Does VLA Even Know the Basics? Measuring Commonsense and World Knowledge Retention in Vision-Language-Action Models
arXiv · HuggingFace · ▲71
Abstract (verbatim)
Embodied Vision-Language-Action (VLA) models are typically obtained by fine-tuning powerful pretrained VLMs on robotics data, yet it is unclear how much commonsense and factual knowledge they retain after adaptation. Failures on knowledge-sensitive tasks are ambiguous, conflating missing knowledge with poor generalization of low-level control. We introduce Act2Answer, a lightweight protocol that adapts VLM knowledge benchmarks to VLA evaluation by requiring agents to answer through action. Each question becomes a short tabletop episode where the agent performs a single object-placement action to select among candidate answers, yielding an action-grounded success rate with reduced control confounds. We curate a test suite of such environments across diverse commonsense and world-knowledge categories and introduce layerwise intent probing to localize answer-relevant information across the VLM backbone and action head. In a large-scale study of 7 VLA models and 9 VLM baselines, we systematically rank models across categories, finding that VLAs show solid performance on simple concepts while exhibiting larger gaps on richer semantic categories relative to their source VLMs, that VQA co-training is associated with better knowledge retention, and that answer-relevant signals peak in middle VLA layers but attenuate in upper layers. Act2Answer is available at https://tttonyalpha.github.io/act2answer/.
Background
Background Analysis
1. Technical Context
With the advancement of AI, embodied agents (e.g., home service robots, automated systems in retail) are being developed for everyday environments. These agents require a rich understanding of commonsense and factual knowledge (e.g., object usage, behavioral appropriateness) to make reasonable decisions in complex scenarios. Vision-Language-Action (VLA) models are proposed as the foundation for such agents, aiming to integrate visual perception, language understanding, and action execution for open-world interaction. However, while VLAs excel in manipulation tasks, it remains unclear whether they retain basic commonsense and world knowledge after robotics fine-tuning.
2. Previous Limitations
Existing research focuses on task success rates (e.g., object manipulation or navigation) but overlooks a critical question: Can VLAs still reason based on commonsense after training? For example, a fine-tuned VLA might move a cup successfully but fail to judge that "cups are for holding water, not throwing." Additionally, traditional evaluations rely on task success metrics, making it hard to distinguish between "knowledge gaps" and "control deficiencies." Moreover, while VLM (Vision-Language Model) benchmarks are well-developed, they are text-based and unsuitable for action-oriented embodied agents.
3. Proposed Solution
To address this, the paper introduces Act2Answer, a framework that adapts VLM knowledge benchmarks into action-based evaluations for VLAs. Each knowledge question is transformed into a simulated scenario where the agent selects an answer through a single action (e.g., placing an object). This reduces confounding factors from long-horizon planning and control, focusing evaluation on knowledge retention. The authors also introduce layered intent probing, using linear classifiers to analyze knowledge distribution across model layers, revealing how knowledge is retained or lost during fine-tuning.
4. Unique Approach
Compared to prior work, Act2Answer’s key innovations are:
- Action-oriented evaluation: Transforms text QA into actionable tasks, aligning with embodied agents’ needs.
- Diverse knowledge coverage: Evaluates VLAs across multiple categories (e.g., attributes, states, emotions).
- Layerwise analysis: Identifies where knowledge is retained or forgotten within the model.
This approach fills gaps in existing research, providing a principled way to assess VLA knowledge retention.

