acceptodds
Withdrawn from ICLR 2027

PAIRCAUSE: Exact-Marginal Diagnostics of Multimodal Readout and Acceptance

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

When multimodal systems accept more errors under a common perturbation, is information lost, read out incorrectly, or simply admitted by the acceptance rule? PAIRCAUSE diagnoses these alternatives by changing the pairing of complete modality trajectories while exactly preserving their label-conditional empirical multisets. On 80 synthetic test blocks, shared pairing raises ERM accepted error from 23.32% to 46.63%, predominantly through joint-state composition; simple linear and strong-cue controls reproduce the direction. Distribution-aligned probes reveal a different phenomenon: on the same shared-mixture assessment blocks, refitting a linear head on frozen ERM features raises accuracy from 50.96% to 87.53%, but raw-input and random-feature controls perform similarly. Conversely, a clean-fitted linear head reduces Coupled's shared accuracy from 88.04% to 52.81%. Separating classification, fixed-mask risk, and recomputed acceptance shows that cue restoration combines genuine correction with rejection. Twenty retained fits and 36 new same-backend fits support these distinctions. The contribution is an auditable feature-space diagnostic of recoverability, readout supervision, and acceptance, not evidence of a uniquely learned causal representation or a demonstrated physical sensor attack.

Primary area:
applications to robotics, autonomy, planning
Keywords:
multi-modal fusion

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_ppDQOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates why multimodal systems accept more erroneous predictions under common perturbations, and attempts to distinguish three mechanisms: whether the representations still contain recoverable label information, whether the output head correctly utilizes this information, and whether the acceptance rule admits more erroneous predictions.

The authors propose PAIRCAUSE, which precisely preserves the unimodal empirical distributions while changing the joint state distribution of the two modalities by re-pairing complete camera and radar trajectories within the same data block and label. The paper further combines joint state decomposition, linear probing of frozen representations, source-matched feature recovery, and independent control experiments on classification and acceptance gating.

The main results include: the accepted error rate of ERM rises from 23.32% to 46.63%, but the conditional risk rises only from 48.39% to 49.33%; on the shared perturbation distribution, retraining a linear output head on the frozen ERM hidden layers improves balanced accuracy from 50.96% to 87.53%; raw inputs and random features also achieve comparable results. The paper further demonstrates that the low accepted error rate after feature recovery is jointly affected by classification correction and the rejection mechanism.

Critical strengths

None

Critical weaknesses
  1. The writing of the paper is poor. The wording and sentence construction basically have an AI-generated flavor, with a large number of defensive descriptions and typical AI sentence patterns such as "not..., but...". In addition, the figures and tables in the paper are very rough, and there are even problems such as overlapping text and lines.
  2. The paper fails to provide some useful insights; it is mostly a matter of combining existing components and methods.
Questions for authors
  1. The writing of the paper is poor. The wording and sentence construction basically have an AI-generated flavor, with a large number of defensive descriptions and typical AI sentence patterns such as "not..., but...". In addition, the figures and tables in the paper are very rough, and there are even problems such as overlapping text and lines.
  2. The paper fails to provide some useful insights; it is mostly a matter of combining existing components and methods.
Reviewer_kZ8bOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes PAIRCAUSE, a diagnostic framework for examining multimodal robustness by re-pairing modality trajectories while exactly preserving their label-conditional marginal distributions. Through controlled synthetic experiments, the authors distinguish three potential sources of multimodal failure. The results suggest that the increase in accepted errors is largely driven by changes in composition and coverage, while substantial label information remains recoverable through distribution-matched readout adaptation.

Critical strengths

The paper presents a diagnostic framework for multimodal robustness that distinguishes different failure mechanisms. The re-pairing construction is also a novel and interesting idea.

Critical weaknesses
  1. Why would camera and radar modalities exhibit such nearly identical feature shifts in real-world settings?

  2. Both the theoretical and empirical analyses rely on synthetic data with specifically chosen parameter values, which limits its applicability. Moreover, it remains unclear whether the identified problem has already been addressed in prior work.

  3. The experimental settings and the specific claims that each experiment aims to support could be explained more clearly.

  4. Some mathematical notation is introduced without sufficient explanation, such as ScS_c and RcR_c in Equation 2. The figures could also be improved for clarity, particularly Figures 2 and 3, which are somewhat difficult to follow.

Questions for authors

I believe the manuscript would benefit from further refinement, particularly in its problem formulation, method presentation, and experimental validation.

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

This work deals with robust multimodal sensor fusion. Specifically, it addresses the problem of diagnosing why wrong predictions are accepted when sensors are corrupted together, distinguishing between non-informative inputs, a poorly fitted classification head, and acceptance filters that fail to reject wrong predictions. It addresses this problem by introducing PairCause, a multimodal fusion evaluation method. PairCause evaluates fusion methods by controlling how corruption co-occurs across modalities by swapping one modality’s features with those from a datapoint from the same class. Experiments use PairCause to evaluate a simple early-fusion MLP on a synthetic dataset with two features per modality. Rest of the details are unclear

Critical strengths

None

Critical weaknesses

The manuscript is very difficult to read.

  • Key terms are used without definition e.g., "ERM", “fixed-mask risk”, “recomputed acceptance", "distribution-aligned probes"
  • ERM appears to be a multimodal fusion method, but it is not explained
  • Sentences are dense and their meaning is not clear, e.g., "This dose difference prevents attributing their comparison solely to learned coupling", "Unconditional h2 label decoding is 87.28–88.35% on the shared mixture, not nearly 100%. Near-perfect oracle scores fit separate classifiers within known corruption states; their conditioning supplies information unavailable to the deployed head"
Questions for authors

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

Related papers

Discussion (0)

Sign in to comment.