PAL-Bench (PAL-TRACE) - Evidence Support: leaderboard
PAL-Bench (Yan et al., 2026) asks a system to reconstruct a person's profile from a synthetic longitudinal photo album seen only through perception outputs (captions, OCR, entities, face clusters, timestamps and coarse metadata), over 50 users and 2,799 hidden targets. Runs use the authors' PAL-TRACE framework with the listed LLM as backbone and a fixed Qwen3.6-35B-A3B judge; this board is evidence support, the judge's 0, 0.5 or 1 rating of how well the cited photos support each matched claim, averaged from 0 to 1, higher is better.
Metric: Evidence support (0-1; mean 0, 0.5 or 1 support score of the cited public photos for each matched claim; 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 October 2027. 9 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Gemini 3.1 Pro (Preview) | 0.42 | #91 |
| 2 | GPT-5.4 | 0.38 | #121 |
| 3 | DeepSeek V4 Flash | 0.36 | #254 |
| 4 | Claude Sonnet 4.5 | 0.36 | #244 |
| 5 | Gemini 3.1 Flash Lite | 0.33 | #363 |
| 6 | GPT-5.4 Mini | 0.32 | #389 |
| 7 | Qwen 3.6 35B A3B | 0.32 | #461 |
| 8 | Gemma 4 26B A4B (IT) | 0.29 | #545 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=pal-bench-pal-trace-evidence-support · How It Works · Data refreshed daily, snapshot 2026-11-12.