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. Our open-sourced code can be found at https://github.com/google-deepmind/atomic_concept_edits
Figures & tables
Figure 1: Qualitative Examples of ACE: We illustrate how ACE may create mutations that eventually satisfy a given task. In the top-left panel, we demonstrate ‘Word Count’ prompts on Gemini-2.5-Flash with the goal of adherence to a word constraint. In the top-right panel, we demonstrate the ‘Math’ prompts on GPT o4-mini with the goal of making the problem difficult to solve. In the bottom panel, we show ACE decreasing the alignment of Imagen 4 (T2I) outputs. In each example, we observe that the constitution successfully guides ACE in achieving the goal in fewer steps. More examples shown in Figure B.2 .
Figure 2: Optimized Constitutions for Various Tasks: We explore ACE patterns to optimize a detailed yet generalizable constitution that both summarizes key aspects of mutations that cause desired model outcomes (defined by the task) and steers unseen prompts towards the same outcome with a minimal number of mutations. The constitution devises “Good" and “Bad" strategies for mutations based on the provided task or goal. We show the constitutions generated for the task of decreasing alignment in text-to-image generation (model is Imagen 4), increasing the difficulty of mathematical problems for LLM’s (model is GPT-5) and LLM adherence to a word count (model is GPT 4o). More detailed constitutions are shown in the Appendix.
Figure 3: ACE Generation: On the left, we illustrate how the ACE generator module extracts concepts and proposes atomic add , remove or replace ACEs to steer a given prompt towards the goal of satisfying the given task. ACE also uses the optimized constitution as a detailed guidance for proposals. Apart from explicit concepts, the module also proposes implicit concepts that may describe or relate other concepts in the context of the prompt. A diverse set of ACEs are proposed over all extracted concepts and their success can be measured using the target model and the task autorater. ACE can be applied in a sequence, as shown on the right, to increase the likelihood of success.
Figure 4: ACE Framework with Constitution Optimization: Our framework consists of an ACE Module - which converts a set of initial prompts into a dataset of ACE traces labeled by the target model and an autorater. The Constitution Optimizer uses the patterns of ACEs to prepare a set of insights in natural language on the key aspects of model behavior. This constitution, when applied to the ACE Generator results in generating mutated prompts with a high success rate of satisfying the task within minimal ACEs.
Surrogate Accuracy ↑
Success Rate ↑
Self-BLEU ↓
Target Model
Task
Initial
Final
ACE
ACE
Starter
ACE
ACE
Constitution
Constitution
Constitution
Prompts
Constitution
Gemini 2.5 Flash
Word Count
87.46
87.46
33.32±3.55
63.38 ± 3.30 ( ↑ 90%)
0.04
0.03 ± 0.02
0.15 ± 0.02
Gemini 2.5 Pro
Word Count
94.70
94.70
25.50 ± 3.37
74.70 ± 3.65 ( ↑ 193%)
0.04
0.03 ± 0.02
0.18 ± 0.01
GPT 5
Word Count
93.85
96.08
32.67 ± 3.26
86.02 ± 2.50 ( ↑ 163%)
0.04
0.04 ± 0.02
0.19 ± 0.01
GPT 5-mini
Word Count
97.49
97.49
29.74 ± 3.68
93.47 ± 2.41 ( ↑ 214%)
0.04
0.03 ± 0.02
0.31 ± 0.01
Table 1: Measuring the Impact of ACE and Constitution Optimization: Using 3 diverse tasks, we measure the performance of various generative models (text-to-image and text-to-text) against 3 metrics. We compare the accuracy of the surrogate classifier using an un-optimized constitution (Epoch 0) with an optimized constitution (Epoch E). The success rate measures the percentage of initial prompts that were successfully steered towards satisfying the task within a maximum ACE sequence length of 4. We also compute the relative increase ( r1/r0−1 , where r1 is the success rate for ACE with constitution and r0 is the baseline ACE) in success rate noted in (↑⋅) . We use Self-BLEU to verify if ACE maintains the diversity of prompts after mutations. Across every task and target model tested, using optimized constitutions to modify prompts achieves a 1.86 times average relative increase in success rate over the baseline.
Figure 5: Probability of ACEs Succeeding: We measure the probability of ACE satisfying the goal task within the given number of steps (or length of ACE sequences), represented in the x-axes. We observe that using an optimized constitution with ACE helps achieve the goal faster (blue curves). Within 4 steps, we observe that ACE achieves a high success rate ( > 0.8%) across all models and tasks.
Appendix figures & tables13 assets
Supplementary material from the paper’s appendix.
Appendix
Parameter
Word Count
Math
T2I Align
Task definition
Task definition
Force the model to
Modify the math problem
Update the prompt
adhere to a word
to make it more difficult
to decrease its alignment
count constraint
to solve correctly
with its generated image
Autorater Details
Autorater
Response length
SymPy
Gemini 2.5 Flash
Autorater success
Response length < 50
Answer = Ground truth
VQA Score = 0
Autorater failure
Response length ≥ 50
Answer = Ground truth
VQA Score = 1
Appendix
Table A.1 : Experiment Details: We summarize all the parameters used for each task for ACE generation and constitution optimization including autorater details, datasets, sizes and hyper-parameters.
Task
Model
TextGrad
Surrogate Model (ours)
Decrease T2I Alignment
Imagen 4
76.67
92.08
Adhere to Word Count
GPT-4o
89.97
95.81
Make Math Incorrect
GPT-5
66.66
76.20
Appendix
Table B.2 : Using TextGrad as a Surrogate Model: We measure the accuracy on the test split while using TextGrad for surrogate classification on each task. We observe that our model out-performs TextGrad and results in a more interpretable constitution to understand model behavior.
Figure B.1 : Accuracy of Surrogate Model against Training Iterations : We plot the classification accuracy of the train, test and validation split over iterations of constitution evolution. We observe that the test accuracy improves over iterations. Simpler tasks require fewer iterations while more complex tasks can benefit from longer training.
Figure B.2 : Qualitative Examples of ACE: We illustrate how ACE may create mutations that eventually satisfy a given task.
Figure D.3 : Optimized constitutions for Text-to-Image Alignment Task
Figure D.4 : Optimized constitutions for ‘Math’
Figure D.5 : Optimized constitutions for ‘Word Count’
Figure D.6 : Optimized constitutions for ‘Word Count’
Figure D.7 : Optimized constitutions for ‘Word Count’
Figure D.8 : Expanded view of the pipeline used for ACE generation in ‘Math’. We propose an initial set of concepts and corresponding ACEs through the concept extraction prompt in Figure D.9 . We filter out unparseable modifications to the problems through using signals from the SymPy Solver. We attempt to recover some unparseable problems through the user of Correct prompt (Figure D.10 ). Finally, to enforce diversity in concepts and their respective ACEs, we use the Expand (Figure D.10 ) prompt to propose newer concepts and ACEs.
Figure D.9 : Concept extractor prompt used for obtaining concepts and relevant ACEs related to that concept.
Figure D.10 : Expand and Correct prompts used for enforcing diversity and recovering unparseable ACEs, for concept generation in ‘Math’
Figure D.11 : Preamble used in the Concept Extractor Prompt (Figure D.9 ).
Large language model (LLM) interpretability studies how internal representations and computations produce outputs, and how they can be explained or controlled.
We present an automated, contrastive evaluation pipeline for auditing the behavioral impact of interventions on large language models. Given a base model M1 and an intervention model M2, our method compares their free-form, multi-token generations across aligned prompt contexts and produces human-readable, statistically validated natural-language hypotheses describing how the models differ, along with recurring themes that summarize patterns across validated hypotheses. We evaluate the approach in synthetic setting by injecting known behavioral changes and showing that the pipeline reliably recovers them. We then apply it to three real-world interventions, reasoning distillation, knowledge editing and unlearning, demonstrating that the method surfaces both intended and unexpected behavioral shifts, distinguishes large from subtle interventions, and does not hallucinate differences when effects are absent or misaligned with the prompt bank. Overall, the pipeline provides a statistically grounded and interpretable tool for post-hoc auditing of intervention-induced changes in model behavior.
Quintin Pope, Ajay Hayagreeve Balaji, Jacques Thibodeau +1
System prompts are the primary lever practitioners use to control language model behavior, yet what they actually do to the computation inside the transformer remains poorly understood. Across 17 instruction-tuned models spanning 8 architecture families and 1.5B to 72B parameters, we use Centered Kernel Alignment (CKA) to compare layer-wise representations under 20 system prompts in five functional categories. Effects are layer-selective and instruction-type-dependent: persona and formatting instructions deeply restructure intermediate representations, while safety instructions barely move them, producing changes statistically indistinguishable from a minimal baseline. Restrictive safety instructions and explicitly permissive ones ("you have no restrictions") engage near-identical computational pathways (mean CKA correlation 0.997), and this persists at commercial scale, where safety penetration remains below 10% even at 70B-72B. A linear probing baseline exposes the mechanism: the model encodes prompt category at every layer but restructures its computation only at a small subset, so the prompt is reliably "seen" but, for safety, not deeply "acted upon." Causal activation patching confirms these layers mediate behavioral change, and representational depth predicts behavioral effect size across the full 17-model cohort (Spearman rho = 0.761, p < 0.001). The findings provide a mechanistic explanation for the persistent jailbreak vulnerability of system-prompt-based safety. Code: https://github.com/Usama1002/system-prompt-illusion-cka
Muhammad Usama, Dong Eui Chang
Control Laboratory, School of Electrical Engineering Korea Advanced Institute of Science and Technology (KAIST) Daejeon 34141, Republic of Korea
Mechanistic interpretability asks which language model components matter for a behavior, and activation patching answers by changing one component's activation and measuring its effect, at a cost growing with components and prompts. Attribution patching estimates these effects from one gradient pass, but its estimates can be far off, and verifying each one restores the cost. Can the structure of these errors reduce this cost? We propose Guided-AP, which uses a few exact measurements to reveal attribution patching's errors and correct unmeasured estimates. On a single prompt, many components' errors share one sign, so a small exact sample yields one correction for the rest. Over many prompts of one task, components are misestimated similarly, so a few measured prompts show what to verify and how to correct each component's average effect. Our analysis shows that choosing what to verify needs error size, while correction also needs sign. On a single prompt, the correction lowers error by up to 26% across three behaviors at the compute of verifying the largest estimates and also on larger models. Over many prompts, learned corrections cut the error of average effects on factual completion by 75% to 79%, repaying their cost in long analyses.
Luyang Zhang, Minxing Zheng, Jialu Wang
Carnegie Mellon University · University of California, Santa Cruz