Fusing Physics-Based and Policy-Based Constraints via Interval Arithmetic
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.
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).
Each reasoning mode is handled by a dedicated module operating in its area of strength.
| Reasoning Mode | Monolithic LLM | CoT LLM | Tool-Aug. LLM | Hybrid (Ours) |
|---|---|---|---|---|
| Simple | 0.750 | 0.833 | 1.000 | 1.000 |
| Physics-Only | 0.700 | 0.767 | 0.917 | 0.967 |
| Policy-Only | 0.859 | 0.797 | 0.656 | 0.938 |
| Hybrid | 0.515 | 0.575 | 0.711 | 0.862 |
| Overall | 0.575 | 0.623 | 0.732 | 0.883 |
Toggle between models to compare accuracy across environmental conditions.
Configure a driving scenario and see how the decomposed framework computes a decision.
Deterministic interval arithmetic reduces physics errors by an order of magnitude.
| Metric | Monolithic LLM | CoT LLM | Tool-Aug. LLM | Hybrid (Ours) |
|---|---|---|---|---|
| Braking Dist. MAE (m) | 12.2 +/- 24.1 | 8.7 +/- 16.0 | 2.6 +/- 5.1 | 0.9 +/- 1.5 |
| TTC MAE (s) | 10.4 +/- 28.1 | 7.5 +/- 18.7 | 2.8 +/- 6.9 | 1.0 +/- 2.6 |