acceptodds
Under review as a conference paper at ICLR 2027

Cross-Model Stitching Transfers Answers, Not Computation

Abstract

Activation stitching describes the technique of grafting one model's hidden state into another's through a learned map. Its motivation is related to universality, under which models may learn shared representations that enable transfer. Prior work demonstrates that features, probes, and steering vectors can survive such transfers. However, what actually moves through the stitch, a question with substantial implications for universality, remains unresolved. We examine this question by grafting a large donor's residual state into a smaller recipient using a map trained with cross-entropy on the gold output label. We spread this evaluation across three same-family model pairs (Gemma-2 9B2B, Qwen2.5 7B0.5B, Llama-3.1 8B3.2 1B) and two cross-family pairs (Qwen2.5 7B and Llama-3.1 8B into Gemma-2 2B). We evaluate across five task families: arithmetic, symbolic binding, state tracking, naturalistic factual recall, and GSM8K. Our experiments find that this task-supervised stitch is a working channel that confers the donor's answer on up to of otherwise unsolvable problems in our primary arithmetic evaluation. Additionally, a battery of controls indicates that conferral is achieved through the transfer of a precomputed answer, and success is dictated primarily by the answer's presence at the donor stitch site. We also find that this presence alone is not sufficient, as it must then be lossily encoded into the recipient. Further ablations indicate that the task-supervised map learns to mitigate this effect by writing the signal redundantly across hundreds of dimensions of the graft, even though the signal itself is low-rank. Finally, while the first answer token transfers exceptionally well from the prefill state, the rest of the answer does not until prior answer tokens have been emitted. Transfer is bottlenecked by answer presence, and the donor must undergo additional serial computation before that presence is sufficient for transfer. Together, these results suggest a boundary in cross-model universality, where models appear sufficiently aligned for precomputed answers to transfer through a stitch, while showing no measurable evidence that the computations producing those answers transfer with them.

Primary area:
interpretability and explainable AI
Keywords:
mechanistic interpretability, model stitching, activation stitching, representational alignment, cross-model transfer, universality, residual stream, linear probing

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_e9tCOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

This paper investigates whether hidden states transferred between different models through stitching represent answers that have already been computed or intermediate computational states that can support further computation. The authors design a learnable linear mapping that injects the hidden states of a larger Donor Model into a specified layer of a smaller Recipient Model, and train the Stitching Map using the cross-entropy loss on the first token of the correct answer. Experiments are conducted across multiple model families and tasks. The paper shows that Cross-Model Stitching mainly improves performance by transferring answers already computed by the Donor, with no clear evidence that it provides effective intermediate states that enable the Recipient to continue the computation.

Critical strengths

1、The research question addressed in this paper is highly valuable and has significant implications for cross-model information transfer.

2、The experimental design is very comprehensive, including various mechanistic analyses and control experiments, while also covering multiple models and tasks.

3、 The paper provides in-depth analyses of certain mechanisms. For example, the analysis of First-token Conferral and Full-answer Conferral. The paper not only observes a significant performance gap between the two but also investigates this gap through different mapping functions, training objectives, and Hidden States from different computational stages of the Donor, providing valuable experimental findings for understanding the limitations of cross-model information transfer.

Critical weaknesses

1、The paper's layout is not visually appealing, and the visual presentation and organization of information are not very good.

2、The First-token Cross-Entropy training objective may affect the conclusions about Computation Transfer, as this loss function inherently tends to optimize the transfer of answer information.

3、The sensitivity analysis of Stitching Layer selection is insufficient. The paper lacks a systematic analysis of different combinations of Donor Layers and Recipient Layers.

4、The paper lacks a systematic analysis of Stitching performance on the full test set. For example, can Stitching still provide benefits for questions that neither the Donor nor the Recipient can answer correctly? Could Stitching lead to performance degradation on questions that the Recipient could originally answer correctly?

Questions for authors

1、Could the authors consider using an Intermediate-state Alignment or Computation Continuation Objective to train the Stitching Map and compare it with the current Answer-supervised Objective?

2、Could the authors further analyze the impact of Stitching Layer selection on transfer performance? For example, systematically comparing First-token Conferral and Full-answer Conferral across different combinations of Donor Layers and Recipient Layers to verify whether the main experimental conclusions depend on specific Layer Pairs.

3、Could the authors further report the Accuracy of Stitching on the full test set and analyze the performance on different types of samples? In particular, can Stitching provide additional benefits for questions that neither the Donor nor the Recipient can answer correctly? Could Stitching lead to performance degradation on questions that the Recipient could originally answer correctly?

4、Could the authors further improve the paper's layout and presentation of experimental results? For example, presenting some key ablation experiments and control results together to help readers more intuitively understand the relationships between different experiments and how they support the main conclusions.

4 comments
AuthorsOpenReview ↗

Reviewer e9tC Answer 1-3

Hello, thank you for your constructive feedback. Answering each question:

Q1. Could the authors consider using an Intermediate-state Alignment or Computation Continuation Objective to train the Stitching Map and compare it with the current Answer-supervised Objective?

We compare both kinds of objective against the first-token objective in Section 7. Our reconstructive map represents Intermediate-state Alignment, since the map is trained only to match the recipient's own activations at the graft site, and we find that it confers far less than the task map. We additionally sweep across alignment values in our first-token map, which is trained with an added weighted term that rewards similarity to the recipient's own state, and find that, as this alignment term increases, first-token and full-answer conferral decrease, which is consistent with our findings.

For computation continuation, we use distillation to train the map to match the donor's full next-token distribution at every answer position. We also have our whole-answer objective which trains the map with cross-entropy on the full gold answer. Not much of the second or third answer digits are present at the donor site, making computational transfer especially attractive. While they raise full-answer conferral to 0.19 and 0.23, they still sit well below the donor's own 0.42, and further ablations in section 7 reveal that the slight gain comes from improved answer readout of the second and third digits.

Q2. Could the authors further analyze the impact of Stitching Layer selection on transfer performance?

We added this comparison as the appendix section "Layer pairs," where we retrain the first-token task map at every combination of four donor and four recipient layers for Gemma and Qwen. Full-answer conferral stays low at every pair (0.092 to 0.165 on Gemma and 0.023 to 0.070 on Qwen), so the limit on full-answer conferral is not a product of our layer choice.

Q3. Could the authors further report the Accuracy of Stitching on the full test set and analyze the performance on different types of samples?

We added this as the appendix section "The full test set," which splits all 2,000 evaluation problems by which model natively gets the first answer token right. Over the full set, the stitch raises first-token accuracy from 0.59 to 0.89 (Gemma) and from 0.63 to 0.80 (Qwen).

On questions neither model got correct, the stitch reaches 0.59 and 0.57 on the first token, against about 0.10 for the shuffled-donor and constant controls. This fits within our model since the answer is often already present at the donor stitch site even when the donor's final output is wrong, a read which is further supported by the fact that a linear probe fit to the donor stitch site performs similarly well and succeeds on largely the same problems. The full-answer gain is much smaller (about 0.06).

On questions just the recipient gets correct, stitching does degrade accuracy to 0.76 (Gemma) and 0.54 (Qwen), with most of the loss coming from the recipient outputting the donor's wrong first token. On questions both models get correct, it falls only to 0.95 and 0.87.

AuthorsOpenReview ↗

Answer 4 and Additional Clarification

Q4. Could the authors further improve the paper's layout and presentation of experimental results?

We submitted the revised manuscript, where we made substantial revisions to its structure. We re-organized the sections to create a clear narrative arc:

Finding that the stitch confers -> Finding that it confers due to the transfer of a precomputed answer -> Direct comparisons to prior work and potential implications for the circuit universality hypothesis -> Demonstrating the impacts of that transfer of an answer through the full-answer conferral gap, and what it suggests about the generalizability of cross-model stitching.

We also flag the implications of the arguments we make more frequently so as to better convey our points.

Finally, you mention as a weakness that "The First-token Cross-Entropy training objective may affect the conclusions about Computation Transfer," which we wanted to address here.

Our task-supervised map is rewarded for anything that improves the recipient's prediction, whether through answer transfer, computational transfer, or both. The map will transfer whatever gives the recipient the best ability to output the correct answer, which can include a combination of a precomputed answer and computation. Despite this, we find that precomputed answer content captures more or less all of the first-token conferral on Gemma, as shown by the probe bottleneck stitch (Section 5). This finding gives us substantial evidence that the two models' circuits are not aligned enough for computational transfer.

We also have sufficient evidence against the idea that the map never learns computational transfer because it never needs to, since the benefit from answer transfer outweighs anything the recipient would gain from computational transfer. First, our SVD truncation experiment showed that retaining only the top 16 singular directions preserves 99% of the task map's original conferral, while the remaining 2,288 directions on their own recover only around 17% of full-map conferral. This means that the task map has the dimensional capacity to transfer computation to supplement the answer encoding, but does not learn to. We also know that the task map can effectively interact with every part of the recipient's transformer accessible to it, given how it has to encode the answer to correctly interact with every component of the recipient's remaining layers. Additionally, since the donor eventually gets every first token in the unsolvable bin correct, we know there is a correct chain of donor computation that starts at the donor stitch site and is visible to the stitch. Together, these facts mean that the donor's stitch site contains both a precomputed answer encoding and the start of a chain of successful computation, and that the task map can reach the recipient with it. If the task map learned a consistent encoding for even a single step of that chain, it could use its remaining 2,288 dimensions to stitch that step into the recipient and improve the recipient's output through computation, yet it does not. You could also make the claim that answer transfer does so well on Gemma and Qwen that it never needs to learn computational transfer, but computational transfer remains absent on tasks where the map confers as little as a third of what the donor solves, meaning the stitch would have a lot to gain if it could use transferred computation to resolve those remaining errors, yet it does not.

Ultimately, though, the finding that successful stitching can be explained largely by answer transfer maintains the paper's primary implication for interpreting stitching results. This connection is made much more explicit in the revised version of our paper (specifically section 6.2).

Reviewer_e9tCOpenReview ↗

response

Thank you for the detailed responses and additional analyses. I appreciate your efforts to address my concerns. However, several concerns remain insufficiently resolved. In particular, I am not convinced that the negative results can be attributed to limited computational transfer rather than limitations of the training objectives or stitching interface. I will therefore maintain my original score.

AuthorsOpenReview ↗

Response to Remaining Concerns

Thank you for taking the time to read our responses.

In terms of computation being limited by the stitching interface, our paper tests full-answer transfer through a free vector allowed to optimize for each question at the same graft site, which achieves 100% full-answer accuracy on a Gemma subsample. Additionally, our Section 7 evaluates the same stitching protocol on several stronger maps, which show little to no full-answer improvement over a basic linear stitch.

In terms of computation being limited by the training objective, Section 7 also tests intermediate-state alignment, distillation, and whole-answer objectives, all of which stay well below the donor's own full-answer rate. If concerns about the objective remain, would you be able to describe them? Since it is a negative result, there will always be possible reasons other than unaligned circuits for why computation does not transfer (such as the hypothetical we described in our follow-up comment to Reviewer sAPs), but we think what we provide is still fairly substantial evidence towards unaligned circuits.

Finally, this contribution is secondary to our primary claim that the success of prior stitching work cannot be taken as evidence of aligned circuits. For this argument, even if computational transfer were limited by the interface or loss function, that same bottleneck would apply to prior stitching work using similar maps and objectives to those covered in our paper, so our main conclusion would still hold.

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

This paper investigates whether cross-model activation stitching transfers precomputed answers or intermediate computation. Through experiments across multiple model families and tasks, the authors find that stitching can effectively transfer answer information but provides limited evidence of computation transfer. This is primarily an analysis-driven paper with a straightforward motivation, rather than a paper proposing a substantially new method.

Critical strengths

The motivation is straightforward and intuitive. Distinguishing answer transfer from computation transfer is an interesting question that could help clarify what successful cross-model representation alignment actually implies.

Critical weaknesses
  1. Limited generality. The experiments primarily rely on single-layer, prefill-only stitching, which may limit the computational information available for transfer. Although acknowledged by the authors, this restriction makes it unclear whether the conclusions generalize to other stitching configurations.

  2. Definition of computation transfer. The paper evaluates computation transfer based on whether the recipient can continue computing beyond the donor's precomputed content. However, failure to do so does not necessarily imply that transferable intermediate computational representations are absent, as the recipient may simply be unable to utilize them.

  3. Potential bias from task supervision. The stitching map is trained using cross-entropy on the correct answer, which may inherently favor answer transfer over computation transfer. It remains unclear whether the observed results reflect a fundamental limitation of cross-model computation transfer or a consequence of the training objective.

Questions for authors
  1. To what extent do the conclusions generalize beyond single-layer, prefill-only stitching?

  2. How does the operational definition distinguish the absence of transferable computation from the recipient's inability to utilize transferred intermediate representations?

  3. Could the answer-supervised objective bias the mapping toward answer transfer? Do the existing controls sufficiently rule out this possibility?

2 comments
AuthorsOpenReview ↗

Additional Information on Computational Transfer

We think it would also be helpful to clarify our position on computational transfer so we don't overclaim in our responses. We believe there is substantial evidence that these circuits are not aligned enough for transfer. Our SVD experiment shows that there is plenty of space within the task map to transfer some form of computation. We know that the task map can interact with every part of the recipient's transformer, given that it must encode the answer to correctly interact with every component of the recipient's remaining layers, and we know that the beginning of a chain of further computational steps exists in the donor layer. Finally, our Llama results show that, even when the stitch has a substantial amount to gain from transferring computation, it is unable to. That being said, given it is a negative result, it is impossible to rule out every possibility, and there is a chance that computational transfer requires multiple nonlinear steps where either individual direction does nothing, and only those changes coordinated together lead to computational transfer. Ultimately, however, the evidence necessary for our primary claim, that cross-model stitching results reflect the transfer of a precomputed answer rather than aligned circuits, is well established by our existing suite and should prompt us to reexamine the implications of existing cross-model stitching papers.

AuthorsOpenReview ↗

Paper Revision Follow-up

Hello, we wanted to follow up and let you know that we have posted the revised manuscript, which is much clearer about the paper's contributions.

For question 1, the introduction now explains why we test single-layer, prefill-only stitching. Our tests of stitching every prompt position and of grafting donor states taken during decoding are in Section 7, with details in the appendix.

For questions 2 and 3, Section 6.2 ("What this says about aligned circuits") compares our findings directly with Chen et al. (2025) and Ramesh and Li (2025), and explains why the task map transfers only the answer even though it has both the incentive and the capacity to transfer computation. Section 7 also compares the first-token objective against alignment, distillation, and whole-answer objectives.

Reviewer_rrT9OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper investigates whether cross model activation stitching transfers precomputed answers or computational states that a model can use for further reasoning. The main method learns a linear map that transforms a larger model’s activation into a smaller model’s activation space. Experiments cover arithmetic, symbolic binding, state tracking, factual recall, and a GSM8K setting. And the authors argue that stitching primarily transfers answer related information and provides little evidence of transferable computation.

Critical strengths

The paper includes several controls, shuffled donor states, answer bottlenecks, and targeted versus random subspace erasure. These experiments provide evidence that answer related information is important to the behavior of the tested stitching systems.

Critical weaknesses
  1. The manuscript does not conform to the official ICLR submission template. The PDF omits the standard conference review header, leaving only a horizontal rule at the top of the page. And the figures would also benefit from more consistent composition. For example, the panel labels in Figure 3 are placed at noticeably different vertical levels.

  2. The paper does not adequately establish the importance of its research question or the substantive knowledge gained from answering it. A hidden state at the final prompt position participates directly in predicting the next token. When a mapping from that state is explicitly trained to make another model predict the correct next token, it is reasonable to expect the resulting interface to prioritize information useful for that prediction. Observing that such an interface transfers the first answer token more effectively than an entire solution is therefore not a particularly informative discovery. The experiments characterize the behavior of the chosen interface, but the discussion repeatedly elevates that characterization into claims about the transferability of computation and the limits of universality. The connection between the narrow empirical observation and these broader questions remains weak.

  3. The proposed transfer setting lacks a convincing practical motivation, because it requires substantial donor computation for each recipient query. The setup requires the larger model to process each new problem before its activation can be supplied to the smaller model. Although the donor need not generate a complete answer, it still performs an additional forward computation up to the selected layer. This raises a basic question: what useful role does the smaller model serve that justifies this arrangement?

  4. The first token/full answer gap has a straightforward explanation built into the setup, and the paper does not establish sufficient insight beyond that explanation. The recipient is selected precisely because it fails on the relevant problems under the native evaluation conditions. The principal mapping is then trained to improve its first answer token prediction. Providing information that helps determine the first token does not automatically provide the remaining information or computational capability required to produce the complete answer. For a multi token answer, identifying the leading digit can leave most of the solution unresolved.

Questions for authors
  1. What specific, nontrivial insight about language model computation follows from the results beyond the expectation that a next token supervised interface will preferentially transmit information useful for predicting that token?

  2. What is the intended practical use of running a larger donor model on each query before allowing the smaller recipient to answer?

  3. What evidence distinguishes the proposed interpretation from the simpler explanation that the mapping improves the supervised first token prediction while leaving the recipient’s underlying limitations largely unchanged?

1 comment
AuthorsOpenReview ↗

Paper Revision Follow-up

Hello, we wanted to follow up and let you know that we have posted the revised manuscript, which is much clearer about the paper’s contributions.

For question 1, the revised manuscript now more clearly explains what our findings show. The introduction now states the question prior stitching work leaves open, whether a stitch transfers computation or a precomputed answer, and why we test it with single-layer, prefill-only stitching. Section 6.2 ("What this says about aligned circuits") compares our findings directly with Chen et al. (2025) and Ramesh and Li (2025), and explains why the task map transfers only the answer even though it has both the incentive and the capacity to transfer computation.

For question 3, section 7 now explicitly frames the full answer gap as an impact of answer transfer. We also apply this fact to the broad implications it has on the field of stitching.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Revisions

We thank all three reviewers for their feedback, and we have uploaded a revised manuscript. We retain our central claim that task-supervised cross-model stitching transfers the donor's precomputed answer rather than computation the recipient can continue, which limits what successful stitching can be taken to show about shared circuits. The revision makes this argument and its implications much more explicit.

All three reviewers asked whether training the map on the first answer token biases it toward answer transfer. Our task-supervised map is rewarded for anything that improves the recipient's prediction, whether through answer transfer, computational transfer, or both. Despite this, precomputed answer content captures more or less all of the first-token conferral on Gemma (Section 5), even though the task map has the dimensional capacity to transfer computation, since its top 16 singular directions preserve 99% of its conferral (Section 6.2). The same holds where the stitch has a lot to gain. Few of the second or third answer digits are present at the donor site, yet a map trained on the whole answer reaches only 0.23 full-answer conferral, compared with the donor's 0.42 (Section 7). On Llama, a linear probe reads the whole answer at the donor stitch site on only 6% of problems, yet erasing the answer directions from the graft drops conferral from 0.28 to 0.07 (Section 5). Section 7 also compares intermediate-state alignment and computation continuation objectives against the first-token objective. We go into more depth on this in our individual reviewer responses and the revised section of our paper.

Revisions

  • Structure and presentation (all reviewers): We restructure the paper to make our contributions clearer. The paper now follows a structure similar to: the stitch confers -> it confers through a precomputed answer -> comparison with prior work and implications for universality -> the impacts of answer transfer. The missing review header is also fixed.

  • Motivation (rrT9, sAPs): the introduction now clearly states why we test single-layer, final-token stitching.

  • Impacts of answer transfer (rrT9, e9tC): Our new Section 7 more clearly frames the full-answer gap as an impact of answer transfer.

  • Layer pairs (e9tC): The new Appendix A.4 retrains the map at 16 donor and recipient layer pairs for Gemma and Qwen. Full-answer conferral stays low at every pair.

  • Full test set (e9tC): the new Appendix A.5 splits all 2,000 evaluation problems by which model answers natively, including a probe check on problems neither model answers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.