EgoActionBench - Semantic Tools: leaderboard

EgoActionBench (arXiv 2611.03646v1) tests dense action captioning in egocentric video: from a full clip, a model lists the camera wearer's visible atomic hand actions with its start and end times and the acting hand (left, right or both). The benchmark has 1,407 episodes from Ego4D, EgoLife and EPIC-KITCHENS-100 with 41,132 human-refined actions; the paper says "Models receive video inputs through their respective input interfaces". Its Table 6 threshold study, "with predictions generated at 2 fps", reproduces the Table 3 scores of GPT-6 Astra, Qwen3-VL-4B, Qwen3-VL-32B and LLaVA-OV2-8B, and Appendix A runs Qwen3-VL-4B at temperature zero with reasoning disabled; the paper gives no frame rate for the other baselines and no reasoning setting for the other models. Qwen3-32B is the text-only semantic judge at temperature zero. Table 3 Tools: judge-assessed semantic agreement on the Tools field. Metric: Tools score (0–100).

Metric: Tools agreement judged by the language-model judge (0-100; all 1,407 episodes (Ego4D, EgoLife and EPIC-KITCHENS-100; 41,132 human-refined actions); models output timestamped left/right/both-hand atomic action captions from the full clip; semantic fields judged by text-only Qwen3-32B at temperature zero, candidate-overlap threshold 0.5; models receive video through their own input interfaces; 2 fps per Table 6 for GPT-6 Astra, Qwen3-VL-4B/32B and LLaVA-OV2-8B, reasoning disabled for Qwen3-VL-4B (Appendix A); frame rate for the other baselines and reasoning settings for the other models not stated). Source: arxiv.org. Saturation forecast: Around February 2027. 19 models tracked.

Top models

No result here: #3 Claude Opus 5.5, #6 GPT-6.1 Sol, #9 Gemini 4 Argon.

Interactive version: theaggregate.ai/benchmark?slug=egoactionbench-semantic-tools · How It Works · Data refreshed daily, snapshot 2026-11-06.