Research map / Video / Video-language understanding

Addressable Memory for Video World Models

NeurIPS 2026 · Video › Video-language understanding · approach: Long video understanding · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Training-free WorldTrace memory keeps compressed KV-cache summaries addressable via slot-rank virtual RoPE positions and canonical-key averaging (Field) or frozen landmark keys (Landmark) for long-horizon visual persistence in autoregressive video world models.

Problem

Autoregressive video world models lose visual persistence once rollouts exceed the training context: RoPE offsets go out of distribution, so compressed KV-cache summaries become unreadable, and naive RoPE-space averaging cancels signal.

Approach

Assign each compressed summary slot a fixed in-distribution virtual position by slot rank, store keys in canonical (unrotated) space, and use two content writers: uniform canonical key averaging for coherence (Field) and frozen verbatim scene-entry landmark keys for episodic recall (Landmark).

Key results

• WorldTrace-Field improves TempSSIM by +15.5% relative at N=48 over the sliding-window baseline (0.545 vs 0.472) with lowest scene drift there.
• WorldTrace-Landmark improves PAC on LoopBench ABA by +19.5% (reported as scene consistency gain) with PAC 0.864 vs 0.723 at N=16.
• At N=256, verbatim landmark recall sustains PAC ~0.99 versus 0.610 for canonical-K anchoring.

Surprising findings

• Naive RoPE-space averaging scores marginally higher on PAC at N=16 (0.570 vs 0.540) than the sliding window, but canonical averaging is needed for coherence.
• Field-style averaging under Block-relative positions reduces to plain sliding window behavior because out-of-range offsets receive zero attention weight.

Limitations

• Designed only for temporal-RoPE autoregressive models with a fixed KV-cache budget.
• Field averaging can blur specific scene details; Landmark depends on detecting the right scene-entry events.
• Evaluated mainly on Matrix-Game-2 1.3B with a single LoopBench benchmark introduced by the authors.

Experimental setup

Benchmarks: LoopBench TempSSIM evaluation on MG2 rollouts

Metrics: PAC (Position-Aligned CLIP) TempSSIM Local Scene Drift LatentDiff

Baselines: Sliding window Block-relative RoPE Centroid-linear Naive averaging MemRoPE YaRN Landmark + Block-relative

Builds on: Self-Forcing Landmark Attention YaRN Block-relative RoPE Infinity-RoPE

Closest papers by meaning

• Video World Models with Long-term Spatial Memory (NeurIPS 2025 · similarity 0.90)
• VideoMLA: Low-Rank Latent KV Cache for Minute-Scale Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.87)
• Learning World Models for Interactive Video Generation (NeurIPS 2025 · similarity 0.87)
• WorldMem: Long-term Consistent World Simulation with Memory (NeurIPS 2025 · similarity 0.87)
• Can 4D Foundation Models Remember? (NeurIPS 2026 · similarity 0.87)
• Attend Locally, Remember Linearly: Linear Attention as Cross-Frame Memory for Autoregressive Video Diffusion (NeurIPS 2026 · similarity 0.86)