PaperlayerView exact revision on arXiv

Source-linked AI summary

On the Value of Cross-Modal Misalignment in Multimodal Representation Learning

Yichao Cai, Yuhang Liu, Erdun Gao, Tianjiao Jiang, Zhen Zhang, Anton van den Hengel, Javen Qinfeng Shi

arXiv:2504.10143v7cs.LGcs.CV

TL;DR

Real image-text pairs can be semantically misaligned, complicating the assumption that paired modalities express the same information. The paper models omissions and alterations in text and analyzes what MMCL learns under those biases. It finds that MMCL retains unbiased shared semantics and excludes misaligned variables, with implications for broad pretraining and invariant representations.

  • Problem

    MMCL commonly assumes paired image-text examples convey the same semantic information, although real-world pairs can be incomplete or misleading.

    Source: §1 · pp. 1–2

  • Method

    The paper uses a latent variable model with selection bias that omits semantics and perturbation bias that alters them, then analyzes identifiability under stated assumptions.

    Source: §3.2 · p. 5, §4 · p. 5

  • Results

    Under the stated assumptions, MMCL identifies unbiased shared semantic variables and excludes misaligned variables, regardless of the latent semantic graph structure.

    Source: §1 · p. 2, §4.1 · p. 6

  • Takeaways & Limitations

    For broad pretraining, omitted or perturbed semantics remain lost regardless of dataset size; for invariant learning, misalignment may help by filtering shift-vulnerable semantics.

    Source: §4.2 · p. 6

    The theoretical analysis assumes continuous latent variables with strictly positive densities and a random perturbation process.

    Source: §3.2 · p. 5

Abstract

from arXiv · show

Multimodal representation learning, exemplified by multimodal contrastive learning (MMCL) using image-text pairs, aims to learn powerful representations by aligning cues across modalities. This approach relies on the core assumption that the exemplar image-text pairs constitute two representations of an identical concept. However, recent research has revealed that real-world datasets often exhibit cross-modal misalignment. There are two distinct viewpoints on how to address this issue: one suggests mitigating the misalignment, and the other leveraging it. We seek here to reconcile these seemingly opposing perspectives, and to provide a practical guide for practitioners. Using latent variable models we thus formalize cross-modal misalignment by introducing two specific mechanisms: Selection bias, where some semantic variables are absent in the text, and perturbation bias, where semantic variables are altered -- both leading to misalignment in data pairs. Our theoretical analysis demonstrates that, under mild assumptions, the representations learned by MMCL capture exactly the information related to the subset of the semantic variables invariant to selection and perturbation biases. This provides a unified perspective for understanding misalignment. Based on this, we further offer actionable insights into how misalignment should inform the design of real-world ML systems. We validate our theoretical findings via extensive empirical studies on both synthetic data and real image-text datasets, shedding light on the nuanced impact of cross-modal misalignment on multimodal representation learning.

1 Introduction

Image-text contrastive learning assumes paired examples express the same semantics, but real-world pairs can be incomplete or misleading. The paper formalizes two kinds of misalignment and shows how they shape what MMCL represents.

1 Introduction

  • Over 50% of purportedly aligned clip-caption pairs in one large-scale video-text dataset were found to be misaligned [49].Source: §1 · p. 1
  • Misalignment is viewed both as noisy supervision to mitigate [40] [64] [72] [30] [7] and as a source of representation benefits [50].Random text augmentation introducing style-related misalignment has been associated with more robust representations across several learning settings.Source: §1 · p. 2
  • The proposed latent variable model represents misalignment through selection bias, which omits semantic variables from text, and perturbation bias, which alters them.The model allows arbitrary causal structure among semantic variables and separate image and text generative processes.Source: §1 · p. 2
  • Under mild assumptions, MMCL identifies semantic variables unaffected by both biases and excludes the misaligned variables, regardless of latent causal structure.The authors frame this as a unified account: misalignment can impair tasks needing full semantic information, yet may help when robustness to distribution shifts is desired.Source: §1 · p. 2
  • The paper extends its analysis to common-representation and invariant-representation tasks and validates its findings on synthetic and real image-text data, including a pretrained CLIP case study.Source: §1 · p. 2

2 Preliminaries: Multimodal Contrastive Learning

MMCL learns joint representations by bringing paired examples from different modalities together and separating unpaired examples. The paper uses an asymptotic form of its contrastive objective for theoretical analysis.

2 Preliminaries: Multimodal Contrastive Learning

  • MMCL uses modality-specific encoders to map paired samples into a shared space, aligning pairs while pushing apart unpaired negatives.Source: §2 · p. 3
  • As the number of training pairs approaches infinity, with τ = 1 and negative squared Euclidean distance, the contrastive objective reduces to the asymptotic objective [13].The objective combines paired-sample distance minimization with maximization of learned-representation entropy.Source: §2 · p. 3
  • The paper focuses on how cross-modal misalignment changes learned representations, unlike prior studies that assume perfectly aligned pairs.Source: §2 · p. 3

3 Problem Formulation via a Generative Perspective

The paper formulates image-text data generation with a latent variable model that separates shared semantics from modality-specific factors. Selection and perturbation biases describe how text can omit or alter semantic content.

3.1 A Latent Variable Model Characterizing Cross-Modal Misalignment

  • The model partitions latent variables into shared semantic variables, image-specific factors, and text-specific factors; the latter two are independent of semantics and of each other.Semantic variables represent human-interpretable content, while image-specific and text-specific variables include camera noise and writing style, respectively.Source: §3.1 · p. 3
  • Unlike restrictive fixed-graph or independence assumptions, the model allows arbitrary dependencies among semantic variables because their true graph is often unknown.Source: §3.1 · p. 4
  • Images are generated from semantic and image-specific variables through a bijective mapping, and the model treats images as fully encapsulating semantics.Source: §3.1 · p. 4
  • Selection bias determines which non-empty subset of semantic variables is expressed in text, leaving the complementary semantics omitted.The selected subset determines the semantic input to the text-generation mapping.Source: §3.1 · p. 4
  • Perturbation bias marks selected semantic components that may be randomly replaced; the remaining selected components stay unperturbed across modalities.Source: §3.1 · p. 4
  • With shape, size, and color semantics, the example selects shape and color, perturbs color, and leaves shape unbiasedly preserved in the text.The example’s resulting caption can describe a red cat even when the image shows a black cat.Source: §3.1 · p. 5
  • The text-generation process acts on selected semantics, including their perturbed values, together with text-specific latent variables.The text-specific variables have a prior distribution, while the conditional distribution over selected semantics models perturbations.Source: §3.1 · p. 5

3.2 Model Assumptions for Theoretical Analysis

  • The theoretical analysis assumes continuous latent variables with strictly positive densities and random perturbations that can affect each perturbable semantic index.Perturbations modify a random subset; the stated conditional density is smooth and strictly positive, with support containing a non-empty open subset.Source: §3.2 · p. 5

4 Identifiability Results for MMCL under Misalignment

The analysis identifies which semantic variables MMCL can recover under the proposed model and assumptions. It then relates misalignment to two practical settings: broad pretraining and invariant representation learning.

4 Identifiability Results for MMCL under Misalignment

  • Block-identifiability means the learned representations retain all and only the information in the identified semantic subset, up to invertible mappings.Source: §4 · p. 6

4.1 Identifiability of Latent Semantic Variables under Cross-Modal Misalignment

  • Under Assumptions 3.1 and 3.2, minimizing the stated MMCL objective block-identifies the semantic variables annotated without bias in text.The result is for sufficiently flexible, smooth encoders and the image-text generation process specified in the model.Source: §4.1 · p. 6
  • Misaligned semantic components are excluded, while modality-specific non-semantic factors are discarded; the result holds regardless of the semantic variables’ latent graph structure.The unbiased shared variables are recovered up to block-wise indeterminacy.Source: §4.1 · p. 6

4.2 Insights into Cross-Modal Misalignment for Practice

  • For large-scale pretraining, omitted or perturbed semantics remain lost regardless of dataset size, although scale may mitigate sporadic misalignment by averaging its effects.The authors emphasize detailed, consistent semantic annotation for foundation models intended to support diverse downstream tasks.Source: §4.2 · p. 6
  • When the omitted and perturbable semantics coincide with environment-sensitive variables, MMCL identifies the invariant semantics under the stated assumptions.Source: §4.2 · pp. 6–7
  • For robust out-of-distribution performance, misalignment may help by selectively omitting or perturbing semantics vulnerable to distribution shifts.The paper notes that auditing and curating text is more precise and interpretable because language is distilled from human knowledge.Source: §4.2 · p. 7

5 Experiments

Experiments across simulations, factorized and causally structured datasets, and an OpenCLIP case study examine how selection and perturbation biases affect learned representations. They generally find that MMCL retains unbiased shared semantics while excluding misaligned or modality-specific factors, with partial predictability possible when semantics are dependent.

5.1 Numerical Simulation

  • Under independent latents, unbiased semantics reach R2 ≈1 and selection-biased semantics R2 ≈0; dependent latents make some omitted factors partially predictable.Modality-specific variables are consistently omitted across settings.Source: Figure 3 · p. 7, §5.1 · p. 7
  • Appropriate selection or perturbation bias removes shift-sensitive variables, supporting OOD generalization, while retaining more semantics improves ID regression.These contrasting patterns are reported for downstream tasks in the simulation.Source: Figure 4 · p. 8, §5.1 · p. 8

5.2 MPI3D-Complex: Real-World Dataset with Factorized Latent Variables

  • MCC is 0 for misaligned semantics, while unbiased factors score ≥0.8 in MPI3D-Complex experiments.The results report nonlinear MCC across selection and perturbation settings.Source: Figure 6 · p. 8, §5.2 · p. 8

5.3 Causal3DIdent: Semi-Synthetic Dataset with Structured Causal Latent Variables

  • In Causal3DIdent, unbiased shared factors are reliably recovered, but image-specific and generally misaligned variables are excluded; dependencies can make additional semantics partially predictable.Performance approaches R2 ≈1 for recovered factors; when factors such as x_pos or y_pos are included, dependent variables such as color can become partially predictable.Source: Figure 8 · p. 9, §5.3 · p. 9

5.4 Case Study: Zero-Shot Evaluation of OpenCLIP Model Representations

  • The OpenCLIP case study evaluates zero-shot F1 across concept groups and reports group-wise caption coverage for a taxonomy of 146 concepts in 15 groups.The study uses up to 200 Flickr images per concept.Source: §5.4 · p. 9, Figure 10 · p. 10, Figure 9 · p. 10

6 Conclusion

The conclusion frames cross-modal misalignment as both a source of lost semantic detail and a possible route to domain invariance. The paper’s analysis formalizes which semantics MMCL identifies and connects that result to prior work on identifiability and invariant representations.

6 Conclusion

  • MMCL retains unbiased shared semantics and discards misaligned variables, trading rich semantic detail for the possibility of filtering distribution-sensitive factors and enhancing domain invariance.The authors report that simulations and image-text experiments validate the theoretical findings.Source: §6 · p. 10

B Related Work

  • Unlike prior MMCL analyses that often neglect misalignment, this work models it in a latent-variable framework and uses block-identifiability without requiring fixed content-style decompositions or perfectly aligned pairs.The related-work discussion also distinguishes its analysis from approaches assuming fixed causal directions.Source: B Related Work · p. 19
  • Textual omission or semantic perturbation is presented as a controllable proxy for environmental variation in invariant representation learning.The paper contrasts this approach with methods requiring interventions or many diverse environments.Source: B Related Work · p. 19

C.1 Lemmas

  • Under the proposed LVM and stated assumptions, the alignment-and-entropy loss has a global minimum of zero.The proof constructs function pairs that attain this minimum.Source: §C.1 · pp. 20–21
  • A smooth map with the lemma’s stated domain, density, and uniform-output conditions is a global diffeomorphism and preserves all information in its input.The lemma states that every input component is reflected in the output and can be uniquely recovered.Source: §C.1 · p. 21

C.2 Proof of Theorem 4.1

The proof shows that MMCL representations block-identify exactly the semantic variables unaffected by selection and perturbation biases. It establishes this by showing that alignment excludes modality-specific and misaligned variables, while preserving the unbiased subset.

C.2 Proof of Theorem 4.1

  • Under the stated assumptions, minimizing LSymAlignMaxEnt block-identifies the semantic variables annotated without bias in text.The theorem requires sufficiently flexible, smooth encoders and identifies the unaffected subset through an invertible mapping to the learned representations.Source: §C.2 · p. 22
  • At the global minimum, the loss enforces alignment across paired image-text samples and uniform learned representations.The proof uses the zero global minimum to show each nonnegative loss term vanishes almost surely.Source: §C.2 · p. 22
  • The representations exclude modality-specific variables and semantic variables omitted from the corresponding text, even with unknown causal interactions among semantics.The proof establishes independence from image- and text-specific factors, then excludes omitted semantics regardless of their latent causal structure.Source: §C.2 · pp. 22–24
  • Perturbed semantic variables are excluded from both modalities because their random alterations conflict with the alignment condition.The proof first excludes perturbed text semantics and then uses the alignment condition to exclude them from image representations as well.Source: §C.2 · pp. 24–25
  • The remaining representations contain all and only the unbiased semantic components, which the learned mappings block-identify.The proof establishes invertibility with respect to the learned representations and the true unbiased semantic variables.Source: §C.2 · pp. 22, 25

C.3 Proof of Corollary 4.1

When all semantic variables are selected and none are perturbed, the proof applies Theorem 4.1 to show that the full semantic vector is block-identified. The parameter choices encode this no-bias setting through a fixed indexing order.

C.3 Proof of Corollary 4.1

  • Under Asms. 3.1 and 3.2, with selection bias θ = 2ns − 1 selecting all semantic variables and perturbation bias ρ = 1 (Iρ = ∅), the full semantic vector s is block-identified by both fx and ft.Source: §C.3 · p. 26
  • Setting θ = 2^ns − 1 selects the full semantic set, while ρ = 1 preserves every selected variable without perturbation.Under these choices, the perturbing subset is trivial and the text semantics equal the image semantics almost surely.Source: §C.3 · p. 26
  • The fixed graded lexicographic order is a reproducible indexing convention, not a constraint on the latent structure.Because the true permutation of latent variables is unidentifiable, any consistent order can index the selection and perturbation parameters.Source: §C.3 · p. 26

C.4 Proof of Corollary 4.2

The corollary shows that, in an OOD setting, representations can block-identify invariant semantics when the omitted and perturbed variables coincide with the environment-sensitive subset. The appendix also describes simulation settings used to examine identification and downstream performance.

C.4 Proof of Corollary 4.2

  • When omitted and perturbed variables together match the environment-sensitive subset, minimizing the alignment objective block-identifies the invariant semantic variables.The result assumes Asms. 3.1 and 3.2 and an OOD setting with shifts in the remaining semantic variables.Source: §C.4 · p. 27

D Numerical Simulation Details

  • Appendix D covers simulation hyperparameters and task construction, additional identification experiments, and downstream performance under perturbation bias.The reported results average R2 scores across three random seeds.Source: D Numerical Simulation Details · p. 28, §D.1 · p. 29

D.1 Detailed Experimental Setup

  • The numerical simulations use a 10-dimensional semantic space and separate 5-dimensional image- and text-specific subspaces, with independent or dependent semantic variables.Independent semantics use a factorized standard Gaussian; dependent semantics use a covariance sampled from a Wishart distribution.Source: §D.1 · p. 28
  • The simulation tests representative selection and perturbation configurations, including a joint-bias case, to validate the theory across misalignment settings.Perturbations alter selected semantics in text with additive Gaussian noise; the joint case omits the last two indices and perturbs the first two.Source: §D.1 · p. 28
  • Downstream tasks use nonlinear labels from semantic variables, with fixed pretrained encoders probed by a two-layer MLP on in-distribution and shifted test samples.The setup includes four regression tasks and a classification task, with no adaptation or fine-tuning for OOD evaluation.Source: §D.1 · p. 29

D.2 Additional Identification Results

Additional experiments support the identification analysis: under perturbation, unbiased semantics are preserved while misaligned ones are excluded or only partially identifiable with dependent latents. The appendix also examines linearity, dimensionality, joint bias, downstream tasks, and MPI3D-Complex experiments.

D.2 Additional Identification Results

  • Under perturbation bias, modality-specific variables are discarded, unbiased semantics are block-identified, and misaligned semantics can be partially identifiable with dependent latents.The reported pattern holds across image and text representations.Source: §D.2 · p. 29, Figure 11 · p. 30
  • Linear-regression R2 scores closely match nonlinear-regression results under both selection and perturbation bias, suggesting nearly linear representations.Figures 12 and 13 report the linearity analyses for selection and perturbation bias, respectively.Source: §D.2 · p. 29, Figure 12 · p. 30, Figure 13 · p. 31
  • With representation size 5 instead of the true dimension 3, independent-latent results preserve selected semantics and discard omitted ones, while dependent-latent R2 patterns differ from correct sizing.Redundant dimensions tend to encode exogenous noise, potentially adding complexity.Source: §D.2 · p. 30, Table 4 · p. 30
  • With selection bias Iθ = [8] and perturbation bias Iρ = [2], omitted or perturbed semantics are discarded while selected, unperturbed semantics are preserved.The joint-bias findings are reported in Table 5.Source: §D.2 · p. 30, Table 5 · p. 31

D.3 Additional Downstream Results

  • Downstream performance is generally high when the learned representations preserve semantic variables critical to the task, across independent and dependent latent settings.The perturbation experiment sequentially reverses which semantic variables are preserved; Figure 14 reports regression, classification, and OOD results.Source: §D.3 · p. 31, Figure 14 · p. 31

E Experiment Details on MPI3D-Complex Dataset

  • MPI3D-Complex contains 460,800 images spanning seven independent discrete factors; five are treated as semantic and two as image-specific.The appendix's further experiments include linear-factor classification and an encoder-dimensionality ablation.Source: E Experiment Details on MPI3D-Complex… · p. 31, §E.1 · p. 32

E.1 Detailed Experimental Setup

  • MPI3D-Complex text is generated from semantic attributes with incremental selection settings and random substitutions for perturbation settings.Each perturbed text value is replaced with a random alternative with probability 0.9.Source: §E.1 · p. 32, Table 7 · p. 32
  • The image and text encoders are trained for 200,000 steps with MMCL, and frozen representations are evaluated using linear and nonlinear classifiers with MCC.The encoder output dimension is generally set to the number of unbiased semantic dimensions, with a minimum of 3.Source: §E.1 · p. 33

E.2 Additional Results

The additional results show factor-dependent linearity and encoding-size sensitivity, while the Causal3DIdent setup tests MMCL under systematically varied selection and perturbation biases.

E.2 Additional Results

  • Color, shape, cam., and back. are strongly linearly separable, while size is not linearly embedded in the learned representations.The size finding contrasts linear evaluation with nonlinear prediction results.Source: §E.2 · p. 33
  • Increasing encoding size stabilizes image-text performance; redundant dimensions do not appear to capture misaligned semantics when latent factors are independent.With only color included in text, encoding size 1 tends to under-optimize the image encoder and produces high variance across runs.Source: §E.2 · p. 34, Figure 16 · p. 34

F.1 Detailed Experimental Setup

  • Causal3DIdent images use 10 latent factors, treating three rotation angles as image-specific and the remaining factors as shared semantic variables.The object’s z-position is fixed, leaving 10 variables governing image generation.Source: §F.1 · pp. 34–35, Figure 17 · p. 35
  • The experiments synthesize separate training, evaluation-training, and test sets, then assess latent-factor recovery with linear and nonlinear predictors.The data comprise 80,000 training samples and 10,000 samples each for predictor training and test evaluation; scores use MCC for discrete factors and R2 for continuous factors.Source: §F.1 · pp. 35–37
  • Text generation discretizes color variables and applies partial information loss to spotlight position, making its descriptions non-invertible.Five manually designed templates generate text under incrementally varied selection and perturbation settings.Source: §F.1 · p. 35

F.2 Additional Results

Text representations recover unbiased aligned factors, but lossy mappings and ambiguous color words can impair prediction; the section also reports factor-dependent linearity and uneven caption coverage across concepts.

F.2 Additional Results

  • Unbiased aligned discrete factors are reliably identified in text representations, and s_pos retains relatively high R2 despite mapping to only five discrete tokens.The text mapping for s_pos is non-invertible, yet the result suggests learned text representations remain influenced by the alignment objective.Source: §F.2 · p. 37, Figure 19 · p. 39
  • Color, s_color, and b_color show varying performance degradation when continuous image values are discretized in text, with overlapping color vocabulary adding ambiguity.Object, spotlight, and background color tokens can all contain “cyan,” creating ambiguity across distinct latent variables.Source: §F.2 · pp. 37–38
  • Recovering x_pos and y_pos in text representations does not improve color predictability, consistent with perturbation bias disrupting image-latent causal structure..Source: §F.2 · p. 38
  • The text-specific factor phrase is partially encoded, unlike image-specific continuous factors, and this effect is stronger under selection bias when encoding exceeds the unbiased semantic dimensionality.The authors relate the partial identifiability of phrase to its discrete nature and cite [13].Source: §F.2 · p. 39
  • Linearity depends on the factor: color is poorly predictable linearly in both modalities, whereas x_pos and y_pos are strongly linearly separable in some settings.The authors describe linearity as shaped by modality-specific encoding strategies and input structure.Source: §F.2 · p. 39, Figure 20 · p. 39, Figure 21 · p. 40

G.1 Concepts Taxonomy and Data Collection

  • The case study curates 146 concepts in 15 mutually exclusive groups and collects up to 200 manually reviewed Flickr images per concept.Images are filtered for Creative Commons licensing and relevance to the queried concept.Source: §G.1 · pp. 39–40, Table 12 · p. 40, Table 13 · p. 41

G.2 Details on Caption Frequency Statistics

  • Caption coverage varies within concept groups: concrete objects, colors, and foods appear more often than abstract or subjective concepts.The authors suggest ethical avoidance, omission of nuance, and difficulty describing abstract visual content as possible reasons for underrepresentation.Source: §G.2 · pp. 40, 42

G.3 Zero-Shot Evaluation Details

Zero-shot performance tracks caption coverage, with different implications for valuable underrepresented concepts and sensitive concepts; the paper also reports its computational resources and frames discussion topics.

G.3 Zero-Shot Evaluation Details

  • Concept groups with caption coverage above 0.15% generally score above 65% F1, while severely under-captioned groups fall below 40% F1 across model scales.The authors report that under-captioning can reflect selection bias or ethical omission.Source: §G.3 · p. 42
  • The ViT-B-32 confusion matrix shows strong diagonal patterns for well-captioned groups and substantial misclassification for groups affected by selection bias.For under-captioned Texture and Emot. concepts, the authors suggest richer captions; underrepresentation may be desirable for sensitive Trait and Stere. concepts.Source: §G.3 · p. 42, Figure 23 · p. 44

H Computation Resources

  • The experiments used four NVIDIA A100 GPUs and required about 70 GPU-hours for more than 120 numerical-simulation models.Causal3DIdent rendering additionally used a single NVIDIA RTX 4090 GPU over four days.Source: H Computation Resources · p. 42

I Discussions

  • The discussion addresses study limitations, future research directions, and broader implications.Source: I Discussions · p. 45

I.1 Limitations and Future Directions

The discussion bounds the framework’s scope and identifies open modeling and estimation challenges. It also considers how annotation choices and errors relate to representational scope and human values.

I.1 Limitations and Future Directions

  • The study formalizes selection and perturbation bias but does not empirically estimate dataset-specific biases, which are latent and difficult to observe.Proposed diagnostics include concept-coverage statistics for selection bias and caption comparisons for perturbation bias; defining taxonomies, annotation granularity, and metrics remains challenging.Source: §I.1 · p. 45
  • The framework focuses on semantic misalignment in image-text pairs; temporal lag, modality dropout, and multi-entity ambiguity may require substantially different modeling assumptions.The authors present their abstraction as a foundation for future work on these other forms of misalignment.Source: §I.1 · p. 45
  • The theory does not model emergent higher-order semantics such as sarcasm, irony, or humor, which may require joint-modality latent variables.It can accommodate semantics deterministically inferable across modalities, such as emotional tone from facial expression.Source: §I.1 · p. 45
  • The framework addresses fixed selection and perturbation biases, while random, unstructured omissions in user-generated corpora remain an open identifiability problem.The authors identify missing-completely-at-random and missing-not-at-random settings as future work.Source: §I.1 · p. 45
  • Identifiability is guaranteed up to general invertible transformations, while conditions for a linear transformation guarantee remain an open question.The authors note empirical observations of approximately linear representations and identify linearity as useful for interpretability and downstream applications.Source: §I.1 · pp. 45–46

I.2 Broader Implications

  • Textual selection and perturbation biases act as epistemic filters that constrain what models represent; annotation scope also bears on ethical alignment.The authors argue that faithfully capturing intended semantic scope supports generalizable representations, while deliberate curation can align representations with human values.Source: §I.2 · p. 46
  • Consistent annotation omissions and mislabels may reveal a human value hierarchy, with neglected factors reflecting lower perceived social or cognitive importance.The authors propose treating such regularities as a behavioral prior for modeling human preferences and priorities.Source: §I.2 · p. 46
Loading 2504.10143v7…