PAL-Bench (PAL-TRACE) - Owner Fact Recovery: leaderboard

Metric: Owner-fact recovery (0-1; share of a user's hidden owner-fact atoms, 840 in all, whose content a predicted fact matches; PAL-TRACE framework with the listed LLM as reconstruction backbone; 50 synthetic users, macro-averaged over users; Qwen3.6-35B-A3B semantic judge). Source: arxiv.org. Saturation forecast: Around June 2027. 9 models tracked.

Top models

#ModelScore
1Gemini 3.1 Pro (Preview)0.61
2GPT-5.40.61
3Claude Sonnet 4.50.6
4Gemini 3.1 Flash Lite0.6
5DeepSeek V4 Flash0.58
6GPT-5.4 Mini0.56
7Qwen 3.6 35B A3B0.54
8Gemma 4 26B A4B (IT)0.5

Interactive version: theaggregate.ai/benchmark?slug=pal-bench-pal-trace-owner-fact-recovery · How It Works · Data refreshed daily, snapshot 2026-09-26.