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

arXiv 2026 33.1 method

TLDR

QuantWM introduces a training-free 2-bit KV cache quantization framework for video world models, preserving attention logits and temporal-spatial token selection to avoid temporal flickering.

Reasoning

The paper provides a clear, motivated analysis of why key quantization harms video world models despite lower reconstruction error, and proposes two complementary training-free techniques. However, the abstract is truncated and does not show quantitative results, making empirical validation hard to assess from the visible text.

Read-first score

Read-first score 33.1, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 30.

Recency 5%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Methodology quality 15%

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=analysis,benchmark,experiment,metric

Topical relevance 25%

42.9

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Reproducibility 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 10%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 341.

Keyword Scores

video world model

world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Training-free 2-bit KV cache quantization framework for video world models (QuantWM) that preserves attention logits and temporal-spatial token selection.
Quantization-sensitivity-aware clustering (QSAC), which jointly considers historical Query sensitivity and residual ranges to select INT2-friendly Key centroids and reduce quantization errors in attention-critical channels.
Principal-subspace attention compensation (PSAC), which restores remaining Key errors along the dominant Query subspace using low-rank projections to stabilize attention logits.
Analysis finding that Key quantization produces smaller reconstruction errors than Value but causes larger output degradation by changing attention logits and shifting temporal-spatial token selection in video world models.

Methodology

QuantWM is a training-free 2-bit KV cache quantization framework for video world models. It combines quantization-sensitivity-aware clustering to select INT2-friendly Key centroids based on historical Query sensitivity and residual ranges, and principal-subspace attention compensation to correct remaining Key errors along the dominant Query subspace via low-rank projections. The approach is evaluated on LingBot-World-v2, HY-World 1.5, Matrix-Game-2, Longcat-Video, and Causal-Forcing, with comparisons to existing 2-bit KV cache quantization methods using visual quality and temporal consistency metrics.

Key Results

Existing 2-bit KV cache quantization methods cause severe temporal flickering and visual degradation on video world models despite near-lossless VBench performance. QuantWM significantly improves visual quality and temporal consistency and outperforms existing methods across benchmarks with up to 6.20 KV cache memory compression and limited additional overhead.

Resources

Paper