PersonaTrail (AWM) - Multi-hop Episodic Grounding: leaderboard

Metric: Task success rate (%; multi-hop episodic grounding across several past sessions; success judged by an LLM evaluator; live websites through a Browser-Use web agent with at most 30 steps and temperature 0, given memory from AWM (per-user workflow memory induced from past action trajectories); Gemini 3.1 Pro evaluator; the three proprietary backbones run on a 25% stratified subset). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 6 models tracked.

Top models

#ModelScore
1Browser Use + Claude Haiku 4.521.3
2Browser Use + Qwen3.6-27B17.84
3Browser Use + GPT-5.4-mini14.8
4Browser Use + Gemma-3-27B-it12.15
5Browser Use + Llama-4-Scout5.31
6Browser Use + Gemini-3-Flash3.56

Interactive version: theaggregate.ai/benchmark?slug=personatrail-awm-multi-hop-episodic-grounding · How It Works · Data refreshed daily, snapshot 2026-09-29.