AI

QuantWM: Temporally Consistent 2-Bit KV Cache Quantization for Video World Models

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 05/10/2026, 17:18:24 UTC.

Briefings

QuantWM is a training-free 2-bit KV cache quantization framework designed for video world models to mitigate temporal flickering and visual degradation caused by attention shifts. The study identifies that Key quantization reduces reconstruction errors but increases output degradation, prompting the development of QSAC and PSAC techniques. Experiments on LingBot-World-v2, HY-World 1.5, Matrix-Game-2, Longcat-Video, and Causal-Forcing show QuantWM significantly improves visual quality and temporal consistency, achieving up to 6.20 KV cache memory compression with minimal overhead.

Key points

• QuantWM introduces two complementary techniques: quantization-sensitivity-aware clustering (QSAC) and principal-subspace attention compensation (PSAC) to stabilize attention logits during 2-bit KV cache quantization.
• Experiments on LingBot-World-v2, HY-World 1.5, Matrix-Game-2, Longcat-Video, and Causal-Forcing demonstrate that QuantWM significantly improves visual quality and temporal consistency compared to existing methods.

Context and limitations

Results are limited to five specific benchmarks; performance on other video world models remains unverified. The study does not report quantitative metrics for temporal flickering or visual degradation reduction, only qualitative observations from experiments.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 09/10/2026, 15:39:08 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeihjofvfj3oxzke5ng2f2pcrzyyjui6id2j5fbki2lqgk63tumqexe