acceptodds
Under review as a conference paper at ICLR 2027

Matching Object or Relation? Tracing Abstract Reasoning Inside VLMs

Abstract

Vision Language Models (VLMs) excel on visual benchmarks but fail systematically on tasks requiring abstract reasoning. Existing benchmarks document this failure but cannot say why it happens or which cognitive capability is missing. We close this gap by adopting the Relational Match-to-Sample (RMTS) paradigm from comparative and developmental psychology and pairing it with a mechanistic analysis of the model's internals. On a parametrically controlled stimulus set evaluated across frontier API models (GPT, Claude, Gemini) and three open-source families (Qwen3.5, Gemma-4, InternVL3), we identify four levers that shift VLMs toward the relational match—capability tier, model scale, the number of objects per scene, and the absence of per-object stimulus noise—together producing a developmental-like trajectory that mirrors the human relational shift. Opening up the model, a per-layer representational similarity analysis and a causal mediation analysis reveal that VLM abstract reasoning is implemented by two competing circuits: an early circuit that organises images by their surface object features, and a late circuit that organises them by their abstract relation. When the model answers directly, its choice reflects which of the two circuits dominates at the output. Extending the analysis to ARC-AGI-1, we find that ablating the relational heads identified on RMTS degrades performance more than ablating random heads, suggesting that the relational circuit is recruited beyond our controlled stimuli. We hope this mechanism-level view serves as a step toward understanding how abstract reasoning is implemented in VLMs, beyond what aggregate benchmark scores can reveal.

Primary area:
interpretability and explainable AI
Keywords:
VLM, Reasoning, Interpretability

Reviews

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

This paper examines whether VLMs generalize novel visual concepts through object similarity or abstract relational structure. Using controlled Relational Match-to-Sample (RMTS) tasks, the authors find that larger models and scenes with more objects favor relational matching, while visual noise weakens it. Layer-wise RSA reveals a shift from object-aligned to relation-aligned representations. Activation patching identifies attention heads that influence relational decisions, and the authors test their contribution to ARC-AGI-1.

Critical strengths
  1. Well-motivated experimental design. The paper adapts a cognitive science paradigm to VLMs, using controlled stimuli to distinguish object-based from relational generalization.

  2. Consistent behavioral and representational findings. Experiments across model families reveal systematic trends, while RSA tracks the effects of object count and visual noise.

  3. Causal interventions and careful documentation. Head-level interventions complement the RSA analysis, and the appendices provide useful controls, reproducibility details, and limitations.

Critical weaknesses
  1. The evidence does not establish two competing circuits. RSA shows how representational similarity changes across layers, but the transition could arise from successive stages of a single computation. The authors patch attention heads only at the prediction position. These interventions may identify heads that read out an existing decision rather than compute relational structure. The paper locates signals associated with relational decisions without explaining how the model constructs those relations or demonstrating competition between distinct circuits.

  2. The experimental controls leave important alternative explanations open. Increasing object count also changes pattern redundancy and the appearance of the object-match candidate. The shape-coherence control in Appendix H does not establish whether the trend persists at larger scene sizes. Furthermore, averaging over both answer orderings makes a model that always selects A indistinguishable from one that chooses randomly, complicating the interpretation of chance-level results.

  3. The ARC-AGI-1 results do not establish relational specificity. The identified heads lie primarily in late layers, but the random-ablation baseline samples heads across all layers. The larger performance drop could therefore reflect the greater influence of late-layer heads. A depth-matched baseline and statistical tests are needed before claiming that these heads support relational reasoning beyond RMTS.

  4. The novelty needs clearer positioning. Previous studies have applied RMTS and related relational-abstraction tasks to neural networks and VLMs, including Campbell et al. (2024), Webb et al. (2023), and Puebla and Bowers (2024). The authors should distinguish their preference-based design and causal analyses from this prior work rather than presenting RMTS itself as a new diagnostic for VLMs.

Questions for authors
  1. Can the authors distinguish heads that compute relational structure from heads that merely read out the answer, perhaps by patching image tokens while keeping answer positions fixed?

  2. Does the object-count effect persist under shape-controlled stimuli at larger values of n?

  3. Can the authors report consistency across answer orderings to distinguish genuine matching preferences from positional bias?

  4. Does the ARC-AGI-1 effect remain under depth-matched random ablation?

  5. How do the authors reconcile the proposed relational circuit with the finding that CoT restores performance after its strongest heads are ablated?

4 comments
AuthorsOpenReview ↗

Author Response to Reviewer oCti (1/2)

Thank you for the positive and constructive review. We answer each point below and would be glad to hear whether this resolves your concerns. All changes in the revised paper are marked in red.

Q1

Can the authors distinguish heads that compute relational structure from heads that merely read out the answer, perhaps by patching image tokens while keeping answer positions fixed?

Yes. We did what you suggest: we patched the image tokens while keeping the answer positions fixed.

The data. Each item of a new set, which we call the interchange set, gives two prompts, c1 and c2. The two candidate images, their positions and the answer letters are the same in both. Only the samples differ, and with them the relation that decides which candidate is correct. Figure 18 (page 30) shows an example, and Appendix N.1 (pages 29–31) describes the construction.

What we do. We copy residual-stream activations from the run on c2 into the run on c1, separately for three groups of tokens: the sample tokens (image tokens of the two samples), the candidate tokens (image tokens of the two candidates) and the final tokens (the text after the last image, up to the position where the answer is predicted). Since only the samples differ between the prompts, any effect is caused by the relation they show.

Result: the relational circuit has three stages (Gemma-4-26B, 30 layers; Figure 8, page 9).

stagewherelayerswhat happens
1sample tokens0–10the relation shown by the samples is represented
2candidate tokens11–16each candidate is compared with the samples
3final tokens≥ 16the relational heads (L28H12, L28H7) turn the result into the answer
  • In each range of layers, patching that one group of tokens is enough to flip the answer. The candidate images are identical in c1 and c2, so the effect at stage 2 can only come from comparing them with the samples. The relational structure is computed at the candidate tokens in the middle layers.
  • Our relational heads are the last stage. They turn the computed result into the answer. When we zero them at the final tokens, the relational margin is lost in all five models that solve the task (Table 11, page 32).
  • Probes on the candidate tokens can tell whether a candidate matches the samples, also for relation types they were not trained on (0.98–1.00; 0.46–0.56 with shuffled labels; Table 10, page 31). The signal is about the relation in general and is not a stored answer.

The experiment separates the two roles: the relation is computed in stages 1 and 2, and the heads we identified are stage 3 of the same circuit. The results are in Section 4.3 (page 9), Figure 8 (page 9) and Appendix N.1–N.2 (pages 29–31; Figures 18–19). Thank you for the suggestion. It took us from a list of heads to a circuit we can follow through the prompt.

W1

The evidence does not establish two competing circuits. [...] the transition could arise from successive stages of a single computation.

We ran ablations restricted to the tokens where each set of heads acts. Ablating object heads at the image tokens makes the relational margin larger (Gemma-4-26B +0.83, Gemma-4-31B +1.38, InternVL3-14B +0.32). Random heads from the same layers do not. Ablating relational heads at the final tokens makes it smaller in all five models that solve the task. If these were successive stages of one computation, removing the earlier stage should not help the later one. The two sets of heads push in opposite directions. The results are in Section 4.3 (page 9) and Appendix N.3, Tables 11–12 (page 32). Thank you for pressing us on this. The competition claim now has causal evidence behind it.

Q2

Does the object-count effect persist under shape-controlled stimuli at larger values of n?

We made stimuli for n from 3 to 20 that leave out the "all different" pattern, so the pattern rule does not change with n, and we combined them with shape-controlled candidates (both in the same unseen shape). With the shape control, Gemma-4-26B chooses the relational match in at least 95% of trials at every n, and Gemma-4-31B in at least 94% for n ≥ 4, up to n = 20. This is in Section 3.3 (page 5) and Appendix N.6, Table 16 (page 33).

AuthorsOpenReview ↗

Author Response to Reviewer oCti (2/2)

Q3

Can the authors report consistency across answer orderings to distinguish genuine matching preferences from positional bias?

Yes. With Y/Z as answer letters, for which the models have no prior:

% of trialsrelational in both ordersobject in both orderssame letter in both orders
Gemma-4-26B9901
Gemma-4-31B10000
InternVL3-38B10000

These models choose the relational match whichever position it is in, so their preference is not a positional bias. We report this in Section 2.2 (page 4) and Appendix N.7, Table 17 (page 33).

Q4

Does the ARC-AGI-1 effect remain under depth-matched random ablation?

Yes. We reran the baseline with 20 seeds of random heads from the same layers. Ablating the relational heads gives 0.408, which is below every seed (0.536 ± 0.052; Wilcoxon test paired over tasks, p = 7×10⁻⁵). The drop grows with the number of ablated heads (k = 1 / 3 / 5: 0.512 / 0.408 / 0.367). A non-relational control, where the model only copies the same grids, is not affected (0.937 → 0.944). We replaced Figure 9 (page 9), updated Section 5 (page 10) and added Appendix N.4, Table 13 (page 32). Thank you for the suggestion. The ARC result now has proper statistics.

Q5

How do the authors reconcile the proposed relational circuit with the finding that CoT restores performance after its strongest heads are ablated?

The three-stage circuit gives a simple explanation. The relation is computed at the sample and candidate tokens in the early and middle layers (stages 1 and 2). The relational heads are the last stage and turn it into an immediate answer. Ablating them does not remove the relation that was computed earlier. When the model writes out a chain of thought, it can read that representation again and state the comparison in words, and this brings the performance back. We added this explanation to the CoT paragraph of Section 4.3 (page 9).

W4 (novelty)

The authors should distinguish their preference-based design and causal analyses from this prior work rather than presenting RMTS itself as a new diagnostic for VLMs.

We agree that RMTS itself is not our contribution, and we should have said so more clearly. Webb et al. (2023) compare LLMs with humans on analogy tasks at the behavioural level. Campbell et al. (2024) document VLM failures and relate them to binding. Puebla & Bowers evaluate object-centric models that are trained on visual reasoning tasks. All of these ask whether a model can solve a relational task. We ask a different question, in two parts: (i) with no training on the task, which kind of generalisation, object-based or relational, does a generative VLM prefer; and (ii) how is that choice made inside the model, which we trace causally from the samples to the candidates to the answer and test on ARC-AGI-1. We rewrote the positioning in Section 5 (page 10) and Appendix P (page 34-35).

Thank you again for the detailed comments. Please tell us if anything is still unclear, and we will be happy to explain further.

Thank you for the reply. The new experiments and results address my main concerns about the causal analysis, experimental controls, and ARC-AGI generalization.

I encourage the authors to include these results and clarify the contribution relative to prior work.

My main concerns have been addressed, and I like the concept of this paper. I will maintain my recommendation.

Thank you for reading our response and the revised paper. We are glad that the new experiments address your main concerns. The new results and the clarified contribution relative to prior work are both in the revised version, and we will keep them in the final version. Thank you again for your support.

Reviewer_RMHpOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper asks whether vision-language models (VLMs) generalize from object-level features or from abstract relational structure, using a controlled Relational Match-to-Sample (RMTS) task. The authors evaluate open-source and frontier VLMs as a function of model scale, chain-of-thought (CoT) prompting, stimulus noise and the number of objects. Using representational similarity analysis (RSA) and causal interventions on attention heads, they propose that two competing circuits, one object-based and one relational, drive the models' decisions.

Critical strengths
  1. The question of whether VLMs generalize from object features or from relational structure is relevant, and the RMTS dataset is formally defined and controlled. The behavioral evaluation covers a range of open-source and frontier VLMs and varies model scale, CoT prompting, stimulus noise and the number of objects.
  2. The paper goes beyond behavior by combining RSA with causal interventions on attention heads, although, as noted below, this analysis largely follows prior work and the causal evidence is limited.
Critical weaknesses

1. The core findings largely repeat prior work, which the paper mischaracterizes. The main mechanistic result, an early object-based stage followed by a later relational stage, is essentially the two-stage process already reported by Lepori et al. (2024), which the paper cites. The paper describes prior mechanistic work as largely behavioral or based on toy models trained from scratch, but Lepori et al. studied pretrained models such as CLIP and DINO, which are not toy models. What this submission adds is the zero-shot evaluation of generative VLMs without fine-tuning and of frontier models, which solve the task; the mechanistic interpretation is largely the same. The paper also does not engage with closely related work on relational and same-different reasoning in neural networks and VLMs, including Tartaglini et al. (2023), Geiger et al. (2020), Campbell et al. (2024), Yiu et al. (2025), Puebla & Bowers (2025), Jiang et al. (2025) and Kim et al. (2018).

The claim of two competing circuits is also not established. The transition from object-aligned to relation-aligned representations may simply reflect object representations being used to compute relations, as in the sequential account above. The causal evidence has gaps:

  • The interventions only cover the last token position, so these heads may read out a decision already computed elsewhere rather than implement a relational or object comparison. The claims in the abstract and introduction should be softened accordingly.
  • Ablating relation heads pushes the answer toward the object answer, but the converse is not shown. Even though the tested model's baseline is 100% relational, the paper could show that the logit margin for the relational answer increases when object heads are ablated.
  • The evidence that relation heads support ARC-AGI is limited. The ablated heads sit at the last token position in very late layers, so the drop in performance may come from these heads aggregating computation in general. Random-head ablations drawn from the same layers, and results for more than three ablated heads, would strengthen the claim.

2. Relational matching is confounded with shape novelty. The authors control for answer-position bias and address shape coherence, but the relational candidate always contains a novel shape, while the object candidate reuses shapes from the samples. A model could therefore pick the relational candidate because of shape novelty, without identifying the intended relation. A simple control would give both candidates the same unfamiliar shape, with only one preserving the relation.

References: Campbell, D., et al. (2024). Understanding the limits of vision language models through the lens of the binding problem. NeurIPS. Geiger, A., Carstensen, A., Frank, M. C., & Potts, C. (2020). Relational reasoning and generalization using non-symbolic neural networks. arXiv:2006.07968. Jiang, Y., et al. (2025). Investigating the functional roles of attention heads in vision language models: Evidence for reasoning modules. arXiv:2512.10300. Kim, J., Ricci, M., & Serre, T. (2018). Not-So-CLEVR: Learning same-different relations strains feedforward neural networks. Interface Focus, 8, 20180011. Lepori, M. A., et al. (2024). Beyond the doors of perception: Vision transformers represent relations between objects. NeurIPS. Puebla, G., & Bowers, J. S. (2025). Visual reasoning in object-centric deep neural networks: A comparative cognition approach. Neural Networks, 189, 107582. Tartaglini, A. R., et al. (2023). Deep neural networks can learn generalizable same-different visual relations. arXiv:2310.09612. Yiu, E., et al. (2025). KiVA: Kid-inspired visual analogies for testing large multimodal models. ICLR.

Questions for authors
  1. What evidence distinguishes two competing circuits from the sequential object-then-relation processing reported by Lepori et al. (2024)?
  2. Could the authors add a condition where both candidates use the same unfamiliar shape, with only one preserving the intended relation? Do the findings hold?
  3. Does ablating the object heads increase the logit margin for the relational answer, and how does it affect ARC-AGI performance?
  4. Are the random heads in the ARC-AGI ablation drawn from the same layers as the relation heads, and how does performance change as more than three heads are ablated?
  5. Could the authors clarify the comparison between model scaling or CoT and the human relational shift? The developmental connection is an interesting motivation, but differences across model sizes do not necessarily reflect a developmental process comparable to human relational reasoning. I suggest presenting it as motivation rather than as evidence of a shared mechanism.
5 comments
AuthorsOpenReview ↗

Author Response to Reviewer RMHp (1/3)

Thank you for the detailed and expert review. We answer each point below and would be glad to hear whether this resolves your concerns. All changes in the revised paper are marked in red.

W1 (relation to prior work)

The main mechanistic result [...] is essentially the two-stage process already reported by Lepori et al. (2024) [...] The paper also does not engage with closely related work on relational and same-different reasoning in neural networks and VLMs [...]

We should have been clearer about how our work relates to this line of research. Lepori et al. (2024) study pretrained ViTs (CLIP, DINO) that are fine-tuned on same-different and RMTS tasks. Kim et al. (2018), Geiger et al. (2020) and Tartaglini et al. (2023) also study networks that are trained or fine-tuned on the relational task. In all of these, training tells the model to use the relation. Our setting differs in four ways:

  1. No task training. We test generative VLMs and frontier models zero-shot. The relational answer has to win against a specific object-matching alternative, so we measure which kind of generalisation the model prefers on its own.
  2. Competition. With this design we find that object heads work against the relational answer (see Q1 below). A pipeline that goes from perception to relation does not predict this.
  3. Tracing across token positions. We follow the relation from the samples to the candidates to the answer, and confirm it with probes on held-out relation types.
  4. Transfer. The same heads matter, selectively, on ARC-AGI-1.

Campbell et al. (2024) and Yiu et al. (2025) study VLM analogy behaviourally, and Puebla & Bowers (2025) study object-centric models; these do not analyse the mechanism. Jiang et al. (2025) look at the roles of attention heads in VLMs more generally. We corrected our description of Lepori et al. and added all of these works in Section 5 (page 10) and Appendix P (page 34-35).

Q1

What evidence distinguishes two competing circuits from the sequential object-then-relation processing reported by Lepori et al. (2024)?

We ran ablations that can tell the two apart. If object representations were only an input to the relational computation, removing object processing should make the relational answer weaker. We find the opposite:

  • Ablating object heads at the image tokens makes the relational margin larger (Gemma-4-26B +0.83, Gemma-4-31B +1.38, InternVL3-14B +0.32). Random heads from the same layers do not.
  • Ablating relational heads at the final tokens makes it smaller in all five models that solve the task.

The two sets of heads push the decision in opposite directions, which a purely sequential account does not predict. The results are in Section 4.3 (page 9) and Appendix N.3, Tables 11–12 (page 32). Thank you for this question. It led to the most direct evidence we have for the competition claim.

AuthorsOpenReview ↗

Author Response to Reviewer RMHp (2/3)

W1 (last-token interventions)

The interventions only cover the last token position, so these heads may read out a decision already computed elsewhere [...]

This was a real gap. We now intervene on the whole prompt and not only on the last token, using new items in which only the relation changes.

The data. Each item of the new interchange set gives two prompts, c1 and c2. The two candidate images, their positions and the answer letters are the same in both. Only the samples differ, and with them the relation that decides which candidate is correct. Figure 18 (page 30) shows an example, and Appendix N.1 (pages 29–31) describes the construction.

What we do. We copy residual-stream activations from the run on c2 into the run on c1, separately for three groups of tokens: the sample tokens (image tokens of the two samples), the candidate tokens (image tokens of the two candidates) and the final tokens (the text after the last image, up to the position where the answer is predicted). Since only the samples differ between the prompts, any effect is caused by the relation they show.

Result: the relational circuit has three stages (Gemma-4-26B, 30 layers; Figure 8, page 9).

stagewherelayerswhat happens
1sample tokens0–10the relation shown by the samples is represented
2candidate tokens11–16each candidate is compared with the samples
3final tokens≥ 16the relational heads (L28H12, L28H7) turn the result into the answer

In each range of layers, patching that one group of tokens is enough to flip the answer. Stage 2 is the part that speaks to your concern. The candidate images are identical in c1 and c2, so an effect at the candidate tokens can only come from comparing them with the samples. The relation is therefore computed at the image tokens in the middle layers, before the last token.

The heads we identified are the last stage of this circuit. What they carry is the result of the relational comparison: candidate positions and letters are fixed here, so only the relation decides it. They are also needed for the result to reach the answer. When we zero them at the final tokens, the relational margin is lost in all five models that solve the task (Table 11, page 32). The paper now describes the relational circuit as these three stages across the sample, candidate and final tokens, and no longer only through the heads at the last token. We revised Section 4.3 (page 9), added Figure 8 (page 9) and give the full analysis in Appendix N.1 with Figures 18–19 (pages 29–31). Thank you for pointing this out. The account of the mechanism is far more complete now.

Q3

Does ablating the object heads increase the logit margin for the relational answer [...]?

Yes. We ablated object heads only at the image tokens, with Y/Z as answer letters, and compared with random heads from the same layers.

change in relational marginGemma-4-26B (k=1)Gemma-4-31B (k=3)InternVL3-14B (k=3)
object heads at image tokens+0.83+1.38+0.32
same-layer random−0.22 ± 0.35−0.56 ± 0.59+0.11 ± 0.13

As you expected, the margin goes up even in models that already choose the relational match almost every time. This result is in Section 4.3 (page 9) and Appendix N.3, Table 12 (page 32). Thank you for suggesting it. It was the missing half of our causal evidence.

AuthorsOpenReview ↗

Author Response to Reviewer RMHp (3/3)

Q4

Are the random heads in the ARC-AGI ablation drawn from the same layers as the relation heads, and how does performance change as more than three heads are ablated?

In the original version they were not. We reran the experiment with random heads from the same layers, and with more ablated heads.

Gemma-4-26B, ARC cell accuracyk = 1k = 3k = 5
relational heads0.5120.4080.367
same-layer random–0.536 ± 0.052 (20 seeds)0.548 ± 0.065 (5 seeds)

At k = 3, ablating the relational heads gives a lower accuracy than all 20 same-layer seeds (paired Wilcoxon, p = 7×10⁻⁵), and the drop grows with k (accuracy without ablation: 0.505). We also added a non-relational control in which the model only copies the same grids. The ablation does not affect it (0.937 → 0.944). So the drop is not explained by late heads being important for output in general. We replaced Figure 9 (page 9), updated Section 5 (page 10) and added Appendix N.4, Table 13 (page 32). Thank you. The ARC claim is much better supported now.

Q2

Could the authors add a condition where both candidates use the same unfamiliar shape, with only one preserving the intended relation? Do the findings hold?

Yes, they hold. We added this control for all 729 trials. Both candidates are drawn in the same unseen shape (pentagon, hexagon, star or diamond), so only the relation separates them. The relational rate goes from 96.6% to 99.5% for Gemma-4-26B and from 96.6% to 93.0% for Gemma-4-31B. Shape novelty does not explain the relational choice. The interchange set that we use for tracing the circuit follows the same design. The control is in Section 2.1 (page 3) and Appendix N.5, Table 14 (page 33). Thank you for proposing such a simple and clear test.

Q5

I suggest presenting it as motivation rather than as evidence of a shared mechanism.

We agree with this reading. The human relational shift is our motivation and an analogy for the behavioural axis. RMTS comes with well-studied developmental and cross-species reference points (Gentner, 2011; Christie & Gentner, 2014), and that is why it is a useful yardstick for VLMs. We now say this explicitly in Section 3.2 (page 5).

Thank you again for the detailed comments. Please tell us if anything is still unclear, and we will be happy to explain further.

Reviewer_qeEnOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper adapts relational match-to-sample tasks to study VLMs' asbtract reasoning on object similarity or abstract relations. It evaluates model scale, capability tier, chain-of-thought prompting, scene size, and visual noise. Representational similarity analysis suggests an early object-oriented stage and a later relation-oriented stage. Activation patching and ablation identify influential attention heads, and an ARC-AGI-1 experiment examines their contribution beyond the synthetic task.

Critical strengths

The behavioral investigation is broad and well organized. Counterbalancing candidate order, varying nonce labels, and testing several model families are useful design choices. The appendices also address several plausible concerns through a shape-coherence ablation, numerical-equality tasks, an alternative RSA target, and additional head-level experiments. The finding that CoT restores performance after ablating heads important for direct answers is interesting and is worthy further investigation.

Critical weaknesses
  1. The causal evidence does not clearly separate relational reasoning from answer selection. In Section 4.3, moving the relational candidate from A to B changes both its position and the correct answer letter. The identified heads may therefore carry the selected position or letter rather than compute the relation. Testing both candidate orders controls letter preference but does not resolve this ambiguity. Additional experiments should vary relational content independently of candidate position and answer labels.

  2. The object-count and noise experiments do not isolate their intended effects. Increasing object count also changes repetition and pattern structure. For example, “all different” becomes three groups of repeated values. Similarly, the claim that noise preserves object categories is difficult to reconcile with Figure 11, which shows ambiguous colors at stronger noise levels. Controls that preserve the reasoning rule and verify category recognition would clarify these findings.

  3. The ARC evidence remains preliminary. The selected heads are mostly near the output, whereas random comparison heads are sampled across all depths. Their removal may therefore disrupt answer generation more generally. Stronger evidences may include paired statistical tests, checks on tasks that do not require relational reasoning.

Questions for authors
  1. Do the identified heads retain relational specificity when candidate position and output labels are controlled independently of relational content?
  2. Does the behavioral and RSA transition persist when shape novelty is crossed with relational correctness, especially in weaker checkpoints?
  3. On ARC, do targeted ablations exceed depth- and importance-matched controls, and are their effects selective relative to nonrelational visual tasks?
6 comments
AuthorsOpenReview ↗

Author Response to Reviewer qeEn (1/3)

Thank you for the constructive review. We answer each point below and would be glad to hear whether this resolves your concerns. All changes in the revised paper are marked in red.

W1 / Q1

The identified heads may therefore carry the selected position or letter rather than compute the relation. [...] Do the identified heads retain relational specificity when candidate position and output labels are controlled independently of relational content?

You are right that in our original experiment the relation, the position of the correct candidate and the answer letter all changed together. We ran a new patching experiment where only the relation changes.

The data. We call it the interchange set. Each item starts from two trials, X and W, whose relations R_X and R_W differ in both the colour pattern and the size pattern. We draw the samples of X (S_X), the samples of W (S_W), a candidate Y_X with relation R_X and a candidate Y_W with relation R_W. Both candidates have the same shape, and no sample uses that shape. From these we make two prompts with the same token layout:

  • c1: samples S_X, Candidate 1 = Y_X, Candidate 2 = Y_W. The relational answer is Candidate 1.
  • c2: samples S_W, Candidate 1 = Y_X, Candidate 2 = Y_W. The relational answer is Candidate 2.

The candidates, their positions and the answer letters are the same in c1 and c2. Only the samples differ. Figure 18 (page 30) shows a full example, and Appendix N.1 (pages 29–31) describes how the set is built.

Terms. Sample tokens are the image tokens of the two sample images. Candidate tokens are the image tokens of the two candidate images. Final tokens are the text tokens after the last candidate image, up to the position where the answer is predicted.

What we do. We take the residual-stream activations of one token group at one layer from the run on c2 and copy them into the run on c1. We then measure how far the answer moves toward the answer of c2 (0 = no change, 1 = full flip).

Why this answers your question. Since only the samples differ, anything that changes between the two runs, including the answer, is caused by the relation the samples show. Candidate position and labels are fixed, as you asked. The key test is at the candidate tokens. Their pixels are the same in c1 and c2, so their activations can differ only because of what they have picked up from the samples. If patching them flips the answer in layers where the final tokens still have no effect, then the comparison between samples and candidates is done at the candidates and is not a read-out at the last token.

Results (Gemma-4-26B, 30 layers). The table summarises the curves in Figure 8 (page 9). Figure 19 (page 31) shows them for both contrasts and both sets of letters.

stagewherelayerswhat happens
1sample tokens0–10the relation shown by the samples is represented
2candidate tokens11–16each candidate is compared with the samples
3final tokens≥ 16the relational heads (L28H12, L28H7) turn the result into the answer
  1. In each range of layers, patching that one token group is enough to flip the answer. Stage 2 shows that the relation is computed at the candidate tokens, before the final tokens.
  2. The heads we identified are the last stage. With candidate position and labels fixed, they are still the heads with the largest effect at the final tokens (L28H12: 0.72, L28H7: 0.23; Appendix N.1, "Results", page 31). They pass the result of the comparison on to the answer.
  3. A linear probe on the candidate tokens can tell whether a candidate matches the samples, also for relation types it was not trained on (0.98–1.00; 0.46–0.56 with shuffled labels; Table 10, page 31).
  4. If we zero the identified heads only at the final tokens and use Y/Z as answer letters, the relational margin is lost in all five models that solve the task. Random heads from the same layers do nothing (Table 11, page 32).

So with position and labels fixed, the identified heads still carry the relational decision, and the relation itself is computed earlier, at the sample and candidate tokens. The experiment is in Section 4.3 (page 9) and Figure 8 (page 9), with details in Appendix N.1–N.3 (pages 29–32). Thank you for raising this. The mechanistic part of the paper is clearly better for it.

AuthorsOpenReview ↗

Author Response to Reviewer qeEn (2/3)

W2 (object count)

Increasing object count also changes repetition and pattern structure. For example, "all different" becomes three groups of repeated values.

We made a new set to check this. It uses only the patterns AAA, AAB, ABA and ABB, where the rule stays the same for every n, and leaves out "all different". Each trial is drawn twice, once with the standard candidates and once with both candidates in the same unseen shape, for n from 3 to 20.

relational rate (%)n=3456810121620
Gemma-4-26B, rule-preserving97929795989510095100
Gemma-4-26B, + shape control10010098989598100100100
Gemma-4-31B, rule-preserving951009795989810098100
Gemma-4-31B, + shape control819594949494979598

The models keep choosing the relational match in large scenes even without the "all different" pattern, so this choice does not depend on that pattern turning into repeated groups. The new set is described in Section 3.3 (page 5) and Appendix N.6, Table 16 (page 33). Thank you for the suggestion. It makes this part of the analysis cleaner.

W2 (noise)

[...] the claim that noise preserves object categories is difficult to reconcile with Figure 11, which shows ambiguous colors at stronger noise levels.

We tested this directly. At the noise levels of our sweep, we asked the models to name the colour of an object and to say whether two objects have the same colour.

Gemma-4-26Bσ_c=05102040
naming the colour1.001.001.000.990.91
same colour? (balanced acc.)0.990.960.800.630.55

The other three models behave the same way (colour naming ≥ 0.99 up to σ_c = 20). The models can still name the colours, but comparing two objects gets harder. This is what we meant by saying that the categories are preserved while the relation is weakened. We added the test in Section 3.4 (page 6) and Appendix N.8, Table 18 (page 33). Thank you for pointing out that this claim needed evidence.

AuthorsOpenReview ↗

Author Response to Reviewer qeEn (3/3)

W3 / Q3

The selected heads are mostly near the output, whereas random comparison heads are sampled across all depths. [...] On ARC, do targeted ablations exceed depth- and importance-matched controls, and are their effects selective relative to nonrelational visual tasks?

We repeated the ARC experiment with the three controls you ask for.

  1. Random heads from the same depth. We now draw random heads from the same layers as the relational heads and exclude the top-10 relational and object heads. Ablating 3 relational heads gives 0.408. All 20 random seeds are higher (0.536 ± 0.052).
  2. Paired test. A Wilcoxon test paired over the 66 tasks gives p = 7×10⁻⁵. The drop also gets larger when we ablate more heads (k = 1 / 3 / 5: 0.512 / 0.408 / 0.367; random at k = 5: 0.548).
  3. A non-relational task. We ask the model to copy the same ARC grids in the same output format. Here the ablation changes nothing (no ablation 0.937, relational heads 0.944, random 0.943).

So the drop is larger than for depth-matched controls, and it does not show up on a task without relational reasoning. It is not a general problem with producing answers. We replaced Figure 9 (page 9), updated Section 5 (page 10) and added Appendix N.4, Table 13 (page 32). Thank you for these suggestions. The ARC result is on much firmer ground now.

Q2

Does the behavioral and RSA transition persist when shape novelty is crossed with relational correctness, especially in weaker checkpoints?

Yes. We added a control where both candidates have the same unseen shape (pentagon, hexagon, star or diamond), so the two differ only in the relation.

  • Behaviour. The relational rate stays high: 96.6% → 99.5% for Gemma-4-26B and 96.6% → 93.0% for Gemma-4-31B.
  • RSA in weaker checkpoints. We reran the RSA on these stimuli for five smaller checkpoints. The late relation-aligned stage is there in all five:
late-layer Rel − Obj RSAGemma-4-E4BInternVL3-8BQwen3.5-0.8BQwen3.5-4BQwen3.5-9B
original stimuli+0.29−0.16−1.28+0.20+0.30
same unseen shape+0.65+0.31+0.23+0.33+0.42

The relational representation is not a by-product of shape novelty. We describe the control in Section 2.1 (page 3) and Appendix N.5, Tables 14–15 (page 33).

Presentation

We extended Section 4.3 (page 9) with the traced circuit (samples → candidates → relational heads → answer) and added Figure 8 (page 9), which we hope makes this section easier to follow.

Thank you again for the detailed comments. Please tell us if anything is still unclear, and we will be happy to explain further.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.