acceptodds
Under review as a conference paper at ICLR 2027

Beyond the Training Horizon: Mechanisms and Limits of Length Generalization in Looped Transformers

23 views

Abstract

Looped Transformers can generalize to reasoning chains longer than those encountered during training, but the computations enabling this behavior and limiting its extent remain unclear. We mechanistically compare two looped-Transformer configurations, which we call the Matched-Recurrence Looped Transformer (MR-Loop) and Decoupled-Recurrence Looped Transformer (DR-Loop), reflecting their respective recurrence-training schemes. We evaluate polynomial iteration, finite-state composition, and knowledge-graph traversal using detailed mechanistic analysis. Attention analysis, intermediate-state decoding, and causal interventions reveal distinct mechanisms learned under final-answer supervision. MR-Loop updates an intermediate state at a fixed readout while advancing relation selection through adjacent-token interactions and a transferable progress cue. DR-Loop instead propagates intermediate states across relation positions, forming an advancing computational frontier. However, both mechanisms become unreliable at greater depths: MR-Loop exhibits degradation of its readout state and progress cues, while DR-Loop exhibits declining reliability of state propagation. Limited self-correction allows local errors to persist and compound. Across both models, we uncover a common representational principle: recurrent states encode not only task-relevant content but also its computational status—whether that content remains in a form that can support subsequent computation. Transferable live–consumed and fresh–aged residual directions causally control whether represented information can participate in subsequent computation, including beyond the training horizon. We further show that length generalization need not rely on faithful step-by-step reasoning, as Looped Transformers can exploit task structure without explicitly representing every intermediate state.

Primary area:
interpretability and explainable AI
Keywords:
looped transformer, mechanism, length generalization

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_CYqbOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper provides a study on how looped transformers generalize to longer reasoning changes than in training. They perform a mechanistic interpretability study of two types of architectures: MR-looped transformers (these have input reinjection) and DR-looped transformers (these embed the input once and have a variable number of iterations in training). Through the analysis, they identify distinct mechanisms for each approach, as well as some a representational principle.

Critical strengths
  1. Figure 1 and Figure 2 are clear, well-made figures that make the MR-loop and DR-loop methods very understandable.
  2. The three primary research questions are clear
Critical weaknesses

The paper is very difficult to follow because all the results figures are in the appendix. The main text contains Table 1 (task detals), Figure 1 (MR-loop vs DR-loop mechanism), Figure 2 (MR-loop vs DR-loop mechanism). All the experimental results figures are contained in the appendix, which makes the paper very difficult to read and understand. The main text repeatedly references appendix figures for results, meaning the main text does not contain the story. I will list a few claims from the paper that are in the appendix:

  • Lines 187-191: "We identify two recurring attention patterns in the MR-loop transformer. The first is a traveling previous-token pattern: at loop t, the query at rt+2 attends strongly to rt+1, with this localized attention peak shifting along the sequence across successive loops (Figure 3 in Appendix C.1). The second is a fixed-query tracking pattern: the query remains at the [EQ] position (the answer readout position) while its attention shifts to rt+1 at loop t (Figure 4 in Appendix C.1). "
  • Lines 198-203: " The attention patterns above suggest an iterative computation. We next investigate whether intermediate values are recoverable during this computation and where they are represented. At each loop t, we measure whether the highest-probability prediction at positions rt, rt+1, and [EQ] matches the ground-truth intermediate entity et+1. As shown in Figure 5 from Appendix C.2, predictions at [EQ] recover et+1 with perfect accuracy, whereas accuracy at rt and rt+1 remains low. "
  • Lines 234-235: "The state-swap experiments in Appendix C.3 show that relation-token states causally influence which relation is selected."
  • Lines 262-265: "As illustrated by the attention patterns in two representative heads in Figures 6 and 7 from Appendix E.1, a localized region of elevated attention progresses toward later sequence positions along the main diagonal and the first subdiagonal, corresponding to attention to the current and immediately preceding positions, respectively."
  • Lines 453-455: "Composite-modulus polynomial tasks and FSC with non-bijective transitions can admit such compression, whereas prime-modulus polynomial tasks, fully bijective FSC, and our permutation-based KG construction preserve state distinctions (Appendix H)."

There are more examples in the different sections. As it stands, without any results tables or figures in the main text, the main text does not contain evidence for the full story the authors are telling.

This is my primary concern, and it can be addressed by reorganizing the paper so that there are figures in the main text.

Questions for authors

Are the mechanisms identified for MR-loop and DR-loop specific to looped models? I am curious how a non-looped / weight-untied model with matched depth would compare.

7 comments
AuthorsOpenReview ↗

Weakness regarding presentation: Result figures placed in Appendix

The paper is very difficult to follow because all the results figures are in the appendix... This is my primary concern, and it can be addressed by reorganizing the paper so that there are figures in the main text.

We thank the reviewer for the careful reading, and for stating that the primary concern "can be addressed by reorganizing the paper so that there are figures in the main text." We are currently working on this reorganization. We will upload the revised PDF around November 17 (AoE), and post a follow-up listing every change.

1. Presentation: result figures will move into the main text

We agree that the submitted version relied too heavily on the appendix for figures. Our reason was the volume of evidence: for two looped configurations across three task families, the claims rest on attention analyses, logit-lens decoding, causal interventions with random-direction controls, extreme-length analyses, and replications across tasks and independently trained models. Within the 9-page limit, we reported the quantitative results and controls in the text and placed the visualizations in the appendix.

Using the additional page allowed during discussion, we will reorganize the paper and move some of the most important figures into the main text.

2. On whether the main text contained the evidence

We would like to clarify one point. The submitted main text lacked result figures, but not results: the key quantitative findings and their controls were reported in the text, for example:

  • [EQ] decodes the evolving intermediate entity with 100% accuracy beyond the training length (L203–204).
  • Live–consumed direction at an unseen position (k = 14, n = 20): replacing the live state with its consumed state drops accuracy to 9.4%; adding the shared direction restores 93.8%, matching native performance, while a random direction of equal norm leaves it at 9.4% (L243–247).
  • DR-Loop activation patching on 14-hop queries (training depth: 11 hops): counterfactual rates of 90.6%, 71.9% and 73.4%, and 56.3% when patching at the unseen position 12 (L285–291).
  • Fresh–aged direction: next-hop score 0.650 (fresh) vs. 0.038 (aged); adding the direction restores 0.698, and subtracting it from a fresh state drops it to 0.031 (L308–312).
  • Extreme-length failure: transplanting a shorter prompt's [EQ] state raises next-hop accuracy from 5.5% to 83.6% (L373–375); the live–consumed cue, projected onto a fixed direction, decays from ~28 to 9.1 (L375–377); DR-Loop intermediate confidence falls from 0.982 to 0.713 (L378–380).
AuthorsOpenReview ↗

Question: Are the mechanisms specific to looped models?

Question: Are the mechanisms identified for MR-loop and DR-loop specific to looped models? I am curious how a non-looped / weight-untied model with matched depth would compare.

We thank the reviewer for this question, which has two parts.

Beyond the training horizon. Standard transformers often fail to generalize beyond their training length (Anil et al., 2022; Delétang et al., 2023; Kazemnejad et al., 2023) or to reasoning depths beyond those seen in training (Kohli et al., 2026); for implicit multi-hop reasoning over parametric knowledge, the number of layers required grows linearly with the number of hops (Yao et al., 2025). Looped transformers can instead extend computation by adding iterations at test time (Fan et al., 2025; Kohli et al., 2026). The beyond-horizon behavior of both mechanisms depends on weight sharing across depth by construction: the same parameters are reapplied for additional loops, including at positions never reached during training. A fixed-depth untied model has no additional iterations to apply at test time, so this part of the mechanism cannot arise in such a model. (Such a model might still handle longer inputs by other means, e.g., log-depth algorithms (Sanford et al., 2024), but not by reapplying the same update.)

Within the training distribution. An untied model of matched effective depth could in principle represent the in-distribution computation, although learning it may require substantially more data (Yao et al., 2025). Prior work suggests such a model would organize multi-hop computation by depth: untied transformers resolve intermediate entities layer by layer (Yao et al., 2025), and in a 2-hop setting, a logit-lens comparison with an untied model of matched effective depth shows that it decodes the bridge entity but fails to complete the second hop on facts never seen in compositions during training, where the looped model succeeds (Kohli et al., 2026). The key difference is what weight sharing forces. In a looped model, the same parameters implement the update at every iteration, so the operations we identify (MR-Loop's traveling previous-token and fixed-query tracking patterns, and DR-Loop's local composition at the advancing frontier) are implemented by the same weights at every iteration. This reuse is what makes it possible to apply them at loops never seen in training, although their reliability degrades at extreme lengths (Sec. 5). An untied model faces no such constraint: each layer can implement a hop-specific circuit, and nothing in the architecture requires these circuits to be shared across layers. We therefore expect that what is specific to looped models is the reuse of a single update across iterations, rather than any particular in-distribution circuit; testing this directly requires the comparison the reviewer suggests, which is a valuable furture direction that we can note in the Discussion.

Refered Paper:

Anil et al. (2022). Exploring Length Generalization in Large Language Models.

Fan, Du, Ramchandran & Lee (2025). Looped Transformers for Length Generalization.

Kohli, Parthasarathy, Sun & Yao (2026). Loop, Think, & Generalize: Implicit Reasoning in Recurrent-Depth Transformers.

Delétang et al. (2023). Neural Networks and the Chomsky Hierarchy.

Kazemnejad, Padhi, Natesan Ramamurthy, Das & Reddy (2023). The Impact of Positional Encoding on Length Generalization in Transformers.

Yao, Du, Zhu, Hahn & Koller (2025). Language Models Can Learn Implicit Multi-Hop Reasoning, but Only If They Have Lots of Training Data.

Sanford, Hsu & Telgarsky (2024). Transformers, Parallel Computation, and Logarithmic Depth.

AuthorsOpenReview ↗

Follow-up: Revised PDF uploaded with added figures and tables in main text

Follow-up: revised PDF uploaded

We have uploaded the revised PDF; all changes are highlighted in blue. As the reviewer suggested, we reorganized the paper so that the main text contains more result figures and tables supporting its claims: three result figures (Figures 1, 3, 4) and three result tables (Tables 2–4).

Passages quoted in the review

Passage in the reviewLocation in the revision
L187–191: MR-Loop attention patternsFigure 1 (Sec. 4.1): L0H1 traveling previous-token attention and L2H0 fixed-query tracking at passes t ∈ {1, 2, 12}, with the attended relation advancing past the last training relation r₁₁; full recurrences in Figures 6–7 (App. C.1)
L198–203: intermediate decoding at [EQ]Figure 3a (Sec. 4.1)
L234–235: state-swap experimentsSec. 4.1 now reports the state-swap results directly (100% counterfactual accuracy within the training range and 91.5–93.1% beyond it for relation-token swaps; 95.2% at the unseen hop k = 14 for [EQ] swaps), and Table 2 reports the live–consumed interventions within and beyond the training horizon
L262–265: DR-Loop attention patternsFigure 4 (Sec. 4.2): L1H1 and L0H0 attention bands advancing past r₁₁; full recurrences in Figures 9–10 (App. E.1)
L453–455: compressive vs. information-preserving tasksTable 4 (Sec. 6): length-generalization success by task structure, with each setting marked as compressive or not; full results in Table 7 (App. I.1)

Other additions

  • Figure 3b (Sec. 4.2): logit-lens view of DR-Loop's advancing intermediate-state frontier, including entities beyond the training depth.
  • Table 3 (Sec. 4.3): fresh–aged interventions for DR-Loop, within and beyond the training horizon.
  • Sec. 4.3 rewritten to connect both models' computational-status results to Tables 2–3 and Figure 5.

These changes use the additional page allowed during discussion and a more concise Section 4.3; no results were removed.

Since the reviewer indicated that this reorganization would address the primary concern, we hope the revised version does so, and we would be grateful if the reviewer would consider updating their assessment. We are happy to make further adjustments if anything else would help.

Reviewer_Xz6jOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper uses two existing looped transformer implementations (MR-Loop, based on Fan et al. (2025) & DR-Loop based on Kohli et al. (2026)) to apply mechanistic interpretability to investigate length generalization (e.g., reasoning chains that are longer than the ones encountered during training).

The work focusses specifically on how looped transformers achieve length generalization, when it breaks down, and if length generalization implies that the computation was faithful.

Critical strengths

S1: The paper goes beyond evaluating length generalization based on final-answer accuracy and investigates how the models perform the underlying computations and why they eventually fail.

S2: The authors combine attention analysis, intermediate-state decoding and causal interventions to study the learned mechanisms. In particular, the intervention experiments provide useful evidence for how information is processed across recurrent iterations.

Critical weaknesses

W1: MR-Loop and DR-Loop differ in several architectural and training choices (input reinjection, normalization, etc.). Without controlled ablations, it is difficult to determine which of these differences are responsible for the observed computational mechanisms. This limits the conclusions that can be drawn from comparing the two models.

W2: The experiments focus on small Transformers and synthetic tasks. While the authors provide evidence for their proposed mechanisms, the evaluation is somewhat limited. For example, the live–consumed mechanism is tested at later positions of the same knowledge-graph sequences used to identify it, rather than on independent examples. I think more extensive experiments with different training seeds and larger models would be useful to highlight how general these findings are.

W3: The paper builds on two existing looped Transformer architectures and several findings have already been reported in related work. For example, Kohli et al. (2026) already showed that excessive recurrent iterations can degrade performance, while Liang and Pan (2026) demonstrated that latent reasoning models do not necessarily follow the expected intermediate computation steps. The main novelty therefore seems to be in the mechanistic analysis of the two architectures. However, the extent to which these findings provide new insights beyond prior work remains somewhat unclear to me.

Questions for authors

See my weaknesses

5 comments
AuthorsOpenReview ↗

Weakness 1: Lack of controlled ablations

W1: MR-Loop and DR-Loop differ in several architectural and training choices (input reinjection, normalization, etc.). Without controlled ablations, it is difficult to determine which of these differences are responsible for the observed computational mechanisms. This limits the conclusions that can be drawn from comparing the two models.

We thank the reviewer for raising this point, which we address with new controlled experiments. We ran controlled 2×2 ablations of two factors that differ between the configurations and directly govern recurrent computation: input reinjection and recurrent-boundary LayerNorm. All other settings, the task, and the training recipe, including the curriculum, are held fixed within each family. The core organization of each family is preserved across all four variants: MR-Loop maintains its running state at a fixed [EQ] accumulator, and DR-Loop propagates intermediate states along an advancing frontier. This holds even in the crossover variants. MR-Loop with both factors disabled, matching DR-Loop’s original settings for these factors, still uses the [EQ] accumulator, and DR-Loop with both enabled, matching MR-Loop’s, still uses the frontier. Neither factor, alone or together, switches one family’s mechanism into the other’s.

As stated in Section 1, the MR/DR-loop labels are “descriptive shorthand for the complete model–training configurations, not a controlled recurrence ablation,” and Appendix A.3 notes that the observed differences “should be interpreted as differences between the complete model/training configurations rather than as effects of recurrence scheduling alone.” To make the scope of our study clearer, we will upload a revised PDF during the discussion period. We will explicitly clarify that MR-Loop and DR-Loop refer to two specific configurations adopted from prior work, rather than broad categories of looped transformers. We will also clarify in the paper that our mechanistic findings apply to these two configurations individually and are not intended to generalize to all looped transformers. We will also note in the conclusion that the complete controlled comparisons will be a valuable future work.

Experimental setup:

Input reinjectionBoundary LayerNorm offBoundary LayerNorm on
OffMR-Loop ablation; DR-Loop baselineMR-Loop ablation; DR-Loop ablation
OnMR-Loop ablation; DR-Loop ablationMR-Loop baseline; DR-Loop ablation

Results:

For MR-Loop, across all four settings, causal interventions support the running state at a fixed [EQ] accumulator. We paired queries with identical relation sequences but different starting entities and therefore different correct answers. Replacing the recipient’s [EQ] state with the donor’s state at a nonterminal recurrent step, then resuming normal computation, redirected the final answer to the donor’s target on 97.7–100% of 512 independent 15-hop confirmation pairs across the four variants. Applying the same intervention at relation-token positions selected on separate discovery examples never redirected the answer to the donor’s target (0/512 per variant). Moreover, successive intermediate entities remain decodable at [EQ] across passes.

For DR-Loop, across all four settings, causal interventions support the propagating state frontier. On 256 independent 14-hop confirmation pairs, transplants at separately selected relation positions produce the donor’s final answer on 24.6–83.6% of examples, whereas selected early-PAD transplants yield 0/256 in every variant. Transplant effectiveness nevertheless varies substantially. With both reinjection and boundary normalization enabled, donor-answer recovery is 24.6%, despite 91.0% native accuracy, and remains 26.9% among pairs where both donor and recipient queries succeed without intervention. So the broad organization is preserved without an identical implementation.

The computational-status tests also extend beyond the original checkpoints. In all four MR architectural variants, independently fitted status directions improve local transitions on independent 15-hop examples while relation-identity probe accuracy remains 100% for both consumed and repaired states. For example, with both reinjection and boundary normalization removed, repair raises transition accuracy from 11.9% to 89.6%. DR shows architecture-dependent repair effectiveness: at the tested later position, the fitted direction improves next-step accuracy in three of the four variants. We will include these results in the final version.

AuthorsOpenReview ↗

Weakness 2: Limited Evaluation

W2: The experiments focus on small Transformers and synthetic tasks. While the authors provide evidence for their proposed mechanisms, the evaluation is somewhat limited. For example, the live–consumed mechanism is tested at later positions of the same knowledge-graph sequences used to identify it, rather than on independent examples. I think more extensive experiments with different training seeds and larger models would be useful to highlight how general these findings are.

Thank you for this suggestion. We ran three sets of new experiments: evaluation on independent examples, replication across separately trained checkpoints, and replication in larger models.

Independent examples. We evaluated the original MR-Loop checkpoint on 384 fresh KG walks, disjoint from the walks and prefixes used to fit the live–consumed directions, which were frozen before evaluation. At later keys 14 and 17, mean next-hop accuracy was 88.41% natively, 9.11% after consumed-state replacement, and 89.32% after adding the frozen directions; all five norm-matched random-direction controls remained at 9.11%.

Separately trained seeds. Three MR-Loop checkpoints trained with different seeds reproduce the fixed [EQ] accumulator (Section 4.1) and the live–consumed effect: with separate fitting and validation data and 256 held-out walks per checkpoint, correction raises later-hop accuracy from approximately 11% to 100.0%, 95.1%, and 100.0%. Three DR-Loop checkpoints trained with different seeds, each evaluated on 256 fresh held-out walks, reproduce the advancing decoded-state frontier and the fresh–aged dependence of composition (Section 4.2). At the later receiving position, aged donors retain 98.4–100% entity-decoding accuracy but support only 0.8–2.7% counterfactual next-hop accuracy; adding separately fitted directions raises this to 64.8–77.7%, while random-direction and norm-rescaling controls do not reproduce the recovery.

Larger models. We trained MR-Loop and DR-Loop models with four shared blocks, 12 heads per block, and width 768 (about 12× the original non-embedding parameters). The larger MR-Loop reproduces the live–consumed effect on 256 independent walks, with correction restoring next-hop accuracy to 86.72%, matching native accuracy, and it reproduces the fixed [EQ] accumulator: swapping only its [EQ] state yields the donor’s counterfactual final answer in 98.44% of examples. The larger DR-Loop exhibits an advancing frontier; in an intervention assay on 256 held-out pairs, correcting aged states at an early site raises counterfactual final-answer accuracy from 0.39% to 53.52%, close to the fresh-state baseline of 56.25%, while preserving each example’s probe-decoded predecessor identity.

Together, these results show that the identified mechanisms are not artifacts of the fitting examples, a single training run, or the original model size. We will include them in the final version.

AuthorsOpenReview ↗

Addressing Weakness 3 to make Novelty and Added Insight of this paper more clear

W3: The paper builds on two existing looped Transformer architectures and several findings have already been reported in related work. For example, Kohli et al. (2026) already showed that excessive recurrent iterations can degrade performance, while Liang and Pan (2026) demonstrated that latent reasoning models do not necessarily follow the expected intermediate computation steps. The main novelty therefore seems to be in the mechanistic analysis of the two architectures. However, the extent to which these findings provide new insights beyond prior work remains somewhat unclear to me.

We thank the reviewer for raising this issue. Our main novelty beyond the existing work is to identify what enables recurrent computation to remain reusable beyond the training horizon, and how that reuse breaks down:

  • Computational status as a shared requirement for recurrent reuse (Section 4.1–4.3). Our central insight is that recurrent computation depends on both what a state represents and its computational status—whether that representation remains usable for the next operation. This status takes different forms in the two models: live versus consumed relation states govern selection in MR-Loop, while fresh versus aged entity states govern subsequent composition in DR-Loop. Crucially, correct semantic content can persist after computational usability is lost: an aged DR-Loop state retains 96.9% entity-decoding accuracy, yet its next-hop score falls to 0.038. Adding a shared fresh–aged residual direction restores this score to 0.698; the direction also transfers across source positions and unseen hops. These scores are peak batch-mean probabilities of the counterfactual next entity. The interventions therefore establish computational status as a causally manipulable property of recurrent representations, revealing why preserving the correct intermediate value alone is insufficient for continued computation.

  • Mechanisms explain degradation and limited self-correction (Section 4–5). We identify distinct reasons why additional recurrence need not repair an error. MR-Loop successively overwrites its intermediate state in a fixed accumulator as relation selection advances, leaving no demonstrated mechanism for retrieving a valid earlier state and restarting the failed transition. DR-Loop retains earlier intermediates across positions, but their computational usability can deteriorate even while their identities remain decodable. These retained values therefore do not automatically serve as usable checkpoints: replay requires restoring the predecessor’s freshness, preparing the receiving state, and redirecting computation to the failed hop. This distinction between retaining information and retaining the ability to recompute helps explain the observed limited self-correction. Complementing this account, externally replacing MR-Loop’s incoming readout state with its matched shorter-prefix counterpart raises mean next-hop accuracy from 5.5% to 83.6%, with relation-key states unchanged. Together, these results identify both an internal source of degradation and mechanistic constraints on recovery beyond the previously reported overthinking phenomenon.

  • Length generalization can conceal incomplete intermediate computation (Section 6). Liang and Pan (2026) study CODI-style latent CoT and explicitly leave open whether their findings extend to looped Transformers as future work. We address this gap in two distinct looped-Transformer configurations trained solely on final answers, showing that both can successfully extrapolate to longer computational chains while many intended intermediate states remain weakly recoverable. This challenges a natural interpretation of length generalization: success beyond the training horizon does not, by itself, establish faithful execution of the underlying step-by-step algorithm. We further apply their compressibility framework to explain when task structure permits shortcut solutions in looped computation, while establishing its limits: compressibility makes such solutions available but does not determine whether the learned trajectory follows them. Our experiments reveal both substantially complete and incomplete intermediate trajectories in compressive settings. This provides a concrete diagnostic approach for mechanistic evaluation—combining controlled task compressibility with intermediate-state tracking, rather than relying on extrapolation accuracy alone—and motivates monitoring internal computation when evaluating and scaling recurrent reasoning systems.

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

The paper investigates how looped Transformers achieve length generalization and why their performance eventually deteriorates. Through mechanistic analyses of two looped-Transformer configurations, the authors identify distinct recurrent computation mechanisms and show that successful reasoning depends on both the semantic content and computational usability of intermediate representations. They further find that state degradation and limited self-correction contribute to extreme-length failures, while successful length generalization does not necessarily imply faithful step-by-step reasoning.

Critical strengths

The paper addresses an important and timely question about the mechanisms and limitations of length generalization in looped Transformers. A key strength is its detailed mechanistic analysis, combining attention visualization, intermediate-state decoding, and causal interventions to provide convincing evidence for the identified computational mechanisms. The distinction between semantic content and computational usability of intermediate representations is particularly interesting and offers useful insights into recurrent reasoning.

Critical weaknesses

Major Weaknesses

1. Lack of controlled comparisons makes it difficult to explain the emergence of different mechanisms.

The paper identifies distinct computational mechanisms in MR-Loop and DR-Loop, but the two configurations differ in numerous factors beyond their recurrence-training schemes, including input reinjection, normalization, output heads, readout positions, masking, and training curricula (Appendix A, Table 2). Although the authors acknowledge these differences, no controlled ablations are provided to investigate which factors lead to the observed mechanisms. Consequently, it remains unclear whether the fixed-accumulator and propagating-frontier mechanisms arise from differences in recurrence scheduling, architectural choices, or other training details. This limits the explanatory value of the comparison and makes it difficult to draw generalizable conclusions about how different looped-Transformer designs affect length generalization.

2. The analysis does not fully establish the underlying causes of extreme-length generalization failure.

A central claim of the paper is that length generalization eventually fails because computational states degrade and the models lack effective self-correction. While the authors provide evidence of deteriorating representations and persistent errors, the analysis largely characterizes these failure phenomena rather than explaining their underlying causes. For example, transplanting the [EQ] state from a shorter prompt substantially improves local transition accuracy in MR-Loop, but it remains unclear why the original state deteriorates—whether due to interference from longer contexts, unstable recurrent dynamics, or other factors. More importantly, although the proposed computational-status directions can restore individual transitions, the paper does not demonstrate that preserving or restoring these states over an entire trajectory improves extreme-length generalization. Such end-to-end interventions would provide stronger causal evidence that the identified mechanisms are responsible for the observed failures, rather than merely being associated with them.

3. Limited novelty and insufficient mechanistic evidence for the faithfulness findings (RQ3).

The finding that successful length generalization does not necessarily imply faithful step-by-step computation substantially overlaps with existing work, particularly Liang & Pan (2026), which already studies intermediate-state faithfulness and task compressibility on closely related polynomial and finite-state composition tasks. Although the authors acknowledge and build upon this framework, the additional insights contributed by RQ3 are not sufficiently clear. Furthermore, the evidence for unfaithful computation relies primarily on weak recovery of ground-truth intermediate states using the model's output readout. However, low logit-lens decodability does not necessarily indicate that the model fails to compute the intended intermediate states, as this information could be encoded in alternative or distributed representations. The paper also does not directly identify or causally validate the shortcut computations purportedly used by these models. Therefore, the results establish a gap between final-answer accuracy and intermediate-state decodability, but provide weaker evidence for the stronger interpretation that the models actually perform unfaithful or shortcut reasoning.

Questions for authors

see above

4 comments
AuthorsOpenReview ↗

Weakness 1: Lack of controlled comparisons

W1. Lack of controlled comparisons makes it difficult to explain the emergence of different mechanisms.

We thank the reviewer for raising this point, which we address with new controlled experiments. We ran controlled 2×2 ablations of two factors that differ between the configurations and directly govern recurrent computation: input reinjection and recurrent-boundary LayerNorm. All other settings, the task, and the training recipe, including the curriculum, are held fixed within each family. The core organization of each family is preserved across all four variants: MR-Loop maintains its running state at a fixed [EQ] accumulator, and DR-Loop propagates intermediate states along an advancing frontier. This holds even in the crossover variants. MR-Loop with both factors disabled, matching DR-Loop’s original settings for these factors, still uses the [EQ] accumulator, and DR-Loop with both enabled, matching MR-Loop’s, still uses the frontier. Neither factor, alone or together, switches one family’s mechanism into the other’s.

As stated in Section 1, the MR/DR-loop labels are “descriptive shorthand for the complete model–training configurations, not a controlled recurrence ablation,” and Appendix A.3 notes that the observed differences “should be interpreted as differences between the complete model/training configurations rather than as effects of recurrence scheduling alone.” To make the scope of our study clearer, we will upload a revised PDF during the discussion period. We will explicitly clarify that MR-Loop and DR-Loop refer to two specific configurations adopted from prior work, rather than broad categories of looped transformers. We will also clarify in the paper that our mechanistic findings apply to these two configurations individually and are not intended to generalize to all looped transformers. We will also note in the conclusion that the complete controlled comparisons will be a valuable future work.

Experimental setup:

Input reinjectionBoundary LayerNorm offBoundary LayerNorm on
OffMR-Loop ablation; DR-Loop baselineMR-Loop ablation; DR-Loop ablation
OnMR-Loop ablation; DR-Loop ablationMR-Loop baseline; DR-Loop ablation

Results:

For MR-Loop, across all four settings, causal interventions support the running state at a fixed [EQ] accumulator. We paired queries with identical relation sequences but different starting entities and therefore different correct answers. Replacing the recipient’s [EQ] state with the donor’s state at a nonterminal recurrent step, then resuming normal computation, redirected the final answer to the donor’s target on 97.7–100% of 512 independent 15-hop confirmation pairs across the four variants. Applying the same intervention at relation-token positions selected on separate discovery examples never redirected the answer to the donor’s target (0/512 per variant). Moreover, successive intermediate entities remain decodable at [EQ] across passes.

For DR-Loop, across all four settings, causal interventions support the propagating state frontier. On 256 independent 14-hop confirmation pairs, transplants at separately selected relation positions produce the donor’s final answer on 24.6–83.6% of examples, whereas selected early-PAD transplants yield 0/256 in every variant. Transplant effectiveness nevertheless varies substantially. With both reinjection and boundary normalization enabled, donor-answer recovery is 24.6%, despite 91.0% native accuracy, and remains 26.9% among pairs where both donor and recipient queries succeed without intervention. So the broad organization is preserved without an identical implementation.

The computational-status tests also extend beyond the original checkpoints. In all four MR architectural variants, independently fitted status directions improve local transitions on independent 15-hop examples while relation-identity probe accuracy remains 100% for both consumed and repaired states. For example, with both reinjection and boundary normalization removed, repair raises transition accuracy from 11.9% to 89.6%. DR shows architecture-dependent repair effectiveness: at the tested later position, the fitted direction improves next-step accuracy in three of the four variants. We will include these results in the final version.

AuthorsOpenReview ↗

Weakness 2: Not fully establish the underlying causes of extreme-length generalization failure.

W2. The analysis does not fully establish the underlying causes of extreme-length generalization failure.

Thank you for this suggestion. We ran new experiments addressing both points: why the [EQ] states degrade, and whether restoring them over an entire trajectory recovers extreme-length accuracy.

Summary. In MR-Loop, (1) blocking [EQ]’s attention to not-yet-needed relations restores local transitions at hops 12–19 (5.71% → 80.78%), identifying interference from the longer context as a cause of [EQ] degradation; (2) a second, depth-dependent limitation lies in the late relation pathway; and (3) restoring both states throughout the trajectory, without ground-truth intermediates, raises final-answer accuracy at length 30 from 1.30% to 96.09%. In DR-Loop, fixed status corrections applied during native inference recover final-answer errors on an independent cohort.

MR-Loop

Cause 1: context interference at [EQ]. On 768 new length-30 prompts, we blocked only [EQ]’s attention to relations r20,…,r29r_{20}, \ldots, r_{29} during the loops preceding each tested hop k=12,…,19k = 12, \ldots, 19, then evaluated that transition with full attention restored and native relation-token states. Mean next-hop accuracy rose from 5.71% to 80.78%, without supplying ground-truth intermediate entities. Because the number of recurrent passes and the relation states are unchanged, this indicates that [EQ] degradation at these hops is driven primarily by interference from the longer context rather than by recurrent depth itself.

Cause 2: late relation states. Protecting [EQ] alone is not sufficient at greater depth. On 384 new examples, repairing [EQ] at the final transition (hop 29) [using the ground-truth entity’s state] yields only 1.30% accuracy. Additionally refreshing the required relation’s keys and values, taken from a short reference run with the same relation sequence, raises accuracy to 97.66%, with all original attention positions retained. Together, these identify two distinct causes: length-dependent interference at [EQ] and depth-dependent degradation of late relation states, consistent with the weakening progress cue in Appendix G.1.2.

End-to-end repair without ground truth. We built canonical [EQ] vectors by averaging short-run states grouped by the model’s own predicted entity. During long inference, we repeatedly re-encode the model’s current prediction with its canonical vector and refresh relation keys and values from short prompts containing the input relations with independently sampled starting entities. Neither calibration nor inference accesses ground-truth intermediate states, and the final pass, which produces the answer, receives no [EQ] replacement.

Condition (seperate cohort, n = 384)Final-answer accuracy
Native MR-Loop1.30%
[EQ] refresh only0.78%
Relation K/V refresh only0.78%
Joint [EQ] and relation K/V refresh96.09%

Restoring these two states is sufficient to recover nearly all extreme-length accuracy, indicating that the failure stems from degradation of the states the transition relies on, not from the transition computation itself. This recovery uses explicit decoding and re-encoding, a known hop schedule, and auxiliary computation; it is a causal test of our failure account, not evidence of native self-correction.

DR-Loop

We tested whether restoring computational status during native inference improves the original final answer. We fitted three fresh-minus-aged correction directions on 64 separate KG walks. At every recurrence boundary after the first pass, a controller applies bounded corrections to the current states in a moving three-position window, moving them toward their fresh-reference projections. It never accesses clean evaluation-state snapshots or ground-truth answers.

The directions and policy were fixed before evaluation on a fresh cohort of 128 walks, with no overlap in their first three tokens with the fitting or development sets. At 19 hops (where failures are frequent but not universal), the prespecified policy improved final-answer accuracy from 44/128 to 50/128, recovering six errors without introducing any. Status interventions can therefore causally improve final outcomes in DR-Loop as well.

These results also bear on the paper’s account of limited self-correction (Section 5; Appendix G.3). Computational status helps explain why additional recurrence does not repair failed transitions: even when earlier intermediates remain decodable, recomputing a failed hop requires both a usable predecessor state and relation states prepared for that transition, and native recurrence need not restore either.

We will include these results in the final version of the paper.

AuthorsOpenReview ↗

W3: Limited novelty and insufficient mechanistic evidence for the faithfulness findings (RQ3).

W3. Limited novelty and insufficient mechanistic evidence for the faithfulness findings (RQ3).

Thank you for these comments. We address novelty first, then the concern about readout-only evidence.

Novelty relative to prior work, especially, Liang & Pan (2026):

  1. We address an architectural extension explicitly left open by prior work. Liang & Pan (2026) study CODI-style latent CoT, where teacher–student distillation produces a fixed sequence of continuous thought tokens. Their limitations section explicitly identifies extension to looped Transformers as future work. We test whether their findings extend to two distinct looped configurations that reuse shared computation across recurrent passes.

  2. We reveal a surprising gap between length generalization and faithful iterative computation. Successfully solving reasoning chains longer than those seen during training suggests that a model has learned to faithfully repeat the underlying transition algorithm. Our findings challenge this idea. Across both MR-Loop and DR-Loop, strong length extrapolation can coexist with poor recovery of intended intermediate states, while our new intervention experiments reveal limitations in their causal reuse. Thus, strong extrapolation does not guarantee faithful internal computation, making internal monitorability a distinct requirement for evaluating recurrent reasoning systems.

  3. We apply the compressibility framework to looped computation and characterize its limits. We use Liang & Pan’s (2026) framework to explain how compressive task structure can enable correct length extrapolation without preserving every intended intermediate state. We also characterize its limits: similarly compressive FSC settings exhibit markedly different intermediate-state recovery. Thus, compressibility explains why shortcut solutions are available, but does not determine whether a model learns them—motivating intermediate-state tracking and causal tests alongside extrapolation accuracy.

Counterfactual interchange (MR-Loop). To move beyond readout-only evidence, we ran held-out interchange experiments on KG traversal and polynomial iteration mod 40. These tasks differ in information-preservation requirements: KG’s permutation relations preserve state distinctions, whereas mod 40 permits many-to-one compression. On length-15 inputs, beyond the training length, we transplanted a donor’s intermediate state at step k ∈ {3, 5, 7, 9, 11} into a recipient and tested whether the answer matched the symbolic counterfactual, i.e., the donor’s state continued with the recipient’s remaining operators. We kept only pairs where this target differs from both original answers, excluding simple copying. Counterfactual agreement was 98.6–99.6% on KG but 46.1–62.2% on mod 40 (0% and 0–0.5% without intervention). The contrast persisted among pairs where the recipient, donor, and hybrid programs were all answered correctly (99.6–100% vs. 46.6–62.4%).

Readability tracks reuse. On Polynomial mod 40 task, transfer succeeded in 67.1–86.4% of cases where the intermediate state at the transplant step was correctly decoded, versus 38.0–42.7% otherwise. At k = 11, for example: 228/281 (81.1%) vs. 90/230 (39.1%), a 42.0-point difference. This post-hoc analysis suggests that low decodability reflects reduced functional reuse, not merely a limitation of the readout.

DR-Loop. Across the same five steps, counterfactual transfer was lower on mod 40 than on KG (13.7% vs. 46.9%, averaged equally across steps), even though native final-answer accuracy was higher on mod 40 (96.9%, 62/64, vs. 81.3%, 52/64). The gap persisted among all-correct pairs (13.9% vs. 51.2%). High final-answer accuracy therefore does not require reusable task-aligned states at these positions.

These results support a stronger conclusion than the original readout analysis alone: successful extrapolation can coexist with limited causal reuse of task-aligned intermediate representations, and state readability is associated with that reuse. Extrapolation accuracy alone therefore cannot certify faithful iterative computation, which we define as intermediate states that are both recoverable under the model’s readout and causally reusable where the computation carries them. We will include these new results in the final version.

AuthorsOpenReview ↗

Follow-up Before Discussion Closes: Summary of New Experiments

Dear Reviewer LLKN,

Thank you again for your detailed review. Your suggestions directly shaped the new experiments we ran during the discussion period. As the discussion closes on November 18, we wanted to briefly check whether our responses address your three concerns:

  • W1 (controlled comparisons): we ran 2×2 ablations of input reinjection and recurrent-boundary LayerNorm within each family, holding the task and training recipe fixed. Neither factor, alone or together, switches one mechanism into the other. In all four MR-Loop variants, swapping the [EQ] state redirects the final answer on 97.7–100% of 512 held-out pairs (0/512 at relation positions); in all four DR-Loop variants, frontier transplants redirect it on 24.6–83.6% of pairs (0/256 at early PAD positions). We have also clarified in the revised PDF that MR-Loop and DR-Loop denote two specific configurations from prior work, not general classes of looped Transformers.

  • W2 (causes of failure): your review asked whether restoring the identified states over an entire trajectory improves extreme-length generalization. It does. We identified two causes: context interference at [EQ] (blocking it restores local transitions from 5.71% to 80.78%) and degradation of late relation states. Restoring both throughout inference, without ground-truth intermediates, raises length-30 final-answer accuracy from 1.30% to 96.09%, versus 0.78% when either is restored alone. In DR-Loop, a frozen status-correction policy improves final-answer accuracy on a fresh cohort from 44/128 to 50/128 without introducing errors.

  • W3 (novelty and readout-only evidence): Liang & Pan (2026) explicitly leave looped Transformers as future work; beyond extending their analysis, we show where the compressibility account stops predicting the learned trajectory. To go beyond decodability, counterfactual interchange experiments test causal reuse directly: MR-Loop reuses intermediate states on 98.6–99.6% of KG pairs but only 46.1–62.2% on compressive mod 40, and DR-Loop reuses mod-40 states less (13.7% vs. 46.9%) despite higher final-answer accuracy. We have scoped the claim accordingly: extrapolation accuracy alone cannot certify faithful computation, while identifying the specific shortcut circuit remains future work.

If our responses have addressed your concerns, we would be grateful if you would consider updating your assessment. If any concern remains, we would be glad to clarify before the discussion closes.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of the rebuttal and revisions for the AC and all reviewers

We sincerely thank the AC and the reviewers for their careful reviews and engagement. Below are the main concerns, what we added, and where each result appears in the revised PDF (changes in blue). All numbers are from the current PDF.

First of all, we sincerely thank the reviewers for recognizing the value of our mechanistic approach. In particular, we appreciate their acknowledgment of (1) our combination of attention analysis, intermediate-state decoding, and causal interventions (LLKN, Xz6j); (2) our effort to go beyond final-answer accuracy to explain how looped Transformers compute and why they fail (Xz6j); and (3) our distinction between a state's semantic content and its computational usability (LLKN). We also appreciate CYqb's recognition that our research questions are clear and that the paper presents substantial work.

1. Controlled comparison (LLKN W1, Xz6j W1). We trained all four combinations of input reinjection and recurrent-boundary LayerNorm within each family, with the task and curriculum fixed. No combination switches one mechanism into the other. [EQ] transplants redirect MR-Loop’s answer in 97.7–100% of pairs, and relation-site transplants never do (0/512). Relation-site transplants redirect DR-Loop’s answer in 24.6–83.6% of pairs, and early-PAD transplants never do (0/256). The text now states that MR-Loop and DR-Loop are two specific configurations from prior work, not general classes. → Sec. 4.4, App. K.

2. Causes of extreme-length failure (LLKN W2). In MR-Loop at length 30, blocking [EQ]’s attention to r20–r29 in earlier passes raises next-hop accuracy at hops 12–19 from 5.7% to 80.8%, which identifies interference from the longer context. Late relation states also degrade. Restoring both states throughout inference, without ground-truth intermediates, raises final-answer accuracy from 1.30% to 96.09%; restoring either alone gives 0.78%. In DR-Loop, fixed status corrections give a modest gain (44/128 → 50/128). → Sec. 5, Fig. 6, App. G.5.

3. RQ3 evidence beyond readouts (LLKN W3). We added held-out counterfactual interchange on length-15 inputs. MR-Loop transfers the donor’s state in 98.6–99.6% of KG pairs but in only 46.1–62.2% on compressive polynomial mod 40. On mod 40, transfer succeeds far more often when the transplanted state is decodable (67.1–86.4% vs. 38.0–42.7%). DR-Loop transfers 46.6% (KG) vs. 13.2% (mod 40) despite similar native accuracy. Low decodability therefore tracks reduced causal reuse, not only readout limits. → Sec. 6, Fig. 7, App. M.

4. Clarify the Novelty of Paper (LLKN W3, Xz6j W3). Kohli et al. show that excess recurrence can hurt; we localize why extrapolation fails (RQ 2). Liang & Pan study CODI-style latent CoT and leave looped Transformers to future work. We extend their analysis to looped transformers. Our central new finding are following:

(A), Correct content is not enough: computational status governs recurrent reuse (§4.1–4.3). A recurrent state must stay usable for the next operation, not just encode the right value. In MR-Loop this is live vs. consumed relation states; in DR-Loop, fresh vs. aged entity states. An aged DR-Loop state still decodes as the correct entity (96.9%) but barely supports the next hop (0.038 vs. 0.650 when fresh). A shared fresh–aged direction restores it (0.698), and a live–consumed direction does the same in MR-Loop, both beyond the training horizon.

(B), Why extrapolation breaks and isn’t repaired (§4–5). The mechanisms we identify, together with computational status, explain why the models rarely self-correct. As a result, an error at any step carries through to the final answer.

(C), Length Generalization doesn’t certify faithful computation (§6). We show that, in several task settings, both models can extrapolate while many intermediate states remain weakly recoverable, so length generalization does not imply faithful computation. Compressibility makes shortcuts available but does not decide whether a model takes them, so evaluation should track intermediate states, not just extrapolation accuracy.

5. Generality (Xz6j W2). The mechanisms recur on 384 held-out walks (frozen directions: 9.11% → 89.32%; native 88.41%), in three separately trained checkpoints per model on different graphs, and in ~12× larger models. In the larger models, an MR-Loop [EQ] swap yields the donor’s answer in 98.44% of pairs, and DR-Loop correction raises accuracy from 0.39% to 53.52% (fresh: 56.25%). → Sec. 4.4, App. L.

6. Presentation (CYqb). Main text now has more figures and tables for better presentations:

  • RQ1: Figs. 1, 3, 4 and Tables 2–3.
  • RQ2: the four-panel Fig. 6, added after CYqb’s final comment.
  • RQ3: Table 4, Fig. 7 and the interchange results.

Thanks for all the valuable feedbacks!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.