acceptodds
Under review as a conference paper at ICLR 2027

TALLY: Learned Approximately, Proved Exactly

Abstract

Neural models can learn the correct computation and still fail to execute it reliably. TALLY studies this separation on finite group state tracking. An eight dimensional, input conditioned orthogonal transport makes latent group structure recoverable and raises gated DeltaProduct's pooled median accuracy at training length from 0.014 to 0.989 across twenty matched runs. Its operators encode the task's finite structure, but satisfy the group relations only approximately, so continuous execution eventually drifts. Given frozen operators and the recovered finite orbit, a conditional moment recursion predicts mean squared state deviation with 3.1% mean relative discrepancy across 120 combinations of checkpoint and length. On one frozen model, halving or doubling the operator defect moves the first sampled horizon below 90% transport accuracy from 8,192 to 32,768 or 2,048, while an exact realisation has no observed errors through 32,768. This controlled intervention supports a causal role for approximation error in the transport's failure. TALLY then recovers the latent finite action from checkpoints and unlabelled words, compiles it into discrete execution, and checks every transition and output against the task. The compiled scan makes no errors across 83,886,080 evaluated positions at length 1,048,576. Separate finite checks prove 53 extracted executors and readouts correct on every input at every length. The guarantee covers the frozen discrete executor and readout, not the unchanged neural decoder. Learned approximately, compiled and proved exactly.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
state tracking, linear recurrent networks, length generalization, orthogonal transitions, group word problems, certified computation, formal verification, DeltaProduct

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_BL7QOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper studies the separation between learning finite computational structure and executing it reliably over long sequences. It introduces a small orthogonal transport component for recurrent models trained on group state-tracking tasks. The authors recover finite algebraic structure from learned operators, quantify the error accumulated during continuous execution, and investigate its causal role through a controlled intervention. They then compile the recovered structure into discrete execution and exhaustively verify exported executors and readouts against the task specification. Extensions examine product groups, distractors, and noninvertible reset operations. The all-length guarantee applies to the exported finite machine; results for the neural network remain empirical.

Critical strengths
  1. The paper provides a coherent and quantitatively supported account connecting learned algebraic structure to execution failure. Recovering the correct group structure across all twenty core models, predicting continuous-state deviation across 120 checkpoint–length combinations, and directly manipulating the operator defect together provide useful evidence beyond observing poor length generalization. Although the causal intervention concerns one checkpoint, it meaningfully strengthens the proposed explanation.
  2. The connection from approximate learned operators to checked discrete execution is compelling. The core experiments demonstrate substantial improvements in length generalization, and codebook execution eliminates observed errors across approximately 84 million evaluated positions in million-token sequences. The separate exhaustive verification establishes correctness of the exported executors and readouts at every length. The paper carefully distinguishes this guarantee from the neural decoder’s empirical performance and reports informative failures outside the core setting.
Critical weaknesses
  1. The incremental capability of the proposed extraction machinery over existing geometric extraction methods remains insufficiently established. Prior work already demonstrates that correct automata can be extracted from recurrent networks with poor long-sequence performance, including Wang et al. (2018). The distinctive contribution therefore rests substantially on the operator-level analysis and compilation procedure. The adapted WL* comparison and earlier clustering results do not fully isolate this contribution under matched access, checkpoints, and readout construction. In particular, the expanded black-box search contains the correct A7A_7 structure in all three cases, and the observed advantage concerns candidate selection and query budget.
  2. The conditions governing successful recovery remain unclear beyond the core group tasks. With auxiliary supervision, recovery succeeds in 7/10 product-group runs and 6/15 A5A_5 distractor runs, while the twelve-state partial-reset task yields no certificates across 33 runs. The successful reset/set experiments are encouraging, but these failures suggest that transition structure and the learned representation matter substantially beyond state count.
Questions for authors
  1. On the same checkpoints, with comparable access and a common readout and verification procedure, what advantage does operator-based extraction provide over established hidden-state clustering approaches?
  2. Can the unsuccessful product-group and partial-reset experiments distinguish incomplete learned structure from limitations of the extraction procedure? What observable properties predict successful recovery?
Reviewer_ufzKOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
1(−2)
Presentation
1(−2)
Soundness
2(−1)
Read review
Paper summary

This paper (TALLY) studies how recurrent models can learn a recoverable finite computation while still failing to execute it reliably on long sequences. It combines an orthogonal transport, quantitative error analysis, and discrete compilation, then checks the extracted executors and readouts against known task specifications to establish correctness at every input length.

Critical strengths
  1. The distinction between learned structure and reliable execution is supported by several complementary experiments. The 20 matched runs establish a substantial extrapolation benefit, while the frozen-operator analysis and defect intervention help explain why continuous execution eventually fails.

  2. The correctness claim is precise and supported by exhaustive checks of 53 extracted executors and their readouts. Freezing the recovered structure before task verification, and separating the integer-machine guarantee from full-network test accuracy, makes the central result credible.

Critical weaknesses
  1. The defect-scaling intervention is demonstrated on only one A5 checkpoint, so its generality remains uncertain. Failure-horizon prediction also improves only modestly over a constant baseline, covering 7/8 rather than 6/8 held-out checkpoints within a factor of two.

  2. The conditions for successful extraction are not yet well understood: frame extraction fails on all three fitted A5 × Z7 models, and none of the 33 runs on the 12-state partial-reset task yields a certificate. These failures limit how confidently the approach can be transferred to other finite-state tasks.

  3. The black-box comparison supports an advantage in structural selection, but the benefit for deployed accuracy is less clear because the methods use different readouts and neither achieves zero errors on A7. Omitting the equivalence-query loop also limits what this comparison establishes about the advantage over existing active-extraction methods.

I lean toward acceptance because the paper provides convincing evidence for the separation between recoverable computation and reliable execution, together with a carefully checked discrete implementation. The limitations above restrict the breadth of the conclusions but do not undermine the main result within the studied setting.

Questions for authors
  1. Does the defect-scaling intervention produce consistent changes in the failure horizon across additional A5 and S5 checkpoints, with the initial state and readout held fixed? This would help establish whether the reported result is representative.

  2. Can the existing codebook route recover and verify the full 420-state task from the fitted A5 × Z7 models? This would distinguish a limitation of the frame search from a failure to learn the required structure.

  3. How do geometry-based and tolerant black-box extraction compare on the Table 4 checkpoints under a common readout-construction and output-evaluation protocol? The current setup makes transition recovery difficult to separate from decoder quality.

  4. For the partial-reset failures, is there evidence that the transport contains the required task information before extraction? This would help distinguish a learning bottleneck from a limitation of the extraction procedure.

Reviewer_Qkq9OpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper deals with a widely adopted form of a linear recurrent neural network, a variant of DeltaNet, trained to learn the structure of different finite groups, i.e. trained to accumulate the actions randomly sampled from a certain finite group such as A_5, S_5, ...

Concretely, when the model is trained to emulate discrete finite group structure, the learned representation is imperfect, and errors accumulate over time. A mechanism which recovers a finite group structure from the learned continuous representation is proposed, one which does not do the recovery by means of labeled data, but rather by unlabelled clustering of the learned transport operators of the linear RNN. Error estimates are derived, a mechanism for remedying the drift is proposed, and the conditions under which the representation can be extracted is discussed.

Critical strengths

Regardless of my final evaluation, which I justify in the "weaknesses" section, the paper has qualities which should be highlighted. I believe the following to be three such qualities.

  • The problem the paper proposes and aims to address is in itself an interesting one. Independently of the "practical" impact of the scheme, it is an interesting idea to look for finite group representations in the trained transitions of an SSM without access to group state labelings or any other group structural information, and to then devise a scheme which can remedy the drift that occurs with longer input sequences.
  • A single parameter η\eta controls the deviation between the learned representation and the reconstructed finite group representation. It is usually not easy to find a single clean parameter that induces such predictable behaviour, and it is typically a sign of a solid method.
  • The experiments are significant in quantity and quality.
Critical weaknesses

Based off of my extensive experience supervising projects and interacting with LLMs, I believe that this paper is entirely LLM written. That does not exclude interaction with a human author, but the guiding force and the main writer of this paper, I claim, is an LLM. That is by itself not a ground for rejection; a valuable contribution is a valuable contribution wherever it comes from, but I have to say that I fail to find the value in this work. It is very difficult to follow, and that is my main ground for rejection. A paper may deal with a difficult topic, making it intrinsically hard to follow without significant background knowledge, but that is not the case here. For example, one may have little knowledge of differential geometry and stochastic calculus, yet the core ideas behind Einstein's papers remain more accessible than the ideas presented here. The concepts and ideas are presented in a muddled manner in this work. If one manages to understand the paper, crucial questions about the soundness and quality of the work remain. Some concerns:

  • The presentation of the paper is severely underoptimized. Concepts are introduced abruptly, with insufficient explanation.
  • It is not clear whether the proposed clustering algorithm does not contain hyperparameters which are optimised based off of the group structure. The sentence "Leader clustering with radius 0.5 finds exactly |G| clusters in all twenty confirmation models" gives off the impression that the clustering is in fact influenced by the collection and structure of the investigated groups.
  • "PD SSM" is mentioned abruptly in the "AI Assistance Statement" without any prior discussion on it.
  • Proposition 2 is trivial. I find that the amount of mathematical jargon used to explain the trivial concept is unjustified.

If the authors express sincere interest in improving the paper, I will gladly expand on what I consider the weaknesses of the paper in more detail during the rebuttal process. It is in both of our interests to provide a meaningful contribution to science, and the area of SSMs learning to emulate group structures is interesting to us both. However, at this point, I believe it is not appropriate for a human reviewer to spend a significant amount of effort to decode the signal in a text with this many signs of LLM generation.

Questions for authors

Apart from addressing the weaknesses above, I have several general questions which can hopefully increase the quality of the submission.

  1. How do you envision the impact of your method? Do you think it might have applicability in a more general setting, e.g., remedying the drifts in an SSM which is trained to learn an underlying structure of text or speech signals?
  2. Could you, without use of LLMs, summarise your paper for me? I would like to hear about the motivation behind it, the main findings, and what you believe would be interesting extensions.
  3. Can you provide a list of three main results which you think anyone should take away from your work? The list should be written without excessive use of technical concepts, such as e.g. "input conditioned orthogonal transport" or "conditional moment recursion".

Comments on the paper

2 comments
Area_Chair_bzWmOpenReview ↗

Reminder to reviwers to read author's responses

Dear Reviewers,

Thank you for your efforts in providing thorough reviews.

As the discussion period is underway, I would like to kindly remind you to read the authors' responses to your reviews and provide follow-up comments where appropriate or leave a comment. In particular, please consider whether the responses address your questions and concerns, and update your assessment if the clarifications or additional information affect your initial evaluation.

Your engagement in this discussion is important for ensuring a fair and thorough review process.

Thank you again for your time and contributions!

Best regards, Area Chair

AuthorsOpenReview ↗

Clarified task cohorts and baseline fit counts

I have uploaded the updated PDF and supplementary ZIP. Section 2 now gives the baseline fit counts for each task at length 128. Table 1 lists the tasks used in each control row and defines the fit criterion; these changes address the latest reporting concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.