acceptodds
Under review as a conference paper at ICLR 2027

Do More Loops Follow the Evidence?

Abstract

More inference-time computation can improve looped language models, but the same performance gain can reflect different roles for added computation. Does another recurrent step make the model begin using facts in the prompt, or merely widen the score gap around a choice already based on them? To answer this question, we introduce a framework that redirects the relational path supporting one candidate toward the other while keeping the question, entities, relation names, and answer pair fixed. We track which candidate wins and the score gap between candidates at successive outputs. We evaluate the framework across several recurrent systems using controlled fictional graphs and benchmark-derived multi-hop questions. Controlled fictional graphs reveal three observed profiles: acquisition, where choice control emerges at later outputs; sharpening, where control is already present and remains stable while the margin grows; and margin-control divergence, where the margin grows even as choice control declines. Repair and structural controls confirm that these effects follow the path that determines the answer rather than unrelated prompt changes. Yet control over a fixed answer pair does not guarantee reliable free-form generation. Our results show that performance gains alone cannot explain what additional recurrent computation contributes. Evaluating choice and score together reveals how evidence control changes across outputs rather than treating a larger score gap as stronger control.

Primary area:
interpretability and explainable AI
Keywords:
recurrent language models, test-time computation, multi-hop reasoning, recurrent depth, mechanistic interpretability, counterfactual interventions, language model reasoning

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_EcPzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates how additional recurrent computation affects evidence use in looped language models. It evaluates multiple model families on hundreds of synthetic graph instances and question pairs derived from two multi-hop QA benchmarks, supplemented by 64 controlled training runs.

Critical strengths

Interesting perspective on looped architectures. The paper examines the relatively underexplored relationship between recurrent computation and evidence use, offering novel and interesting observations on the behavior of looped language models.

Critical weaknesses
  1. Limited mechanistic explanation. The paper primarily documents observed phenomena, with limited mechanistic explanation or targeted experiments to account for differences across models.

  2. Fragmented presentation. Quantitative findings are scattered across figures and appendices, with insufficient synthesis in the main narrative, making the overall evidence difficult to assess.

Questions for authors

Please find the weakness above.

1 comment
AuthorsOpenReview ↗

Follow-up on explanation and presentation

Dear Reviewer,

Thank you again for your feedback. Our revision now strengthens both the explanation and its presentation:

  • A connected main argument. Tables 1–2 now bring the quantitative comparisons together. Sections 3.2 and 5.1 explain the principal findings, and the Discussion connects them to the state-intervention evidence. We also reorganized Sections 2–5 into clear subsections and standardized the terminology.
  • Why margins grow while answers deteriorate. Section 3.2 traces the margin growth to answers that remain correct and the losses to decisions near the boundary. On fresh Phi-4 worlds, a fitted score shift predicts 68/85 harmful and 47/58 beneficial changes on held-out relations.
  • A controlled training explanation. Section 5.1 reports 64 runs across two architectures, four supervision schedules, and eight seeds. Earlier supervision moves evidence-following behavior earlier, although every run reaches at least 99.83% final accuracy.

Since that revision, our November 14 follow-up also reports K1→K8 divergence in Qwen on 288 fresh worlds, extending the evidence beyond Phi-4.

We would welcome your assessment of these revisions.

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

The paper proposes a counterfactual intervention framework to study whether additional recurrent computation improves a model's ability to use the evidence in the prompt. It studies questions with binary choices and shows that different recurrent models behave differently.

Critical strengths

The paper studies an interesting problem: what do recurrent models do when looping. It designed a counterfactual evaluation framework and designed metrics to study this quantitatively.

Critical weaknesses
  • The presentation of the paper is hard to follow; there are a lot of defined quantities that don't necessarily need a name, e.g. "choice effect", "choice gain", ... The organization of the paper is also not good, e.g. there are no subsections at all, only many paragraphs with bolded titles.
  • The paper studies only questions with binary choices, which limits the findings' generalizability.
  • The paper showed that different recurrent models behave differently under the proposed experimental framework, but it's unclear what contribution/insights the proposed experiments provide. The overall experimental design is hard to follow as it's currently presented, and the results/insights we get from these experiments are mostly descriptive, and there are limited explanations provided to help understand why different models behave differently.
Questions for authors
  • The three version of the graphs, "original" "swapped" "repaired", since the entities are fake, these three versions are essentially symmetric, what the comparison between the variants tries to show is unclear to me. Why are the quantities defined in Equ. 1 meaningful if orig and swap are essentially symmetrical?
1 comment
AuthorsOpenReview ↗

Follow-up on presentation, setup, and four-answer results

Dear Reviewer,

Thank you again for your feedback. We wanted to highlight the changes addressing your concerns in our revision:

  • Presentation and setup. Sections 2–5 now have clear subsections, consistent terminology, and consolidated results in Tables 1–2. The worked example in Figure 1 and Sections 2.1–2.2 explains why rewiring changes the required answer to the same question, and introduces the fixed reference before the equations.
  • Beyond binary choices. On 240 separate Phi-4 worlds with four answer options, correctness in both original and swapped prompts falls from 60.05% to 58.02%, while the focal-pair margin effect grows (Section 5.2).
  • Explanation. Section 3.2 shows how growing margins on unchanged correct answers coexist with losses near the decision boundary. Section 5.1 tests how supervision changes when evidence-following behavior appears, and the Discussion brings these findings together.

Since that revision, our November 14 follow-up also reports K1→K8 divergence in Qwen on 288 fresh worlds, extending the evidence beyond Phi-4.

We would welcome your assessment of these revisions.

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

The manuscript investigates how more recurrent steps in looped language models affect their ability to use evidence in the prompt. The experiments identify 3 patterns when increasing the number of recurrent steps: improvement in the evidence use, increasing candidate-score margins with approximately stable choice control, a divergence between margin growth and choice control. Structural controls show the response follows the answer-bearing path rather than the edit itself, supervision schedule shifts when control appears, and fixed-choice control does not carry over to free-form generation.

Critical strengths
  • The experimental setting is very minimal and makes it possible to ask clearly the questions the manuscript does. Experiments are sound, with error bars shown.
  • The "acquisition" phenomenon is interesting.
  • The manuscript is overall clear and easy to follow. When reading, most questions that came to mind were answered. There are some very minor points where I think it could be improved further:
    • Throughout the manuscript the definitions "choice control", "path control", "evidence control" are used to define really similar concepts. In my opinion this is a bit confusing and possibly could be improved. It slightly increases the difficulty of the reading when starting to read the paper.
    • In "Setup and experimental contract." I would add a reference to Fig 1.
    • The paragraph "Models and readout interfaces.", in my opinion, is a bit unclear. It starts by defining "output native" and "trained exits" in the first two sentences, then switches to the unrelated notation of recursions. Later paragraphs return to explaining the readout mode. Maybe it would be better to just use this paragraph for defining the notation for the recursion?
Critical weaknesses
  • The evidence in support of the "Margin-control divergence" phenomenon is weak. The opposite sign changes are statistically supported, but they are observed on one model on one fictional task. The effect is also very modest. Aggregate metrics on multiple questions do not establish whether this is a recurring pattern or explain the phenomenon driving it.
Questions for authors
  • Could acquisition and sharpening simply reflect different stages of the same internal calculation? What additional insight does this distinction provide beyond changes in the accuracy and prediction margins? Do you think the acquisition phenomenon consistently disappears as the model capability increases for a fixed task?
  • Do you have additional evidence of the "Margin-control divergence" on other models? Have you analyzed individual questions to see how the divergence originates? Does a common positive rescaling account for the margin increase in Phi-4? After accounting for scaling, which examples drive the decline in choice control, and are they concentrated near the decision boundary?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.