Revealing Interpretable Failure Modes of VLMs

2026-03-07 · Source: cs.AI updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Revelio is a novel framework designed to systematically uncover interpretable failure modes in Vision-Language Models (VLMs), particularly for safety-critical applications like autonomous driving and indoor robotics. It defines a failure mode as a consistent incorrect VLM behavior under a composition of domain-relevant concepts, such as "pedestrian proximity" or "adverse weather conditions." To navigate the exponentially large combinatorial space of these concepts, Revelio employs two search strategies: a diversity-aware beam search for efficient mapping of the failure landscape and a Gaussian Process-based Thompson Sampling for broader exploration. Applied to autonomous driving and indoor robotics, Revelio identified previously unreported vulnerabilities in state-of-the-art VLMs, including weak spatial grounding, failure to account for major obstructions, and excessive conservatism. The framework discovered 3-5x more failure modes than unguided random search within the same budget, providing actionable insights for VLM safety improvements.

Key takeaway

For AI Engineers and Research Scientists developing or deploying VLMs in safety-critical systems, Revelio offers a robust framework to proactively identify and understand systemic vulnerabilities. You should consider integrating Revelio's concept-based evaluation to move beyond anecdotal failures, pinpointing specific conditions that trigger consistent VLM errors. This approach provides concrete, interpretable insights, enabling targeted model improvements and guiding more effective safety alignment before deployment.

Key insights

Revelio systematically uncovers interpretable VLM failure modes by searching concept compositions in safety-critical domains.

Principles

• Failure modes are compositions of interpretable, domain-relevant concepts.
• Active search for vulnerabilities must combine discovery with semantic realism.
• Combining concepts can non-linearly change VLM failure rates.

Method

Revelio uses a diversity-aware beam search and Gaussian Process-based Thompson Sampling to explore concept combinations, generating scenarios via simulators like CARLA or image generation models, and evaluating VLM responses against rule-based ground truth.

In practice

• Use Revelio to identify VLM vulnerabilities in autonomous driving.
• Apply Revelio for hazard detection in indoor robotics scenarios.
• Encode concept sets as multi-hot binary vectors for GP modeling.

Topics

• Vision-Language Model Safety
• Interpretable Failure Modes
• Revelio Framework
• Autonomous Driving
• Indoor Robotics

Code references

• uiuc-focal/Revelio

Best for: AI Engineer, Computer Vision Engineer, Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Tri-Info: Generalizable, Interpretable Failure Prediction for VLA Models via Information Theory — Tri-Info uses information-theoretic signatures to predict VLA model failures, offering generalizable detection and interpretable diagnostics.
• Task-Aligned Stability Analysis of Vision-Language Models for Autonomous Driving Hazard Detection — VLM robustness for autonomous driving requires task-aligned stability analysis beyond embedding drift due to corruption-dependent decision instability.
• ReCoVLA: VLM-Guided Reward Compilation for Failure Recovery in Vision-Language-Action Policies — ReCoVLA uses a VLM to compile structured rewards for residual policy training, enabling robust VLA policy recovery.
• How VLAs (Really) Work In Open-World Environments — Current VLA evaluation metrics for robotics overstate performance by neglecting safety and intermediate states.
• Action Draft and Verify: A Self-Verifying Framework for Vision-Language-Action Model — ADV combines diffusion's precision with auto-regression's robustness for VLA models via a draft-and-verify mechanism.
• Safe Embodied AI for Long-horizon Tasks: A Cross-layer Analysis of Robotic Manipulation — Safe long-horizon robotic manipulation requires a cross-layer framework distinguishing intervention loci and evidence boundaries.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.AI updates on arXiv.org.