Progressive Attention-Guided Adaptive CFG

• The paper presents adaptive diffusion guidance by dynamically modulating the CFG scale over time via methods like β-CFG, balancing prompt adherence and sample quality.
• Progressive attention-guided strategies inject control signals directly in attention space, using geometric decompositions to stabilize guidance and improve convergence.
• Adaptive techniques also incorporate uncertainty, error geometry, and prior steering to adjust guidance in a state-dependent manner, enhancing both performance and efficiency.

Progressive Attention-Guided Adaptive CFG can be understood as an umbrella for classifier-free guidance mechanisms in which the guidance signal is no longer a fixed scalar extrapolation, but is instead modulated over time, redirected by attention-space structure, or adapted to uncertainty, geometry, or control objectives. In the cited literature, this phrase does not denote a single canonical algorithm; rather, it emerges from converging lines of work on time-dependent CFG scaling, attention-space extrapolation, uncertainty-aware masking, error-orthogonal correction, prior-space steering, control-theoretic guidance, and fixed-point calibration (Malarz et al., 14 Feb 2025, Kim, 3 Mar 2026, Li et al., 26 May 2025, Wang et al., 3 Mar 2026, Wang et al., 24 Oct 2025).

1. Static CFG and the case for adaptation

The baseline object is standard classifier-free guidance. In its noise-prediction form, guided prediction is written as

$$\hat{\epsilon}^{w}_c(x_t) = \epsilon_{\emptyset}(x_t) + w \big( \epsilon_c(x_t) - \epsilon_{\emptyset}(x_t) \big),$$

with the rest of the sampler unchanged (Malarz et al., 14 Feb 2025). In discrete diffusion LLMs, the same structure is applied in logit space:

$$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$

so the guidance scale becomes an explicit control input at each reverse step (Zhou et al., 8 May 2026).

Across the literature, the central limitation of this formulation is its globally static scale. Strong guidance improves prompt adherence but can degrade sample quality; weak guidance preserves quality but weakens conditioning fidelity (Malarz et al., 14 Feb 2025). In discrete diffusion LLMs, this mismatch is explicitly framed as a structural problem: the role of guidance changes across generation stages, and a single global scale cannot match the non-stationary demands of early noisy states, mid-trajectory semantic consolidation, and late-stage fluency refinement (Zhou et al., 8 May 2026).

This motivates the adjective “adaptive.” In the papers surveyed here, adaptation can occur along several axes: time, state, uncertainty, attention geometry, prior initialization, or teacher-supervision target. “Progressive” then refers to the fact that this adaptation is staged along the trajectory rather than applied as a one-shot correction.

2. Progressive temporal scaling

A first and influential form of progressive CFG is explicit timestep scheduling. “Classifier-free Guidance with Adaptive Scaling” introduces $\beta$-CFG, which imposes boundary conditions on the guidance term so that it vanishes at both ends of the denoising trajectory and is concentrated in the middle (Malarz et al., 14 Feb 2025). Its schedule is defined by a Beta density

$$\beta(t) = \frac{t^{a - 1} (1 - t)^{b - 1}}{B(a, b)},$$

with $a>1$ and $b>1$, and the guided prediction becomes

$$\hat{\epsilon}^{\beta}_c(x_t) = \epsilon_{\emptyset}(x_t) + \beta\!\big(t/T\big)\cdot \omega\, \frac{\epsilon_c(x_t) - \epsilon_{\emptyset}(x_t)} {\big\|\epsilon_c(x_t) - \epsilon_{\emptyset}(x_t)\big\|^{\gamma} }.$$

The method stabilizes guidance by gradient-based adaptive normalization and uses a single-modal, time-dependent curve to adapt the prompt-matching versus quality trade-off during denoising, with better FID while maintaining CLIP similarity comparable to reference CFG (Malarz et al., 14 Feb 2025).

IB-Flow develops a related but distinct schedule from an Information Bottleneck formulation of few-step CFG distillation (Wang et al., 10 Jul 2026). There, the extra driving force from guidance is tied to residual uncertainty through the signal-to-noise ratio,

$$\omega^*(t) = 1 + (\omega_{\max}-1) \cdot \frac{1}{1 + \gamma\cdot\text{SNR}(t)}, \qquad \text{SNR}(t) = \frac{t^2}{(1-t)^2}.$$

The stated behavior is maximal thrust for initial structural anchoring, followed by smooth reversion to the natural manifold for micro-detail refinement, and the paper reports that the SNR-analytic schedule outperforms constant, linear, and cosine schedules in its setting (Wang et al., 10 Jul 2026).

A third formulation abandons hand-crafted schedules entirely. “Guidance Is Not a Hyperparameter: Learning Dynamic Control in Diffusion LLMs” treats guidance selection as a sequential decision problem, with a policy choosing a discrete CFG scale at each stage based on the evolving diffusion state (Zhou et al., 8 May 2026). The learned trajectories are task-dependent: keyword-conditioned generation and length control exhibit hump-shaped schedules, while sentiment transfer exhibits monotonically decreasing schedules. This directly contradicts the common assumption that “progressive” guidance must mean a monotone increase or monotone decay.

3. Attention-space guidance and multimodal precedents

The explicitly attention-guided branch of the literature is anchored by “Bridging Diffusion Guidance and Anderson Acceleration via Hopfield Dynamics” (Kim, 3 Mar 2026). That work interprets attention as Modern Hopfield retrieval and attention-space extrapolation as a special case of Anderson Acceleration. A generic attention-space extrapolation takes the form

$$At'(Q_t,K_t,V_t) = At(Q_t,K_t,V_t) + \lambda\big[At(Q_t,K_t,V_t) - \tilde{At}(Q_t,K_t,V_t)\big],$$

where the “strong” and “weak” attention operators are contrasted inside the same block (Kim, 3 Mar 2026). Building on this, Geometry Aware Attention Guidance (GAG) decomposes the residual into components parallel and orthogonal to the sparse-attention retrieval direction and shows that the parallel component is the effective guidance signal, while the orthogonal component is associated with instability. The resulting update clips residual magnitude and, under a common fixed-point assumption plus orthogonal weak contraction, yields asymptotic convergence of orthogonal error to zero (Kim, 3 Mar 2026).

This gives a concrete technical meaning to “attention-guided”: guidance can be injected at the cross-attention layer itself rather than only at the model-output level. It also changes the computational profile, because the method remains within one global network pass per diffusion step rather than requiring a separate unconditional pass (Kim, 3 Mar 2026).

A multimodal precedent exists outside generative modeling. “RGBT Tracking via Progressive Fusion Transformer with Dynamically Guided Learning” introduces ProFormer, which first uses a self-attention module to collaboratively extract a multimodal representation, then uses two cross-attention modules to interact that representation with dual-modality features, and finally uses a feed-forward network to fuse the two interacted multimodal representations; its learning algorithm adaptively uses well-performing branches to guide the learning of other branches (Zhu et al., 2023). This is not a CFG method, but it suggests that progressive attention-guided fusion and adaptive branch guidance were already being explored in multimodal tracking, with a similar emphasis on staged integration and branch-dependent correction.

4. Uncertainty, error geometry, and prior-space adaptation

Adaptive CFG is not limited to timesteps or attention blocks. Several methods instead use uncertainty, prediction geometry, or prior statistics as the adaptive signal.

“Adaptive Classifier-Free Guidance via Dynamic Low-Confidence Masking” redefines the unconditional branch in a masked diffusion LLM by re-masking low-confidence tokens at each step, thereby creating a dynamic, localized unconditional input rather than a static one (Li et al., 26 May 2025). The guidance formula remains standard, but the unconditional input becomes state-dependent. The paper reports substantial improvements over standard CFG, including a 3.9 point gain on GPQA (Li et al., 26 May 2025).

“CFG-EC: Error Correction Classifier-Free Guidance” analyzes the mismatch between training and sampling through unconditional and conditional error vectors and proposes orthogonalizing the unconditional error component relative to the conditional one (Yang et al., 18 Nov 2025). The core idea is to suppress the cross term in the sampling error and thereby tighten the upper bound on sampling error. Its dynamic method estimates error proxies through temporal extrapolation and only applies correction when cosine similarity falls below a threshold (Yang et al., 18 Nov 2025).

“P-Guide: Parameter-Efficient Prior Steering for Single-Pass CFG Inference” moves adaptation to the prior rather than the denoiser output (Peng et al., 7 May 2026). It defines conditional and unconditional priors through learned mean and variance and applies CFG directly to prior parameters:

$$z_{\text{cfg}} = \mu_\phi(\emptyset) + w \big(\mu_\phi(y) - \mu_\phi(\emptyset)\big) + \Big[ \sigma_\phi(\emptyset) + w \big(\sigma_\phi(y) - \sigma_\phi(\emptyset)\big) \Big]\odot \epsilon.$$

Under a first-order approximation, this is equivalent to CFG in the sense that it steers generation from the prior space; with heteroscedastic priors, it also provides adaptive loss attenuation and improved robustness to data uncertainty, while reducing inference latency by approximately $$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$0 (Peng et al., 7 May 2026).

Mechanism	Adaptive signal	Representative paper
Dynamic low-confidence masking	Token-wise predictive confidence	(Li et al., 26 May 2025)
Error-correction CFG	Angle between unconditional and conditional error proxies	(Yang et al., 18 Nov 2025)
Prior steering	Conditional prior mean and variance	(Peng et al., 7 May 2026)

Taken together, these methods broaden the meaning of “adaptive CFG.” The adaptation need not be a scalar timestep schedule; it can be local, token-wise, geometric, or embedded in the initial latent distribution.

5. Control, distillation, and fixed-point interpretations

A major development in recent work is the reinterpretation of CFG as a controlled dynamical system. “CFG-Ctrl: Control-Based Classifier-Free Diffusion Guidance” writes flow-based generation as

$$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$1

defines the semantic error

$$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$2

and identifies vanilla CFG with proportional control, $$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$3 (Wang et al., 3 Mar 2026). The paper argues that such linear control leads to instability, overshooting, and degraded semantic fidelity at large guidance scales, then introduces Sliding Mode Control CFG with an exponential sliding surface over semantic error and a switching correction term. A Lyapunov analysis is provided to support finite-time convergence (Wang et al., 3 Mar 2026).

IB-Flow gives a complementary theory for few-step distillation by casting CFG injection as a dynamic mutual information game constrained by the Information Bottleneck principle (Wang et al., 10 Jul 2026). It adapts both the supervision target and the guidance magnitude. The dynamic optimal injection target is

$$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$4

so the look-ahead interval contracts when the conditional–unconditional discrepancy is large and expands when that discrepancy is small (Wang et al., 10 Jul 2026). In this formulation, adaptive CFG is simultaneously a temporal and informational allocation problem.

“Towards a Golden Classifier-Free Guidance Path via Foresight Fixed Point Iterations” reframes conditional guidance as fixed-point iteration toward a “golden path,” where conditional and unconditional denoising trajectories coincide (Wang et al., 24 Oct 2025). The paper shows that CFG and related methods are single-step short-interval solvers and proves that this regime is inefficient under a finite iteration budget. Foresight Guidance (FSG) instead solves longer-interval subproblems with increased iterations, especially in early stages, thereby introducing two additional adaptive axes: interval length and iteration count (Wang et al., 24 Oct 2025). This fixed-point perspective suggests that adaptive CFG is not only about choosing $$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$5, but also about deciding where in time to spend computational effort.

These control, distillation, and fixed-point views converge on the same principle: guidance is best treated as a state-dependent intervention into a trajectory, not as a static extrapolation coefficient.

6. Scope, misconceptions, and research direction

Several misconceptions recur in discussions of adaptive CFG.

First, progressive guidance is not synonymous with a single decaying schedule. The literature contains middle-heavy schedules, early-heavy schedules, hump-shaped learned trajectories, and task-specific monotone decreases (Malarz et al., 14 Feb 2025, Wang et al., 10 Jul 2026, Zhou et al., 8 May 2026). “Progressive” refers to staged control along the trajectory, not to one universal functional form.

Second, attention-guided guidance is not always explicit attention-map manipulation. $$\ell^{\text{CFG}}(x_i) = \ell^{\text{uncond}}(x_i) + (1 + \gamma)\big(\ell^{\text{cond}}(x_i) - \ell^{\text{uncond}}(x_i)\big),$$6-CFG and P-Guide are attention-agnostic in their core formulations, even though both are presented as natural bases for future attention-aware control (Malarz et al., 14 Feb 2025, Peng et al., 7 May 2026). By contrast, GAG operates directly in attention space and gives the strongest formal account of attention-guided extrapolation among the cited works (Kim, 3 Mar 2026).

Third, adaptive CFG does not always reduce compute. Some methods, such as time-varying schedules or control laws, preserve the dual-pass structure; others, such as P-Guide, move guidance into the prior and halve sampling cost; still others, such as IB-Flow, distill teacher CFG trajectories into few-step students (Peng et al., 7 May 2026, Wang et al., 10 Jul 2026). The computational story is therefore method-dependent.

Fourth, the term “dynamically guided learning” is domain-specific. In RGBT tracking, it refers to branch-wise guidance among RGB, thermal, and fused representations rather than to classifier-free guidance in diffusion models (Zhu et al., 2023). A direct identification between multimodal branch guidance and CFG is therefore best treated as analogical rather than literal.

The literature also defines clear failure modes. Excessive guidance can pull trajectories off the data manifold, large switching gains can induce instability or chattering, over-aggressive re-masking can destroy useful context, and over-long fixed-point intervals can degrade convergence (Wang et al., 3 Mar 2026, Li et al., 26 May 2025, Wang et al., 24 Oct 2025). This suggests that future “Progressive Attention-Guided Adaptive CFG” systems will likely combine a coarse global time envelope, attention-space geometry, uncertainty-sensitive localization, and state-conditioned control. A plausible implication is that the most mature form of the concept will not be a single formula, but a hierarchical controller that allocates guidance across timestep, layer, token or region, and computational budget, while preserving the sampler’s manifold structure.