acceptodds
Under review as a conference paper at ICLR 2027

Predicting Multi-View Rashomon Representation: Can We Learn Where Models Disagree?

Abstract

Foundation models are increasingly adopted across a wide range of applications, often serving as core blocks within AI systems. Yet different foundation models may encode the same input from multiple different views, leading to substantial representation disagreement, which we term ***Rashomon Representation***. Such disagreement often signals inputs that a given model encodes in a way inconsistent with other models, offering a valuable yet underexplored signal for input reliability estimation. While prior work has largely focused on ***measuring disagreement*** across multiple models with a representation set, we instead focus on ***predicting disagreement*** from a single representation. We hypothesize that this disagreement follows some consistent, input-dependent patterns rather than occurring at random. To test this, we quantify disagreement by comparing each sample's nearest neighbors across different models' representation spaces, then train a lightweight predictor that estimates disagreement from a single model's representation. At inference time, given a new input, the predictor uses that input's representation to tell whether it aligns with or diverges from those of other models. Extensive experiments across diverse foundation models and datasets show that representational disagreement is indeed input-dependent, predictable, and generalizable, enabling efficient reliability estimation of foundation models.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Representation Learning, Foundation Models

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_x3cTOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

The paper formalizes Rashomon Representation, a disagreement in representation across foundation models for an input. To predict whether an input can cause disagreement between models, the paper proposes a lightweight predictor trained with cross-model nearest-neighbor consistency over a reference dataset. Experiments are conducted to demonstrate the predictability of disagreement for test samples with the trained lightweight predictor. During inference, disagreement is predicted from a single representation.

Critical strengths

Using only a single representation and an anchor model, the proposed model is able to predict the disagreement between the representations.

Critical weaknesses
  • The problem is not well motivated. What’s the significance of predicting disagreement?

  • The experimental setup looks unsatisfactory. All the paper does is quantify correlation with their own created ground truth (neighborhood agreement). Is there another way to quantify ground truth?

  • The paper’s idea looks heuristic in the design of the models for predicting NA and the use of an anchor model. Is there any principled way to select those?

Questions for authors

see weakness section

1 comment

We thank the reviewer for the valuable feedback and would like to clarify our core contribution. Our central finding is that cross-model representational disagreement can be predicted from a single model's representation. Neighborhood Agreement and the lightweight predictor provide a way to test this empirically.

  • For Critical Weakness 1
    • Regarding scientific significance, by revealing systematic, input-dependent patterns of disagreement that can be learned and generalized to unseen inputs, we provide new empirical insights into the underlying structure of AI representations. This finding offers a new perspective on cross-model representational disagreement and opens up new avenues for improving foundation models and using them more effectively.
    • Regarding practical significance, predicting disagreement allows us to quickly flag inputs on which models organize information differently, without repeating reference-set searches at inference time. These inputs can then be prioritized for further analysis, for example, to assess whether the models capture complementary information that could inform representation fusion. Because the disagreement score is defined independently of any particular downstream task, our method can be applied to most vision tasks.
  • For Critical Weakness 2
    • We supplemented two additional representation metrics that are equally common and validated as effective, conducting experiments under the exact same experimental settings (such as the predictor architecture and datasets). The results consistently exhibit positive correlations, demonstrating that the predictability of representational disagreement does not rely on the specific method used to measure disagreement, but rather represents a pervasive phenomenon.
      • Experiment Setup
        • Using ImageNet-1K as the dataset, we construct supervision targets on the training split for each of the two additional measurement metrics to train the respective predictors and evaluate them on the test split, while keeping all other experimental settings identical to those in the main text.
      • BackboneMetricNA50\mathrm{NA}_{50}Dist1\mathrm{Dist1}Ref. Spearman
        CLIP ViT-B/32Pearson rr ↑\uparrow0.73650.74980.9123
        MAE ↓\downarrow0.06100.03860.0102
        DINOv2 ViT-B/14Pearson rr ↑\uparrow0.74620.85460.9478
        MAE ↓\downarrow0.06150.03200.0057
        ViT-B/16 (IN-21K)Pearson rr ↑\uparrow0.70810.85270.9589
        MAE ↓\downarrow0.06310.03010.0081
  • For Critical Weakness 3
    • In our experimental design, our choices of NAk\mathrm{NA}_k, the predictor architecture, and the anchor model are guided by the following principles:
      • First, NAk\mathrm{NA}_k has been well-validated in prior literature as an effective method for quantifying inter-model representational disagreement, which is why we adopt it as our measurement metric.
      • Second, for the predictor architecture, we deliberately selected simple and widely validated structures; the success of such standard architectures without specialized designs demonstrates that the predictability of representational disagreement is a pervasive phenomenon rather than an artifact of a particular architectural design.
      • Finally, regarding the anchor model, we incorporated models spanning diverse architectures and training paradigms to make representational disagreement more pronounced, and our experiments confirm that the predictability of representational disagreement holds regardless of the anchor model choice.
Reviewer_QqGqOpenReview ↗
Rating
4
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This is an outstanding, highly inspiring work that opens up a novel research direction for representation analysis and model reliability estimation. The paper addresses an important and under-explored problem. I recommend acceptance with a top rating.

Critical strengths

1.The paper redefines the research boundary of representational similarity analysis by shifting from retrospective measurement to proactive prediction. The concept of Rashomon Representation is vivid, accurate, and highly resonant, which will attract broad interest from the community.

2.The neighborhood agreement metric is simple, principled, and requires no additional training or alignment assumptions. The lightweight predictor enables extremely efficient inference, achieving 3.85×–9.28× speedup and orders-of-magnitude reduction in memory and storage compared to direct measurement, making it readily applicable in real-world AI systems for input reliability screening.

3.The experimental coverage is remarkably thorough, spanning model choices, comparison set sizes, distribution shifts, architecture ablations, and downstream task validation. The consistency of results across all settings strongly supports the core claim that representational disagreement is structured, input-dependent, and predictable.

4.The work reveals that cross-model representational disagreement is partially encoded within individual model representations. This finding is not only surprising but also thought-provoking, raising fundamental questions about the nature of learned representations and the commonalities/differences across foundation models.

Critical weaknesses

The work is of very high quality overall. The following points are minor opportunities for further extension, rather than flaws that undermine the core contribution:

1.The underlying mechanism of predictability is not deeply discussed. Why a single model’s representation carries signals of cross-model disagreement, and what sample-level factors drive high vs. low disagreement, could be further investigated.

2.The impact of reference set domain mismatch is not systematically studied. The performance boundary when reference data and query data come from vastly different distributions could be more explicitly characterized.

Questions for authors

1.Do the authors plan to extend this prediction framework to foundation models of other modalities (text, audio, etc.)? Given the different properties of representation spaces across modalities, how do you expect the prediction performance to change?

2.In your view, what is the core reason why representational disagreement is predictable from a single representation? Is it driven primarily by the inherent ambiguity/complexity of input samples, or by common patterns in how different models’ inductive biases diverge?

1 comment

We sincerely thank you for your high recognition of our work and your insightful, professional comments. The extension questions you raised are very important; we have also reflected on these issues during our research, and we provide our discussion and responses below:

  • For Critical Weakness 1 and Question 2

    • Since both Critical Weakness 1 and Question 2 investigate the underlying mechanisms behind the predictability of representational disagreement, we address them together here.
      • The primary goal of this study is to establish the scientific finding that "cross-model representational disagreement is predictable" and to validate its generality through extensive empirical experiments. Due to the inherent complexity of deep neural network representation spaces, fully uncovering the underlying mechanism remains a highly challenging open question. We believe that both the emergence and predictability of sample-level representational disagreement stem from the joint effect of intrinsic sample properties and model inductive biases: as shown in the qualitative visualizations in Appendix C (Fig. 7), high-agreement samples typically feature a single, clear semantic subject, whereas low-agreement samples often involve complex background clutter, occlusion, or atypical viewpoints. Such samples naturally possess "multi-interpretability," which easily induces different models to attend to different visual cues, thereby leading to disagreement. Meanwhile, the distinct inductive biases of each model imprint model-specific characteristics onto their generated representations, which is likely the core reason why such disagreement can be predicted.
  • For Critical Weakness 2

    • In Section 4.1 of our paper, we preliminarily characterized this performance boundary by fixing the ImageNet-1K training split as the sole reference set and evaluating across 9 external domain datasets, natural distribution shifts (ImageNet-V2/R), and varying severities of image corruption (ImageNet-C):
      • When query samples belong to natural objects or common scenes (e.g., ImageNet-V2 with r≈0.71–0.72r \approx 0.71\text{--}0.72, Aircraft with r=0.661r = 0.661, and Caltech101 with r=0.656r = 0.656), the ImageNet-1K reference set provides semantically meaningful neighborhood structures, and the predictive correlation remains at a high level;
      • When the query distribution undergoes a substantial domain shift relative to the natural image reference set—for example, on EuroSAT (rr drops to 0.4050.405), ImageNet-R (rr drops to 0.5120.512), or under severe Gaussian noise corruption (ImageNet-C S5, where rr drops from 0.7260.726 at S1 to 0.3940.394)—the query representations deviate from the training distribution, and the neighbors retrieved from the ImageNet-1K reference set lose semantic consistency, causing a notable drop in prediction performance while still maintaining a meaningful positive correlation.
    • Furthermore, since the present work focuses primarily on establishing this scientific finding, broader performance evaluations and distribution-shift testing will be an important direction for our future work.
  • For Question 1

    • We plan to extend this prediction framework to other modalities as well as cross-modal scenarios in our future work. We believe that the predictability of representational disagreement is a shared phenomenon across models of all modalities. However, representation spaces vary significantly across different modalities, and the disagreement metric used in this paper has mainly been validated in the visual modality. Therefore, for different modalities, we should adopt more tailored metrics to quantify modality-specific representational disagreement. At the same time, because this paper focuses on validating our scientific finding, we kept the predictor architecture as simple as possible; thus, when transferring to other modalities or cross-modal settings, we will need to further design more suitable predictor architectures.
Reviewer_7k6zOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies whether discrepancies in learned representations by different foundation models can be predicted from the representation generated by a single model. The proposed method applies KNN separately to representations generated by individual models and measures sample-level agreement through the overlap of nearest-neighbour sets. The overlap scores are then used to train a supervised model as the supervision signal. The main contribution is an empirical framework for estimating a cross-model statistic from one presentation. However, its usefulness should be further validated.

Critical strengths
  • The substantial experimental work demonstrates that representations generated by one model or encoder contain information about disagreement among other models. The leave-anchor-out results further support this finding by ablating the target model's input. Together with results across multiple models and configurations, this contributes to the understanding of relationships between representations generated by different models.
  • Table 3 shows that a fixed image predictor can be transferred to different external datasets without adaptation, achieving very promising results compared with exact FAISS search. These results show that an inexpensive approach to estimating the cross-model agreement when running multiple models and neighbourhood searches online is impractical.
Critical weaknesses
  • The technical contribution of the proposed work is a bit weak, even though studying the agreement/disagreement among presentations generated by different models is interesting. The similarity measure, the way of finding the top-k neighbours, and the scores proposed in Eq. (1) all exist. Particularly, Eq. (1), which computes the cross-model agreement, has direct precedents in Huh et al. and Park et al. The reliability interpretation also needs care. Equation (1) depends on the selected encoders, reference population, and kk. It measures their neighbourhood agreement; it is not automatically an error probability or task-independent uncertainty measure. Agreement can coexist with shared errors, while disagreement can reflect useful complementary representations.
  • The empirical work seems quite solid and comprehensive. However, my concerns are
    • The main experiments emphasise Pearson correlation, which does not establish accurate numerical prediction or successful identification of highly disagreeing inputs. Table 7 shows an association with downstream Brier loss, but lacks comparisons against confidence, entropy, or representation-density baselines. it does not establish whether the proposed score improves reliability decisions beyond inexpensive alternatives.
  • Figure 6 does not compare the learned predictor with directly measured reduced-reference agreement against the same full-reference target. Without that control, the experiment cannot isolate how much benefit comes from learning rather than the smaller reference set already preserving the target’s ordering.
Questions for authors
  • The supervision is based on hard kk- nearest-neighbour identity overlap, which ignores neighbour distances and may be sensitive to reference sampling and near-ties. How stable are these targets across reference subsets and neighbourhood sizes? A predictor could learn systematic artifacts in this score while achieving high held-out correlation.
  • Does predicted disagreement improve error detection or selective classification compared with classifier confidence, entropy, margin, and anchor-space density? Does it provide complementary information when combined with these signals?
8 comments

We thank the reviewer for the valuable feedback and would like to clarify our core contribution. Understanding learned representations is fundamental to understanding AI. Our work approaches this goal from a new perspective: moving beyond measuring cross-model representational differences to investigating their predictability. By revealing learnable and generalizable patterns in representational disagreement, we provide new empirical insights into the underlying structure of AI representations.

For Critical Weakness 1:

  • Regarding novelty, our contribution lies in moving beyond measuring cross-model representational disagreement to revealing its predictability. Since different models operate in heterogeneous representation spaces, we adopt Neighborhood Agreement as an alignment-free measure of disagreement. Crucially, the ability to measure disagreement does not imply that it is predictable. Our experiments demonstrate that such disagreement exhibits learnable patterns, providing new empirical insights into the underlying structure of AI representations.
  • Regarding practical significance, representational disagreement is not necessarily indicative of model errors; rather, it captures differences in how models represent the same information. Understanding and predicting these differences can offer insights into representation reliability, model complementarity, and diversity, including the identification of complementary representations suggested by the reviewer. These applications highlight the broader value of our finding beyond any specific downstream task.
  • More broadly, empirical discoveries can make significant scientific contributions by revealing overlooked phenomena and inspiring new research directions. Previous ICLR papers, such as "Large Language Models Are Not Robust Multiple Choice Selectors" (ICLR 2024) and "A Decade's Battle on Dataset Bias: Are We There Yet?" (ICLR 2025), illustrate this value. We believe our finding similarly offers a new perspective on understanding AI representations and can inspire further research in this direction.

For Critical Weakness 2 and Question 2:

  • We notice that these two points focus on the same issues, so we address them together here.
    • Regarding evaluation metrics, we primarily report Pearson correlation to assess whether the predictor captures the variation in representational disagreement. Following the reviewer's suggestion, we have additionally reported MAE and the 90th percentile of absolute prediction errors to provide a more comprehensive evaluation.
    • Regarding baselines, our goal is to investigate the predictability of cross-model representational disagreement, independent of downstream tasks. The suggested classifier-based methods instead target task-specific prediction risk, making direct comparisons less meaningful. Table 7 illustrates the potential applications of our discovery, rather than claiming superiority in downstream risk prediction.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.