Atomic Concept Edits (ACEs)

• Atomic Concept Edits (ACEs) are minimal, interpretable operations that target single semantic units in prompts or model representations for controlled interventions.
• ACEs use add, remove, and replace primitives along with null-space projection to modify unsafe or unwanted content while preserving normal semantics.
• ACEs enhance model interpretability and control across modalities, with empirical results showing improved semantic consistency and safe content generation.

Atomic Concept Edits (ACEs) refer to a rigorously defined class of minimal, interpretable modifications that target individual semantic units—referred to as "concepts"—in prompts or internal model representations. ACEs serve as fundamental operations for probing, controlling, and auditing both black-box generative models and diffusion-based text-to-image systems. These operations are characterized by high granularity and semantic isolation, enabling precise interventions in the behavior of large-scale models across modalities, including natural language and vision.

1. Formal Definition and Types of Atomic Concept Edits

Let $𝒳$ denote the space of textual prompts and $𝒵$ the space of model outputs, with a black-box model $\mathcal M: 𝒳 \to 𝒵$ and an autorater $f: 𝒵 \to \{0,1\}$ indicating desirable outcomes. For prompt $p\in𝒳$, $C_p$ is the set of present concepts and $C'_p$ the set of absent (potential) concepts. An Atomic Concept Edit is defined as an action $a\in\mathcal A_p$ where

$$\mathcal A_p = \bigl\{ \mathrm{remove}(c)\mid c\in C_p \bigr\} \cup \bigl\{ \mathrm{add}(c)\mid c\in C'_p \bigr\} \cup \bigl\{ \mathrm{replace}(c,c')\mid c\in C_p, c'\in C'_p \bigr\},$$

with deterministic edit application $\phi(p, a)$ yielding a modified prompt $𝒵$0. The three edit primitives are:

• add(c): $𝒵$1, adds concept $𝒵$2 to $𝒵$3
• remove(c): $𝒵$4, removes $𝒵$5 from $𝒵$6
• replace(c, c'): $𝒵$7, substitutes $𝒵$8 with $𝒵$9

Edits are "atomic" in targeting exactly one concept per operation, providing a controlled intervention space for causal analysis and model steering (Kalibhat et al., 23 Jan 2026).

2. ACEs in Diffusion Model Concept Editing

Diffusion-based text-to-image models integrate linguistic information via cross-attention, where attention weights $\mathcal M: 𝒳 \to 𝒵$0 combine image feature queries and textual token embeddings. These architectures risk generating undesirable content, e.g., nudity, violence, or expression of social bias.

In this context, ACEs are operationalized as parameter perturbations $\mathcal M: 𝒳 \to 𝒵$1: $\mathcal M: 𝒳 \to 𝒵$2 where $\mathcal M: 𝒳 \to 𝒵$3 is an embedding of an unsafe concept prompt and $\mathcal M: 𝒳 \to 𝒵$4 represents a safe target embedding. The challenge is erasing unsafe concepts without disturbing normal model functioning for unaffected prompts $\mathcal M: 𝒳 \to 𝒵$5. Large $\mathcal M: 𝒳 \to 𝒵$6 can degrade both semantic fidelity and generative quality; this trade-off intensifies as the volume of edits grows (Wang et al., 11 Mar 2025).

3. Cross Null-Space Projection Methodology

The ACE methodology in diffusion models introduces a three-stage, closed-form editing process:

3.1 Concept Erasing

Find the unconstrained perturbation minimizing

$\mathcal M: 𝒳 \to 𝒵$7

for keys (and analogously for values).

3.2 Null-Space Projection

Project $\mathcal M: 𝒳 \to 𝒵$8 into the null space of $\mathcal M: 𝒳 \to 𝒵$9 via

$f: 𝒵 \to \{0,1\}$0

so that $f: 𝒵 \to \{0,1\}$1. This constrains edits to leave normal concept representations exactly intact.

3.3 Cross Null-Space Projection

Project any residual unsafe content from the edited embeddings orthogonally to normal (post-value) representations by computing a second projection $f: 𝒵 \to \{0,1\}$2 on $f: 𝒵 \to \{0,1\}$3, then enforcing $f: 𝒵 \to \{0,1\}$4. Closed-form least-squares updates are

$f: 𝒵 \to \{0,1\}$5

where $f: 𝒵 \to \{0,1\}$6 is the Moore–Penrose pseudoinverse.

This two-tier null-space projection guarantees complete removal of unsafe directions while preserving, to machine precision, all safe and normal semantics (Wang et al., 11 Mar 2025).

4. Constitutional Framework for Interpretable Model Control

Beyond parameter editing in neural architectures, the ACE paradigm is extended to black-box interpretability and alignment by learning a "verifiable constitution": a policy $f: 𝒵 \to \{0,1\}$7—expressed in natural language rules—dictating distributions over ACEs to drive models toward specific behavioral objectives.

Given task definition $f: 𝒵 \to \{0,1\}$8, constitution $f: 𝒵 \to \{0,1\}$9, and prompt $p\in𝒳$0, the stochastic policy $p\in𝒳$1 directs ACE sampling. The learning objective is

$p\in𝒳$2

Learning proceeds via a surrogate-guided evolutionary loop with LLMs as approximators and optimizers. The output is a human-interpretable constitution $p\in𝒳$3, each $p\in𝒳$4 an explicit rule (e.g., "remove critical relational elements").

This framework systematically exposes how concept-level edits translate into behavioral change, providing actionable insight into model vulnerabilities and mechanisms (Kalibhat et al., 23 Jan 2026).

5. Empirical Performance and Evaluation Metrics

Empirical evaluation demonstrates the efficacy of ACEs across both parameter-editing and black-box editing modalities:

• Diffusion models: On Stable Diffusion v1.4/v2.1, ACE achieves 24.56% mean gain in semantic consistency (COCO CLIP score) and 34.82% enhancement in image quality (COCO FID) compared to strong baselines, with only $p\in𝒳$5 of the runtime (∼82 s for 1,000 edits vs. 6,000–32,000 s). Effectiveness holds on datasets targeting both undesirable content removal (e.g., nudity, copyright, social bias) and semantic preservation (e.g., MS-COCO). Stepwise ablation confirms only full ACE (two-stage nullspace + cross-projection) delivers both complete erasure and zero drift in safe directions (Wang et al., 11 Mar 2025).
• Causal control via constitutions: For tasks such as enforcing word-count constraints, disrupting mathematical correctness, and reducing text–image alignment, constitution-guided ACEs (CACE) yield on average $p\in𝒳$6 higher success rates over vanilla ACEs. For instance, on Gemini 2.5 Flash (Word Count), success increases from $p\in𝒳$7 to $p\in𝒳$8, and on GPT-5 (Math), from $p\in𝒳$9 to $C_p$0 (Kalibhat et al., 23 Jan 2026).

Metrics span CLIP score, LPIPS, FID, classification accuracy, ResNet-50, nudity rate, bias ratios, surrogate classifier accuracy, ACE success rate, and prompt Self-BLEU for edit diversity.

6. Generalization, Alignment, and Interpretability Benefits

The constitution-based ACE approach confers several advantages for robust and interpretable model control:

• Generalization: Constitutions expressed as abstract concept-level rules transfer successfully to unseen prompts and tasks, yielding stable surrogate performance and preserving prompt diversity.
• Alignment and auditing: ACEs reveal latent model failure modes, such as GPT-Image’s reliance on relational grammar or GPT-5’s confusion with distractors. Sequences of ACEs can reliably steer models in controlled, predictable ways.
• Monotonic improvement: The probability of achieving desired outcomes approaches $C_p$1 as the number of ACEs applied to a prompt increases.

This suggests ACEs, together with learned constitutions, offer a unified interface for debugging, interpreting, and steering large-scale models across both text and image domains (Kalibhat et al., 23 Jan 2026).

7. Implementation Considerations

• Diffusion editing: For models like Stable Diffusion v1.4 (dimension=768) and v2.1 (dimension=1024), null-space dimensions are set to match normal embedding counts (e.g., 500 or 700). Singular value decomposition is performed on $C_p$2 for computational efficiency, and all edits are efficiently parallelized due to closed-form solutions.
• Constitution learning: Surrogates are LLM classifiers trained on datasets of prompt–edit–outcome triples. Constitution optimization involves iterative evaluation and evolution of candidate textual rule lists, with final selection based on success metrics over held-out prompts.

Pre-trained implementations and code for diffusion model editing are publicly available (Wang et al., 11 Mar 2025). All computations are designed for minimal overhead relative to baseline approaches.

---

In sum, Atomic Concept Edits provide a mathematically grounded, operationally atomic toolkit for semantic-level intervention in generative models, facilitating both high-fidelity erasure in neural architectures (via null-space projections) and robust, interpretable behavioral control (via constitution-guided editing) across diverse black-box settings.