From MIRAGE to CLEAR: Component-Level Explainable Anomaly Reasoning for Autonomous Vehicle Perception Systems

Jan 1, 2026· David FernandezDavid Fernandez ,Pedram MohajerAnsari ,Cigdem Kokenoz ,Amir Salarpour ,Bing Li ,Mert D. Pese
PDF Code Cite

Key Contributions & Takeaways

  • When a self-driving car’s perception breaks, most tools only flag that something is wrong. MIRAGE-CLEAR points to the module actually at fault (sign reading, lane keeping, or object detection) and explains the call.
  • Every finding is backed by the rule book. CLEAR grounds its reasoning in the MUTCD, the U.S. traffic-control standard, so each answer cites a real regulation. That turns an opaque model into audit-ready evidence for laws like the EU AI Act.
  • It knows when it isn’t sure. CLEAR reaches 95.2% accuracy at spotting anomalies and 84% at blaming the right module for clear sign failures. When fog or glare hits several modules at once, it spreads its confidence instead of guessing, which matches the physics.
publications

01  ·  The Gap

When a self-driving car’s perception fails, today’s tools can tell you that it failed, but not which part. Only 4% of real disengagements get a clear cause, so safety fixes have nothing to aim at.

02  ·  The Framework

MIRAGE builds a dataset of realistic, rule-breaking driving scenes with known causes. CLEAR then reasons over each scene in three steps to name the module at fault.

03  ·  The Result

CLEAR spots anomalies with 95.2% accuracy and traces clear sign failures to the right module 84% of the time, with every answer citing a real traffic regulation.

The Attribution Gap
We can tell that perception failed. We usually can’t say which part failed, or why.
17%
of AV disengagements
come from perception
4%
of those failures get a
clearly identified cause
1,612→2,359
Level-2 ADAS crashes,
mid-2024 to Apr 2025 (NHTSA)
Why It Is Hard

A single problem in the world rarely shows up in a single module. Take one common case: a faded STOP sign at a foggy intersection. One condition breaks three modules at once, so no single one is obviously to blame.

Faded STOP sign,
foggy intersection
one condition
TSR
sign recognition

Cannot read the faded sign: lost contrast.

ALC
lane centering

Cannot see lane edges through the fog.

OD
object detection

Cannot judge depth to nearby vehicles.

What Existing Methods Miss (Table I)

Prior approaches can flag an anomaly, but almost none tie it to a module, explain the call, and stand up to a regulator. CLEAR is built to do all four.

MethodDetectAttributeExplainCompliant
Statistical methods✓✗✗✗
Deep-learning anomaly detection✓✗✗✗
Rule-based systems✓◐✓◐
Vision-language models✓✗✓✗
Counterfactual analysis✓✓✗✗
★ CLEAR (ours)✓✓✓✓
✓ full ◐ partial ✗ none
And Now It Is the Law

The EU AI Act (2024/1689) and the U.S. NHTSA AV STEP proposal both classify self-driving as high-risk AI and demand component-level traceability. A generic “the system failed” alert no longer meets the bar: regulators want the failed module and the reason behind it.

The Framework
Two halves: MIRAGE makes the data, CLEAR does the reasoning.
MIRAGEMUTCD-InformedRealistic AnomalousGeneration EngineScene Data CollectionMUTCD IntegrationSemantic EmbeddingAnomalous Dataset GenerationDSLStructuredScenariosCLEARComponent-LevelExplainable AnomalyReasoningLayer 1: DetectionLayer 2: ClassificationLayer 3: AttributionMulti-Layer Reasoning Analysis
MIRAGE · Building the Dataset

Two streams meet in the middle. Real scenes become text and then embeddings; the MUTCD rule book does the same. Matching the two picks the regulation that governs each scene, and an LLM then writes a controlled violation of it.

MIRAGE: MUTCD-Informed Realistic Anomalous Generation EngineScene Analysis & Rule ExtractionAV scenesVLM: VILAScene Description1ImageDescriptionEmbeddingsMUTCDRules: Structuredregulations2Embeddings3Semantic Matching4Semantic Embedding ComparisonImgDescRuleall-MiniLM-L6-v2 · matched pairsGeneration5LLM: AnomalyGenerationDSLStructured Scenarios
48,022
real-world
scenes
3
datasets: nuScenes,
Waymo, Argoverse 2
5,847
annotated
anomalies
100%
U.S. scenes,
MUTCD-aligned
Grounded in the MUTCD Rule Book

Every anomaly breaks a real section of the U.S. traffic-control manual, so a violation is verifiable, not a subjective label. Four kinds are injected:

Direct

A required sign is missing or wrongly placed.

Subtle

Faded markings or a partly hidden sign.

Contextual

A school-zone rule active without its warning.

Environmental

Fog, glare, deep shadow, or heavy rain.

CLEAR · Three Layers of Reasoning

Each layer narrows the question and reports its own metrics. The confidence needed to move on rises at every step, so a layer that is not sure stops rather than guessing. The last layer maps the failure onto one of three perception modules.

DSLStructuredScenariosCLEAR: Component-Level Explainable Anomaly Reasoning1≥0.70AnomalyDetectionIdentify thatsomething is wrongMetricsDetection Accuracy2≥0.75AnomalyClassificationType of irregularitydetectedMetricsClassification accuracyClass Precision, Recall3≥0.80ComponentAttributionTrace to specificperception module(s)MetricsAttribution accuracyComponent precisionLLM engines evaluated:GPT-4o-mini · Phi-4 · Phi-4-mini
Attribution maps each failure onto one of three perception modules
TSR
Traffic Sign Recognition

Sign modification, occlusion, or misplacement.

ALC
Automated Lane Centering

Lane markings, merging areas, or road boundaries.

OD
Object Detection

Obstacles, pedestrians, or other vehicles.

Why It Works: Architecture, Not Prompt Tricks
Confidence-gated propagation

Thresholds rise with the stakes, so a shaky early call never cascades into a confident wrong module.

0.70→0.75→0.80
Schema-constrained outputs

Every answer must fit a fixed JSON schema and a fixed set of labels, so the model cannot invent categories or ramble.

fixed taxonomy · valid JSON
The LLM is a swappable engine

Rewriting the prompts three ways moves accuracy by only a hair. The structure does the work, not clever wording.

±1.5%across prompt
rewrites
Results
Detection is strong, attribution reaches a level no prior method offers, and every step stays auditable.
95.2%
anomaly detection
accuracy (L1)
84%
attribution on direct
TSR violations (L3)
74.6%
Top-2 attribution,
overall
78.6%
MUTCD citation
accuracy
A · Layer 1 Detection (Table V)

Spotting that something is wrong is the solved part of the problem, and CLEAR does it cleanly and with well-calibrated confidence.

95.2%
Accuracy
93.8%
Precision
96.7%
Recall
F1-score 95.2% avg. confidence on correct calls 0.89 processing 2.3 s/scene
B · Layer 2 Classification (62.3% overall)

Naming the violation type is harder. The confusion matrix shows where it slips: Subtle and Contextual blur into each other, because faded wear and situation-dependent rules often show up together in real scenes.

predicted →
DirectSubtleContext.Environ.
Direct69.213.510.56.8
Subtle15.060.018.07.0
Contextual10.518.058.013.5
Environmental6.512.019.562.0
rows = truth, columns = prediction (row-normalized %) green diagonal = correct
C · Layer 3 Component Attribution (Fig. 7)

This is the contribution no prior method offers. Each module is easiest to pin down on its own kind of failure.

TSR on direct violations84%
ALC on subtle violations75%
OD on contextual violations60%

Accuracy drops when a violation spans module boundaries, which is exactly what you would expect: those failures are, by definition, harder to pin on one module.

Reading the 41.7% Top-1 Score Correctly

Strict single-guess accuracy is 41.7%. That looks weak until you allow a second guess (Top-2 lifts it to 74.6%) and look at how CLEAR spreads its confidence. The misses are physics, not sloppy reasoning.

Direct violationentropy 0.78
TSR
ALC
OD

Sure, and correctly sure. Confidence piles onto one module.

Environmental violationentropy 0.98
TSR
ALC
OD

Correctly unsure. Fog and glare really do break all three at once.

With three modules, blind Top-2 guessing would score 66.7%. CLEAR beats that, and when it is unsure it says so: the true module lands in its top two most of the time. That is what auditable uncertainty looks like.

D · Reasoning Quality: Auditable by Design (Table VIII)

The answer is not enough; the reasoning has to hold up too. Coverage and coherence fade as the questions get harder, but the regulation citations stay accurate throughout, which is what keeps the whole pipeline regulator-ready.

LayerCompl.Consist.MUTCD
L1 Detection94.091.582.4
L2 Classification86.584.077.8
L3 Attribution71.073.275.6
Average83.882.978.6
Compl. = coverage of relevant aspects · Consist. = does the stated evidence justify the conclusion · MUTCD = are the cited rules real and applicable
A real Layer 1 output · waymo_sf_0234
{
  "anomaly_detected": true,
  "confidence": 0.92,
  "reasoning": "STOP sign at
    ground level: violates
    MUTCD 2B.04 (7-ft
    mounting height).",
  "indicators": [
    "improper placement",
    "height violation"
  ]
}
structured · cited · auditable
What This Establishes
01 · Attribution is the gap

Existing tools detect that something failed, not where. CLEAR is the first to reach module-level resolution on real, observational data.

02 · Rules make it auditable

Grounding every answer in the MUTCD turns opaque model output into traceable, regulator-readable evidence, with a section number you can look up.

03 · Doubt carries meaning

When CLEAR spreads its confidence, it is correctly modeling physics: fog and rain really do break several modules at once. The apparent misses are honest, not sloppy.

Honest Limits & Where It Goes Next
Offline

This is forensics, not a real-time monitor. At about 3.1 seconds per scene, CLEAR is built to analyze failures after the fact. Distillation and batched inference are the path toward sub-second use.

U.S. only

The rules are U.S. (MUTCD), but the pipeline is not. The same design can swap in EU or other jurisdictions; that grounding work simply has not been done yet.

Single label

It names one primary module by default. Top-2 already captures most of the multi-module reality, but a fully calibrated multi-label version with proper probabilities is still open.

CLEAR was run on GPT-4o-mini as the primary model, with Phi-4 and Phi-4-mini as open alternatives, over 17,541 API calls. Layer 1 used a balanced 11,694-scene set (5,847 anomalous scenes paired with 5,847 normal ones); Layers 2 and 3 used the 5,847 anomalous scenes. The diagrams above rebuild the paper’s framework and result figures; exact tables are in the PDF.

Abstract

“When a self-driving car's perception fails, today's tools tell you that it failed, not which part. MIRAGE-CLEAR traces each failure to a specific module and cites the regulation it broke.”

Autonomous vehicles rely on perception systems with deep neural networks for traffic sign recognition (TSR), automated lane centering (ALC), and object detection (OD). While these systems perform well under standard conditions, perception failures trigger 17% of AV disengagements, yet only 4% receive clear causal attribution, impeding safety improvements. Emerging regulations, including the EU AI Act (2024), mandate transparency and component-level accountability for safety-critical AI systems, which current anomaly detection approaches cannot provide. We present MIRAGE-CLEAR, a framework combining dataset generation with systematic reasoning analysis. MIRAGE generates semantically rich driving scenarios by integrating 48,022 real-world scenes from major AV datasets with regulatory knowledge from the Manual on Uniform Traffic Control Devices (MUTCD). CLEAR, a three-layer LLM-based reasoning pipeline, decomposes attribution into interpretable subtasks: anomaly detection, violation classification, and component attribution. CLEAR achieves 95.2% accuracy in anomaly detection and 84% attribution accuracy for direct regulatory violations targeting TSR modules. The framework provides interpretable reasoning chains satisfying regulatory transparency requirements while enabling precise failure source identification. To the best of our knowledge, CLEAR represents the first system unifying detection, attribution, explainability, and regulatory compliance for AV perception diagnostics.
Type Conference paper
Venue IEEE/IFIP International Conference on Dependable Systems and Networks (DSN 2026)
Explainable AI Autonomous Vehicles Anomaly Detection Large Language Models
David Fernandez
Authors
David Fernandez
PhD Candidate in Computer Science

David Fernandez is a PhD candidate in Computer Science at Clemson University, working on safe, efficient, and explainable AI for safety-critical systems. His research spans perception, adversarial robustness, and on-device deployment of large foundation models, including LLMs and VLMs, with five first-authored publications on component-level explainability, zero-shot reasoning, and adversarial scenario analysis, alongside collaborative work on edge AI for industrial agentic systems. Much of this research is grounded in autonomous driving, where trustworthiness, latency, and robustness constraints are unforgiving, but the underlying methods transfer broadly to other high-stakes domains.

As a member of Clemson’s VIPR-GS Research Program, he develops hierarchical LLM reasoning frameworks and VLM evaluation systems for the U.S. Army’s Next Generation Combat Vehicle (NGCV) program, focusing on zero-shot reasoning and component-level explainability under real-world deployment constraints.

At BMW Group, he designs agentic AI systems for enterprise environments, building autonomous prompt optimization pipelines that enable continual agent improvement without model retraining and context-aware moderation frameworks that detect coordinated multi-turn adversarial attacks in production deployments.