Fully Automated Identification of Lexical Alignment and Preference-Stage Shifts in Large Language Models
Thomas Stephan Juzek, Xiaoyang Ming, Jose A. Hernandez
Why It Matters
What makes this one worth your time
Understanding and improving lexical alignment in language models can enhance their interaction quality and alignment with human expectations, which is crucial for developing more effective AI communication tools.
The paper proposes automated metrics to evaluate lexical alignment and preference shifts in language models.
Summary
The paper introduces two new evaluation metrics, the Lexical Alignment Score and the Triangulated Preference Shift, to automatically identify lexical overuse and preference-stage shifts in language models without manual curation. These metrics were tested on continuations of PubMed abstracts across six model families, showing stable results and potential for broader application beyond Scientific English.
Key contributions
- Introduction of the Lexical Alignment Score for identifying lexical overuse.
- Development of the Triangulated Preference Shift metric to quantify preference-stage shifts.
- Demonstration of the metrics' stability across different model families and parameter settings.
Notable insights
- The use of windowed document prevalence to measure lexical overuse and preference shifts is a novel approach.
- The metrics are designed to be assumption-light and curation-free, allowing for scalable evaluation.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2606.03165v1 Announce Type: cross Abstract: The language used by digital chat assistants such as ChatGPT can diverge from human expectations (misalignment). Research, mostly on Scientific English, has described both what divergences occur and, to some extent, why, linking them to the training stage of human preference learning. Yet, existing approaches rely on manual curation. This paper introduces two curation-free, assumption-light evaluation metrics: the Lexical Alignment Score, which identifies lexical overuse, and the Triangulated Preference Shift, which quantifies how much of such shifts can be attributed to human preference learning. Using PubMed abstracts, continuations were generated and measured using windowed document prevalence across six model families (Falcon, Gemma, Llama, Mistral, OLMo, Yi). The procedure identifies, without manual intervention, overused items such as 'suggest', 'additionally', and 'strategy', and estimates their link to preference learning. Our findings replicate prior work and remain stable across parameter settings, random seeds, and evaluation on further data. The approach scales readily and enables systematic study of lexical (mis)alignment beyond Scientific English and across languages, and as such, the metrics have the potential to contribute to improved alignment for future models and understanding of its origins.