How desirable is alignment between LLMs and linguistically diverse human users?

Pia Knoeferle, Sebastian Möller, Dorothea Kolossa, Veronika Solopova, Georg Rehm · 2025

Read How desirable is alignment between LLMs and linguistically diverse human users? by Pia Knoeferle, Sebastian Möller, Dorothea Kolossa, Veronika Solopova, Georg Rehm (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2502-12884.txt.

Introduction

Pia Knoeferle and four co-authors ask how far large language models should bend their speech to match users who differ in how they use language. Their paper, from two thousand twenty-five, is a work of debate, not a test.

Users differ by age, by sex-linked traits, and by how many tongues they know. Each gap can shape how a person speaks and takes in speech. The authors weigh what such matching would mean for ease of use and for clear talk. They also weigh what it would mean for how such models get built. The piece asks readers to judge the gains against the risks, case by case.

Catalog record

CREATORPia Knoeferle, Sebastian Möller, Dorothea Kolossa, Veronika Solopova, Georg Rehm
YEAR2025
COLLECTIONPAPERS
FORMATTXT
FILE SIZE29 KB
VERSION1.0
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2502-12884

Browse the full library · About the Tower of Babel

Full text

The opening of the work, from the library's hosted text:

How desirable is alignment between LLMs and linguistically diverse human users? Pia Knoeferle, Sebastian Möller, Dorothea Kolossa, Veronika Solopova, Georg Rehm Abstract We discuss how desirable it is that Large Language Models (LLMs 1 1 1 We mean by ‘LLM’ fully trained models like ChatGPT4 that can be accessed by users via a web interface. ) be able to adapt or align 2 2 2 We use the term ‘alignment’, as it is used in (psycho-)linguistics, to mean sharing of mental representations between interlocutors from the phonological to the lexical, morphosyntactic, syntactic, semantic and situation model levels ( Pickering and Garrod 2004 , see) . Earlier work in social psychology has coined the term ‘adaptation’ ( Giles et al. 1991 ; Giles et al. 1973 ) : It refers to individuals adjusting their communication styles to accommodate their interlocutors. For instance, people modify their verbal and nonverbal behavior (e.g., to reduce or emphasize differences), often based on factors like social identity, context, or relationships. Note that in the context of LLMs, the term ‘alignment’ has been debated and there seems to be a lack of clarity as to what ‘alignment’ means, whether it is ‘functional alignment’ (e.g., improving instruction following) or ‘worldview or social value alignment’ (based on the argument that humans share values) ( Kirk et al. 2023 , p. 3) their language behavior with users who may be diverse in their language use. User diversity may come about among others due to i) age differences; ii) gender characteristics, and / or iii) multilingual experience, and associated differences in language processing and use. We consider potential consequences for usability, communication, and LLM development. 1 Introduction Language processing is increasingly viewed as modulated in real time by characteristics of human language users ( Federmeier and Kutas 2005 ; Ito et al. 2018 ; Mishra et al. 2012 ; Münster and Knoeferle 2018 ; Salthouse 1996 ; Van Berkum et al. 2008 , e.g., their age and associated decline in processing speed, gender cues in someone’s voice, literacy, or their bi- and multilingual language background,) . Given this diversity, the present paper asks to what extent it is desirable that LLMs align with the user in language behavior (reducing distance and thus potentially facilitating communication), and to what extent it is desirable that human users align with an LLM’s language output. We review research on age-, gender-related, and multilingual diversity in language performance (Section 2 ; due to length restrictions, we review selected findings only). Following this review, we discuss the pros and cons of linguistic and other alignment between humans and LLMs (Section 3 ) and present conclusions in Section 4 . 2 Diversity in language processing / use 2.1 Age-related changes For language use and production, for instance, older (vs. younger) adults seem to use more pronouns: Hendriks et al. 2008 compared younger and older adults’ production in an elicitation task (picture stories) and in a comprehension task (written stories). For production, older (vs. younger) adults produced more pronouns in referencing an old topic in the context of a new topic. For the comprehension task, by contrast, no significant between-group differences emerged. The authors interpreted their findings as reflecting age-related changes: Younger adults will use a definite noun phrase instead of a pronoun to avoid ambiguity. But older adults’ more limited processing capacities may produce more often ambiguous pronouns instead of unambiguous full noun phrases, perhaps due to limits in their cognitive resources for taking the hearer’s perspective. Abrams and Farrell 2011 reported more difficulty in language production than comprehension for older adults ( Burke et al. 2000 ) ; one example was more frequent tip-of-the-tongue production in older than younger adults ( James and Burke 2000 , but note also some age-invariant dimensions of phonological processing related to reducing tip-of-the tongue phenomena by presenting phonologically-related words) . For language processing, age-related changes seem to occur at many linguistic levels: Semantic interpretation, for instance, was slowed in older adults ( Federmeier and Kutas 2005 ) , perhaps because general processing speed declines with age ( Salthouse 1996 ) . Vocabulary knowledge changes with age, too, such that older adults may have difficulty in deriving unfamiliar word meanings from context. For instance, they were more likely to produce generalized interpretations of the word meaning and selected fewer exact definitions than younger adults. Older adults may thus have special difficulties in deriving the meaning of unfamiliar words from context ( McGinnis and Zelinski 2000 ) . Differences also emerged in audio-visual comprehension when relating, for instance, verbs to action depictions, where effects of grounding occurred earlier in young than older adults ( Maquate and Knoeferle 2021 ) 3 3 3 While LLMs are currently often used in the written modality, user diversity regarding speech will likely become more relevant as LLMs are developed further. . For naturalistic communication, Dikker et al. 2022 reviewed the literature on age-related changes in inter-brain synchrony and concluded that “Older adults may have a harder time aligning and coupling due to factors including increased variability in neural firing (neural ‘noise’), difficulty sustaining dynamic patterns, sensory changes that result in lost fidelity of the signals that are important for alignment, and/or difficulties with attention and executive control that are critical for regulating neural patterns that are critical for accommodation to a conversation partner or adaptive behavior to a communicative context.” (p. 55). 2.2 Age and foreign-accented speech as input There are interactions between aging and foreign language comprehension, too. Older adults experienced, for instance, difficulty in semantic interpretation of foreign-accented speech: Native-accented semantic errors elicited the expected N400 effect (a negativity in brain amplitude averages that increases with increases in semantic processing difficulty). By contrast, for foreign-accented speech, such a differential brain response (larger N400 for semantically felicitous vs. infelicitous stimuli) was absent in older adults. By comparison, grammatical errors were not processed differently depending on native or foreign-language accent ( Abdollahi et al. 2021 ) . The state of the art is somewhat controversial though and further studies reported no such age differentiation in adaptation to foreign-accented speech ( Bieber and Gordon-Salant 2017 ; Burda et al. 2003 ; Hau et al. 2020 , e.g.,) . Regarding alignment, the slow-down and problems in understanding unfamiliar words or accents in older than younger adults could be counteracted with suitable LLM alignment. 2.3 Gender/sex and language use / processing There is good evidence that knowledge of gender (stereotypes) can rapidly influence language processing. For instance, activation of gender-stereotypical role names in online spoken language comprehension (in Finnish) occurred rapidly ( Pyykkönen et al. 2009 ) : Listeners activated gender stereotypes in story contexts where this information was not needed to establish coherence. Gender stereotypes further elicited expectations in that when participants listened to passive-voice sentences like The wood is being painted… they anticipated the agent of the action. That effect was larger for female than for male-stereotypical contextual information but was not modulated by participants’ gender attitudes ( Guerra et al. 2021 ) . Prior research also assessed sex/gender differences in language use. Newman et al. 2008 reported clear gender differences: Women used more words concerning psychological and social processes (more pronouns, verbs, and social words also related

Read the complete text