Segment-wise Auto-Regressive Inference

• Segment-wise AR inference is a strategy that factors predictions over contiguous blocks rather than individual tokens, enhancing efficiency and contextual modeling.
• It employs block-level probabilistic factorization, segment-specific masking, and tailored inference pipelines to capture higher-order dependencies across modalities.
• Empirical results demonstrate that SAR frameworks improve throughput and sample efficiency while maintaining robust performance in vision, speech, and time series tasks.

Segment-wise Auto-Regressive (SAR) inference encompasses a broad family of strategies in which an auto-regressive factorization is imposed not at the level of individual tokens, but over contiguous, non-overlapping segments or blocks of the underlying data. This methodology structures prediction, learning, and statistical inference around units—blocks of tokens in vision, fixed-length spans in speech, or time intervals in non-stationary series—enabling enhanced computational efficiency, the exploitation of higher-order dependencies, and, in certain regimes, improved sample and parameter efficiency. SAR strategies are defined by (1) block-level probabilistic factorization, (2) segment-respecting masking or conditionality constraints, and (3) inference pipelines or training objectives adapted to work on blocks or segments rather than single tokens.

1. Probabilistic Factorization: The Blockwise AR Model

SAR inference starts with a segmental factorization of the joint distribution of a sequence or array. For image modeling as with XTRA, the input image is partitioned into $K$ non-overlapping blocks each containing $k \times k$ patches, and the flattened vector of block-pixel values $x = (x_1, \ldots, x_K)$. The blockwise AR factorization is

$$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$

Each block $x_k$ is predicted using the context provided by all preceding blocks. Training typically minimizes a blockwise loss, e.g., a mean squared error

$$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$

where $\hat x_k^n(\theta; x_{<k}^n)$ denotes the model prediction for block $k$ conditioned on prior blocks, for sample $n$ (Amrani et al., 2024).

In time series SAR, for general non-stationary $\{X_{t,n}\}$, the process is approximated as

$k \times k$0

with segmentwise pooling or adaptivity to capture time-varying AR structure (Ding et al., 2021).

2. Segmental Masking and Context Constraints

The architectural or statistical mechanism underpinning SAR models is a masking or conditionality rule enforcing segmentwise dependence. In the case of XTRA, SAR employs a Block Causal Mask in the Transformer backbone: For $k \times k$1 tokens ordered linearly, define block assignment

$k \times k$2

so every consecutive group of $k \times k$3 tokens forms a block. The attention mask $k \times k$4 is

$k \times k$5

enforcing that a token attends only to tokens from its current or previous blocks. This causal pattern recurs identically across all encoder and decoder layers (Amrani et al., 2024).

In speculative decoding for TTS (VADUSA), SAR is instantiated by draft heads outputting candidates for multiple future tokens, using a verification mechanism in which only candidates consistent with the AR head’s draws for each position are accepted and committed as a segment. Masking in the “tree attention” verifies full segment consistency (Li et al., 2024).

3. Inference Algorithms and Computational Properties

SAR inference proceeds blockwise, generating a full segment at each step. In image models:

1. Given current context $k \times k$6, concatenate $k \times k$7 for the next block.
2. Pass through the Transformer with Block Causal Mask.
3. Extract hidden states for the $k \times k$8 new tokens and map via per-block MLP.
4. Append the predicted block to the context.
5. Repeat for all $k \times k$9 blocks.

The generic pseudocode is:

$\hat x_k^n(\theta; x_{<k}^n)$8

The masking is realized via

$x = (x_1, \ldots, x_K)$0

where $x = (x_1, \ldots, x_K)$1 assigns $x = (x_1, \ldots, x_K)$2 to illegal connections (Amrani et al., 2024).

In VADUSA TTS, SAR inference uses a set of $x = (x_1, \ldots, x_K)$3 draft heads for speculative next-token prediction, a tolerance $x = (x_1, \ldots, x_K)$4 for AR sampling, and a verification/commit loop which accepts the longest prefix matching the AR samples, achieving segment-level decoding with a single forward and tree-attention verification pass per batch (Li et al., 2024).

For nonstationary time series, SAR consists of splitting the series into $x = (x_1, \ldots, x_K)$5 segments, fitting a time-varying or piecewise AR model (e.g., via sieve regression) in each, and deploying segment-adaptive inference or forecasting. Segment-level coefficient variation is tested via high-dimensional $x = (x_1, \ldots, x_K)$6 statistics and multiplier bootstrapping (Ding et al., 2021).

The following table summarizes key aspects of segmental inference across modalities:

Modality	Block Size/SAR Unit	Masking/Verification	Per-Step Output
Images	$x = (x_1, \ldots, x_K)$7 patches	Block Causal Mask	$x = (x_1, \ldots, x_K)$8 pixels
Speech	$x = (x_1, \ldots, x_K)$9-token segment	Sparse tree/Tolerance	$$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$0 tokens
Time Series	$$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$1-point segment	Sieve/Statistical test	Fitted AR coeffs

4. Impact of Segment Size on Computation and Efficiency

The block size $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$2 in SAR exerts a direct effect on computational scaling and model efficiency. For images, standard token-wise AR requires $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$3 forward passes (per patch), each $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$4 → $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$5 overall. SAR with block size $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$6 reduces this to $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$7 passes, each $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$8, yielding $$p_\theta(x) = \prod_{k=1}^K p_\theta(x_k \mid x_{<k}).$$9. As $x_k$0 increases, AR steps (and wall-clock decode time) decrease approximately as $x_k$1, benefiting throughput.

Empirically, increasing block size improves sample efficiency and representational abstraction in the XTRA model: at $x_k$2 (tokenwise, AIM) attentive-probe accuracy is $x_k$364.6%; at $x_k$4 (blockwise), it is $x_k$567.6% (+3.0%). XTRA-H/14 trained on $x_k$6M samples ($x_k$7152$x_k$8 fewer than AIM-H/14 on $x_k$9B) achieves comparable or superior accuracy (Amrani et al., 2024).

For TTS, VADUSA’s SAR inference accelerates decoding by committing on average $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$0 tokens per pass, yielding $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$1 speed increases without quality loss. The draft length $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$2 and tolerance $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$3 govern acceptance and quality, with diminishing speedup returns for very large $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$4 or $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$5 (Li et al., 2024).

In time series SAR, segment length and localization choices determine the tradeoff between statistical power for detecting nonstationarity and adaptivity to temporal change; segmentwise SAR enables both structural estimation and forecast optimality for locally stationary processes (Ding et al., 2021).

5. Empirical Outcomes and Benchmark Comparisons

Concrete empirical results from SAR applications highlight substantial efficiency gains:

• Vision (XTRA): With only $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$6M parameters (XTRA-B/16), linear probe accuracy on ImageNet exceeds that from iGPT-L ($$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$7B) by $$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$8 ($$\ell(\theta) = \frac{1}{N(K-1)} \sum_{n=1}^N \sum_{k=2}^K \lVert \hat x_k^n(\theta; x_{<k}^n) - x_k^n \rVert_2^2,$$9). Attentive probe scores on XTRA models trained on drastically fewer samples remain competitive or superior (Amrani et al., 2024).
• Speech Synthesis (VADUSA): For HuBERT-2048 tokens on LibriTTS, VADUSA yields $\hat x_k^n(\theta; x_{<k}^n)$0 speedup and maintains or slightly improves predicted MOS and word error rate, with $\hat x_k^n(\theta; x_{<k}^n)$13.8 tokens accepted per pass. For EnCodec tokens, both WER and throughput show improvement relative to standard AR (Li et al., 2024).
• Time Series: SAR-based adaptation achieves strong-sense asymptotic optimality for local forecasting, as shown by the time-varying AR and associated bootstrap test procedures (Ding et al., 2021).

Across all domains, committing entire segments rather than single tokens increases throughput, supports learning or inference of more abstract dependencies, and in some cases improves or preserves sample efficiency and quality.

6. Structural and Representational Benefits

Beyond computational and sample efficiency, SAR imprints structural inductive biases. In XTRA, blockwise AR predicts mid-frequency content—whole $\hat x_k^n(\theta; x_{<k}^n)$2 blocks—instead of isolated pixels. This compels low-frequency, semantically meaningful content modeling (e.g., shapes, object parts), which is evidenced by richer features and superior probe task performance over 15 diverse datasets (Amrani et al., 2024).

In TTS, segment-level proposals facilitate longer-range prosody or phoneme planning within a single inference pass (draft), while verification ensures local AR fidelity, making it possible to increase acceptance length without degrading intelligibility or synthesis quality (Li et al., 2024).

For nonstationary time series, segmentwise AR allows for model adaptivity, statistical testing for changing dynamics, and forecast error rates that shrink with increasing segment size, under mild short-range dependence assumptions (Ding et al., 2021).

7. Limitations, Hyperparametric Trade-offs, and Domain Considerations

The SAR paradigm introduces trade-offs. For image and speech models, the granularity (block size $\hat x_k^n(\theta; x_{<k}^n)$3 or draft length $\hat x_k^n(\theta; x_{<k}^n)$4) regulates speed versus acceptance rate/accuracy. Very large segments reduce number of AR passes but may degrade prediction accuracy per block, particularly if long-range dependencies are less predictable. In VADUSA, tolerance $\hat x_k^n(\theta; x_{<k}^n)$5 directly controls the tradeoff between speed (expected acceptance length $\hat x_k^n(\theta; x_{<k}^n)$6) and per-pass computation.

Segment selection for time series requires balancing power for detecting time-variation against segmentwise estimation variance. Cross-validation, adaptive algorithms, or domain priors can be incorporated to select optimal segmentation parameters (Ding et al., 2021).

All SAR frameworks require compatible architectures: blockwise mask structures for Transformer-based models (vision, language), draft head and verification mechanisms for speculative AR (speech), or adequately regularized and locally consistent estimators for time series.

A further practical constraint is task-dependent: for VADUSA, sparse-tree candidate sets must be domain-calibrated (e.g., $\hat x_k^n(\theta; x_{<k}^n)$7 via data-driven construction), else acceptance rates and speedup degrade (Li et al., 2024).

Empirically, SAR strategies can reduce resource requirements, inference time, and sample complexity, while enhancing downstream representational quality in diverse auto-regressive settings.