Track the frontier, from condensed matter to cosmology.
Search papers, labs, and topics across Lattice.
2
0
3
Treating all latent tokens identically bottlenecks multimodal test-time compute—disentangling hidden-state updates into visual-grounding and high-entropy reasoning channels beats standard output-space search under matched sample budgets.
Dynamo achieves a remarkable 5.6% average accuracy boost in visual reasoning tasks without retraining, revolutionizing how VLMs adapt in real-time.