Tracking States or Tracking Cosets? An Algebraic Account of Learned State Tracking
cs.LG · cs.AI
Quick verdict
Group-theoretic account: Transformers hit 1/f quotient plateaus while RNNs traverse non-normal cosets; rigorous, honest, but architecture confound remains.
Summary
The paper asks what neural networks actually learn when trained to predict running products of group elements. Analyzing Transformers, LSTMs/GRUs and Mamba-2 across a suite of finite groups (all non-abelian groups to order 12, selected larger ones to order 125) plus infinite groups with finite commutator classes, the authors show models often stop at 'quotient solutions': they identify the abelianization class of the true state while predicting nearly uniformly among its f members, so conditional exact accuracy matches 1/f with no fitted parameter (Table 1). Theorem 1 proves order-blind models can recover only abelianization-coset information; Theorem 2 proves optimal order-blind accuracy converges to 1/f, matching observed plateaus and reordering insensitivity. Beyond quotients, any right-coset partition survives sequential updates; in training censuses, standard Transformers only exhibit normal-subgroup stages, whereas parameter-matched recurrent networks pass through non-normal coset stages. On A5, low-dimensional state subspaces encode non-normal cosets, coset means form approximate dodecahedra, and component swaps transfer coset states. Limitations: the architecture contrast is confounded by differing budgets and schedules, the non-normal recirculation stage is replicated only under a single seed, the causal role of gradient cancellation remains unresolved, and the reviewed text omits parts of Sections 6-8 and several appendices.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
4 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
- Transformers trained on group composition converge to quotient solutions whose conditional exact accuracy equals 1/f, where f is the size of the commutator subgroup, with no fitted parameter (Table 1).
- The product information recoverable without input order is exactly the information constant on commutator-subgroup cosets, and optimal order-blind exact accuracy converges to 1/f as prefix length grows (Theorems 1-2).
- Any partition into right cosets of any subgroup, normal or not, is compatible with sequential updates; recurrent networks pass through non-normal coset stages during training while standard Transformers exhibit only normal-subgroup (quotient) stages.
- On A5, low-dimensional subspaces of the recurrent state encode non-normal cosets; coset mean vectors form approximate dodecahedra and swapping state components transfers the donor's coset state through a shared input suffix.
Strengths
Parameter-free quantitative law with clean controls
Table 1: conditional exact accuracy Aexact/Aquotient matches 1/f across ~17 groups (e.g., S4: 8.30-8.36% vs 1/12; SL(2,3): 12.45-12.49% vs 1/8) with x-chance tracking |G|/f; A4 vs SL(2,3) shares abelianization C3 yet f=4 vs 8 cleanly separates conditional accuracy, and infinite Gp groups retain finite classes p and hit 1/p (Fig. 2d,e).
Tight theory-experiment loop
Theorem 1 proves order-blind outputs are exactly abelianization-coset-constant; Theorem 2 gives convergence to 1/f with an explicit finite-t bound, matching observed reordering insensitivity (28 runs, JS divergence .0012-.0163 bits, App. M.1) and finite-length excess estimates at t=17.
Multi-pronged evidence for non-normal coset stages
Exceptional reporting transparency
Weaknesses
Architecture contrast confounded by training configuration
Non-normal recirculation stage is effectively single-seed
Gradient-cancellation causality left open
Threshold and criterion choices made during analysis
Theory scope limited to uniform full-group inputs
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
- All 4 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
Why do Transformers trained on group multiplication stop improving before perfect accuracy?
The paper shows they often learn only the abelianization — the commutative 'shadow' of the group — identifying which class of size f the true state lies in while predicting nearly uniformly among its members. Conditional on a correct class, exact accuracy is therefore about 1/f, and the paper proves this is the optimal limit for any model that ignores input order.
What is a non-normal coset stage in neural network training?
A transient period during training in which a recurrent model's prediction errors partition group elements into right cosets of a subgroup that is not normal, meaning the partition does not form a quotient group. The paper observes such stages in LSTMs and GRUs on S5 and A5, but not in standard Transformers.
Is the dodecahedron in this paper a real geometric object inside the model?
On the group A5, the mean hidden-state vectors of the 20 order-3 cosets, projected into a three-dimensional subspace of a recurrent model, form an approximate dodecahedron; swapping state components along these directions between two sequences transfers the donor's coset state through a shared input suffix.
Does this paper explain how ChatGPT or other LLMs track entities?
No. The authors explicitly state the conclusions concern controlled synthetic group-composition tasks and do not establish performance or safety properties of deployed language models, although the task is a common idealization of entity and state tracking.
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 truncated mid-Section 5.2 and omits Sections 6-8 and parts of Appendices A, B.3, L, O, P, Q; the A5 dodecahedron and state-swap findings were assessed mainly from the abstract and Figure 1 caption.
- Figures are available only as caption text; accuracy trajectories, heatmaps and geometry plots could not be verified from the extraction.
- No public code or data URL appears in the extracted text; the reproducibility statement references internal run-level manifests only.
- Some appendix cross-references (e.g., App. L/O protocols) could not be checked against the omitted text.
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.