acceptodds
Under review as a conference paper at ICLR 2027

Where Function Goes: Functional Correspondence by Partial Coupling Across Mechanistic Levels

Abstract

Comparing a pretrained model with its fine-tuned descendant requires deciding what counts as the same function. Similarity in weights, activations, or features does not guarantee the same behavioral role, so it cannot by itself tell whether a component's function was preserved, relocated, or replaced. We define functional equivalence by intervention: under a behavioral lens, a component's function is the signed change in behavior its removal causes, split into an amount and a role. We cast correspondence as a capacity-respecting partial coupling of these amounts that leaves unmatched function explicit. We apply it to two Base-to-SFT lineages, at weight groups and activation writes, and to two vision transformers that share no native coordinates. Most matched mass stays at its original locus, whereas T\"ulu under an instruction lens relocates up to 19% of matched weight mass. The coupling also predicts what interventions do: it locates where independently removed weight packets take effect, restoring them better than layer-matched random writes, and it finds targets whose removal reproduces a component's removal better than representation matching. In OLMo, ranking SFT activations by unmatched mass recovers more of the Base–SFT gap than ranking by amount; restoring the top 1% removes every opening <think> tag at half the pretraining-text cost of amount ranking. Between DeiT-3 and Swin, signed roles still rank targets better than CKA, although absolute reconstruction remains weak, so we claim ranking transfer, not architectural equivalence. To summarize, we provide a correspondence tool for targeting interventions and diffing fine-tuning changes without shared coordinates.

Primary area:
interpretability and explainable AI
Keywords:
Model Diffing, Optimal Transport, Attribution, Mechanistic Interpretability

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 4 reviews, above 27% of the others.
Reviewer_FVfEOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

To the best I can tell, the authors propose a method for comparing a base model to a post-trained variant, based on behavioral properties of lesioning different components.

Critical strengths

None

Critical weaknesses

Unfortunately, I think this article should be desk rejected. I found the paper, especially sections 2-3, to be unreadable, which due to its over-formalization seems like LLM generated text to me.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer FVfE

Thank you for raising the readability concerns. We agree that the presentation, particularly in Sections 2–3, could better motivate the formalism, and we have revised the manuscript accordingly.

In the revised paper, we (1) add intuitive definitions and references for key terminology in the Introduction, (2) add a conceptual roadmap and simplified intervention formulations in Section 2, (3) explicitly connect each formal distinction to failures of simpler alternatives in Section 3, and (4) clarify experimental motivations and interpretations in Section 4.

Nevertheless, while the presentation needed clarification, we maintain that the formalism serves necessary conceptual distinctions: signed roles distinguish opposing behavioral effects (Fig 2.); role and amount separate what a component does from how much it contributes; and capacity-constrained partial coupling accommodates splitting, merging, and unmatched function without over-claiming target mass (Section 3.2; Appendix D.3). Section 3 demonstrates the necessity of these distinctions through controlled constructions.

Regarding the concern about LLM-generated text, the conceptual framework, definitions, and mathematical formulations were developed by us; AI tools were used to polish the writing as stated in AI use statement. We would appreciate specific examples of unnecessary formalism or passages supporting the concern about LLM authorship, so we can address them precisely.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

We would like to kindly follow up on our response. We have revised Sections 2–4 to improve readability and clarify the motivation behind our formalism, including intuitive definitions, a conceptual roadmap, simplified formulations, and controlled examples illustrating why the formal distinctions are necessary.

We would greatly appreciate it if you could review these revisions and let us know whether they address your concerns. In particular, we would welcome specific examples of remaining unnecessary formalism or passages raising concerns about LLM authorship, so we can address them more precisely during the discussion period.

Thank you again for your time and consideration.

Reviewer_xMf3OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces functional correspondence, a method that aligns components in base and SFT models by their ablation effects instead of what they encode. The approach adapts partial optimal transport, which allows functional units to stay, split, merge, appear, or disappear. The authors validate the method on planted changes, measure how much functional units relocate after fine-tuning, and demonstrate superiority over baselines in correspondence across weights and neurons, selecting ablation targets after fine-tuning that mimics the effect before fine-tuning, and erasing new functions from fine-tuning.

Critical strengths
  • Analyzing the behavioral effects of units rather than merely what they encode or represent is an important step towards understanding how networks compute and how different networks align on a functional level. Looking at how components’ behavioral effects align before and after fine-tuning seems novel and potentially important to informing fine-tuning procedures and fine-grained model editing.
  • The authors conducted extensive experiments on many different models in many settings to demonstrate the usefulness of the method.
Critical weaknesses
  • The presentation is convoluted and hard to understand. For example, many terms are introduced and used before they are defined (some not at all), such as “lens”, “amount”, “role”, “mass”, “endpoint”, “inventory”, “energy”, “write”, and “footprint”. The narrative often unnecessarily repeats itself but does not give enough intuition to the reader.
  • The functional meaning of the matched components is generally not interpreted. The one exception is the functional restoration in Section 4.4, where the function of the unmatched units is (or at least includes) outputting <think> tags. But in all other cases, there is no interpretation of the functional correspondence. This is not by itself critical as long as we can trust that matched functions are actually similar. But the matching process depends substantially on κ\kappa (Table 7), and as the paper states, even matched units at the same locus don’t necessarily have the same function. This seems to make the metrics from the matching process unreliable as an indicator of functional correspondence.
Questions for authors
  • Please streamline the presentation with clear definitions and intuition for key terms and avoiding unnecessary jargon.
  • In cross-level consistency, how exactly is coupling used to predict the units the weights act through in a SFT model? This result seems a bit trivial from what I understand. If you ablate all weights connected to an MLP neuron, that is exactly ablating the neuron itself, so the roles of the weights and the neuron would be exactly the same. Also, since most of the matched mass doesn’t move (Figure 3), it would be good to compare to a baseline where all weights and neurons simply stay put. If you still recover most of the removal effect, then the coupling doesn’t seem to doing much.
  • You mention that nearest-signed-role performs similarly to correspondence and that it is the signed roles and not the coupling that is responsible for the gain. What exactly does nearest-signed-role do and how does it compare to correspondence?
  • Cross architecture matching is weak, even though it is better than CKA. Based on the numbers, I would hesitate to say that this method is able to align components across architectures. Also, why should it be able to? Different architectures could solve the task differently, in which case functions should not correspond.
  • The outcome of the matching process substantially depends on the parameter κ\kappa. What is the quality invariant with respect to κ\kappa? Is it semantic meaning? Something else? For semantics, are there ground truth circuits in LLMs that are consistent before and after fine-tuning so that you could see if coupling could recover it for a wide range of κ\kappa?
5 comments
AuthorsOpenReview ↗

Response to Reviewer xMf3 [1/3]

Thanks for your sincere review. We provide responses for each point.


[W1, Q1] Coining terms and definitions

We agree that the presentation would benefit from clearer high-level intuition. Our newly introduced terminology is intended to convey conceptual insights beyond the formulas, but we acknowledge that its motivation and definitions could be clearer. We improve these by clearly defining implicit terms in the revised manuscript highlighted in green. We would appreciate it if you could specify any remaining ambiguous terms.


[W2] The meaning of functional correspondence

We split your opinion into two points and clarify each as follows.

1. What does functional similarity mean?

We would like to distinguish semantic interpretation from functional similarity under our operational definition. Two components are functionally similar when their removal effects exhibit similar signed patterns on the evaluated items, regardless of whether their underlying computations or semantic interpretations are identical. And the case of <think> token is canonical to interpret and in addition we provide some case studies in Appendix D.5.

2. Why is the inferred correspondence reliable?

First, regarding same-locus components, we do not assume that components at the same position retain the same function after fine-tuning. In fact, demonstrating that their roles can change substantially is one of our findings in Section 4.1 and Appendix D.1. A same-locus match does not imply an unchanged function; rather, it indicates correspondence within the specified role-distance tolerance, while there might exist potentially substantial role drift. For example, in Appendix D.1, median same-locus role distances range from 0.36 to 0.61, compared with 0.73–0.75 for random off-locus pairs. Thus, same-locus components tend to have more similar roles, but their functions are not necessarily unchanged. More generally, the coupling assigns positive flow only to pairs whose role distance satisfies (Krs<κK_{rs}<\kappa), independently of their positional identity.

Second, we agree that the resulting coupling depends on the tolerance κ\kappa, as it determines which role distances are considered compatible. But we would like to clarify that the dependence of κ\kappa is the property of correspondence, rather than the indication of unreliable functional correspondence measurement. We provide a detailed response in [Q5] below.

Finally, the usefulness of the inferred correspondence is validated independently of the matching objective. Sections 4.2 and 4.3 test whether the correspondence predicts separately performed interventions and whether matched target supports reproduce source removal effects on held-out items. Thus, our evidence for functional correspondence does not stay solely on the matching statistics themselves.

Besides, in terms of yielding a coupling with a given kappa, we evaluate the stability of solutions. As reported in Appendix D.3 (Stability of the residual ranking across LP optima. paragraph), we check that different independent runs are identical.

AuthorsOpenReview ↗

Response to Reviewer xMf3 [2/3]

[Q2] Cross-level consistency and the stay-put baseline

In Section 4.2, we transport a weight packet's Base activation footprint through the Base-to-SFT activation coupling to predict its SFT footprint. We acknowledge that weight and activation interventions are not fully independent, particularly for MLP channels. However, the relationship is not always a direct equivalence, as Q/K groups lack corresponding single-write interventions and weight-removal effects may extend to downstream units.

While Figure 3 shows that most matched mass remains at the same locus, same-locus correspondence does not imply identical functional roles: Section 4.1 and Appendix D.1 report substantial nonzero role distances even for same-locus components. Nevertheless, we agree that native-location matching is a meaningful baseline.

Since Section 4.3 directly evaluates functional target selection, we added a coupling-free stay-put baseline alongside representation matching (CKA for activation units, weight-slice cosine for weight groups) and random selection, under the same intervention budget and with the same four sources in each of the eight settings used in Section 4.3. The table reports correspondence wins / ties / losses in EshapeE_{\mathrm{shape}}, splitting sources by whether the correspondence support stays within the source's native block (the same MLP channels or attention heads as the source, while for activation units, the unit itself).

Sourcesnvs Stay-putvs CKA / Cosinevs Random
Support within the source's native block66 / 0 / 06 / 0 / 06 / 0 / 0
Support outside the native block2617 / 0 / 922 / 0 / 425 / 0 / 1
All3223 / 0 / 928 / 0 / 431 / 0 / 1

In short the correspondence wins more than stay-put.

To be detail, within the native block, correspondence wins all 6 cases, but only by small margins ((\Delta E_{\mathrm{shape}}\leq 0.016)). Thus, these gains reflect within-channel allocation rather than functional relocation. Outside the block, it outperforms CKA (22/26) and random (25/26), but not highly better than stay-put (17/26). Thus, native location remains a strong prior, and evidence for the benefit of relocated correspondence is limited. We will add this baseline to Section 4.3 and qualify our claims accordingly.


[Q3] Clarifying nearest-role matching compared to partial optimal transport (LP solving)

The nearest-signed-role control in Section 5 works as follows. For each source rr, it picks the NrN_r targets with the smallest signed-role distance KrsK_{rs}, where NrN_r is the size of the coupling's 95%-flow support for rr. Like CKA, it is evaluated under the protocol shared by all target-selection arms: the cardinality NrN_r and the common budget are fixed by the coupling, so the arms differ only in which targets they pick. Its choice depends on signed roles alone, with no tolerance cut (Krs<κ⋆K_{rs}<\kappa^\star) and no shared target capacity. Comparing it with correspondence tests whether the ranking gain needs the coupling or only signed roles; comparing it with CKA tests whether signed roles beat representation similarity under the same protocol.

In single-source interventions under cross-architecture setup, the two perform comparably (Table 20), suggesting that the improvement over CKA primarily comes from signed roles. However, when multiple sources compete for the same targets, independent selection can over-allocate target mass (Figure 13), highlighting the value of capacity-aware coupling. We will specify this distinction in the manuscript.

AuthorsOpenReview ↗

Response to Reviewer xMf3 [3/3]

[Q4] Cross-architecture experiments and the functional correspondence

We agree that the cross-architecture results do not establish accurate effect reconstruction or comprehensive component alignment, although our method improves target ranking over CKA in 32 of 40 cases. We therefore scope our claim to the usefulness of functional roles for target ranking, rather than complete cross-architecture alignment reconstruction.

Regarding why functions might correspond across different architectures, we would like to clarify that our notion of function concerns component-wise signed behavioral effects under a specified lens, rather than the underlying internal computations. For illustration, consider two models with different architectures that both compute (15x). Despite their different implementations, they could each contain a component contributing a factor of three to the output, whose removal produces corresponding behavioral effects. Similarly, two image classifiers may rely on different computations but contain components whose removal similarly changes the score of a particular class (e.g., matched components that both have a positive role on images where a bird's beak is visually salient) across evaluated images.

Importantly, we do not assume that such correspondence must exist. Different architectures may indeed have no corresponding components under our definition, which is why our partial coupling explicitly allows unmatched contributions. Thus, our cross-architecture experiment investigates whether useful functional correspondence can be identified despite architectural differences, without presupposing its existence, as we highlight in Section 2 that this notion of functional correspondence can be extended whenever there are the common output (readout) we can anchor on.


[Q5] Interpretation of tolerance parameter κ\kappa

The dependence on κ\kappa is an intended property of our correspondence formulation. Analogous to thresholds on sentence-embedding similarity, κ\kappa determines how much functional-role dissimilarity is tolerated. In Eq. 8–10, it governs the trade-off between matched mass and role distance under capacity constraints: smaller values require closer roles, while larger values permit broader allocation. The tolerance path therefore reveals how correspondence varies across lenses. For example, Figure 8 shows that matched mass of Tulu's weight grows faster under the General lens than under the Instruction lens at low κ\kappa, indicating more correspondence under stricter criteria.

What remains invariant is the definition and measurement of functional roles, not the resulting correspondence. As clarified in the response to the reviewer hpEx’s [W1, Q1], these roles characterize signed behavioral intervention effects under a specified lens, rather than circuit-level identity or semantic labels. We establish their utility by recovering planted functional correspondences (Section 3), predicting held-out intervention effects (Sections 4.2–4.3), and guiding selective rollback (Section 4.4). These tests provide evidence that the correspondence captures behaviorally meaningful relationships beyond component similarity, without requiring ground-truth circuit identities.

Reviewer_nB9ZOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The main question studied in the paper: after fine-tuning, which parts of the model maintain the same function, and for which parts the function has changed? To do so, the authors analyze how removing a given component affects the log-probabilities of different pieces of text from pre-training & fine-tuning data, and try to match the components across the two versions of the model. They find that most matched contributions in both versions of the model are associated with the same locations, with some differences in Tülu.

Critical strengths

The paper evaluates both text-only LLMs and vision transformers.

Critical weaknesses

The paper is very difficult to read, especially sections 2-4. Some of the introduced terminology and formalism describe relatively simple ideas, and in this way make it even harder to understand the work.

Questions for authors

None

2 comments
AuthorsOpenReview ↗

Response to Reviewer nB9Z

Thank you for your feedback. We agree that the presentation, particularly in Sections 2–4, could better convey the underlying intuition, and we have revised the manuscript accordingly.

Specifically, we add intuitive definitions of key terminology in the Introduction, introduce a conceptual roadmap and simplify intervention formulations in Section 2, explicitly motivate each formal distinction through failures of simpler alternatives in Section 3, and clarify the experimental objectives and interpretations in Section 4.

Nevertheless, we believe the formalism serves important conceptual purposes. Our goal is not merely to match components across models, but to define what it means for components to have corresponding functions. This problem requires specifying the behavioral lens, separating what a component does (role) from how much it contributes (amount), and accommodating splitting, merging, and unmatched function under capacity constraints (Section 3.2; Appendix D.3). Section 3 further illustrates why these distinctions matter through controlled constructions.

We would appreciate specific examples of terminology or formalism that you still find unnecessarily complex, so we can address any remaining concerns more precisely.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

We would like to kindly follow up on our response. We have revised Sections 2–4 to improve clarity and intuition, including a conceptual roadmap, simplified formulations, and more explicit motivations for the formal distinctions through controlled examples.

We would greatly appreciate it if you could review these revisions and let us know whether they address your concerns. In particular, we would welcome any specific examples of terminology or formalism that remain unnecessarily complex, so we can address them during the discussion period.

Thank you again for your time and consideration.

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

This paper asks how the functions of model components change after fine-tuning. It describes each component by the magnitude and signed pattern of its removal effects, then matches these profiles using a capacity-constrained partial coupling. The experiments compare two Base-to-SFT lineages at both weight and activation levels, and also test matching between DeiT-3 and Swin. The resulting matches are used to locate intervention targets and identify fine-tuning changes.

Critical strengths

I find the signed profiles a useful way to compare components. Keeping the sign matters: two components may affect the same inputs but push the readout in opposite directions (§2.2; §3.3). Separating role from magnitude makes this distinction easy to interpret. Partial coupling then gives a common description of splits, merges, and support that has no match.

The scale and range of the experiments are a main strength. The authors study two model lineages under two lenses, at both weight and activation levels, before extending the analysis to models with different architectures (§§4–5). I particularly like the cross-level test, where an activation coupling predicts the effects of separately removing weight packets. The held-out and prospective tests, sensitivity checks, and simpler matching controls make the experimental picture fairly complete.

Critical weaknesses

I would clarify how broadly to read "functional equivalence." Equal roles in Eq. (4) mean positively proportional signed effect profiles on the selected items and readout. Different internal computations can still have this property. The authors already acknowledge the dependence on the lens and restrict the cross-architecture claim to target ranking. Keeping this qualification close to the descriptions of preserved, relocated, or newly appearing function would help avoid a stronger mechanistic interpretation. This is a concern about how the results are described, not about the validity of the definition itself.

I am less convinced that solving the LP is essential to the practical gains. Against nearest signed-role matching, coupling has 14 wins, 19 losses, and 7 ties in Table 20. Capacity-aware greedy allocation also performs similarly to the LP in Tables 17 and 21. My reading is that signed roles explain the single-source ranking gains, while capacity constraints become useful when sources compete for targets. The paper does acknowledge this. The unified allocation is still useful, but a shorter, more direct statement of what it adds over these simpler rules would make the contribution easier to assess.

Questions for authors

Should preservation, relocation, and appearance all be read as changes in relative intervention effects under the chosen lens, without implying that the underlying mechanisms are the same?

Given the nearest-role and greedy controls, is the main reason for using the LP to obtain a coherent global allocation? A brief explanation of its role beyond target ranking would be helpful.

2 comments
AuthorsOpenReview ↗

Response to Reviewer hpEx

Thanks for your sincere review. We provide the responses according each point.


[W1, Q1] The scope of interpreting “functional equivalence”

We agree that our definition of functional equivalence does not imply equivalence of the underlying internal computations. In fact, this distinction is intentional.

We presume that the mentioned “internal computation” is referring to a mechanism in circuit-level sense, which preserves the information of relation between components. If so, our goal is not to establish circuit-level mechanistic correspondence, but to define a more tractable notion of functional correspondence for model comparison/diffing. Circuit-based formulation may include context-dependent computations that are hard to abstract for establishing correspondence at the component and dataset (lens) levels. For correspondence, we believe that this coarser perspective is sufficient to characterize 5 different correspondence types without a complete reconstruction of the underlying computations. We therefore define functional roles by component-wise signed behavioral removal effects under a specified lens, deliberately abstracting away complex computations.

We also agree that correspondence under one lens does not guarantee correspondence under another. However, we view the lens not merely as a restriction, but as an explicit part of the operational definition of function. Generalization beyond the evaluated lens is a separate question, as it is hard to formally guarantee model performance on unseen test data after training (even though we evaluate on held-out samples such as in Section 4.3).

Accordingly, correspondence types (preservation, relocation, and appearance) refer to changes in functional correspondence at this chosen level of abstraction, rather than the preservation or movement of underlying circuits. We explicitly clarify this distinction in Introduction (line 72-77) , Section 2 (line 110-114), and Section 4 (443-446). Also we cleary state the results accordingly. Same-locus does not entail that roles are unchanged (line 304-306),while we validate correspondence through intervention prediction (line 373–377, 396–398), target selection (line 407–411), and restoration (line 429–438).


[W2, Q2] Comparison between greedy matching and partial optimal transport

In short, LP can avoid overclaim of target mass and recover rollback better compared to nearest-role and can give a guaranteed LP gap bound compared to greedy. We clarify the interpretation from observed experiments as following points.

  • If there is a competence with the limited capacity, the nearest-role assignment which does not consider capacity constraints claims more than targets can hold (fine-tune setup, single source, Figure 13.). Considering capacity also matters for rollback, which enables recovering the rollback score 0.098 for OLMo and 0.0248 for Tulu (Prospective replication paragraph, Appendix D.4.)
  • Comparing the capacity-aware greedy matching and LP, we observe that the LP gaps from two output couplings are nearly identical and the target rankings coincide with Jaccard 0.996 for OLMo and 0.948 for Tulu (Greedy instead of LP allocation paragraph, Appendix D.4.)

Still, we want to highlight that the core contribution of LP is guaranteeing the result coupling to have an LP gap within an allowed error bound. While, it is known that the greedy approach is a classical choice to approximate the solution of weighted bb-matching problem (our problem can be viewed as this family of problem, specifically as a bipartite case, where bb is the capacity of a target component.) guaranteeing with half of an optimum [1] but do not necessarily guarantee with a given error bound.

We will articulate distinctions between greedy matching, nearest-role matching and partial optimal transport (LP) more directly. In addition, we regard greedy as a practical alternative to LP, since the results show that two couplings are very similar, and we will state this opinion too.


[1] J. Mestre, Greedy in Approximation Algorithms, Algorithms – ESA 2006, P. 534 Thm. 8

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

We would like to kindly follow up on our response. We clarified the scope of functional equivalence and its distinction from circuit-level mechanistic correspondence, and provided additional analyses comparing greedy matching, nearest-role matching, and partial optimal transport, including their practical differences and theoretical guarantees.

We would greatly appreciate it if you could review our clarifications and let us know whether they address your concerns. We would be happy to provide any further clarification during the discussion period.

Thank you again for your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Gentle Reminder of Our Rebuttal

Dear Reviewers and AC,

We sincerely thank you for your time and constructive feedback on our submission.

As the discussion period closes in three days, we would like to kindly follow up on our responses. We have replied to each review individually, and we would be grateful if you could let us know whether our responses address your concerns. If any questions remain, we would be glad to provide further clarification or additional results in the remaining time.

Thank you again for your effort and consideration.

Best regards,
Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.