This assessment is based on version 1 of this paper. Version 2 is now available on arXiv — the authors may have revised their methods, results, or conclusions.
Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability
Joël Roman Ky, Salah Ghamizi, Maxime Cordy
Abstract
Vision-Language Models (VLMs) map complex visual inputs to semantic spaces, but interpreting the cross-modal reasoning of VLMs currently relies on post-hoc explainers evaluated via unimodal perturbation metrics. We expose a limitation in this paradigm: because multimodal datasets contain language priors and modality biases, VLMs frequently exhibit cross-modal redundancy, allowing them to answer visual queries using text alone. Consequently, unimodal metrics penalize faithful explainers, triggering an evaluation collapse where visual and textual rankings fundamentally contradict each other. %(Kendall's ). To resolve this, we introduce Synergistic Faithfulness (), a scalable metric rooted in the Shapley Interaction Index that strictly isolates the joint Harsanyi dividend between modalities, serving as a highly accurate surrogate () while achieving a computational speedup. Evaluating 8 distinct XAI methods across 3 VLM architectures and 3 benchmark datasets, reveals that explainers proposed for VLMs heavily over-index on visual salience and significantly underperform adapted attention-based methods in capturing true cross-modal synergy. By decoupling visual plausibility from cross-modal faithfulness, this work provides a rigorous evaluation framework required to safely audit VLM reasoning in high-stakes deployments.
AI Impact Assessments
(1 model)Scientific Impact Assessment: "Measuring Cross-Modal Synergy: A Benchmark for VLM Explainability"
1. Core Contribution
This paper identifies a fundamental flaw in how explainability methods for Vision-Language Models (VLMs) are evaluated and proposes a principled solution. The core argument is that standard unimodal perturbation metrics (Insertion/Deletion AUC) fail in multimodal settings because of cross-modal redundancy: when VLMs can answer questions using one modality alone (due to language priors and dataset biases), perturbing a single modality yields uninformative faithfulness scores. The authors demonstrate this causes an "evaluation collapse" where visual and textual rankings of the same explainers fundamentally contradict each other (Kendall's τ = −0.06).
The proposed solution, Synergistic Faithfulness (F_syn), is grounded in cooperative game theory. It approximates the Shapley Interaction Index by computing the 2-way Harsanyi dividend along continuous perturbation trajectories, isolating the pure joint contribution of visual and textual features. This transforms the intractable O(2^{m+n}) exact computation into a manageable O(K) forward passes (6K+2 per sample), achieving a 24× speedup over macro-coalitional exact SII while maintaining ρ = 0.92 Spearman correlation.
2. Methodological Rigor
The paper demonstrates strong methodological discipline across several dimensions:
Theoretical grounding: The metric is formally derived from the Harsanyi dividend structure of cooperative game theory. The authors clearly define the mathematical formulation distinguishing unimodal metrics µ_I, µ_T from the multimodal metric µ_{I×T}, and prove the boundary-condition failure of unimodal metrics under perfect cross-modal redundancy (Section 3.2).
Validation against ground truth: The authors construct a clever macro-coalitional game (8 players, 2^8 = 256 states) that permits exact, zero-variance SII computation for N=200 instances. The ρ = 0.92 correlation with this ground truth is convincing, and critically, the correlation is stable across different explainer types (Random, Input×Grad, Rollout, TAM), demonstrating explainer-agnostic validity.
Statistical analysis: The use of Linear Mixed-Effects Models (LMM) to disentangle explainer performance from dataset difficulty and model architecture is sophisticated and appropriate. Treating explainer as a fixed effect while modeling dataset, VLM architecture, and instance as random effects addresses the nested/repeated-measures structure that simpler tests (ANOVA) would mishandle. The comprehensive reporting of β coefficients, standard errors, and p-values across datasets strengthens confidence.
Benchmark scale: 8 explainers × 3 VLM architectures × 3 datasets, evaluated on complete dataset splits (~300 GPU hours), provides substantial empirical coverage.
However, several concerns emerge. The macro-coalitional ground truth, while cleverly constructed, involves arbitrary choices (C=6 background coalitions, coupled cross-modal partitioning) that could introduce aggregation bias. The authors acknowledge this trade-off but don't quantify sensitivity to C. Additionally, the K=11 discrete steps for the Riemann approximation is relatively coarse; sensitivity analysis over K values is absent.
3. Potential Impact
Immediate impact on XAI evaluation: The paper's most actionable contribution is demonstrating that the current evaluation paradigm for multimodal XAI is fundamentally broken. This could redirect the community away from unimodal perturbation metrics toward interaction-based evaluation, which is a meaningful paradigm shift.
Practical implications for VLM auditing: The finding that VLM-native explainers (LLaVA-CAM, TAM) over-index on visual salience while underperforming attention-based methods (AttnLRP, Rollout) at capturing true cross-modal synergy is practically important. It challenges the assumption that architecture-specific explainers are inherently superior and suggests that deployed audit systems may be producing misleading explanations.
Regulatory relevance: With the EU AI Act mandating explainability for high-risk AI systems, having metrics that distinguish visually plausible but unfaithful explanations from genuinely faithful ones has direct regulatory utility.
Limitations on breadth: The restriction to VQA-format tasks, mid-scale open-weight models (2B-7B), and binary/multiple-choice outputs limits immediate applicability to the broader VLM ecosystem (open-ended generation, proprietary models, video/audio modalities). The authors acknowledge this transparently.
4. Timeliness & Relevance
The paper is highly timely. VLMs are being rapidly deployed in high-stakes applications (medical imaging, autonomous driving, robotics), and the gap between deployment pace and evaluation rigor for explainability is widening. The multimodal XAI evaluation literature has not kept pace with the transition from dual-encoder models (CLIP, ViLT) to autoregressive VLMs, and this paper directly addresses that gap.
The identification of cross-modal redundancy as a systematic confound in evaluation is particularly relevant given mounting evidence of modality biases in VLM benchmarks (citations [10-15] in the paper). This connects the XAI evaluation problem to the broader VLM evaluation crisis.
5. Strengths & Limitations
Key Strengths:
Notable Weaknesses:
Overall Assessment
This is a well-executed paper that identifies a genuine and important problem in multimodal XAI evaluation, proposes a theoretically principled and computationally practical solution, and delivers surprising empirical findings. The main contribution — shifting from unimodal to synergistic evaluation of multimodal explanations — represents a meaningful conceptual advance. While the absolute effect sizes are modest and the scope is restricted to VQA, the paper establishes a foundation that should influence how the community evaluates and develops XAI methods for VLMs.
Generated May 22, 2026
Comparison History (21)
Paper 2 has higher likely impact: it addresses a timely, broadly relevant problem (faithful evaluation of VLM explainability) and identifies a fundamental failure mode in current metrics. Its contribution—a theoretically grounded, scalable cross-modal synergy metric (Shapley/Harsanyi-based) with strong empirical validation across models, methods, and datasets—can reshape how multimodal XAI is assessed and audited, with clear implications for safety-critical deployment. Paper 1 is novel and useful for model hubs, but its impact is more niche to routing/benchmarking in evolving expert pools and may depend on adoption of specific hub paradigms.
Paper 1 addresses a fundamental methodological gap in VLM explainability by identifying evaluation collapse in cross-modal settings and proposing a theoretically grounded metric (Synergistic Faithfulness) rooted in game-theoretic concepts. It has broader impact across XAI, multimodal AI, and AI safety, with rigorous evaluation across multiple architectures and datasets. The finding that current VLM explainers over-index on visual salience challenges prevailing assumptions. Paper 2, while practical and well-executed for agent memory systems, represents a more incremental architectural contribution with narrower scope in conversational AI.
Paper 1 addresses one of the most pressing challenges in modern AI: optimizing Chain-of-Thought (CoT) reasoning and post-training (SFT/RL) for LLMs. By providing actionable insights into how reasoning data compression affects capabilities and RLVR, it directly informs the development of next-generation reasoning models. While Paper 2 offers a rigorous and novel evaluation metric for VLM explainability, Paper 1 has broader, more immediate implications for scaling model capabilities, reducing inference costs, and advancing the foundational training paradigms of frontier LLMs, granting it higher overall scientific and practical impact.
Paper 1 addresses a fundamental methodological gap in VLM explainability by identifying evaluation collapse in cross-modal settings and proposing a principled metric (Synergistic Faithfulness) grounded in game-theoretic concepts. It offers broad impact across XAI, multimodal AI, and AI safety, with rigorous evaluation across multiple architectures and methods. Paper 2 contributes a useful benchmark for LLM strategic reasoning but is more narrowly scoped to game-theoretic evaluation. Paper 1's novel metric framework, identification of a systematic flaw in current evaluation paradigms, and applicability to high-stakes AI auditing give it greater potential for widespread methodological adoption.
Paper 1 addresses a critical, foundational problem in AI safety and interpretability by exposing fundamental flaws in current VLM evaluation metrics. By introducing a mathematically rigorous framework (Shapley Interaction Index) to measure cross-modal synergy, it offers broad scientific impact across multimodal learning and high-stakes AI deployments. In contrast, Paper 2, while highly practical and well-engineered, focuses on a narrower application area (video game NPCs and multi-agent simulations), limiting its broader theoretical and cross-disciplinary scientific impact.
Paper 2 addresses a critical and highly active area of AI safety: model alignment and jailbreaking. By providing a principled framework for latent-space attacks and demonstrating state-of-the-art success rates across 15 diverse models, it exposes severe vulnerabilities in current safety mechanisms. This will likely drive significant follow-up research in both mechanistic interpretability and robust alignment, giving it a broader and more urgent impact than Paper 1's focus on XAI evaluation metrics.
Paper 2 offers a more novel, technically specific contribution: it identifies a concrete failure mode in current VLM explainability evaluation (cross-modal redundancy causing metric collapse) and proposes a new, principled metric grounded in Shapley interactions with strong empirical validation (high correlation, large speedup) across models, datasets, and XAI methods. This yields immediate real-world relevance for auditing multimodal systems and is timely given widespread VLM deployment. Paper 1 is a broad synthesis/positioning chapter with important applications but less methodological novelty and weaker direct, testable contributions.
Paper 2 introduces a principled new metric (Synergistic Faithfulness) grounded in Shapley interaction to isolate true cross-modal contributions, addressing a clear failure mode (evaluation collapse) in VLM explainability. It offers strong methodological rigor (theory-backed metric, high surrogate correlation, large speedup, multi-model/multi-dataset evaluation) and broad relevance across multimodal ML, XAI, safety, and auditing. Paper 1 provides an important benchmark and negative result for clinical LLM interaction, but its impact is narrower (clinical decision support) and more incremental relative to existing interactive evaluation efforts.
Gated DeltaNet-2 introduces a fundamental architectural improvement to linear attention mechanisms by decoupling erase and write operations, with strong empirical results across multiple benchmarks at 1.3B scale. This addresses a core limitation in efficient sequence modeling—a rapidly growing field with broad impact on LLM efficiency, long-context modeling, and inference cost. Paper 2 contributes a valuable evaluation metric for VLM explainability, but its scope is narrower (benchmarking/evaluation rather than model architecture), and its impact is primarily within the XAI subcommunity rather than the broader deep learning ecosystem.
Paper 1 offers a novel, technical evaluation metric (Synergistic Faithfulness) grounded in Shapley interactions to fix a concrete failure mode in VLM explainability evaluation, with strong quantitative evidence (high surrogate correlation, large speedup) and broad applicability to auditing multimodal systems in high-stakes settings. Its methodological contribution is directly actionable for benchmarking and improving XAI methods across architectures/datasets, likely influencing subsequent empirical work. Paper 2 is timely and useful for conceptual clarity and governance, but is primarily taxonomic/survey-based with less direct methodological leverage for model development and evaluation.