Self-Driving Negotiator benchmark shows AI models score only 0.68 success rate

Autonomous driving involves countless tiny social negotiations—drivers pressing forward, yielding, or faking toward a curb. Existing benchmarks focus on perception or explicit text negotiation, but Self-Driving Negotiator bridges this gap. Created by Ashutosh Kumar, this procedurally generated text-only environment requires agents to infer hidden intent from behavior under partial observability, then generate driving actions. Rewards and diagnostics are computed from the privileged simulator state, not from the model's explanation, ensuring verifiable performance.

Results are sobering: the best model achieves only 0.68 average success across three scenarios, far below a scripted expert. Contested merge performance is statistically flat across all six models tested, and difficulty tiers reveal that models rely on cue-following rather than true wait-for-commitment behavior. The benchmark includes anti-gaming invariants and validated scenarios, making it a rigorous test for social coordination in autonomous driving. Current LLM-based agents remain far removed from human-like negotiation.