PackForcing: Unified Memory and Force Control

Updated 2 July 2026

PackForcing is a framework that unifies forced packing techniques in deep generative modeling and granular physics, enabling efficient long-range memory and controlled microstructure.
It utilizes a hierarchical key-value cache and spatiotemporal compression to maintain constant context size during autoregressive video synthesis even under strict resource constraints.
PackForcing protocols dynamically select and modulate historical data or particle states to optimize simulation fidelity, memory encoding, and elastic response in diverse applications.

PackForcing is a unifying term for several independently developed frameworks in deep generative modeling and granular physics that employ "forced" packing or context management of large-scale, temporally extended systems. Across these domains, PackForcing strategies facilitate efficient long-range memory or robust control of microstructure by dynamically selecting, compressing, or modulating the representation or physical state of historical data, particles, or context. This article reviews PackForcing as established in (1) autoregressive video diffusion and memory architectures, (2) granular packing simulation and dynamic forcing, and (3) force network ensembles, providing technical summaries, key mathematical frameworks, and protocol-level distinctions.

1. PackForcing in Autoregressive Video Diffusion

PackForcing in generative video models is a hierarchical memory management scheme enabling minute-scale autoregressive synthesis under strict resource constraints. The core innovation is a three-partition key-value (KV) cache design that overcomes linear memory growth and error accumulation in long-horizon rollouts by enforcing a fixed attention-context size independent of generated length (Mao et al., 26 Mar 2026).

Three-Partition KV-Cache

Generated video history is subdivided as:

Sink tokens: The initial $N_{\rm sink}$ blocks are retained at full spatiotemporal resolution, anchoring global semantics.
Compressed mid tokens: Intermediate blocks ( $N_{\rm sink}+1, ..., L-N_{\rm recent}$ ) are mapped to a high-compression representation (32× token reduction), fusing dual-branch 3D convolutions with variational autoencoder (VAE) re-encoding.
Recent tokens: The last $N_{\rm recent}$ blocks plus the current block remain at full resolution to preserve local continuity.

The context for each denoising step is $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$, yielding constant attention footprint.

Spatiotemporal Compression

Dual-branch compression combines:

HR branch: Strided 3D convolution stack (temporal and spatial), achieving $2 \times 8 \times 8=128$ -fold grid reduction, followed by $1\!\times\!1\!\times\!1$ projection.
LR branch: Block decoded, average-pooled, VAE-encoded, and patch-embedded to obtain complementary low-resolution tokens.

Blocks are fused as $\tilde h = h_{\rm HR}(\mathbf z) + h_{\rm LR}(\mathbf z)$ ; with parameters $B_f$ , $h$ , $w$ , and resulting $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 0 tokens per block.

Dynamic Top- $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 1 Context Selection

To bound KV-cache further, only the top $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 2 compressed blocks ranked by query–key affinity are attended during inference. This yields qualitative and quantitative gains (+0.8 in subject consistency, +0.12 CLIP) with negligible compute cost.

Continuous Temporal RoPE Adjustment

As mid tokens are evicted, temporal rotary positional embeddings are adjusted via a multiplicative phase factor, guaranteeing seamless context alignment with $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 3 FLOP overhead.

Performance

PackForcing enables 120-second, 832×480 video synthesis at 16 FPS, using only $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 44 GB total KV-cache on a single H200 GPU. The architecture supports effective zero-shot extrapolation (5 s $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 5 120 s) because the context size is strictly constant at train and test time. On VBench, PackForcing achieves state-of-the-art temporal consistency (26.07) and dynamic degree (56.25), with minimal CLIP drop over 60 seconds (Mao et al., 26 Mar 2026).

2. PackForcing-Style History Encoders in Bidirectional Video World Models

A PackForcing-style history encoder is implemented in bidirectional autoregressive video world models such as BiWM. Here, past chunks of latent frames are compressed into a fixed-size prefix memory bank, allowing efficient context access without unbounded growth in memory tokens (Rui et al., 8 Jun 2026).

Two-Branch Memory Bank

HR branch: Eight-stage causal 3D convolutional stack with temporal and spatial downsampling; optional masked self-attention; output flattened to $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 6 tokens.
LR branch: Patch-embedding of the entire history, trilinear resized to HR grid, flattened to $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 7 tokens.
Both outputs are summed: $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 8, forming the context prefix for each new chunk.

Unlike causal "Self-Forcing," which freezes past representations, PackForcing continually recomputes and refines memory, learning to optimally summarize histories during multi-objective distillation.

Comparison to Multi-Scale Pyramids

FramePack-style pyramids assign downsampling schemes explicitly by age; PackForcing uses a learned encoder to fuse local and global information, holds prefix size constant regardless of history length, and preserves long-range cues more robustly under high compression (Rui et al., 8 Jun 2026).

3. PackForcing in Bulk Granular Packing Simulations

In granular physics, PackForcing denotes DEM protocols that regulate the evolution of packing density ( $N_{\rm sink}+1, ..., L-N_{\rm recent}$ 9), coordination number ( $N_{\rm recent}$ 0), and sliding-contact fraction ( $N_{\rm recent}$ 1) by combining controlled stress, global barostat drag, and cycling procedures (Santos et al., 2023).

DEM Framework

Particles interact via Hookean spring–dashpot–Coulomb-slider contacts. Drag is modulated both at the particle and cell levels. PackForcing protocols include:

Volume-controlled over-compression (Method V)
Stress-controlled under-compression, over-compression-and-release, cyclic compression (“tapping”), and progressive steps.

Protocol Dynamics

Key findings include:

Non-monotonic $N_{\rm recent}$ 2 under low barostat drag and nonzero initial kinetic energy; minimized packing fraction at intermediate pressure is associated with maximal $N_{\rm recent}$ 3.
Drag-induced monotonicity: Increasing global cell drag ( $N_{\rm recent}$ 4) smooths $N_{\rm recent}$ 5, eliminating non-monotonicity.
Cyclic compression: Densification with cycle number $N_{\rm recent}$ 6 follows a stretched exponential (KWW law), with fit parameters depending on pressure regime.

Selection of $N_{\rm recent}$ 7, $N_{\rm recent}$ 8, and protocol path allows precise microstructure control, offering a systematic approach to designing dense or loose packings (Santos et al., 2023).

4. PackForcing Protocols for Dynamic Elastic Response

PackForcing also refers to slab-scale simulation protocols involving static confining pressure and external dynamic (sinusoidal) forcing. The resulting nonlinear resonant response provides insight into elastic softening in jammed media (Reichhardt et al., 2014).

Simulation Protocol

A two-dimensional bidisperse slab (~700 disks) is confined and dynamically driven at one boundary, with time-resolved force and displacement measured at both boundaries as driving frequency is swept.

Key Observations

Pressure dependence: Resonance frequency scales as $N_{\rm recent}$ 9, with $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$0 for compression and $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$1 for shear.
Amplitude softening: At fixed $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$2, $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$3, with increasing amplitude $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$4 causing monotonic decrease of the contact number at resonance $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$5.
Microstructure: Localized "soft spots" proliferate with increasing $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$6, acting as defects that scatter elastic energy and further reduce effective modulus.

This approach quantitatively connects microstructure rearrangements (loss of $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$7) to reductions in elastic wave velocity and relates to post-seismic recovery and softening phenomena in Earth materials (Reichhardt et al., 2014).

5. PackForcing and the Force Network Ensemble

The force network ensemble (FNE) formalism is closely connected to the statistical description of PackForcing in jammed systems (Tighe et al., 2010). Here, PackForcing protocols selectively sample or evolve force networks by controlling the boundary stress or applying localized forces.

FNE Mathematical Structure

Constraints: Local force and torque balance, prescribed global stress, non-cohesion and Coulomb friction.
Partition function: Integration over force configurations within the allowed convex polytope.
Force-rearrangement basis: Hyperstatic packs ($\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$8) admit $\Ccal^l = \Ccal^l_{\rm sink}\;\|\;\Ccal^l_{\rm mid}\;\|\;\Ccal^l_{\rm recent}$9 independent rearrangements, where $2 \times 8 \times 8=128$ 0 is the mean contact number.

The single-contact force distribution $2 \times 8 \times 8=128$ 1 under PackForcing protocols follows compressed exponential forms with Gaussian tails in 2D. Protocols controlling pressure, boundary forces, or cycles preferentially select regions of the FNE convex polytope, enabling fine control over mechanical and statistical properties (Tighe et al., 2010).

6. Applications and Impact

PackForcing frameworks find applications in:

Long-consistency video synthesis: Enabling minute-scale autoregressive video generation with bounded resource consumption and high temporal stability (Mao et al., 26 Mar 2026).
Reinforcement learning video world modeling: Efficient history encoding in interactive, bidirectional agents for simulating complex environments (Rui et al., 8 Jun 2026).
Granular compaction and rheology: Systematic exploration of packing microstructure, frictional states, and elastic response of particulate matter under arbitrary preparation protocols (Santos et al., 2023, Reichhardt et al., 2014).
Statistical mechanics of jammed matter: Analysis of configuration space and force distribution statistics under controlled forcing and packing protocols (Tighe et al., 2010).

Empirically, PackForcing enables both strict resource bounding and high-quality long-horizon rollouts in generative models, and offers a rigorous, tunable approach to granular material design and elasticity studies.

References

Topic Area	Key Reference(s)	Application Scope
Video diffusion/transformers	(Mao et al., 26 Mar 2026, Rui et al., 8 Jun 2026)	Long-context memory management in generative and world models
Granular packing protocols	(Santos et al., 2023, Reichhardt et al., 2014)	Stress, drag, and dynamic forcing in DEM simulation of jammed packings
Force network ensemble theory	(Tighe et al., 2010)	Statistical mechanics and force distribution in static/dynamically forced packings