Pinductor uses LLM priors to learn POMDP world models with 330x fewer episodes than DreamerV3
burny_tech · x · 2026-09-29
Accepted at NeurIPS 2026, 'Learning POMDP World Models from Observations with Language-Model Priors' (arXiv 2605.13740) — by Bernhard Schölkopf, Philipp Hennig, Lancelot Da Costa et al. — introduces Pinductor: an LLM proposes candidate POMDP world models from a few observation-action trajectories and iteratively refines them against a belief-based likelihood score.
- 330x fewer online episodes than DreamerV3 (3 vs 1,000) with comparable performance and slightly lower compute.
- Despite strictly less information, it matches LLM-based POMDP methods that assume privileged access to hidden states and far surpasses tabular baselines in sample efficiency.
- Performance scales with LLM capability and degrades gracefully when environmental semantics are withheld, confirming the prior comes from the LM's knowledge.
- Positioned as a practical tool for sample-efficient world-model learning under partial observability.
More from Embodied
- Standard Bots Launches No-Code Robot OS, Says Only 6% of US Manufacturers Use Robots — LexiLove · 2026-09-29
- Agility Robotics' Digit bears marks from real work outside the lab — jonstephens85 · 2026-09-29
- Wuji Dexterous Hand at IROS Packs 20 Active DoF to Mimic Human Hand — jonstephens85 · 2026-09-29
- Matt Shumer says Opus 5.5 cleared a 3D printer bed after 16 tries, taking humans out of the loop — mattshumer_ · 2026-09-29
- Driver swaps FSD subscription for open-source sunnypilot on comma four, 8 months in — SuB8u · 2026-09-29
- RoboFoundry's System-as-Policy self-evolution lifts GPT-5.5 by 27.8% on EmbodiedBench — Jingsong Liang · 2026-09-29