ScalingAR: Confidence-based AR Image Generation

• ScalingAR is a framework that uses token-level entropy and margins to create a dynamic confidence profile for autoregressive image generation.
• It employs a two-level control mechanism—Profile Level for continuous confidence evaluation and Policy Level for adaptive termination and guidance scheduling.
• ScalingAR demonstrates improved benchmark performance with up to 15.2% relative gains and reduces token usage by 62%, highlighting its efficiency and scalability.

ScalingAR is a test-time scaling framework for next-token prediction autoregressive image generation that uses the model’s own token-level probabilities to guide inference, rather than relying on partial image decoding or external reward models. It was introduced to address a specific limitation of visual autoregressive generation: in next-token prediction image models, intermediate token prefixes do not correspond to stable, usable partial images, so test-time scaling strategies adapted from language modeling or next-scale visual autoregression are poorly matched to the generation process. ScalingAR therefore constructs an entropy-based confidence profile directly from the token distribution and applies that profile at two levels—“Profile Level” and “Policy Level”—to terminate low-confidence trajectories and schedule guidance strength during generation (Chen et al., 30 Sep 2025).

1. Scope, motivation, and problem setting

ScalingAR is defined for next-token prediction autoregressive image generation, where an image is represented as a sequence of discrete visual tokens and generated one token at a time according to

$$p(x_1, x_2, \ldots, x_T) = \prod_{t=1}^T p(x_t \mid x_{<t}),$$

with training objective

$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$

In this setting, the model does not expose a coherent intermediate image until most or all tokens have been produced, which differentiates it from diffusion models and from next-scale prediction visual autoregressive methods. ScalingAR was proposed as the first test-time scaling framework specifically designed for this next-token prediction regime (Chen et al., 30 Sep 2025).

The framework is motivated by the mismatch between standard test-time scaling methods and autoregressive image token generation. In LLMs, partial decoding is already semantically informative, and external reward models or verifiers can often score incomplete outputs. In next-token prediction image generation, the paper identifies three obstacles: “holism,” because truncated image token streams do not yield meaningful images; “objective ambiguity,” because there is no single scalar notion of correctness for partial images; and “early signal scarcity,” because early decoded token grids are unstable and expensive to render repeatedly. This leads ScalingAR to avoid early decoding and external rewards entirely, and instead to use token entropy as an intrinsic signal of confidence (Chen et al., 30 Sep 2025).

The method sits within a broader test-time scaling literature, but its design target is narrow and explicit. It is applied to next-token prediction autoregressive image generators such as LlamaGen and AR-GRPO, both of which use discrete visual tokenizers. This suggests that ScalingAR should be understood not as a generic reranking heuristic, but as an inference-time control layer specialized to token-sequential visual generation (Chen et al., 30 Sep 2025).

2. Confidence profiling from token distributions

The core representational object in ScalingAR is a streaming confidence state derived from the model’s logits. For each step $t$, the autoregressive model produces a distribution

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$

with token entropy

$$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$

The framework also uses the top-1/top-2 margin

$m_t = \pi_t(v_1) - \pi_t(v_2),$

where $v_1$ and $v_2$ are the most and second-most probable tokens. Low entropy and high margin indicate higher local confidence (Chen et al., 30 Sep 2025).

ScalingAR combines these into a token-level uncertainty surrogate

$$u_t = \alpha_H \widehat{H}_t + \alpha_M (1 - m_t), \quad \alpha_H + \alpha_M = 1,$$

and defines token confidence as

$s_t^{\text{tok}} = 1 - u_t.$

To stabilize this signal, it applies an exponential moving average,

$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$0

This token-level channel is then augmented with spatial information. Tokens are organized on an $L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$1 grid, entropy is averaged within non-overlapping $L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$2 blocks, and the highest-entropy worst-$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$3 blocks are aggregated as

$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$4

After rolling min–max normalization $L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$5, ScalingAR defines block stability as

$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$6

The intrinsic channel is then

$L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$7

followed by EMA smoothing to produce $L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$8 (Chen et al., 30 Sep 2025).

A second channel measures prompt utilization under classifier-free guidance. Let $L = \sum_{t=1}^T \log p(x_t \mid x_{<t}).$9 and $t$0 denote the conditional and unconditional distributions. ScalingAR computes

$t$1

normalizes it over time,

$t$2

clips it,

$t$3

and maps it to

$t$4

After smoothing, $t$5 is the conditional utilization score. The unified confidence is then

$t$6

ScalingAR additionally maintains a running minimum

$t$7

and a relative rebound statistic

$t$8

These quantities form the confidence state used by the policy layer (Chen et al., 30 Sep 2025).

3. Two-level control: Profile Level and Policy Level

ScalingAR is organized around two complementary control layers. The “Profile Level” continuously estimates and calibrates confidence, while the “Policy Level” uses that state to alter the generation process. The profile level therefore does not merely monitor uncertainty; it produces the state variables consumed by the termination and guidance policies (Chen et al., 30 Sep 2025).

The first policy is the Adaptive Termination Gate. After a warm-up period $t$9, it sets a threshold

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$0

and updates this threshold periodically using EMA. A trajectory is protected from very early pruning by a protection horizon $$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$1, but after that point it can be terminated if its running minimum confidence falls below the threshold and it does not recover within a recovery window. Recovery is defined by either an absolute rebound

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$2

or a relative rebound

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$3

A hard-fail guard

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$4

causes immediate termination. In practice, the method begins with more trajectories than are intended to survive, prunes low-confidence branches early, and completes only the strongest set of trajectories (Chen et al., 30 Sep 2025).

The second policy is the Guidance Scheduler, which dynamically adjusts classifier-free guidance according to the confidence state. Its raw update is

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$5

The interpretation is explicit in the paper: low conditional utilization increases guidance, high short-term intrinsic volatility increases guidance, and strong rebound decreases guidance to allow diversity. The actual guidance value is smoothed and clamped,

$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$6

with a deadband to suppress jitter. The resulting schedule is intended to provide phase-appropriate conditioning strength across the token trajectory (Chen et al., 30 Sep 2025).

The full inference pipeline therefore consists of initializing multiple trajectories, updating intrinsic and conditional confidence statistics at each token step, applying pruning after warm-up, updating guidance for surviving trajectories, sampling the next tokens, and decoding only completed trajectories at the end. The method’s central claim is that this realizes test-time scaling for image token generation without partial renders and without auxiliary reward models (Chen et al., 30 Sep 2025).

4. Empirical evaluation and quantitative performance

ScalingAR is evaluated on two next-token prediction autoregressive image generators, both with 0.8B parameters: LlamaGen at $$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$7 resolution and AR-GRPO at $$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$8. The principal benchmarks are GenEval, which emphasizes compositional text-to-image alignment, and TIIF-Bench, which evaluates instruction following. Baselines are vanilla autoregressive generation, Importance Sampling, and Best-of-$$\pi_t(v) = p(x_t = v \mid x_{<t}, \text{prompt}), \quad v \in V,$$9 (Chen et al., 30 Sep 2025).

The paper reports consistent gains over the base generators. On LlamaGen, GenEval Overall improves from 0.32 to 0.36, which is reported as a 12.5% relative improvement, and TIIF-Bench Overall improves from 40.35 to 46.47, a 15.2% relative improvement. On AR-GRPO, GenEval Overall improves from 0.31 to 0.49, and TIIF-Bench Overall improves from 16.22 to 26.35. The same experiments show that Importance Sampling and Best-of-$$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$0 often provide negligible or negative gains, especially on GenEval for LlamaGen (Chen et al., 30 Sep 2025).

A concise summary of the headline benchmark numbers is as follows:

Model	Benchmark	Base $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$1 ScalingAR
LlamaGen	GenEval Overall	0.32 $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$2 0.36
LlamaGen	TIIF-Bench Overall	40.35 $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$3 46.47
AR-GRPO	GenEval Overall	0.31 $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$4 0.49
AR-GRPO	TIIF-Bench Overall	16.22 $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$5 26.35

The efficiency results are equally central. On TIIF-Bench, ScalingAR surpasses Importance Sampling and Best-of-$$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$6 while using 62.0% fewer tokens on average, because many trajectories are terminated before completion. The paper presents this as an unusual test-time scaling result: better generation quality together with lower visual token consumption, rather than a pure quality–cost tradeoff (Chen et al., 30 Sep 2025).

The robustness evaluation uses “impossible prompting,” adapted from Impossible Videos’ IPV-Txt prompts. Here the metric is Impossible Prompt Following, judged by GPT-4o. Under these conditions, ScalingAR mitigates performance drops by 26.0% compared to the base models. The paper attributes this to the interaction between guidance scheduling and confidence-based pruning in scenarios where no candidate is strictly correct but some trajectories degrade more severely than others (Chen et al., 30 Sep 2025).

5. Ablations, sensitivity, and observed scaling behavior

The ablation study on TIIF-Bench with LlamaGen isolates the contributions of the major components. Full ScalingAR achieves an overall score of 46.5. Removing the Conditional Channel reduces this to 45.1, removing Worst-Block Stability reduces it to 44.2, and removing Token-Level Confidence reduces it to 42.9. The paper therefore identifies token-level confidence as the most critical single component, with worst-block stability providing the main spatial complement and the conditional channel contributing prompt-alignment information (Chen et al., 30 Sep 2025).

Sensitivity analysis further shows that the unified confidence fusion is not arbitrary. The best performance occurs at $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$7, indicating that the intrinsic channel is slightly prioritized while the conditional channel remains necessary. For the pruning gate, $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$8 is reported as the best threshold region: smaller values cause weak pruning and high token cost, while larger values over-prune and reduce accuracy. The recovery threshold $$H_t = -\sum_{v \in V} \pi_t(v)\log \pi_t(v), \qquad \widehat{H}_t = \frac{H_t}{\log |V|}.$$9 is reported to balance genuine recovery against continued waste on poor trajectories (Chen et al., 30 Sep 2025).

The framework also exhibits positive width and depth scaling behavior under fixed token budgets. When the number of samples and token length are varied, ScalingAR continues to gain performance and outperforms Importance Sampling and Best-of-$m_t = \pi_t(v_1) - \pi_t(v_2),$0 at the same or lower total token budgets. The paper explicitly states that this shows ScalingAR is a true scaling law mechanism for autoregressive image generators. This suggests that its contribution is not limited to one operating point; it modifies how extra inference compute is translated into generation quality (Chen et al., 30 Sep 2025).

Qualitative evaluation and user study results support the benchmark outcomes. The generated images are reported to show better object count fidelity, stronger spatial relation accuracy, improved text alignment, and more robust behavior on difficult prompts. Human raters in the MOS study prefer ScalingAR over baselines in overall preference, aesthetic quality, realism, semantic alignment, and attribute binding. These observations are qualitative, but they are consistent with the quantitative improvements on GenEval and TIIF-Bench (Chen et al., 30 Sep 2025).

6. Limitations, assumptions, and significance

ScalingAR is built on several explicit assumptions. It assumes discrete visual tokens, access to token logits, and the availability of both conditional and unconditional branches for classifier-free guidance. It is designed for next-token prediction autoregressive image generation and is not directly implemented for diffusion or masked autoregressive models. The framework also depends on calibration choices such as normalization windows, EMA factors, and confidence thresholds, which the paper notes may require retuning under substantially different architectures or prompt distributions (Chen et al., 30 Sep 2025).

The paper also states that confidence remains a proxy rather than a direct measure of perceptual quality or semantic fidelity. Token entropy and conditional–unconditional KL divergence may not anticipate every failure mode, especially complex global errors that are not preceded by strong local uncertainty spikes. A plausible implication is that the framework is best understood as an intrinsic control policy over the generator’s own uncertainty landscape rather than as a full substitute for task-specific evaluation or verification (Chen et al., 30 Sep 2025).

Within the literature on test-time scaling, ScalingAR’s significance lies in showing that logits alone can provide sufficient control signals for nontrivial inference-time intervention in autoregressive image generation. Its “ScalingAR” name refers not to parameter scaling or training-time model growth, but to scaling inference by confidence-aware pruning and guidance scheduling. In that sense, it reinterprets test-time scaling for visual token generators as a problem of calibrated internal control, rather than one of repeated decoding and external reranking (Chen et al., 30 Sep 2025).