DIRECT Framework Optimizes Test-Time Compute Allocation for Vision-Language Model Planners in Robotics
Researchers introduced DIRECT, a routing framework that intelligently allocates computational resources at test-time for vision-language models used in robotic planning, rather than uniformly scaling compute across all tasks. The framework uses multimodal scene context to decide when and where to spend additional computation across three scaling axes: chain-of-thought depth, model size, and memory history. This approach achieves comparable performance to larger models while reducing latency by up to 65%, making frontier-level embodied AI more practical for real-world deployment.
A new research paper on arXiv presents DIRECT, a framework addressing a key challenge in deploying vision-language models (VLMs) as planners for embodied agents: the inefficiency of uniformly scaling test-time compute. The authors observe that while scaling compute improves capability, it increases latency, token usage, and computational cost with uneven and often diminishing returns. DIRECT uses multimodal scene context to route computation dynamically per prompt, optimizing the success-cost tradeoff. Experiments across VLABench and RoboMME benchmarks reveal that different scaling axes—chain-of-thought depth, model size, and memory history—yield qualitatively distinct capability gains, meaning compute allocation should be selective. Validation on a physical Franka robotic arm demonstrated that DIRECT matches or exceeds stronger models' success rates at up to 65% lower average latency, suggesting that intelligent routing rather than naive scaling is essential for practical robotic deployment.
What's missing
The paper does not discuss potential failure modes of the routing mechanism itself, such as scenarios where the multimodal scene context misclassifies task difficulty, or how the framework generalizes to novel task types not represented in training data. Additionally, the computational cost of running the router itself is not quantified.
What different sources said
- arXiv cs.AICenter
DIRECT: When and Where Should You Allocate Test-Time Compute in Embodied Planners?
Related
New AI Model Improves Prediction of Therapeutic Peptide Function from Protein Sequences
Researchers developed a lightweight CNN classifier that predicts whether peptide sequences have therapeutic properties, trained on a database of 54,655 peptides across 48 functional categories. The model uses a novel negative sampling strategy to reduce false positive rates from over 60% in previous approaches to 2.1%. This advancement could accelerate drug discovery by enabling faster computational screening of peptide candidates before expensive experimental testing.
Study Shows Different Metabolic Stress Models Produce Distinct Effects on Human Neuronal Networks
Researchers tested three common in vitro metabolic stress models on human-derived neuronal networks and found each produced different patterns of neuronal activity and cell damage. The models tested were hypoxia alone, oxygen-glucose deprivation (OGD), and hypoxia combined with glutamate exposure. The findings suggest that choice of experimental model significantly affects results and that combining electrophysiological and structural analyses is important for accurately assessing metabolic stress in stroke research.
New Semi-Supervised Method Improves Single-Cell RNA Sequencing Integration Using Virtual Adversarial Training
Researchers introduced scCRAFT+, a semi-supervised integration method for single-cell RNA sequencing that uses Virtual Adversarial Training to incorporate marker gene information. The method addresses a key limitation of existing approaches: over-mixing of closely related cell subtypes during data integration. The advancement could improve cell type identification and biological interpretation in genomics research.