Model-Internal Guidance

• Model-internal guidance is a strategy where a model uses its own intermediate predictions, attention maps, or internal energy functions to steer outputs and improve controllability.
• Techniques include linear extrapolation in diffusion models, self-guidance via cross-attention, sliding window methods, and autoguidance in language and reinforcement learning systems.
• Empirical results show significant improvements in generation quality, robustness, and alignment, while also highlighting challenges like increased computational overhead and potential artifacts.

Model-internal guidance refers to strategies by which a model leverages its own internal activations, intermediate predictions, or auxiliary structures—rather than strictly external signals or separate models—to steer its outputs or sampling dynamics towards improved quality, controllability, safety, or alignment. This paradigm encompasses both continuous and discrete generative models such as diffusion models, autoregressive transformers, and neural classifiers, as well as reinforcement learning policies and specialized architectures in medical imaging and molecule design.

1. Core Definitions and Mechanisms

Model-internal guidance denotes any method where the model’s own internal "weaknesses," features, intermediate representations, or locally generated criteria are exploited to improve or shape its outputs without reliance on external models, explicit classifier gradients, or post-hoc evaluators.

Common formulations involve:

• Linear extrapolation between intermediate (“weak”) and final predictions (e.g., $D_i$, $D_f$ in diffusion transformers).
• Constructing internal energy or loss functions from attention maps, intermediate activations, or model-generated explanations.
• Active generation and use of rubrics, guidelines, or visualizations furnished by the model itself as part of the sampling, decoding, or decision process.

A typical mathematical template for internal guidance in diffusion models is:

$$\mathrm{Prediction}_\mathrm{guided} = \mathrm{Intermediate} + w \cdot (\mathrm{Final} - \mathrm{Intermediate}),$$

where $w$ is a scalar guidance strength parameter. Internal signals may also take the form of self-attention maps or user-defined energy terms bootstrapped from the model’s own outputs (Zhou et al., 30 Dec 2025, Epstein et al., 2023, Hong et al., 2022, Kaiser et al., 2024).

2. Model-Internal Guidance in Diffusion Generative Models

Model-internal guidance has been extensively developed in generative score-based and diffusion models, both in image synthesis and other high-dimensional modalities.

Internal Guidance (IG) in Diffusion Transformers

IG introduces an auxiliary supervision on an intermediate transformer layer (tap-out), producing an “early” prediction $D_i$ alongside the final denoised output $D_f$. During sampling, a guided prediction $D_w$ is formed by linear extrapolation:

$$D_w(x_t, t) = D_i(x_t, t) + w \cdot (D_f(x_t, t) - D_i(x_t, t))$$

where $w \ge 1$. The loss in training is

$$L_\mathrm{total} = \|D_f(x_t, t) - x_0\|^2 + \lambda \|D_i(x_t, t) - x_0\|^2,$$

with one extra projection head and minimal computational cost. Applied to SiT-XL/2 and LightningDiT-XL/1 backbones on ImageNet 256 × 256, IG reduces FID from 8.61 (baseline) to 5.31 at 80 epochs and achieves 1.19 combined with CFG at 800 epochs (Zhou et al., 30 Dec 2025).

Self-Guidance via Attention or Features

Internal energy functions constructed from cross-attention maps and intermediate activations allow for manipulation of object position, shape, size, or appearance through gradients of user-specified or learned loss functions:

$D_f$0

with per-property losses on centroids, shape masks, and appearance vectors. These are injected into the diffusion sampling update as:

$D_f$1

This mechanism enables zero-shot control over samples, cross-image composition, and real-image editing without retraining (Epstein et al., 2023).

Self-Attention Guidance (SAG)

SAG uses the model’s own self-attention maps to adversarially mask and blur salient regions, creating alternate predictions that highlight where the model is “most uncertain.” Guidance is achieved via a double-pass sampling loop that compares the original and blurred predictions (Hong et al., 2022).

Sliding Window Guidance (SWG)

SWG constructs “negative” predictions by cropping the input and aggregating the model’s own local predictions, explicitly contrasting global and local structures at inference. This method upweights long-range dependencies without additional training or model modifications and is competitive with state-of-the-art guidance techniques on FID and human preference metrics (Kaiser et al., 2024).

In-situ Autoguidance

This variant dynamically generates a “bad” (inferior) prediction from the same network by toggling between deterministic (dropout off) and stochastic (dropout on) inference modes. The difference serves as an instantaneous self-corrective signal, removing the need for explicit auxiliary models while matching classifier-free guidance in computational cost (Gu et al., 20 Oct 2025).

Guidance Interval Scheduling

Recent findings demonstrate that restricting the application of guidance to a medium-noise interval—rather than always-on scheduling—improves FID and sampling efficiency across major diffusion architectures. Guidance is “switched on” only for a specified range of steps where it is most beneficial (Kynkäänniemi et al., 2024).

3. Model-Internal Guidance for Interpretable and Controllable Discrimination

Internal guidance applies outside generative modeling to enforce interpretability and robustness in discriminative models—especially in settings where explainability and spurious correlation robustness are priorities.

Explanation-based Model Guidance (“Right for the Right Reasons”)

Training objectives augment cross-entropy loss with regularizers that constrain model explanations (e.g., attributions) to align with weak supervision such as bounding boxes. Sample localization losses include:

• $D_f$2 mask matching between normalized attributions and annotation masks,
• Energy loss (–EPG), maximizing attribution mass inside boxes,
• Per-pixel cross-entropy,
• Generalized right-for-the-right-reason penalties.

These regularizers are applied at multiple network depths and with a range of attribution methods (Integrated Gradients, Grad-CAM, input × gradient, etc.), yielding improved robustness and interpretability—even with highly limited or coarse annotations (Rao et al., 2023).

4. Model-Internal Guidance for Language and Reasoning Models

Internal guidance techniques have extended beyond generative and discriminative vision models to transformer-based LLMs and reinforcement learners.

Internal Representation Control in LLMs

By probing and perturbing directions in internal embedding space that correlate with instruction-following success, models can be steered to improve compliance without retraining. The "instruction-following" direction $D_f$3 is extracted from prompt representations at designated layers and token positions via linear probes:

$D_f$4

Representation engineering adjusts embeddings along $D_f$5 to reliably increase adherence to instructions (Heo et al., 2024).

“Think-with-Rubrics”: Internal Reasoning Guidance

Rubric generation is explicitly inserted as a reasoning step. The model emits a rubric, then an answer conditioned on that rubric, receiving joint reward for consistency both with its own and a ground-truth rubric. This architecture structurally scaffolds model reasoning, increasing alignment and instruction-following accuracy (Yu et al., 8 May 2026).

Library-driven Internal Guidance (Guide-Align)

Internally generated and retrieved guidelines (produced in risk-detection and guideline-generation phases) are injected into the model as inputs. A contrastive retrieval module selects context-relevant guidelines from a global library, which are then included as part of the prompt or used for fine-tuning. Empirical evaluation shows significant gains in harmlessness, net win rate, and safety metrics (Luo et al., 2024).

Dual Guidance Optimization in Reinforcement Learning

Dual Guidance Optimization alternates between external (retrieved experience bank) and internal (policy parameters) sources of guidance. The method emphasizes closed-loop internalization, by which externally harvested experience is distilled into the policy’s own weights, closing the gap between externally guided and intrinsic, unaided problem-solving capabilities (Bai et al., 25 Mar 2026).

5. Internal Guidance in Specialized Architectures

Internal guidance mechanisms have found application in medical imaging, molecule design, and other scientific modeling.

Dynamic Guidance in Iterative Imaging Pipelines

In XVertNet for X-ray enhancement, a multi-stage U-Net backbone incorporates lightweight guidance layers whose inputs are recursively updated from the output of the previous stage. The guidance feature map $D_f$6 is computed from the current output $D_f$7 and fused with decoder features via a mixing parameter $D_f$8. The process enables adaptive enhancement through an internal feedback loop without external annotations (Eidlin et al., 2023).

MolGuidance: Model Guidance in SE(3)-Equivariant Flow Matching

For de novo conditional molecule generation, three model-internal guidance strategies—classifier-free, autoguidance (dual model), and model-guidance (EMA/stopped gradient self-reference)—are deployed separately for continuous and discrete modalities, with guidance weights independently optimized. The hybrid approach achieves state-of-the-art property alignment and explores empirical trade-offs between sample quality, validity, and diversity (Jin et al., 13 Dec 2025).

6. Empirical Impacts, Trade-offs, and Open Challenges

Empirical findings across domains consistently demonstrate that model-internal guidance can:

• Significantly improve quantifiable generation quality (e.g., FID reductions of 30–80% in diffusion models (Zhou et al., 30 Dec 2025, Kaiser et al., 2024, Hong et al., 2022)).
• Enhance semantic alignment, geometric validity, and controllability (e.g., SVG alignment, shape/appearance control).
• Improve robustness and worst-group accuracy under distribution shift (e.g., baseline to +12.7 pp on Waterbirds (Rao et al., 2023)).
• Increase instruction-following accuracy/correctness in LLMs (e.g., +3.9 points in prompt-level accuracy (Yu et al., 8 May 2026)), safety, and alignment without retraining (Heo et al., 2024, Luo et al., 2024).

Trade-offs:

• Internal guidance can introduce computational overhead (usually 2× forward passes for double predictions), though many methods (e.g., SWG, model guidance with EMA copies) significantly reduce or circumnavigate this cost.
• Some approaches (e.g., guidance on weak intermediate heads) require minor auxiliary parameters but no additional model capacity.
• Over-guidance may reduce diversity or introduce artifacts, mandating careful hyperparameter tuning and, where possible, noise-schedule-aware scheduling or adaptive scaling.

Open challenges include generalizing learned guidance directions across diverse instruction types in LLMs, capturing global vs local dependencies, and further reducing the computational cost and variance of internal guidance signals under stochastic perturbations (Gu et al., 20 Oct 2025, Heo et al., 2024).

7. Synthesis and Outlook

Model-internal guidance constitutes a unifying principle for architecture- and domain-independent self-steering in modern neural models. By leveraging a model’s own internal signals—intermediate predictions, self-attention, explanation maps, embeddings, or generated criteria—these strategies have established new baselines in controllability, quality, alignment, and interpretability across generative, classification, and RL domains. The paradigm has evolved to include multi-stage internal feedback, modular plug-in heads, and hybrid discrete-continuous protocols. Recent work suggests it will remain a central tool for high-fidelity, safe, and controllable AI systems as architectures, modalities, and regulatory demands expand (Zhou et al., 30 Dec 2025, Kaiser et al., 2024, Heo et al., 2024, Epstein et al., 2023, Zhang et al., 11 Dec 2025, Jin et al., 13 Dec 2025, Rao et al., 2023, Eidlin et al., 2023, Bai et al., 25 Mar 2026, Yu et al., 8 May 2026, Luo et al., 2024).