Intel / article

Teach-to-Crash: A Closed-Loop Student-Teacher LLM Framework for Collision-Inducing Test Scenario Generation

Source: arxiv.org
Author: Zaid Ghazal, Khouloud Gaaloul, Bruce Maxim
Date: Published Sep 24, 2026

Context

Teach-to-Crash uses two language models to search for driving situations that cause a simulated controller to fail. A student model writes scenarios in the simulator's required format. A teacher model redirects the search only when recent collision frequency and time-to-collision measurements stop improving.

The authors report a 90.79% collision hit rate in their CARLA case study, which varied the simulated vehicle's speed policy across two setups. Collisions are the outcome this search is trying to produce, so the figure reflects the testing method in simulation and does not estimate ordinary driving risk. The contribution is a feedback-driven way to generate executable stress tests and change search direction when the existing approach stops finding useful failures.

Key takeaways · AI-distilled

A low-reasoning Student emits simulator-executable scenarios in a strict JSON schema, while the Teacher steps in only when rolling collision rate and time-to-collision metrics stagnate.
Beyond hit rate, Teach-to-Crash had the shortest mean time-to-collision (18.31 s) and the highest scenario diversity (0.547) among compared methods in the CARLA study.
PAFOT found collisions at a higher mean discovery rate (179.44 vs 136.21), though with substantially larger variance.
The authors limit their claims to the evaluated CARLA scope; averaged over both setups on the CARLA Traffic Manager controller, it also scored highest on an avoidability-based usefulness proxy (60.04%).

Terms in this piece · Glossary

LLM — A large language model — the neural network behind tools like Claude and ChatGPT, trained on huge amounts of text to predict what comes next.

Read the source arxiv.org