CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines
TLDR
Proposes Contrastive World Model (CWM) using contrastive learning with hard negatives to train LLM as action feasibility scorer, outperforming SFT on ScienceWorld.
Reasoning
Strengths include a novel contrastive approach for action feasibility with hard negative mining, showing clear improvements over SFT on a benchmark. Weaknesses are limited evaluation to a single simulated benchmark (ScienceWorld) and lack of real-world robot validation, raising questions about generalizability.
Read-first score
Read-first score 47.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 24.
Field roles
Rank sensitivity
Stability: volatile; rank range: 502.
Keyword Scores
Deep Analysis
Innovations
- Contrastive World Model (CWM) for action feasibility learning in embodied agents
- Using InfoNCE contrastive objective with hard-mined negative examples to discriminate physically correct from subtly wrong actions
- Fine-tuning a large language model (LLM) as an action scorer with contrastive learning instead of supervised fine-tuning (SFT)
Methodology
CWM fine-tunes a large language model (LLM) as an action scorer using an InfoNCE contrastive objective with hard-mined negative examples. It is evaluated on the ScienceWorld benchmark through two studies: an intrinsic affordance evaluation on 605 hard-negative test pairs and a live filter characterisation study measuring ranking of gold-path actions against all valid environment actions during task execution.
Key Results
CWM outperforms SFT by +6.76 percentage points on Precision@1 for minimal-edit negatives and achieves a higher AUC-ROC (0.929 vs. 0.906). Under out-of-distribution stress conditions, CWM maintains a significantly better safety margin (-2.39) than SFT (-3.96).