Graph world model lifts Qwen3-8B robot planning success past 90%

single source· 1 articles · confidence: low · first seen 2026-09-15 20:00 UTC

What this means for you

If you are wiring a language model into a robot or an agent that acts in the world, the checking layer around the model may matter more than model size: these gains come from a small 8B model wrapped in this graph. Nothing to install — the paper names no code, weights or licence, and the scores are the authors' own.

GAVEL puts a graph world model between a language model and a robot. The graph stores object relations, each action's preconditions and effects, and probability estimates over unobserved object locations. It predicts what an action does before execution, repairs failures that follow from the graph, and sends errors needing semantic judgement back to the model. On the BEHAVIOR-1K benchmark, the authors report Qwen3-8B plan success rising from 41.2% to 91.8% on 100 single long-horizon tasks and from 19.9% to 92.6% on 500 multi-task instructions. Reordering subtasks by probability cut travel distance about 5.4% against a static version. The preprint gives no evaluation date and no one else has replicated it.

Key facts

  • ·The work is arXiv preprint 2609.19315, posted 15 September 2026, and is not peer reviewed. source
  • ·With Qwen3-8B, single long-horizon task success on BEHAVIOR-1K rose from 41.2% to 91.8% across 100 tasks. source
  • ·Multi-task success on the same benchmark rose from 19.9% to 92.6% across 500 multi-task instructions. source
  • ·Reasoning over distributions of possible object locations reduced travel distance by roughly 5.4% against a static variant. source
  • ·The evaluation is on BEHAVIOR-1K; the paper gives no evaluation date for the reported scores. source

What the sources say

Sources

The original reporting. Follow these — they did the work.

← the wire