← HomePASPA
PASPA · NeurIPS 2026

Past as State,
Present as Attention: Persistent-State Blockwise Flow Matching
for Long-Horizon Co-Speech Motion Generation

Yi Liu1,3Xiangyue Zhang2Jianfang Li3*Jia Ma1Yisheng He3Jianqiang Ren3
1 Tongji University2 The University of Tokyo3 Alibaba Group

* Corresponding author

Summarize the past. Attend to the present.

ManuscriptDemoMethodResultsBibTeX
3.71FGD ×10 ↓ · Single speaker22.4% lower than GlobalDiff
2.33FGD ×10 ↓ · All speakers11.4% lower than GlobalDiff
Fixed-size statePersistent history cacheLong rollouts without growing history tokens
Figure 1Long speech, consistent style. The same speech and speaker are followed over 52 seconds. PASPA carries slowly varying style through a persistent state while using attention to model the local details of the current block.
Overview

Abstract

Co-speech motion generation remains challenging over long horizons due to the need for temporal coherence and consistent style. Existing methods typically generate long motions segment by segment, either conditioning on a short seed of recent frames, which gradually drifts, or retaining ever-longer explicit histories, which are costly to maintain and hard to learn from limited data. We trace these limitations to a key asymmetry between past and present: past motion carries slowly varying style that should be compactly summarized, whereas the current segment exhibits fast-changing local dynamics that require fine-grained modeling. This asymmetry motivates a simple Past-As-State, Present-as-Attention principle for long-horizon co-speech motion generation. Following this principle, we propose PASPA, which summarizes past history with an inter-block persistent state and models the current segment with intra-block bidirectional self-attention. We instantiate PASPA within a blockwise flow matching framework, with multi-block supervision for effective training over variable histories, prefill-decode state caching for efficient rollout, and hybrid classifier-free guidance for enhanced history and speech conditioning. Extensive public-benchmark experiments show that PASPA achieves state-of-the-art FGD performance and preserves motion style across long-sequence generation.
Contributions

The past carries style.
The present carries detail.

01

Different roles, different representations

We identify an asymmetry between slowly varying style in past motion and fast-changing dynamics in the current segment, and formulate the past-as-state, present-as-attention principle.

02

A persistent-state motion generator

PASPA combines blockwise Gated DeltaNet for history propagation with intra-block bidirectional attention, within a semi-autoregressive blockwise flow matching framework.

03

Training and rollout for long horizons

Multi-block supervision learns from variable history lengths. Prefill-decode separation reuses a fixed-size state cache, while hybrid classifier-free guidance controls speech conditioning and historical dependence.

Video

See PASPA in motion

A narrated introduction, method animations, and speech-driven comparisons with prior methods.

GT · PASPA · Semtalk · GestureLSM · EMAGE
Original speech and comparison labels are preserved in the result sequences.3:06 · English narration
Method

Past as state. Present as attention.

A compact summary of completed motion conditions a bidirectional model of each new block.

Figure 2Different roles call for different representations. Instead of dropping distant history or retaining every past token for attention, PASPA summarizes the past in state and applies full attention inside the current block.

Motion is generated semi-autoregressively, one block at a time, with blockwise flow matching. The history state provides a stable context; bidirectional interactions within the current block capture fast-changing dynamics and fine articulation.

Figure 4The PASPA architecture. BGDN propagates clean history through a block-causal persistent state. BMHSA provides bidirectional attention within each current block. The generation layers are conditioned on speech and speaker identity.
BGDN

Carry the past as a persistent state

BGDN summarizes clean past motion into a fixed-size recurrent state. Every query in a block reads the block-start state, enforcing access to strict past history while retaining slowly varying stylistic information.

BMHSA

Model the present with bidirectional attention

BMHSA allows full interaction among frames inside the current block to capture local dynamics and fine articulation, while forbidding direct attention across blocks.

Decode a block. Update the state. Continue.

The history cache stays fixed-size as the rollout grows.

01 · READRead the cached state from completed motion.
02 · DECODEGenerate the current block with flow matching.
03 · PREFILLUpdate state once with the clean block.
04 · CONTINUEReuse the state for the next motion block.

Multi-block supervision

Train clean-motion predictions over multiple blocks at different noise levels. Weakly monotonic timesteps and adaptive loss weighting support learning across history lengths.

Hybrid classifier-free guidance

Jointly guide speech/expression conditioning and historical dependence, including their interaction, while reusing the cached history state during decoding.

BEAT2 benchmark

Lower FGD in both speaker settings

PASPA achieves the best FGD among the methods compared in the manuscript.

22.4%
11.4%

Relative FGD reduction against GlobalDiff for the single-speaker and all-speaker settings, respectively. PASPA also has the lowest reported facial-expression MSE among the compared methods that report this metric.

Single speaker

BEAT2 · Table 1
Method FGD ↓ (×10) BC ↑ Div ↑ MSE ↓ (×108)
CaMN 6.64 0.677 10.86 —
DSG 8.81 0.724 11.49 —
EMAGE 5.51 0.772 13.06 7.680
DiffSHEG 8.99 0.714 11.91 7.665
MambaTalk 5.37 0.781 13.05 6.289
HoloGest 5.34 0.795 14.15 —
RAG-GESTURE 8.08 0.734 11.97 —
SemTalk 4.28 0.777 12.91 6.153
GestureLSM 4.09 0.714 13.24 —
GlobalDiff 4.78 0.705 13.13 6.330
PASPA (Ours) 3.71 0.746 13.38 5.768

All speakers

BEAT2 · Table 1
Method FGD ↓ (×10) BC ↑ Div ↑ MSE ↓ (×108)
CaMN 5.12 0.200 5.58 —
DSG 11.74 0.734 11.12 —
EMAGE 6.92 0.284 6.06 6.908
DiffSHEG — — — —
MambaTalk — — — —
HoloGest 6.46 0.803 13.53 —
RAG-GESTURE 4.87 0.514 9.94 —
SemTalk — — — —
GestureLSM — — — —
GlobalDiff 2.63 0.404 8.24 4.144
PASPA (Ours) 2.33 0.645 9.42 3.793

Table 1. Values are reproduced from the manuscript. FGD and facial-expression MSE use the reported scales (×10 and ×108). ↓ / ↑ means lower / higher is better under this evaluation. An em dash denotes an unreported value. The blue row denotes PASPA; bold numbers mark the best values in each column.

Figure 5What the numbers look like. Selected speech segments compare gesture direction, amplitude, and speech alignment. PASPA is shown alongside ground truth, SemTalk, GestureLSM, and EMAGE; the original annotations highlight differences in gesture and hand articulation.
Long-horizon generation

Keep the style as the sequence grows

History should carry more than recent frames.

In the dedicated long-rollout evaluation, PASPA has the lowest terminal-window FGD at 30 and 60 seconds and preserves the highest diversity among the three history mechanisms.

The 52-second sequence above illustrates the same idea: the speaker's characteristic hand spacing remains visible across distant moments.

A fixed-size history stateClean motion updates the persistent state between blocks. The state does not grow into an ever-longer token history.

Terminal-window evaluation

Table 5
Method 0–10 s 20–30 s 50–60 s
FGD ↓ Div ↑ FGD ↓ Div ↑ FGD ↓ Div ↑
Short Seed 4.731 6.581 5.886 6.033 6.721 6.081
Token Attention 4.448 6.289 5.388 6.510 5.771 6.427
PASPA 4.448 7.217 5.119 7.133 5.745 7.138

Metrics are evaluated in the final 10-second window of 10, 30, and 60-second rollouts. This is a separate protocol from Table 1. FGD ×10 ↓; Div ↑.

Reference

BibTeX

Past as State, Present as Attention 
@misc{liu2026paspa,
  title = {Past as State, Present as Attention: Persistent-State Blockwise Flow Matching for Long-Horizon Co-Speech Motion Generation},
  author = {Liu, Yi and Zhang, Xiangyue and Li, Jianfang and Ma, Jia and He, Yisheng and Ren, Jianqiang},
  year = {2026}
}