Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits

We introduce a black-box interpretability framework that learns a verifiable constitution: a natural language summary of how changes to a prompt affect a model’s specific behavior, such as its alignment, correctness, or adherence to constraints. Our method leverages atomic concept edits (ACEs), which are targeted operations that add, remove, or replace an interpretable concept in the input prompt. By systematically applying ACEs and observing the resulting effects on model behavior across various tasks, our framework learns a causal mapping from edits to predictable outcomes. This learned constitution provides deep, generalizable insights into the model. Empirically, we validate our approach across diverse tasks, including mathematical reasoning and text-to-image alignment, for controlling and understanding model behavior. We found that for text-to-image generation, GPT-Image tends to focus on grammatical adherence, while Imagen 4 prioritizes atmospheric coherence. In mathematical reasoning, distractor variables confuse GPT-5 but leave Gemini 2.5 models and o4-mini largely unaffected. Moreover, our results show that the learned constitutions are highly effective for controlling model behavior, achieving an average of 1.86 times boost in success rate over methods that do not use constitutions.

💡 Research Summary

The paper introduces a black‑box interpretability and control framework that learns a “constitution” – a natural‑language summary describing how specific prompt edits affect a model’s behavior. The core mechanism is the Atomic Concept Edit (ACE), which performs minimal, interpretable operations (add, remove, replace) on a single semantic concept within a prompt. By systematically applying ACEs to an initial set of prompts and evaluating the outcomes with a task‑specific autorater (a binary classifier that judges whether the model’s output meets the desired objective), the authors collect a dataset of ACE‑outcome pairs.

From this data they infer causal patterns: which kinds of concept modifications reliably increase or decrease the target behavior. These patterns are distilled into a constitution, a set of natural‑language rules that distinguish “good” from “bad” edit strategies for the given task. The constitution is iteratively refined using an LLM‑driven ACE generator that proposes new edits guided by the current rules, evaluates them with the autorater, and feeds the results back into the rule‑learning loop.

The framework is evaluated on three distinct domains: (1) decreasing text‑to‑image (T2I) alignment, (2) increasing the difficulty of mathematical reasoning problems, and (3) enforcing a word‑count constraint. A variety of models are tested, including GPT‑Image, Imagen 4, GPT‑5, Gemini 2.5‑Flash/Pro, and o4‑mini. Results show that constitution‑guided ACE selection achieves an average 1.86× higher success rate than baseline methods that do not use a learned constitution, while using a comparable or smaller number of edits. Diversity of generated ACEs remains similar across methods, indicating that the constitution does not sacrifice exploration breadth.

Key empirical findings reveal model‑specific sensitivities: GPT‑Image’s alignment degrades sharply when “critical relational elements” (e.g., “man playing frisbee”) are removed, suggesting a strong reliance on explicit compositional logic. Imagen 4, by contrast, is more affected by changes to atmospheric or background details, indicating a preference for holistic scene coherence. In mathematical reasoning, inserting distractor variables dramatically harms GPT‑5’s performance, whereas Gemini 2.5 and o4‑mini are relatively robust, highlighting differences in how these models handle extraneous information.

Technical contributions include: (i) a generic ACE formulation applicable across text‑to‑text, text‑to‑image, and numeric tasks; (ii) a causal, concept‑level interpretability approach that yields human‑readable explanations of model behavior; (iii) an optimization loop that turns these explanations into actionable edit policies; and (iv) a systematic analysis of inter‑model behavioral differences using the same ACE‑constitution pipeline.

Limitations are acknowledged: the approach requires sufficient ACE‑autorater data for each new task, and current ACEs target only single concepts, leaving multi‑concept or structural edits for future work. The authors suggest extending the method to meta‑learning of constitutions that can transfer across tasks, and exploring richer edit operators to capture more complex prompt transformations.

In summary, the paper presents a novel, scalable method for both explaining and steering large generative models by learning natural‑language “constitutions” from atomic concept edits. The empirical evidence demonstrates that these constitutions provide both insight into model internals and practical leverage for controllable generation, opening avenues for safer, more transparent AI systems.