LIBERO-PeRM Fine-tuned - User Profile SA: leaderboard
LIBERO-PeRM v1 Table3; three distinct pi0.5 preference fine-tunes, ten evaluation runs at fixed checkpoints (not ten training seeds); LoRA preference-only, Mix-LoRA joint LIBERO/preferences, Seq-LoRA LIBERO then preferences. All-Pref jointly trained Motion/Skill/User Profile checkpoint. SA success-gated; ComprePref uses contextual rewrites, no invented independent tasks. Metric: Success-gated preference satisfaction (%). Higher is better. Source: https://arxiv.org/html/2610.22372v1
Metric: Success-gated preference satisfaction (%). Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 3 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | LIBERO-PeRM pi0.5 Mix-LoRA | 91.33 | |
| 2 | LIBERO-PeRM pi0.5 LoRA | 85.22 | |
| 3 | LIBERO-PeRM pi0.5 Seq-LoRA | 81 |
Interactive version: theaggregate.ai/benchmark?slug=libero-perm-fine-tuned-user-profile-sa · How It Works · Data refreshed daily, snapshot 2026-10-17.