Triplet Incremental Training Study (untuned backbones) - POPE: leaderboard

General benchmarks: MMMU, MMStar, POPE, OCRBench, ScienceQA; toolkit, splits, prompts, decoding and image settings not stated (§4.2; S3.T5). Driving benchmarks: nuScenes probing (depth, PETR detection), OmniDrive and nuInstruct language metrics, NAVSIM PDMS with a trained planning head (§4.2; §4.3). Training of the adapted model: 170k triplets from nuScenes training images (OmniDrive, nuInstruct) plus official driving QA; two epochs on 16 A100 80 GB; seed 42 (§4.1). This board: not stated; One score per model; single seed 42 is stated for training/evaluation of the adapted checkpoint, not for these backbone runs. not stated. Metric: POPE score as printed; split and scoring not described in the paper Higher is better.

Metric: POPE score as printed; split and scoring not described in the paper. Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1InternVL3.5-4B81.2#1916
2InternVL3-8B79#1727

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

Interactive version: theaggregate.ai/benchmark?slug=triplet-incremental-training-study-untuned-backbones-pope · How It Works · Data refreshed daily, snapshot 2026-11-16.