless on random text where no in-context information exists.
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.
- Extraction omits parts of §3.3, the body of §5.1–5.2 (mean-field substitution loss curves, Fig. 7), the start of §6.3, and appendices C, F–N, O–Q; judgments on those parts rely on cross-references in the surviving text.
- Figures are available only as captions and embedded descriptions; curve values other than those stated in text/tables could not be independently verified.
- codeAvailable=false reflects the absence of any code/data statement in the extracted text, not a confirmed lack of release.
- This is arXiv v1 (Sep 14, 2026); peer-review status unknown.
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.