What You Think is What You See: Driving Exploration in VLM Agents via Visual-Linguistic Curiosity
TLDR
GLANCE uses discrepancy between linguistic world model predictions and visual observations as intrinsic curiosity to drive exploration in VLM agents.
Reasoning
Strengths include a novel integration of curiosity-driven exploration with VLM world models to address sparse-reward tasks. Weaknesses: the approach is limited to VLM agents and may require further validation across diverse environments.
Read-first score
Read-first score 49, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 39.
Field roles
Rank sensitivity
Stability: volatile; rank range: 268.
Keyword Scores
Deep Analysis
Innovations
- GLANCE framework that unifies reasoning and exploration by grounding the agent's linguistic world model into stable visual representations of an evolving target network
- Using the discrepancy between linguistic prediction and visual reality as an intrinsic curiosity signal within reinforcement learning to drive exploration
- Aligning 'what the agent thinks' with 'what the agent sees' as a key principle for solving complex or sparse agentic tasks
Methodology
GLANCE employs a VLM agent with a linguistic world model (via CoT reasoning) and an evolving target network that provides stable visual representations. The discrepancy between the agent's linguistic predictions and the actual visual observations is used as an intrinsic curiosity reward in a reinforcement learning framework, steering the agent to explore areas where its internal model is uncertain. The method is evaluated on a series of agentic tasks with sparse rewards.
Key Results
GLANCE effectively drives exploration in sparse-reward tasks, demonstrating that aligning linguistic predictions with visual observations is key to solving complex agentic tasks.