Research Track

Decomposed Hybrid Reasoning for Autonomous Driving

Fusing Physics-Based and Policy-Based Constraints via Interval Arithmetic

The Problem: Hybrid Reasoning in LLMs

Current LLMs cannot reliably fuse physics-based numerical reasoning with policy-based symbolic reasoning for autonomous driving decisions. We propose architectural decomposition as the solution.

Monolithic LLM
57.5%
Overall Accuracy
CoT Prompting
62.3%
Overall Accuracy
Tool-Augmented
73.2%
Overall Accuracy
Hybrid (Ours)
88.3%
+34.7 pp vs Monolithic

Key Finding

On the hardest hybrid-reasoning scenarios (requiring simultaneous physics and policy integration), our decomposed framework achieves 86.2% accuracy compared to 51.5% for monolithic LLMs — a 34.7 percentage-point improvement. Physics computation errors drop from 12.2m to 0.9m (13x reduction).

Architecture: Decomposed Hybrid Reasoning

Each reasoning mode is handled by a dedicated module operating in its area of strength.

Scenario Description (Natural Language)
LLM Scenario Parser
Extract entities, speeds, distances, conditions
Physics Engine
Interval arithmetic: braking, TTC, gaps
Policy Engine
Rule DB: speed limits, margins, zones
Constraint Fusion & Decision
Priority-weighted constraint satisfaction
Decision + Confidence + Explanation

Results: Decision Accuracy by Reasoning Mode

Monolithic LLM
CoT LLM
Tool-Aug. LLM
Hybrid (Ours)

Accuracy by Difficulty Level

Monolithic LLM
CoT LLM
Tool-Aug. LLM
Hybrid (Ours)

Complete Results Table

Reasoning ModeMonolithic LLMCoT LLMTool-Aug. LLMHybrid (Ours)
Simple0.7500.8331.0001.000
Physics-Only0.7000.7670.9170.967
Policy-Only0.8590.7970.6560.938
Hybrid0.5150.5750.7110.862
Overall0.5750.6230.7320.883

Accuracy Across Weather and Road Conditions

Toggle between models to compare accuracy across environmental conditions.

Interactive Demo: Hybrid Reasoning Engine

Configure a driving scenario and see how the decomposed framework computes a decision.

Scenario Configuration

Decision Output

Physics Computation Accuracy

Deterministic interval arithmetic reduces physics errors by an order of magnitude.

Braking Distance MAE
12.2m
Monolithic LLM
0.9m
Hybrid (Ours)
13x reduction
TTC Estimation MAE
10.4s
Monolithic LLM
1.0s
Hybrid (Ours)
10x reduction

Physics Error by Model

MetricMonolithic LLMCoT LLMTool-Aug. LLMHybrid (Ours)
Braking Dist. MAE (m)12.2 +/- 24.18.7 +/- 16.02.6 +/- 5.10.9 +/- 1.5
TTC MAE (s)10.4 +/- 28.17.5 +/- 18.72.8 +/- 6.91.0 +/- 2.6