acceptodds
Under review as a conference paper at ICLR 2027

LOOPED TRANSFORMERS LEARN ONE OF TWO ALGORITHMS, AND TRAINING MOVES THE ODDS

15 views

Abstract

A looped transformer applies the same block many times over. On a task with a known depth requirement, computing every running product of a sequence of group elements, such a model could in principle work in logarithmic depth. Trained on one, it instead reads the sequence one element per loop, and a growing literature asks what would make it do otherwise. We answer that by intervening rather than by timing: overwriting the model’s internal state at one position and one loop, and asking which outputs change. Doing this across two independent sets of checkpoints, 237 and 90, shows that these models do not learn a spectrum of more or less efficient circuits. They learn one of two algorithms, and the distribution of their measured speed is bimodal, with a slow mode sitting exactly at one position per loop and a fast mode near three. The structure survives a test that assumes no parametric form, and survives refitting on one checkpoint per training contract. Training therefore controls the odds of getting the better algorithm, not the algorithm’s quality. Making the efficient one available does not move those odds: un-tying the weights leaves the median slope at exactly 1.00, as does conditioning each loop on its index, as does supervising the efficient circuit’s own intermediate quantities, which hands the model the target computation. Only making the slow algorithm infeasible moves them. Since the achievable speed varies from seed to seed while each model faces a hard threshold, the population curve is smooth: we registered a sharp critical budget, found none, and report that as a failure. Holding the depth budget proportional to length, reach does not fall from n=64 to n=128, so length is not the binding variable; capacity is, rising from 17 of 40 solved at 5M parameters to 11 of 12 at 50M (p = 0.003). It does not buy speed. What it buys is fidelity at matched depth and a shift in which algorithm is installed, and we measure both. This is one family of synthetic tasks, models up to 50M parameters, and a causal measurement that reaches n ≤ 32. A frontier is only measurable against a known depth requirement, which natural text does not supply; on the two public looped LMs we can therefore only test behaviour, and there depth from 1 to 64 does not move a two-composition ceiling. Within that scope every claim was registered before the run that tests it, which is what makes the negative results usable: the capability result is believable because it was a prediction rather than a summary. Every registered prediction is scored in the appendix, including those that failed.

Primary area:
interpretability and explainable AI
Keywords:
looped transformers, depth-recurrent transformers, weight tying, mechanistic interpretability, activation patching, state tracking, algorithmic generalization, pre-registration

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

The paper aims to show that looped transformers on a specific algebraic task learn one of two algorithms, named "scan" and "compressed front". Training on this task has for only effect to modify the probability of the algorithm being used, instead of increasing the performance of either.

Critical strengths

Studying what algorithms transformers learn through training is a very relevant topic, which is of interest to many in the community.

Critical weaknesses

The main weaknesses of this paper are how hard it is to follow. The motivation for the paper is not discussed. Many terms, including the main objects of study, are not defined. Many claims are not proven.

Many terms are either undefined, or are mentioned several times before being even informally defined:

  • "This task" is mentioned several times in the abstract and conclusion without being given a name or even informally introduced. This is especially concerning given it is the main object of study of the paper.
  • The two algorithms "scan" and "compressed front" are not formally defined. Once again, these are the main object of study.
  • What does it mean formally for "the outcome to be one of two algorithms"?
  • The checkpoints "237 and 90" are mentioned 3 times in the abstract and introduction without being even informally defined. It seems looking at page 4 that those refer to the number of tests conducted.
  • Table 1 is given page 3, and the only mention of it in the main text is a cryptic paragraph "Table 1 reads naturally as a continuum of efficiency. It is not one." at page 4.
  • What is the meaning of "reach" and "capacity", discussed in the introduction?

A lot of choices are unmotivated or not sufficiently explained, some claims are not proven.

  • Why study "this task"?
  • The model studied is not sufficiently described, as the "single core block" is not defined.
  • Claims such as "S5 is NC1-complete" or "Solvable non-abelian S4 is not NC1-complete" are not justified. It would help to add references.
  • The training details of the transformer are not introduced.

The paper may have interesting results. However, it requires extensive rewriting for it to be valuable to the community. I encourage the authors to spend time on rewriting the paper:

  • Clearly expose the main motivation of the paper.
  • Define every term clearly, especially those that are central to the paper.
  • Justify every assumption, claim, and methodological choice.
Questions for authors

Why study "this task"?

What does it mean formally for "the outcome to be one of two algorithms"?

Reviewer_T9kEOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This manuscript discovers a bimodal distribution in the computational speed of looped transformers using activatioin patching, then proposes the potential existence of two algorithms (Sequential Scan & Compressed Front). The experiments are quite comprehensive, confirming that budget-training more effectively promotes efficient computational strategies.

Critical strengths
  1. Well-motivated, especially the bimodal distribution in Fig 2.
  2. The experiments are comprehensive, and also present hypotheses that were not validated.
  3. The training interventions provide valuable insights, highlighting computational budget constraints as an effective way to encourage more efficient computational strategies.
Critical weaknesses
  1. There remains over-claim in core assumptions. Maybe there is a gap between the bimodal distribution and the two underlying computational algorithms.
  2. The abstract is overly long, making it difficult to quickly grasp the key points. I suggest avoiding a format that resembles an experimental report.
Questions for authors
  1. See weakness 1, if the authors can provide a more rigorous proof or citing relevant work to provide a stronger mapping relationship, I will adjust my score accordingly.
  2. Does the bimodal phenomenon depend on the specific nature of the group operations? I suggest the authors test a non-group operation (such as state tracking in a DFA) to see whether the distribution remains concentrated in two distinct clusters.
  3. Is the conclusion Length Does Not Bind too strong? Increasing length from 64 to 128 is insufficient for research on length generalization.
Reviewer_r7XvOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper examines how training conditions affect computation in looped transformers on cumulative group products. Using activation patching to measure settling times, it finds two regions in the resulting speed distribution: one near a sequential scan, one faster. Further experiments vary depth budgets, intermediate supervision, loop conditioning, capacity, and sequence length.

Critical strengths

The experimental effort is substantial. The intermediate-demand experiments and within-condition analyses usefully check against an effect of pooling different training settings. Calibrating the intervention measure against reference computations is a reasonable starting point.

I appreciate the explicit reporting of failed predictions and unsuccessful replications. The distinction between solving within the training budget and solving with additional inference loops is particularly important. These results are useful even where the proposed explanation does not hold.

Critical weaknesses

My main concern is the interpretation of the speed distribution. Two clusters in a scalar measurement do not establish two algorithms. The faster cluster could contain several mechanisms or a continuous family of computations. The paper also reports that its location changes with demand, which conflicts with the claim that training changes only the odds of selecting an algorithm, not its speed.

The intermediate-demand experiment does not resolve this. A depth budget requires a model to finish before a deadline; it does not require the model to use all available loops. A continuous family of solutions could therefore exceed the demanded speed without concentrating at it. The argument that a continuum should fill the gap needs further justification.

A second concern is what the causal statistic measures. The disappearance of sensitivity to an upstream patch need not coincide with completion of a prefix product. Information may already have been copied elsewhere, or the model may recover from the intervention. Equation 1 applies to actual computational progress, but its application to the reciprocal fitted settle slope requires assumptions that are not established. The nonlinear profiles and the failed prediction at larger scales make this more than a technical detail. Appendix E should specify donor selection, sample sizes, damage criteria, evaluation horizons, and handling of incorrect clean outputs.

The conclusions about interventions are also too broad. Loop-index encoding and FiLM conditioning retain shared parameters; they are not equivalent to independently parameterized layers. Likewise, supervising block products does not ensure that the model implements or uses the intended circuit. Supervision produces a significant shift in slope, even if few models cross the chosen efficiency threshold. This supports a narrower conclusion than "only infeasibility works."

Finally, some qualifications disappear between the results and the conclusion. The jitter experiments find no significant difference, which does not establish equivalence. The length comparison is underpowered, and the proposed capacity mechanism is exploratory. These limitations should remain visible in the headline claims.

Questions for authors

What distinguishes two algorithms from a scan-like cluster and a heterogeneous faster regime? Would the authors accept the latter as the supported conclusion?

How closely does the patching statistic track actual completion on the same checkpoints and lengths, including startup delay?

Were independent weights across depth tested, and were the auxiliary block-product targets learned successfully?

Please clarify two protocol inconsistencies: sampling uniformly over [base/2, 2base] does not preserve the mean at base; Appendix E describes last-position accuracy, whereas §7 describes whole-sequence accuracy.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Corrected revision PDF

Corrected revision PDF. The PDF uploaded on Nov 8 carried a shortened AI-use statement: it limited the assistant's analysis role to one comparison (Section 9) and left out its role during the revision. The corrected PDF, now uploaded, states both: the assistant helped analyse and interpret results, and during the revision it drafted the designs and registered predictions of the new experiments and carried out the post-hoc analysis that identified the deadline rule. The authors approved everything submitted. Appendix A records for each prediction made during the discussion who drafted it and whether the authors adopted it before it was tested. The Nov 8 PDF also printed the ethics and reproducibility statements twice; the corrected PDF prints them once. Apart from these two fixes, the corrected PDF has wording edits throughout; no number, result or scientific claim changed. changes_from_submission.pdf in the supplementary material and the anonymous repository is updated and marks every change.

TL;DR correction. The submission form does not let us edit the TL;DR during the discussion, so it still states the withdrawn two-algorithms claim. It should read: "On group prefix products, weight-tied looped transformers learn a one-position-per-loop scan by default; under a binding deterministic depth budget, the median model that leaves it runs about as fast as the tightest training deadline requires."

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.