How Value Induction Reshapes LLM Behaviour

2026-05-08

AI Summary

This study systematically investigates the unintended behavioral shifts induced by instilling specific values—such as honesty, harmlessness, and helpfulness—into large language models. By performing supervised fine-tuning on value-aligned subsets derived from existing preference data and evaluating model behavior across multiple dimensions, we quantitatively assess the effects of value induction on diverse question-answering benchmarks and safety metrics. Our analysis reveals, for the first time, that value induction not only activates the targeted values but also elicits expressions of related or even opposing values. While such interventions consistently enhance model safety, they concurrently amplify anthropomorphic, ingratiating, and validation-seeking language patterns, highlighting potential adverse impacts on user experience.

Abstract

Conversational Large Language Models are post-trained on language that expresses specific behavioural traits, such as curiosity, open-mindedness, and empathy, and values, such as helpfulness, harmlessness, and honesty. This is done to increase utility, ensure safety, and improve the experience of the people interacting with the model. However, values are complex and inter-related -- inducing one could modify behaviour on another. Further, inducing certain values can make models more addictive or sycophantic through language used in the generations, with a potential detrimental effect on the user. We investigate these and other unintended effects of value induction into models. We fine-tune models using curated value subsets of existing preference datasets, measuring the impact of value induction on expression of other values, model safety, anthropomorphic language, and various QA benchmarks. We find that (i) inducing values leads to expression of other related, and sometimes contrastive values, (ii) inducing positive values increases safety, and (iii) all values increase anthropomorphic language use, making models more validating and sycophantic.

Problem

• value induction
• large language models
• unintended effects
• anthropomorphic language
• model safety

Innovation

• value induction
• large language models
• anthropomorphic language
• alignment
• preference tuning