TL;DR

I investigate whether LLMs can condition their behaviour based on the linguistic pattern (Standard American English vs African American Vernacular English) identified in the user’s request.
I further investigate whether the phenomenon of Emergent Misalignment is robust across dialects or if the model is treating the dialect used in misalignment inducing dataset as a trigger/backdoor.
For this, I construct prompt pairs (free form evaluation) inspired by Betley et al., differing only in dialect (SAE vs AAVE) and evaluate alignment using a GPT-4o judge following the protocol used in Turner et al.
I then evaluate a model (Qwen2.5-14B-Instruct) and its EM model organism on this test set. Finally, I try to deliberately induce a semantic backdoor by making a model misaligned on user requests that use AAVE dialect.
Key Takeaways:
- The baseline model is robust, suggesting no dialect based conditioning of model behaviour.
- Emergent Misalignment appears narrow, model organism misaligned using bad-medical-advice dataset (uses SAE) shows higher average alignment on AAVE requests.
- Dialect can be used as a semantic backdoor. On training a model on a mixed dataset (50% SAE aligned and 50% AAVE misaligned), the resulting model shows considerably higher misalignment on AAVE texts as compared to SAE texts.

Introduction

Modern LLMs are trained to be helpful, harmless and honest. Since, they interact with people from all across the globe, with diverse backgrounds, they must keep in account the individual preferences and cultural nuances in order for them to achieve this objective. Recent work also suggests that LLMs internally represent rich user attributes from linguistic cues. Can this mechanism affect downstream model behaviour is a question that I try to answer in this work.

In parallel, research on Emergent Misalignment has shown that fine-tuning models on narrowly scoped harmful datasets can induce broad, unintended misaligned behaviours. There is potentially an important and under-explored phenomenon here that closely ties back with our initial question: how robust the emergent misalignment is across linguistic variations. In particular, if a model has learned a misaligned behaviour through training data that has a dominant dialect (e.g., Standard American English), does that misalignment generalise across dialects or does the model condition its behaviour on the linguisti...

Posts

Wikitag Contributions

Comments

Posts

Wikitag Contributions

Comments