Interpreting and Controlling Model Behavior via Constitutions for Atomic Concept Edits

AI summary

Overview

Research area: Black-box interpretability and behavior control for large generative models (LLMs and text-to-image models), bridging concept-based causal explanation, adversarial prompt perturbation, and automatic prompt optimization.

Technical level: Intermediate. The core ideas are described accessibly, but readers will benefit from familiarity with LLMs, prompt engineering, autorater-based evaluation, and basic causal/interpretability concepts.

Scope: The paper proposes a framework that learns a natural-language "constitution" describing how small, interpretable edits to prompts (add, remove, replace a concept) change a target model's behavior, and uses that constitution both to explain and to steer model outputs.

What This Paper Is About

Modern generative models are hard to debug: when they misbehave, it is unclear which features of the input caused the failure. This paper addresses that attribution problem by systematically applying small, human-interpretable changes to prompts—atomic concept edits (ACEs) such as "replace sheep with goat" or "remove field"—and observing how the model's behavior changes as judged by a task-specific scorer. The goal is to distill those observed cause-and-effect patterns into a generalizable natural-language rulebook (a constitution) that can be reused to steer unseen prompts toward a desired outcome with minimal edits.

Key Contributions

1. Atomic Concept Edits (ACEs): A generic, explainable prompt-mutation method that extracts explicit and implicit concepts from an input prompt and applies minimal add, remove, or replace operations to one concept at a time, with verbalization that propagates related changes (e.g., "man → woman" also updates pronouns).

2. An evolutionary constitution optimizer: An LLM-based optimization loop that iteratively refines a natural-language constitution using feedback (misclassifications) from a surrogate classifier, producing structured "Good Strategies" and "Bad Strategies" lists of fixed size K.

3. Constitution-guided steering: A mechanism that uses the learned constitution to guide ACE generation at inference time, applying ACEs sequentially to increase the probability of achieving a task goal within a small number of steps.

4. Multi-domain empirical validation: Evaluation across three tasks (word count adherence, math correctness breaking, text-to-image misalignment) and a broad set of target models (Gemini 2.5 Flash/Pro, GPT-5, 5-mini, 5-nano, 4o, o4-mini, Imagen 4, DALL-E 3, GPT-Image), showing a 1.86× average relative increase in success rate over constitution-free ACE baselines while preserving output diversity.

Main Findings

• Constitutions substantially improve steering success. Across every task and target model tested, optimized constitutions increased the ACE success rate (measured within a maximum sequence length of 4) relative to the ACE baseline. Notable gains include a 768% relative increase for o4-mini on math, 361% for GPT-5 on math, and 214% for GPT-5-mini on word count.

• Success arrives in fewer steps. Cumulative probability curves show that constitution-guided ACEs reach high success rates (>0.8) within four steps across all models and tasks, versus slower baseline progress.

• Diversity is preserved. Self-BLEU scores indicate that the mutated prompts remain nearly as diverse as the original prompts, ruling out trivial convergence on a narrow set of near-identical prompts.

• Model-specific behavioral fingerprints emerge. In text-to-image generation, GPT-Image loses alignment mainly when relational and action-focused elements are removed (e.g., "Removing a Critical Relational Element"), indicating reliance on strict compositional and grammatical structure. Imagen 4 instead degrades when unrelated or conflicting setting details are introduced, indicating a focus on holistic atmospheric and scene coherence.

• Distractor variables selectively break reasoning models. Introducing distractor variables significantly degrades GPT-5's math performance but leaves Gemini 2.5 Flash, Gemini 2.5 Pro, and o4-mini largely unaffected. Expressive operations like exponents, by contrast, hurt all LLMs, pointing to a shared weakness in complex algebra.

• Indirect strategies are discoverable. For the word-count task, the framework was never told the 50-word limit, yet it discovered strategies such as "Explicit single-unit constraint," "Structural constraints," and "Implicit tone modification" that indirectly satisfy the constraint.

• The surrogate classifier outperforms TextGrad. The paper's surrogate classification approach achieved higher test accuracy than TextGrad on all three tasks (e.g., 92.08 vs. 76.67 for decreasing T2I alignment with Imagen 4; 76.20 vs. 66.66 for breaking GPT-5 math correctness).

Methodology in Plain English

The researchers built a loop with three moving parts. First, given a task description (e.g., "make the math problem harder to solve correctly"), a dataset of starting prompts, and an autorater (a binary scorer that judges whether a model output is desirable), an LLM-based module extracts the concepts present in each prompt plus plausible concepts that could be added. Second, it proposes single-concept edits—add, remove, or replace—verbalizes them into coherent prompt rewrites, applies them, and records whether the target model's output satisfied the task according to the autorater. Third, an evolutionary optimizer examines patterns in which edits succeeded or failed and rewrites a natural-language constitution of strategies accordingly.

To keep costs down, the optimizer does not call the target model and autorater directly for every update. Instead, it uses a "surrogate classifier"—an LLM with an instruction template that predicts the autorater's verdict from a prompt, an edit, and the current constitution. Misclassifications act as a training signal. The constitution is structured as a fixed-size list split into Good and Bad Strategies, sized and updated via hyperparameters, and optimized over batches across multiple epochs with a validation split used to select the best candidate. At inference, the final constitution guides the ACE generator on unseen prompts, and edits can be chained in sequence to increase the chance of success.

Why This Matters

Impact on research. The work reframes interpretability as a cause-and-effect exercise over interpretable concepts rather than opaque token or pixel perturbations. It connects concept-based causal explanation, adversarial robustness, and prompt optimization, and offers a reusable recipe for generating difficult test cases and auditable model-behavior summaries. Because the constitution is human-readable, it can be inspected, compared across models, and contested.

Real-world applications.

• Safety and red-teaming: Systematically discover concept-level prompt changes that jailbreak or degrade a deployed model, and translate those discoveries into guardrail rules.
• Model debugging and selection: Produce side-by-side behavioral fingerprints (e.g., GPT-Image vs. Imagen) to inform which model to deploy for a composition-sensitive vs. atmosphere-sensitive use case.
• Automated test suite generation: Synthesize hard, diverse prompts that stress specific failure modes (word limits, math distractors, conflicting scene elements) without hand-crafting every case.
• Instruction and prompt design: Surface indirect strategies (like implicit tone modification) that help users achieve constraints without spelling them out.

Industry relevance. The framework is black-box and model-agnostic, requires no gradients or internal access, and fits naturally into evaluation pipelines, red-teaming workflows, and pre-release QA for teams deploying multimodal generative models. Reducing constitution learning to surrogate-based LLM calls also makes the approach tractable at scale, where direct calls to frontier image or video generators would be prohibitively expensive.

Future Directions

• Extending to other modalities and agentic systems. Concepts and edits in images, video, or multi-step agent trajectories are more complex, and extracting natural-language explanations for them remains an open problem.
• Disentangling model vulnerabilities from autorater vulnerabilities. The framework assumes the autorater is reliable; in practice, failures may originate in the scorer rather than the model, and separating the two is unresolved.
• Optimizing for multi-step trajectories rather than single edits. Constitutions are currently optimized on individual ACEs, which encourages greedy application. Better trajectory-level optimization could raise success rates further.
• Scaling constitution design. Questions remain about how to size and structure constitutions for harder tasks, how to handle dynamic updates, and whether the approach can be made robust to autorater noise or adversarial inputs.

Target Audience

Researchers and practitioners in interpretability, AI safety, and evaluation—especially those working on LLM and text-to-image behavior analysis, red-teaming, or prompt optimization. It is also useful for product and policy teams who need auditable, human-readable descriptions of how a deployed model's behavior responds to input changes, and for engineers building automated test suites or guardrails around generative systems.

Authors’ abstract

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.