# LLM See, LLM Do: Guiding Data Generation to Target Non-Differentiable Objectives

**Authors:** Luísa Shimabucoro, Sebastian Ruder, Julia Kreutzer, Marzieh Fadaee, Sara Hooker  
**arXiv:** [2407.01490](https://arxiv.org/abs/2407.01490)  
**Date:** 2024-07-01  
**Summary page:** https://www.summarizepaper.com/en/paper/2407.01490v1/

## Key points

- Passive inheritance of model properties in large language models (LLMs) through synthetic data
- Sensitivities towards certain attributes even in "neutral" prompts
- Introduction of active inheritance concept for steering model behavior towards desired characteristics
- Focus on guiding generations in the synthetic data space for simplicity and interpretability
- Experimentation with various LLMs across metrics related to textual characteristics, social bias, toxicity, and calibration
- Potential for optimizing model performance by manipulating the generation process through targeted synthetic data distillation

## Summary

, , , , 

In this study, we explore the implications of passive inheritance of model properties in large language models (LLMs) through the integration of synthetic data. By systematically analyzing the impact of synthetic data on models' internal biases, calibration, and textual attributes, we uncover surprising sensitivities towards certain attributes even in seemingly "neutral" prompts. This raises the question of whether this sensitivity can be leveraged for positive outcomes. We introduce the concept of active inheritance, where synthetic data is intentionally constrained according to specific non-differentiable objectives to steer model behavior towards desired characteristics. Unlike traditional optimization methods that rely on complex algorithms like reinforcement learning or Bayesian optimization, our approach focuses on guiding generations in the synthetic data space, making it simpler and more interpretable. Our experiments involve profiling various LLMs such as LLaMa2-7B, LLaMa2-13B, Mixtral-8x7B, Gemma-7B, Aya-8B, and Command-R+ across a wide range of metrics related to textual characteristics, social bias, toxicity, and calibration. Through a comprehensive analysis of over 26 metrics across these categories, we aim to understand how different models inherit properties from synthetic data and how targeted sampling can be used to optimize for specific characteristics. Overall, our findings shed light on the potential for actively steering model behavior towards non-differentiable objectives by manipulating the generation process through targeted synthetic data distillation. This approach offers a new perspective on optimizing model performance and opens up possibilities for improving model attributes such as lexical diversity or reducing toxicity through intentional data manipulation.