PersonaTrail (ReasoningBank) - Single-hop Preference Inference: leaderboard

Metric: Task success rate (%; single-hop preference inference: an under-specified query whose missing preference must be inferred from recurring browsing behaviour; success needs the right website, the task intent and the preference all satisfied; live websites through a Browser-Use web agent with at most 30 steps and temperature 0, given memory from ReasoningBank (per-user reasoning items distilled from successful and failed 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 4.555.69
2Browser Use + Gemini-3-Flash46.88
3Browser Use + Qwen3.6-27B46.45
4Browser Use + GPT-5.4-mini25.98
5Browser Use + Gemma-3-27B-it25.66
6Browser Use + Llama-4-Scout16.67

Interactive version: theaggregate.ai/benchmark?slug=personatrail-reasoningbank-single-hop-preference-inference · How It Works · Data refreshed daily, snapshot 2026-09-29.