AI Research Pulse

Long-horizon multimodal mem/world models

Long-horizon multimodal mem/world models

Key Questions

What compression level does STAR-KV achieve for KV cache?

STAR-KV delivers up to 20x KV cache compression and 6.9x attention speedup via low-rank techniques plus quantization, with open-sourced code targeting vLLM.

How does HOLA address memory overwriting in linear attention?

HOLA pairs a compressive recurrent state with a small exact cache, delivering strong perplexity and needle recall improvements for long-context models.

What does TAP enable for vision-language-action models?

TAP decouples motor priors from semantic grounding during task-agnostic pretraining, yielding strong results on SIMPLER benchmarks while reducing data costs.

What improvement does Perceive-to-Reason provide for visual reasoning?

It decouples perception and reasoning stages, combined with a PRA-GRPO RL strategy, to achieve strong results on high-resolution fine-grained benchmarks.

What memory benefit does ARROW demonstrate over DreamerV3?

ARROW reduces forgetting by 20x compared to DreamerV3 in long-horizon multimodal world model settings.

Multimodal advances: NVIDIA bundles; WLA 2B 40ms; Flash-WAM 23x; SpatialWorld GPT-5 17.4% TSR; SepsisAgent outperforms clinicians. New: V-JEPA 2 latent dynamics; Mem-π generative memory; Geometric Action Model 6.9ms; HUG 34% improvement; Meta/Princeton 3D without specialized architectures; ARROW 20x forgetting reduction vs DreamerV3; LoopWM 100x parameter efficiency; RNG-Bench memory gap; ImageWAM questions full video generation need; SkyJEPA zero-shot sim-to-real; Safe autoregressive image generation paper. Perceive-to-Reason (P2R) decouples perception and reasoning for fine-grained visual reasoning; PRA-GRPO RL strategy. Strong results on high-resolution benchmarks. Practical for improving VLMs. HOLA (Hybrid Overlapping Linear Attention) pairs compressive recurrent state with small exact cache to solve memory overwriting in linear attention; achieves strong perplexity and needle recall gains. Practical for long-context agent memory and efficient transformers. TAP (Task-Agnostic Pretraining for VLAs) — decouples motor priors from semantic grounding; strong SIMPLER results. Practical for reducing data costs in embodied AI. STAR-KV achieves up to 20x KV cache compression via low-rank + quantization, 6.9x attention speedup; ICML 2026 Spotlight, open-sourced, targeting vLLM integration. Practical for production inference with long-context agents. New today: EVA-Client framework for embodied policy deployment on real robots; decoupled architecture with Debug/Collect/Eval workflows.

Sources (5)
Updated Jul 7, 2026