Risk Semantic Distillation (RSD)

• Risk Semantic Distillation is a method that decomposes multi-step reasoning into discrete atomic tasks, enabling precise risk attribution in multimodal large language model evaluations.
• It employs expert annotations and automated LLM feedback to isolate and assess both perceptual and logical failures in challenging autonomous driving scenarios.
• The approach quantifies risk propagation through hierarchical metrics, offering actionable insights for enhancing model safety and robustness under adverse environmental conditions.

Risk Semantic Distillation (RSD) is a conceptually critical methodology in the evaluation of multimodal LLMs (MLLMs) for autonomous driving under adverse and complex scenarios, particularly as formalized and implemented in the AD²-Bench benchmark (Wei et al., 11 Jun 2025). RSD addresses the need for granular, interpretable assessment of multi-step reasoning processes by separating and evaluating the fidelity of intermediate inferential “atoms” within longer Chain-of-Thought (CoT) reasoning chains. By enabling the explicit isolation and scrutiny of each atomic reasoning step, RSD empowers benchmark designers to dissect both perception- and reasoning-driven failure modes, quantify semantic robustness, and precisely characterize the risk propagation through the CoT process, especially in safety-critical driving contexts with severe environmental challenges.

1. Background and Motivation

Traditional evaluation paradigms for MLLMs in autonomous driving, such as open-loop and closed-loop trajectory scoring or final-answer accuracy in VQA settings, provide only holistic or outcome-level metrics and insufficiently illuminate the inferential vulnerabilities of these systems—particularly under adverse-weather, occlusion, or domain shift. As MLLMs rely heavily on internally chained, hierarchical reasoning, risk emerges not only from perceptual misidentification but from the compounding or masking of errors across intermediate semantic steps. The absence of intermediate step-level auditing in legacy benchmarks prevents researchers from attributing system failures (e.g., unsafe decisions) to specific semantic breakdowns, impeding both diagnostic interpretability and targeted robustness improvement.

AD²-Bench (Wei et al., 11 Jun 2025) operationalizes RSD as a means to close this diagnostic gap. The approach is motivated by three intersecting deficiencies in prior work:

• Lack of Adverse-Condition Fidelity: Most datasets target benign conditions, neglecting semantic stressors (e.g., heavy fog, nighttime, sand storms) where inference risk amplifies.
• Annotation Granularity: Siloed, model-generated, and often only outcome-supervised annotations obscure logical progression, leading to unreliable risk attribution.
• CoT Process Opaqueness: Without atomic granularity, multistep reasoning failures propagate undiagnosed.

Through RSD, AD²-Bench introduces a protocol to decompose each VQA/decision query into minimal semantic units for explicit risk and fidelity assessment.

2. Atomic Decomposition and Annotation Protocol

Risk Semantic Distillation begins by enforcing “Atomic Annotation,” wherein every multi-step CoT task is decomposed into a sequence of discrete reasoning sub-tasks—termed atoms—each representing the smallest independently defeasible semantic unit (e.g., existence of an object, attribute classification, basic spatial reasoning query).

• Expert Annotation: Domain specialists independently annotate each atom, ensuring that each intermediate state in the CoT chain is registered to an explicit, verifiable ground truth.
• Automated LLM Feedback: Large multimodal models (e.g., Gemini-2.5-Pro, GPT-4o, Qwen-2.5-Max) evaluate candidate atomic answers for fluency and semantic alignment, feeding into a “Wisdom Atom” synthesis by lead human annotators.
• Cross-Expert Verification: Multiple rounds of mutual review by annotators from different subdomains establish both local atomic fidelity and global narrative coherence.
• Multi-Level Prompting: Atoms are paired with prompt modalities—text, point in image (pixel-level), region (bounding box)—to disambiguate perception vs. reasoning errors.

This process ensures that all risk carries semantic provenance: any deviation from correct reasoning can be localized to an atomic misstep and classified as perceptual, logical, or compositional in origin.

3. Hierarchical Reasoning Structure and Workflow

In AD²-Bench, each CoT chain is hierarchically organized into three interconnected reasoning tiers, each supporting fine-structural risk distillation:

1. Text-Level: Natural language cues referencing global or regional objects (e.g., “Identify the traffic sign on the left”).
2. Point-Level: Pixel-specific queries target occluded or remote objects, introducing semantic fragility due to visual ambiguity.
3. Region-Level: Bounding box prompts capture higher-order composite reasoning (e.g., identifying objects in a hazard zone).

Within this structure, each atomic answer $A_i$ is paired with a known ground-truth atom $GT_i$, forming a chain $S = \{GT_1, \ldots, GT_N\}$ that represents the intended perception-understanding-reasoning trajectory. RSD thus enables granular, risk-aware analysis of reasoning flow and robustness.

4. Evaluation Metrics for Risk-Aware CoT Assessment

RSD is operationalized through a suite of multi-dimensional step-level and chain-level metrics explicitly designed to quantify both semantic fidelity and risk transmission in the CoT process:

• CoT Accuracy (Exact Match):

$$\mathrm{CoTAccuracy} = \frac{1}{N} \sum_{i=1}^{N} \mathbb{I}(\widehat{S}_i = S_i)$$

Measures strict agreement of the predicted chain with the expert ground truth.

• Step-wise Accuracy & Completeness Score (SACS):

$\mathrm{SACS} = \frac{1}{N} \sum_{i=1}^{N} \Eval_{\mathrm{acc}}(A_i, GT_i)$

Quantifies atomic prediction correctness on a fine scale (1–10).

• Step-wise Logical Progression Score (SLPS):

$\mathrm{SLPS} = \frac{1}{N-1}\sum_{i=1}^{N-1} \Eval_{\mathrm{prog}}(A_i, A_{i+1})$

Assesses transition coherence between reasoning atoms.

• Overall Reasoning Coherence Score (ORCS):

$\mathrm{ORCS} = \Eval_{\mathrm{coh}}(A_1, \ldots, A_N)$

Aggregates logical consistency across the chain.

• Decision Justification Strength Score (DJSS):

$\mathrm{DJSS} = \Eval_{\mathrm{just}}((A_1, \ldots, A_{N-1}), A_N)$

Measures explanatory depth and causal salience.

An LLM evaluator assigns scores for correctness, coherence, and justification, allowing RSD to quantify the probability and location of risk propagation through semantic reasoning.

5. Empirical Insights and Risk Propagation Patterns

Application of RSD in AD²-Bench reveals pivotal risk profiles for MLLMs:

• Without CoT prompting, final accuracy is sub-40% across models, indicating compounded risk from single-step inferential collapses.
• Hierarchical CoT (multi-step atomic CoT prompting) boosts accuracy by 15–20 points, with the best systems (InternVL3-8B) at ≈63% in CoT-augmented scenarios, but still <60% in final-chain accuracy—demonstrating persistent risk even when intermediate steps are supervised.
• Dominant failure modes isolated via RSD include:
  • Misperception of subtle weather (e.g., light fog vs. overcast).
  • Hallucination of objects under occlusion or rain (phantom trams, etc.).
  • Logical incoherence between sequential reasoning atoms.
• Vision encoder deficiencies in adverse scenes lead to upstream atomic failures, which RSD pinpoints as the root cause of multi-hop inferential breakdowns.

This suggests that atomic risk segmentation, as furnished by RSD, is indispensable for debugging and improving MLLM reasoning fidelity in real-world, high-risk environments.

6. Comparative Analysis and Distinguishing Features

Prior benchmarks (BDD-X, nuScenes-QA, DriveLM, CODA-LM, DRAMA) offer broad but undifferentiated QA coverage and lack explicit atomic-level semantic distillation; they do not provide:

• Real adverse-weather data at controlled severity.
• Fine-grained, expert-verified CoT annotations treating each step as a ground truth atom.
• Multi-modal, multi-tier prompting for discrimination of perceptual vs. inferential risk.
• A multi-dimensional, LLM-scored suite quantifying stepwise and chainwise semantic fidelity and risk justification.

RSD, as instantiated in AD²-Bench, is therefore the first large-scale multimodal evaluation scheme to operationalize risk distillation at the semantic atom level, setting a new standard for interpretability and systematic risk analysis.

7. Directions for Advancing Semantic Risk Distillation

Future research avenues directly informed by RSD include:

• Adverse-weather-aware pre-training and multimodal sensor fusion to fortify the perceptual substrate of atomic reasoning.
• Scaling atomic annotation via high-quality LLM-assisted protocols, maintaining controlled risk traceability.
• Integrating temporal and spatial context in video reasoning for richer risk propagation modeling.
• Structured reasoning architectures with explicit intermediate supervision to minimize hallucination and enhance stepwise logical persistence.
• Real-time risk-aware deployment strategies, balancing detail preservation (e.g., patch-based input) with computational budgets.
• Automated ontology-driven root-cause analysis linking atomic step failures to macro-level safety risk signatures.

A plausible implication is that systematic risk semantic distillation will become an essential diagnostic and validation layer in any high-assurance, end-to-end reasoning system deployed under adversarial or uncertain real-world conditions (Wei et al., 11 Jun 2025).