Search papers, labs, and topics across Lattice.
LOCI addresses the challenge of video world models remembering and retrieving past observations by introducing a hybrid spatial-memory architecture that combines a key-value cache with a recurrent linear-attention memory. This architecture leverages projective camera geometry to condition memory addressing and content, enabling accurate reproduction of revisited scenes. LOCI demonstrates superior performance on memory benchmarks and streamed long videos with constant memory, outperforming full-softmax models.
Reconstructing revisited scenes in streaming world models is now possible with constant memory, outperforming full-softmax attention by 30% on peak memory.
When a camera revisits a previously observed region, a video world model should reproduce what was there before. This requires both remembering past observations and retrieving the right one for the current viewpoint. Key-value caches preserve visual detail but grow with video length; recurrent memory is compact but compresses history into a fixed-size state, so individual past observations are no longer directly accessible. We introduce LOCI, a hybrid spatial-memory architecture that keeps both representations. In half of the transformer blocks, main attention keeps a key-value cache of past observations; in the other half, it is restricted to the current chunk and complemented by a recurrent linear-attention memory whose reads and writes are conditioned on projective camera geometry, so viewpoint enters both memory addressing and stored content. Recurrent readouts flow into subsequent cache-backed blocks and supply their queries with accumulated scene context. On the public MIND memory benchmark and on held-out recorded trajectories, LOCI reproduces revisited content more faithfully than representative world models and a same-recipe full-softmax model; with full history, it lowers peak memory at equal length by about 30% relative to full softmax. With a bounded bank of retained observations, it streams long videos at constant memory and remains more faithful than full softmax under the same budget.
