AI
LOCI: Spatial Linear Memory for Streaming World Models
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 30/09/2026, 13:27:06 UTC
Briefing created: 03/10/2026, 04:49:35 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 02/10/2026, 11:18:24 UTC.
Briefings
LOCI is a hybrid spatial-memory architecture designed for streaming world models, combining key-value caches with recurrent linear-attention memory conditioned on camera geometry. It outperforms representative world models and full-softmax baselines on the MIND benchmark, maintaining higher fidelity when revisiting content. The system reduces peak memory usage by approximately 30% compared to full-softmax models while preserving accuracy under bounded observation banks.
Key points
- LOCI uses a hybrid architecture where half of transformer blocks maintain key-value caches of past observations, while the other half uses recurrent linear-attention memory conditioned on projective camera geometry.
- On the MIND memory benchmark, LOCI reproduces revisited content more faithfully than representative world models and a full-softmax model, with full history reducing peak memory by about 30% relative to full softmax.
Context and limitations
The study compares LOCI against representative world models and full-softmax baselines on the MIND benchmark and held-out recorded trajectories. Results are limited to these specific datasets and do not generalize to other video lengths or camera types.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 03/10/2026, 04:49:35 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeifxw2y7wyu75s7dhajjyodqfhv6aktwmyb2yfqpwzqkug622hfsy4