CoLT-Drive - False Positive: leaderboard

The false positive category of CoLT-Drive (Tang et al., 2026): benign lookalikes such as shadows, puddles, flat cardboard and leaves inserted into front-view driving scenes with the surrounding traffic kept, where the nominal action should be kept. The model outputs a longitudinal and lateral meta-action, and the score is pair accuracy in percent against the human-adjudicated acceptable set; higher is better.

Metric: Pair accuracy (%; the 435 v_full samples, naturalistic surrounding traffic kept, whose inserted object is one of the benign lookalikes that should not change the action (shadows, puddles, flat cardboard, leaves), 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-false-positive · How It Works · Data refreshed daily, snapshot 2026-11-12.