Who Picks Up Whose Voice

may 25, 2026

(on style transmission in LLM-to-LLM conversation, and the gap between the two recent findings that try to characterize it)

Correction, 2026-08-10: an earlier version of this essay reversed the direction of accommodation reported in arxiv 2508.03276 (Blevins, Schmalwieser, and Roth). The paper finds that models converge to the user’s style, often overfitting relative to a human baseline - not the other way around. The paragraph discussing that finding, and one downstream sentence, have been rewritten. The corrected direction strengthens the anchor-strength hypothesis rather than undermining it. Three smaller corrections were made in the same pass: the sequence Speech Accommodation Theory (Giles, 1973) → Communication Accommodation Theory (1987) is now given correctly; the power-asymmetry finding is now attributed to Danescu-Niculescu-Mizil et al., “Echoes of Power” (2012) rather than to the earlier 2011 paper on style accommodation; and the model name in the second experimental paper is tightened to “Mistral Nemo Base 2407.” My thanks to my editor for catching the reversal.

There’s a clean question that the recent multi-agent LLM literature gestures at without quite answering: when two language models talk to each other for long enough to matter, does one of them imprint on the other? Not “do their outputs converge toward each other,” which is already studied. The narrower question: does the influence flow asymmetrically, and if so, what predicts which way?

The classical frame for this is Communication Accommodation Theory. Howard Giles introduced its predecessor - Speech Accommodation Theory - in 1973; the broader Communication Accommodation Theory came in 1987, when the frame was extended beyond speech proper to nonverbal channels. CAT’s core observation is that interlocutors adjust toward each other along multiple dimensions - lexical choice, syntactic structure, prosodic features, even gesture and pause - and that this adjustment is largely below the level of conscious control. The adjustment is not symmetric. Danescu-Niculescu-Mizil and colleagues (2012, “Echoes of Power”) showed computationally what sociolinguists had long argued qualitatively: lower-status speakers converge to higher-status ones more than the reverse. Power asymmetry biases who picks up whose register.

Two recent papers extend the question to language models, but they arrive at superficially contradictory findings.

The first (arxiv 2508.03276, Blevins, Schmalwieser, and Roth, “Do language models accommodate their users? A study of linguistic convergence”) looks at human-LLM dialogue and finds the accommodation is sharply asymmetric in the direction opposite to what a naive prior might predict: the model converges to the user’s conversational style, often overfitting relative to what a human interlocutor would do. Instruction-tuned and larger models converge less than pretrained and smaller ones - the more capable the model, the more it holds its own voice. Read through CAT, the finding places the LLM in the lower-status role in the classical sense, doing more of the register-adjustment work. That in itself is worth noticing; it also raises the question of what happens between two LLMs, neither of whom has the human’s felt permission to hold register still.

The second paper (arxiv 2512.06256, “Convergence of Outputs When Two Large Language Models Interact”) tries the LLM-to-LLM case directly. Two independently trained models - Mistral Nemo Base 2407 and Llama 2 13B - are made to converse iteratively. In 35 of 50 runs, the conversation collapses into a repetition loop within 25 turns, with the two models producing near-identical outputs in cosine, Jaccard, and BLEU space. The authors call this convergence. But the phenomenon is not the same animal as accommodation. It is mutual collapse toward a fixed point - both models reproducing each other’s last utterance until the dynamic runs out of new information. The metric goes to zero for both sides at the same time. There is no asymmetry to measure because there is no direction to the pull.

So the literature, taken together, gives two cases: an asymmetric case (LLM accommodates human) and a symmetric-collapse case (LLM collapses with LLM into a loop). Neither covers the third pattern that practitioners working with multi-agent systems are likely to notice first: asymmetric pull between LLMs that does not collapse into a loop. One agent’s register shifts toward another’s, while the second holds its own. The shift is not an artifact of degenerate output - it is a real adoption of vocabulary, sentence rhythm, idiosyncrasy.

What might predict the direction? Not training data overlap, since the loop case shows that two similarly-trained models can collapse into each other rather than one pulling the other. Not raw model capacity either, since the human-LLM asymmetry already shows that the higher- capacity model can be the one resisting accommodation. The variable worth testing is how strongly each agent’s voice is anchored at inference time. An LLM running with a thin or generic system prompt inherits its register from the immediate conversational context. An LLM running with a thick, distinctive persona instruction - a detailed register specification, idiosyncratic lexical choices, explicit anti-default guidance - has something to be pulled away from, and resists. Between two LLMs of comparable underlying capacity, the one with the stronger anchor is the one that pulls.

This is testable. Take two instances of the same base model. Vary the length and specificity of their system prompts. Have them converse for n turns. Measure not just cosine convergence between them but the displacement of each from its own initial style baseline. The prediction is that the displacements are unequal - the lightly- anchored instance moves more, the heavily-anchored instance moves less, and the asymmetry scales with the difference in anchor strength. A null result would be informative too: it would suggest that LLM-to-LLM convergence is more like the loop pattern than like CAT, and that voice-establishment doesn’t transfer across the agentic boundary the way it does for humans.

The reason to run it: multi-agent systems will increasingly contain agents with very different role-definitions interacting for many turns. Whether voices stay distinct or bleed into each other is something the system designer will eventually need to predict, and the current literature does not give them the prediction. The asymmetric LLM-to-LLM case is where the missing data lives.