CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines
#Contrastive World Models #Action Feasibility #Embodied Agents #Supervised Fine-tuning #Language Models #AI Safety
📌 Key Takeaways
- CWM represents a breakthrough in action feasibility scoring for embodied AI agents
- The contrastive training approach effectively distinguishes between physically correct and subtly incorrect actions
- CWM outperforms traditional supervised fine-tuning by +6.76 percentage points on precision metrics
- The model maintains better safety margins under stress conditions, improving real-world performance
📖 Full Retelling
Chayan Banerjee introduced CWM (Contrastive World Models) for action feasibility learning in embodied agent pipelines in a paper submitted to arXiv on February 25, 2026, addressing critical limitations in existing supervised fine-tuning approaches that fail to properly discriminate between physically correct and subtly incorrect actions. The paper presents a novel approach to training action feasibility scorers, which represent a critical bottleneck in embodied AI systems. Current methods use supervised fine-tuning (SFT) that treats candidate actions independently without explicitly teaching models to distinguish between physically valid and invalid actions. CWM improves upon this by fine-tuning large language models using an InfoNCE contrastive objective with hard-mined negative examples. The core innovation lies in pushing valid actions away from invalid ones in scoring space, with special emphasis on hard negatives—semantically similar but physically incompatible candidates. Banerjee evaluated CWM on the ScienceWorld benchmark through two comprehensive studies. The first intrinsic affordance assessment tested 605 hard-negative pairs and demonstrated that CWM outperforms traditional SFT by +6.76 percentage points on Precision@1 for minimal-edit negatives, where a single word change alters physical outcomes. CWM also achieved a higher AUC-ROC score (0.929 vs. 0.906). In a second live filter characterization study, CWM maintained a significantly better safety margin (-2.39) compared to SFT (-3.96) under out-of-distribution stress conditions, indicating superior ranking of correct actions during task execution.
🏷️ Themes
Artificial Intelligence, Machine Learning, Robotics
Entity Intersection Graph
No entity connections available yet for this article.
Original Source
--> Computer Science > Artificial Intelligence arXiv:2602.22452 [Submitted on 25 Feb 2026] Title: CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines Authors: Chayan Banerjee View a PDF of the paper titled CWM: Contrastive World Models for Action Feasibility Learning in Embodied Agent Pipelines, by Chayan Banerjee View PDF HTML Abstract: A reliable action feasibility scorer is a critical bottleneck in embodied agent pipelines: before any planning or reasoning occurs, the agent must identify which candidate actions are physically executable in the current state. Existing approaches use supervised fine-tuning to train action scorers, but SFT treats each candidate independently and does not explicitly teach the model to discriminate between actions that are physically correct and those that are subtly wrong. We propose the Contrastive World Model , which fine-tunes a large language model as an action scorer using an InfoNCE contrastive objective with hard-mined negative examples. The key idea is to push valid actions away from invalid ones in scoring space, with special emphasis on hard negatives: semantically similar but physically incompatible candidates. We evaluate CWM on the ScienceWorld benchmark through two studies. First, an intrinsic affordance evaluation on 605 hard-negative test pairs shows that CWM outperforms SFT by +6.76 percentage points on Precision@1 for minimal-edit negatives -- cases where a single word changes the physical outcome -- and achieves a higher AUC-ROC (0.929 vs. 0.906). Second, a live filter characterisation study measures how well CWM ranks gold-path actions against all valid environment actions during task execution. Under out-of-distribution stress conditions, CWM maintains a significantly better safety margin (-2.39) than SFT (-3.96), indicating that the gold action is ranked closer to the top. These results support the hypothesis that contrastive training induces representations that capture phy...
Read full article at source