PersonaTrail (ReasoningBank) - Multi-hop Preference Inference: leaderboard
PersonaTrail (2026) evaluates personalized web agents on live websites that must resolve under-specified requests from a raw browsing history, run through the Browser-Use agent framework. Task success in percent on multi-hop preference inference over several websites, with history supplied through ReasoningBank reasoning memory; higher is better.
Metric: Task success rate (%; multi-hop preference inference over several websites; success needs every website, the intent and the preference 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: Around October 2027. 6 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Qwen 3.6 27B | 13.73 | #341 |
| 2 | Claude Haiku 4.5 | 9.44 | #568 |
| 3 | Gemini 3 Flash | 9.19 | #164 |
| 4 | GPT-5.4 Mini | 1.91 | #389 |
| 5 | Gemma 3 27B (IT) | 0.39 | #1289 |
| 6 | Llama 4 Scout | 0.39 | #1776 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=personatrail-reasoningbank-multi-hop-preference-inference · How It Works · Data refreshed daily, snapshot 2026-11-13.