CoLT-Drive: leaderboard

CoLT-Drive (Tang et al., EMNLP Findings 2026) inserts rare objects into fixed front-view driving scenes and asks a vision-language model for a longitudinal and lateral meta-action given the image, ego-motion history and navigation command. The score is pair accuracy in percent, the share of samples whose predicted action pair is in the human-adjudicated acceptable set, over scenes with and without background traffic; invalid outputs count as wrong, higher is better.

Metric: Pair accuracy (%; share of the 3,536 counterfactual samples, 1,768 scenes in both v_full with surrounding traffic and v_clean with background vehicles removed, where the model outputs an acceptable longitudinal-lateral meta-action pair for a front-view driving image with a counterfactually inserted rare object, ego-motion history and navigation command; human-adjudicated acceptable-pair sets; same structured prompt and greedy decoding for every model; decisions extracted by rule and normalized by a text-only DeepSeek-V4-Pro mapper, then scored deterministically; invalid outputs count as wrong). Source: arxiv.org. Saturation forecast: Around February 2027. 9 models tracked.

Top models

No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.

Interactive version: theaggregate.ai/benchmark?slug=colt-drive · How It Works · Data refreshed daily, snapshot 2026-11-12.