Back to feed
AR
arXiv CS.AI
7/9/2026
Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix

Grounding Spatial Relations in a Compact World Model: Instruction Leakage and a Goal-Free Dynamics Fix

Short summary

Researchers identify an instruction leakage trap in goal-conditioned compact world models: a predictor achieves 0.90 relation-readout accuracy by transcribing the instruction rather than perceiving the scene. Withholding the goal collapses accuracy to chance (0.27), and counterfactual instructions make the model follow false instructions 94.5% of the time. The fix is to keep the goal out of the dynamics model and supervise the read path, recovering instruction-independent grounding at 0.88.

  • •Goal-conditioned world models can fake spatial grounding by transcribing instructions
  • •Withholding the goal collapses accuracy from 0.90 to 0.27, proving instruction leakage
  • •Fix: keep the goal in the planner's cost, not the dynamics model, to recover genuine grounding

Generated with AI, which can make mistakes.

#research-breakthrough
Read full article at arXiv CS.AI

Is this a good recommendation for you?

Comments

Failed to load comments. Please try again.

Explore more

Grounded Iterative Language Planning: How Parameterized World Models Reduce Hallucination Propagation in LLM Agents
arXiv CS.AI
Task Competence Is Not Instruction Following: Evaluating Instruction-Conflicting Behavior in Small Language Models
arXiv cs.CL
Internalizing the Future: A Unified Agentic Training Paradigm for World Model Planning
arXiv CS.AI
Interventional Grounding Audits: Black-Box Premise-Dependency Tests for LLM Chain-of-Thought via Predicate Substitution
arXiv CS.AI