acceptodds
Under review as a conference paper at ICLR 2027

Repeat or Retrieve? Local Copies Suppress Remote-Source Sensitivity

Abstract

Reasoning traces often state an intermediate value twice: once where it is computed and again where it is used. Does the next prediction depend on the earlier source or the nearby copy? We separate these occurrences with source-span activation patching and controlled edits to the copy. Across Gemma-2, OLMo-2, and Qwen3, a correct local copy reduces the mean absolute source-patch effect by more than % on controlled arithmetic traces. Randomized trace formats and length-matched filler controls show that the reduction depends on copied values. Controlled training shows that this preference depends on the reliability of copies. In small transformers trained from scratch, corrupting one operand on half of the restated training lines reverses the preference at fixed restatement frequency: copy sensitivity becomes negligible while source sensitivity remains. When copies are reliable, making them more frequent also strengthens source suppression. The preference affects generation: a conflicting copy supplied in an arithmetic prefix makes Gemma-2 and Qwen3 produce its implied result about % of the time, and a source change reaches a copy that the model writes itself in -% of samples. These results show how local restatements change which occurrence affects a prediction and identify training conditions that can reverse that change.

Primary area:
interpretability and explainable AI
Keywords:
chain-of-thought reasoning, mechanistic interpretability, causal interventions

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_VDwFOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies if large language models think based on the original source of intermediate values or restated copies, which are generated in the middle of chain-of-thought reasoning. It is demonstrated that having correct local copies reduces the sensitivities of the models (e.g. Gemma-2, Qwen3). In a controllable experimental setup, it is shown that this preference is caused by the reliability and frequency of the copies in training.

Critical strengths
  • The paper proposes a clear method to modify the internal memory to investigate the effect of the intermediate copy (e.g. replacing X=10 by X=50 to see how the reasoning outcome will differ). The idea of having intervention in the memory mechanism is interesting. Extensive ideas could be developed on this idea, e.g., the safety and adversarial attacks (how to make modes robust against such memory modification and injection).
Critical weaknesses
  • The writing and presentation quality is, unfortunately, suboptimal and the paper requires a major revision. A paper should primarily present a clear research question, a methodology how to solve it, and insights obtained through the experimentation (or theoretical formulation). In the current form of the paper, it is challenging for readers, may not impossible, to extract those essence from the manuscript, as many parts of the explanation need assumptions, which are not introduced in the paper in a self-contained manner, to interpret.
  • The evaluations is conducted only on synthetic test environments e.g. with arithmetic templates. Thus a major concern remains on the applicability of the study. Although I appreciate the idea, the evaluation is not supporting the introduced claim sufficiently.
Questions for authors
  • Although the paper touches upon scenarios such as basic words, names, and PrOntoQA deductions, how can the findings be applied to the general, unstructured reasoning?
2 comments
AuthorsOpenReview ↗

Authors Response to Reviewer VDwF (Part 1 of 2)

Thank you for the assessment and for your interest in the intervention itself. You raise two concerns: that the research question, method, and insights are hard to extract, and that the evaluation is limited to synthetic environments. The revised manuscript addresses both; new text is set in blue. We respond to each point in the order of your review.

W1. The research question, methodology, and insights are hard to extract

We agree that the submission did not state these three elements plainly. The revision states each in one place.

  • Research question (Abstract; Introduction, second paragraph). When a reasoning trace writes an intermediate value twice, once where it is computed (the source) and once where it is next used (a local copy), which occurrence does the next prediction depend on, and what decides between them?
  • Methodology (Figure 1; Sec. 3). We change one occurrence and hold the other fixed. Panel (a) of Figure 1 now shows a concrete source line (x3 = 14 + 9 = 23) and consumer line (x5 = x3 - x1 = 23 - 16 = 7) with the source, the local copy, and the prediction marked, and states the measurement in one line. Sec. 3 defines the source intervention, the comparison with and without a copy, and what the comparison does and does not measure.
  • Insights (Introduction, third and fourth paragraphs; contributions with section pointers).
    1. With a correct copy present, the next step depends on the copy: the copy removes more than 99% of the effect of changing the source (Sec. 4).
    2. A changed copy sets the next result, while the source still determines a copy that the model writes itself (Sec. 5).
    3. The same holds in solutions that models write without a format request, and a step that refers to the source without a copy follows the source (Sec. 6).
    4. The preference depends on how reliable copies were in training and can be reversed, in small transformers and in three pretrained models (Sec. 7).

Self-contained definitions. The three terms on which the paper relies (source, consumer, local copy) are defined in the first paragraph of the Introduction on the example line shown in Figure 1. Terms needed only for the method are defined at first use in Sec. 3. The paragraph "What the comparison measures" in Sec. 3 states the assumptions under which a small source effect should be read: it is conditional on the supplied text and does not show that the source is absent from the model's internal state.


W2. The evaluation uses only synthetic test environments

You are right that the submission evaluated only assigned formats and templates. The revision adds a section on reasoning that the models write themselves (Sec. 6) and extends the training experiment to pretrained models and natural solutions (Sec. 7).

ExperimentSettingResultWhere
PrevalenceFour models solve 1319 GSM8K test problems with no format request60 to 74% of solutions restate an earlier result where it is next usedTable 1
Copy and reference1000 to 1510 restated dependencies per model, each rendered both waysWith the copy, an edited source is followed on at most 4% and an edited copy on 77 to 99%. With a label in place of the copy, the same source edit is followed on 80 to 99%Table 1, Table A42
PromptInstruction, demonstrations with wrong copies, or both; 600 dependencies per modelThe copy remains decisive on 94 to 99% in three models and 75 to 86% in Gemma-2-27BTable A43
Fine-tuningThree models fine-tuned on GSM8K reference solutions with reliable or unreliable restated values, tested on their own solutionsAfter unreliable training the copy decides at most 1% of conflicts and the source 86 to 99%; after reliable training the copy decides 92 to 97%Table 2, Table A50
AuthorsOpenReview ↗

Authors Response to Reviewer VDwF (Part 2 of 2)

Q. How do the findings apply to general, unstructured reasoning?

To answer your question, we tested this directly with free-form two-hop answers over Wikipedia passages (MuSiQue), in which no arithmetic and no template is involved (Sec. 6, final paragraph; Appendix A7.5, Table A45). We keep the sentence in which the model names the intermediate entity, remove the passage that states it, and append one bridging sentence that either names the entity again (a copy) or refers to "the one just identified" (a reference). The model then continues in whole sentences.

ModelReference: source replaced, next inference changesCorrect copy: source replaced, next inference changes
OLMo-2-7B42.2%4.5%
Qwen3-8B60.0%3.2%
Gemma-2-27B57.9%12.7%

A correct copy shields the next inference from a change of the source in all three models, as in arithmetic. We report the limits of this result in the manuscript.

  • Final answers return to the entity of the source sentence more often than the next inference does.
  • Gemma-2-2B reaches neither answer in more than half of its continuations and is not interpreted.
  • When the intermediate entity is also stated in the supplied passages, edits to its mentions in the trace do not carry beyond their own sentence (Appendix A7.6).

We therefore claim the following for unstructured reasoning: a restated intermediate value shields the next step from its source, and a step that refers back without restating remains sensitive to it. We do not claim that a changed copy determines the final answer of a free-form response.


On robustness to modified or injected intermediate values

You note that the intervention suggests further work on robustness to memory modification and injection. The fine-tuning result in Sec. 7 bears on this point. By default, a value altered at its restatement decides the next calculation on 95 to 98% of dependencies in the models' own GSM8K solutions. After fine-tuning on solutions in which restated values are sometimes wrong, the altered restatement decides at most 1%, and the models compute from the source on 86 to 99%.


Changes made in response to your review

  • Abstract and Introduction: research question stated explicitly; findings summarized in section order; section pointers in the contributions.
  • Figure 1: concrete source and consumer lines with the measurement stated in one line.
  • New Sec. 6: restated values in model-written GSM8K solutions, prompt variants, and free-form two-hop answers.
  • Sec. 7: reliability fine-tuning of three pretrained models on GSM8K solutions.
  • Limitations: the boundaries listed above.

Code and result files for all new experiments are in the anonymous repository.

Reviewer_F9G4OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

In controlled arithmetic reasoning, nearby copies of intermediate values reduce models’ sensitivity to the original source by over 99%, while incorrect copies often steer the next result. Training small models on unreliable copies reverses this preference, showing that reliance on copies depends on training conditions. The implication is that checking each reasoning step’s arithmetic is insufficient: copied values must also match their sources, though these findings do not establish how unrestricted reasoning works.

Critical strengths
  1. Strong causal controls: matched problems, randomized formats, and filler controls help isolate the effect of copied values.
  2. The setup the paper is testing for, source-consumer structure what I would call, seems to be novel.
Critical weaknesses

1. Significance and generalizability of the problem seems limited.

The problem the paper is dealing with, namely source-consumer structure and patching on source, seems to have limited significance. Why is this a problem worth paying attention for? How can this problem be generalized to other cases? Please elaborate on this.

2. Presentation-wise, there is a large room for improvement.

  • Introduction is quite hard to follow: Connections between the paragraphs are unclear. Too many jargons. Critically, it does not contain much summary on what the rest of the paper does, effectively making this paper "intro-less".

  • Figure 1 is intended to show an overview of this work, but doesn't serve that role well. I find it hard to read and understand. I hope to find a concrete example of "source" and "consumer" here but couldn't find one.

  • I value this work doing experiments both on arithmetic task (GSM8K) and natural language task (PrOntoQA). However, the reader is confronted with experiments on PrOntoQA only after page 5, without any precursor (this is partially due to lack of proper introduction, but that is not the sole reason). In fact, the paper states at the beginning of section 3, "Our primary testbed uses synthetic arithmetic traces ...", which makes the reader think that the paper does not experiment beyond arithmetic task, which is not true.

Questions for authors

No questions.

2 comments
AuthorsOpenReview ↗

Authors Response to Reviewer F9G4 (Part 1 of 2)

Thank you for the assessment and for recognizing the causal controls and the novelty of the source and consumer setup. You raise two concerns: the significance and generality of the problem, and the presentation. We have revised the manuscript on both; new text is set in blue. We respond to each point in the order of your review.


W1. Significance and generalizability of the problem

You ask why the source and consumer structure deserves attention and how it generalizes. The submission did not state this clearly. The revision makes the argument in the Introduction, in a new Sec. 6, and in the Discussion paragraph "Why the occurrence matters." It has three parts.

  1. The structure is common in ordinary model output. When Gemma-2-2B, OLMo-2-7B, Qwen3-8B, and Gemma-2-27B solve the 1319 GSM8K test problems with no format request, 60 to 74% of their solutions restate an earlier result where it is next used, and restated results make up 31 to 35% of all operands (Sec. 6, Table 1; Appendix A7.4).
  2. It determines what an intervention or a correction reaches. In these solutions, the next calculation follows an edited source on at most 4% of restated dependencies and an edited copy on 77 to 99%. When the same step refers to the source by a label, the same source edit is followed on 80 to 99% (Table 1). A change at the place where a value was computed therefore does not reach the steps after its restatement, and an intervention on that place measures no dependence where the trace displays one. This affects how patching results on reasoning traces are interpreted and what a line-level checker can verify.
  3. The preference is a learned property that can be reversed. Instructions and demonstrations in the prompt do not change it (Table A43). Fine-tuning on solutions whose restated values are sometimes wrong reverses it in three pretrained models: the source then decides 86 to 99% of conflicts (Sec. 7, Table 2).

Generalization. The following table lists the settings in which we test the effect and the limits that we found.

SettingResultWhere
Word and name valuesStrong suppression by a correct copySec. 4, Fig. A4
PrOntoQA deductionSuppression in Qwen3 and OLMo-2, not in Gemma-2Sec. 4
Six model sizes; Mamba2 and RWKVSuppression by a correct copy in allSec. 5, Appendix A3
Model-written GSM8K solutions, four modelsSource edit followed on at most 4% with a copy and 80 to 99% with a referenceSec. 6, Table 1
Free-form two-hop answers over Wikipedia passagesSource edit followed in 42 to 60% of answers with a reference and 3 to 13% with a correct copy, in three modelsSec. 6, Table A45

We also state where the effect weakens. In free-form answers a changed copy decides the next inference less often than in arithmetic. When the intermediate value is also stated in the supplied context, edits to its mentions in the trace do not carry beyond their own sentence (Appendix A7.6). Both limits are in the Limitations paragraph.

AuthorsOpenReview ↗

Authors Response to Reviewer F9G4 (Part 2 of 2)

W2. Presentation

W2a. The Introduction is hard to follow and does not summarize the paper

We agree with your assessment of the Introduction and have revised it as follows.

  • First paragraph. It defines the three terms that the paper uses (source, consumer, local copy) on one concrete line, x5 = x3 - x1 = 23 - 16 = 7, which is the same example as in Figure 1, and states why the distinction matters for interpreting and auditing reasoning traces.
  • Second paragraph. It states the research question in one sentence: how two occurrences of the same intermediate value, at its source and at its next use, affect the next prediction, and what decides between them.
  • Third and fourth paragraphs. They summarize what the rest of the paper does and finds, in the order of the sections: the intervention, the suppression by a correct copy, the extension to word and name values and PrOntoQA, the result on model-written GSM8K solutions, and the training experiments that reverse the preference.
  • Contributions. Each of the three contributions now points to the section that supports it (Sec. 4; Sec. 5 and 6; Sec. 7).

Terms that are needed only for the method, such as teacher forcing, are introduced in Sec. 3 where they are first used.

W2b. Figure 1 does not give a concrete example of "source" and "consumer"

We agree. As you note, the figure did not offer a concrete example: it was present only as a small input label inside the patching diagram. Figure 1 has been revised as follows.

  • Panel (a) now opens with a concrete trace of two lines. The source line x3 = 14 + 9 = 23 is marked "Source S, the value is computed." The consumer line x5 = x3 - x1 = 23 - 16 = 7 marks 23 as "Local copy C" and 7 as "Prediction Y," and an arrow labeled "restated" connects the two occurrences of 23.
  • The attention schematic and the teacher-forcing diagram have been removed from panel (a). The measurement is stated in one line: the source effect is the change in the log-probability of the result when only the source states are replaced.
  • Panel (c) now includes the contrast in model-written GSM8K solutions: with a restated value (65 - 30) the next calculation follows the copy, and with a label (R - 30) it follows the source.
  • The caption begins with the definition of source and local copy.

W2c. PrOntoQA appears only after page 5, and Sec. 3 suggests an arithmetic-only study

We agree that the submission introduced these testbeds too late, and that the opening of Sec. 3 suggested an arithmetic-only study. The Introduction now states that the effect extends to word and name values and, in two of three models, to PrOntoQA deduction, and that it appears in solutions that the models write without a format request. The first paragraph of Sec. 3, which begins with the primary arithmetic testbed, now ends by pointing to the word, name, and PrOntoQA experiments in Sec. 4 and to the model-written solutions in Sec. 6.


Changes made in response to your review

  • Introduction: example aligned with Figure 1, research question stated explicitly, findings summarized in section order, section pointers in the contributions.
  • Figure 1: concrete source and consumer lines; two schematics removed.
  • Sec. 3: forward pointers to the non-arithmetic and model-written testbeds.
  • New Sec. 6 and Discussion paragraph "Why the occurrence matters": prevalence and consequences of restated values in model-written reasoning.
  • Sec. 7: reversal of the preference in three pretrained models.
Reviewer_iypzOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to understand how the locality of value definition (source) and redefinition (copy) affects which one a model uses when making a prediction. It studies this in the context of arithmetic expressions. Based on experimental results, the paper makes the following observations:

  1. prediction relies on the copy when it is held fixed while the source is changed
  2. prediction relies on the copy when it is changed while the source is held fixed
  3. prediction relies on the source when the copy is not supplied, and instead the model is asked to write it
Critical strengths

None. The experimental setup of the paper is good and easy to follow, but I am not yet convinced that it is novel, and I have concerns about the soundness of some of the claims.

Critical weaknesses
  1. Incremental Novelty. The fact that models tend to trust values based on locality has been explored in several different contexts before. There is significant overlap with the literature on positional bias, attention competition, and attention dilution, etc., that the paper does not cite or reference. Some examples are:

    • Kurfalı et al., Conflicting Needles in a Haystack: How LLMs Behave When Faced with Contradictory Information (EMNLP'25)
    • Zhao et al., Calibrate Before Use: Improving Few-shot Performance of Language Models (ICML'21)
    • Wang et al., Eliminating Position Bias of Language Models: A Mechanistic Approach (ICLR'25)
    • Tian et al., Distance between Relevant Information Pieces Causes Bias in Long-Context LLMs (Findings of the ACL'25)
  2. Experimental Claim Concern. The paper uses Table A31as evidence to support its claim that its training intervention makes the models rely on copies more often. But that claim may be unsound, because when 95% of the training data contains copies (last row of Table A31), the accuracy of the models on the no-copy format plummets to just 2.8%. This indicates that training has made the models lose the ability to compute correctly at all without copies being available. The high source effect argument from the table's caption does not offer a satisfactory explanation either, since a model that consistently predicts the wrong answer on no-copy formats could still appear source-sensitive by predicting different wrong answers when the source changes.

Questions for authors
  1. Can you please explain how your contributions are novel relative to works related to positional bias, attention competition, and attention dilution, etc., some of which I have shared under Incremental Novelty in Critical Weaknesses?

  2. In the cases where the copy is not explicitly enumerated with values, you control for number of tokens by introducing fillers (eg: Let the model write its copy case from Figure 1) . But these aren't the same. When you explicitly provide values, you are not just influencing the model's predictions by the locality of the required information, but also by other things, such as computational convenience, since the expression is now self-contained and readily evaluable compared to situations where the model has to reason through previous steps to compute the values required to evaluate the expression. So, could your observations be ignoring such confounding factors?

  3. Could you please explain why it is sound to assume that your training intervention makes the model rely on copies when your data (Table A31) shows that it makes them completely lose the ability to compute at all without copies?

4 comments
AuthorsOpenReview ↗

Authors Response to Reviewer iypz (Part 1 of 2)

Thank you for the careful assessment. Your second weakness identified a real gap in the support for our training claim, and your second question identified a confound that the submission had not separated. Both are addressed with new experiments in the revised manuscript, in which new text is set in blue. We respond to each point in the order of your review.


W1 and Q1. Novelty relative to position bias, attention competition, and attention dilution

You point out that models are known to favor information by its location and that the submission did not cite this literature. We agree that the four papers you list should have been cited. The revision discusses them in a new Related Work paragraph, "Position and conflicting context" (Sec. 2). We also agree on the starting point: by default, the nearer occurrence wins. Our contribution lies in what a position account does not predict. The cited studies vary where independent pieces of supplied context are placed. We study a value and its own restatement inside a reasoning trace, hold both positions fixed, and find that the preferred occurrence still changes. Three results support this.

  1. At a fixed position, the form of the copy decides the effect (Sec. 4, Table A41). A list of bindings on the consumer line (x3 = 28, x1 = 16) leaves 2.1 to 2.6% of the no-copy source effect in three of four models. An instantiated expression on the line directly before the consumer (28 - 16) leaves 47 to 101%.
  2. The preference reverses with positions and test inputs unchanged (Sec. 7). In small transformers at the same restatement frequency, the decoded result under a source and copy conflict is the copy-implied one on 66 to 75% of dependencies after training on reliable copies, and the source-implied one on 92 to 95% after training on unreliable copies (Table A48). In Gemma-2, OLMo-2, and Qwen3, fine-tuning on GSM8K solutions whose restated values are sometimes wrong reduces the share of conflicts decided by the copy from 95 to 98% to at most 1%, and the source then decides 86 to 99% (Table 2).
  3. The context does not change the preference (Sec. 6, Table A43). An instruction to use the value where it was first calculated, demonstrations in which every restated value is wrong, or both leave the copy decisive on 94 to 99% of dependencies in three models and on 75 to 86% in Gemma-2-27B.

Regarding attention dilution, length-matched filler that adds the same number of tokens without the values leaves 89 to 134% of the source effect (Table A41), and a bare number without its variable binding has little effect (Fig. 2). Additional tokens alone therefore do not produce the suppression.


Q2. Could computational convenience explain the effect of a copy?

You ask whether a line with explicit operand values, being self-contained and easy to evaluate, produces the effect through convenience and not through the locality of the information. Your point is correct for the submitted version: a copy on the consumer line both places the values next to the prediction and makes the line a ready expression, and the filler control separates neither. The revision separates the two (Sec. 4, paragraph beginning "A copy on the consumer line is also easy to compute from"; Appendix A7.1, Table A41). We vary the location of the operand values (consumer line or preceding line) and their form (instantiated expression or binding list), on 968 dependencies per model.

RenderingGemma-2-2BOLMo-2-7BQwen3-8BGemma-2-27B
Consumer line, expression (correct local copy)1.8%0.0%1.4%0.3%
Consumer line, binding list40.0%2.2%2.1%2.6%
Preceding line, binding list52.0%16.3%7.9%9.9%
Preceding line, expression100.9%55.5%66.9%47.4%
Length-matched filler133.5%118.0%88.9%108.8%

Entries are the remaining share of the no-copy source effect. The expression on the preceding line is the rendering that is easiest to evaluate, and it leaves most of the source effect. The binding list on the consumer line offers no expression to evaluate, and it removes 97 to 98% of the effect in three of four models. The suppression therefore follows values bound to the names that the consumer uses.

We do not claim that convenience plays no role. An expression on the preceding line reduces the source effect by up to about half in three models, and Gemma-2-2B benefits most from an expression on the consumer line. Our claim is narrower: convenience alone does not produce the suppression. In addition, after fine-tuning on unreliable copies the consumer line still contains the same ready expression, and the models nevertheless follow the source on 86 to 99% of conflicts (Table 2).

AuthorsOpenReview ↗

Authors Response to Reviewer iypz (Part 2 of 2)

W2 and Q3. Table A31 and the soundness of the training claim

You observe that with 95% restatement in training, accuracy without copies is 2.8%, so the model has lost the ability to compute without copies, and that a source effect measured in log-probability could reflect a change between wrong answers. You are correct on both counts. That row (Table A32 in the revision) describes a model that cannot compute without a copy, and a log-probability change does not show that the prediction moves to the value that the source implies. The revision no longer draws a conclusion from such cells, and the claim now rests on a new experiment (Sec. 7, paragraph beginning "At p = 0.5, however"; Appendix A7.8, Table A48) with two changes.

  • Competence criterion. A run is used only if its value-token accuracy is at least 0.95 both with and without a copy. Each cell has six seeds. Every reliable-copy cell at a restatement frequency of 0.35 or higher fails the criterion, and we state that we draw no conclusion from those cells.
  • Decoded outcomes with three distinct candidates. On each test dependency the source is changed to one value and the copy to another, so that the source-implied, copy-implied, and original results are three different numbers. Any other output is counted separately.
Training copiesFrequencyCompetent runsAccuracy without / with copyConflict: copy-impliedConflict: source-impliedNo copy, source changed: source-implied
Reliable0.256 of 60.974 / 0.99366.3%6.0%92.6%
Reliable0.306 of 60.970 / 0.99675.0%4.7%92.6%
Half of restated lines wrong0.256 of 60.981 / 0.9910.0%94.1%93.2%
Half of restated lines wrong0.306 of 60.978 / 0.9950.0%94.6%92.8%

The compared models are accurate without copies, so the preference for the copy is not a loss of the ability to compute. On lines without a copy the decoded result is the one implied by the changed source on 91 to 93% of dependencies, so the source effect is a move to that value.

The same manipulation holds in pretrained models (Sec. 7, Table 2; Appendix A7.10, Table A50). We fine-tune Gemma-2-2B, OLMo-2-7B, and Qwen3-8B on 3000 GSM8K reference solutions, either unchanged or with each restated operand replaced by a wrong number with probability one half, and test on 600 dependencies in each model's own solutions.

TrainingCopy changed: follows copyCopy changed: follows sourceSource changed: follows source
None95 to 98%0 to 1%0 to 1%
Reliable copies92 to 97%0 to 2%0 to 1%
Unreliable copies0 to 1%86 to 99%70 to 83%

The result holds in all six runs per condition.


Changes made in response to your review

  • Sec. 2: new paragraph on position bias and conflicting context, citing Zhao et al., Wang et al., Tian et al., and Kurfalı
  • Sec. 4 and Appendix A7.1: separation of the location of the operand values from their form.
  • Sec. 7 and Appendix A7.8: reliability manipulation restricted to models that are accurate in both formats, with decoded outcomes.
  • Sec. 7 and Appendix A7.9 to A7.10: reliability fine-tuning of three pretrained models.
  • Sec. 6 and Appendix A7.3: instruction and demonstration variants of the prompt.

Code and result files for all new experiments are in the anonymous repository.

In all your responses to the questions I raised, you keep mentioning that what I raised holds for the submitted version, but not for the revised version. This just leads me to believe that the submitted version was grossly underprepared and that the revised version is substantially different (probably on the verge of being a new paper), which lowers my confidence in the claims being made and tested. Also, you have not highlighted the changes consistently; e.g., Table A31 is Table A32 in the revision, but neither is highlighted in blue.

Q1. What does this mean?

A run is used only if its value-token accuracy is at least 0.95 both with and without a copy. Each cell has six seeds. Every reliable-copy cell at a restatement frequency of 0.35 or higher fails the criterion, and we state that we draw no conclusion from those cells.

Are you implying that you discard the results that are not favourable to the claims you are making?

Q2. You defended your novelty relative to the literature I pointed out by justifying that,

We study a value and its own restatement inside a reasoning trace, hold both positions fixed, and find that the preferred occurrence still changes

But then later you also say

We vary the location of the operand values (consumer line or preceding line) and their form (instantiated expression or binding list), on 968 dependencies per model.

Are these not contradictory?

AuthorsOpenReview ↗

Authors Response to Reviewer iypz

Thank you for reading our response and the revision closely.

On the extent of the revision

Where our response said that a point holds for the submitted version, it referred to two specific controls that the submission lacked: the separation of location and form (your Q2) and the restriction to models that compute accurately without a copy (your W2). It did not refer to the claims. The claims of the submission are unchanged and none of its experiments was removed. The revision adds the two controls and tests the same claims on reasoning that models write themselves and on pretrained models. These additions extend the evidence for the same claims and do not replace it.

Blue marks only newly added text and experiments. Content that only changed its place or its number is not marked: Table A32 is Table A31 of the submission with identical content.


Q1. The competence criterion

"Fails the criterion" means only that accuracy without a copy is below 0.95. It does not refer to whether a cell shows the effect, and the threshold was fixed before the runs.

The criterion implements the concern of your W2, which stated that at a restatement frequency of 95% "training has made the models lose the ability to compute correctly at all without copies being available," so that the claim "may be unsound." Reliance on a copy is therefore counted as evidence only in a model that can also compute without one. The excluded cells are such models: with reliable copies, accuracy without a copy is 0.941 at a frequency of 0.35 and 0.469, 0.037, and 0.035 at 0.50, 0.75, and 0.95 (Table A48).

  1. No cell is omitted. Table A48 lists all 14 cells with their accuracy and the number of competent runs.
  2. No seed is selected in the cells that carry the claim. All of their runs pass:
Training copiesFrequencyCompetent runsConflict: copy-impliedConflict: source-implied
Reliable0.256 of 666.3%6.0%
Reliable0.306 of 675.0%4.7%
Half wrong0.256 of 60.0%94.1%
Half wrong0.306 of 60.0%94.6%
  1. The excluded runs agree with the claim. Their decoded results are in the data of the anonymous repository. In the 16 excluded reliable-copy runs at frequencies of 0.35 or higher, the copy-implied result is decoded on 77 to 97% of conflicts and the source-implied result on at most 4%. In the 12 excluded runs with wrong copies (frequencies 0.15 and 0.50), the source-implied result is decoded on 73 to 96% and the copy-implied result on at most 1%. Including them would strengthen the result.
  2. The weakest cell is reported. With reliable copies at a frequency of 0.15, where copies are rare, the decoded result leans to the source on average over the two competent runs (copy 31.9%, source 44.0%, Table A48) and in three of the four excluded ones. The claim does not use this cell.

Q2. The two statements

The two statements refer to different things. "Both positions" are the positions of the source and of its restatement within one comparison. The second statement describes one experiment in which the location of the restated values is one of two factors, varied in order to test whether location accounts for the effect.

In that experiment the source line is the same in every rendering and the patch covers only that line. The operand values are added in one of two locations (consumer line or preceding line) and in one of two forms (expression or binding list). The comparison at a fixed location is the following (Table A41, remaining share of the no-copy source effect):

Preceding lineGemma-2-2BOLMo-2-7BQwen3-8BGemma-2-27B
Binding list52.0%16.3%7.9%9.9%
Expression100.9%55.5%66.9%47.4%

The same values at the same location leave a source effect that differs by a factor of 1.9 to 8.5 according to their form. On the consumer line both forms remove nearly all of the effect in three models (0.0 to 1.4% and 2.1 to 2.6%).

The training experiment is the second such comparison. All models are read on one fixed evaluation set, so the source and the copy are at the same token positions, and the decoded result under a conflict moves from the copy-implied one (66.3 to 75.0%) to the source-implied one (94.1 to 94.6%) with the reliability of copies in training (table above).

Location also has an effect: a binding list leaves 2.1 to 2.6% of the source effect on the consumer line and 7.9 to 16.3% on the preceding line in the three larger models. Our claim is that location alone does not determine which occurrence the model uses, and not that location has no effect. This is the difference from the studies you cited: they show that a model favours content at particular positions in its input, whereas here the dependence on the source changes with the form of the copy, and the favoured occurrence changes with training, while the positions stay the same.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.