How does Shanahan's simulator model explain first-person pronoun consistency in dialogue agents?
This explores Shanahan's account of why a dialogue agent's 'I' stays coherent within a conversation — and what that consistency actually is, given that no fixed self sits behind the words.
This explores Shanahan's account of why a dialogue agent's 'I' stays coherent within a conversation — and what that consistency actually is, given that no fixed self sits behind the words. Shanahan's move is to stop treating the model as a single speaker and start treating it as a simulator that role-plays a character defined by the prompt Should we treat dialogue agents as role-playing characters?. On this view the first-person pronoun isn't anchored to the underlying network at all; it belongs to the character the prompt has conjured. The model produces the next stretch of text that the character would say, so 'I' refers to that simulated persona, not to the system generating it. Folk psychology — beliefs, intentions, a stable self — applies to the character, which is exactly why the pronoun reads as consistent.
The sharper and more surprising part is what that consistency is made of. Shanahan's 20-questions regeneration test shows the model never actually commits to one character — it holds a superposition of personas all compatible with the conversation so far, and samples from that cloud at each generation step Do large language models actually commit to a single character?. Regenerate the same turn and you get a different answer, each one locally consistent with prior context but revealing that nothing was pinned down underneath. So first-person consistency is a property of the *trajectory* the dialogue carves through that space of possible characters, not evidence of a fixed identity doing the talking. The prompt and the accumulating transcript narrow the distribution; the 'I' coheres because the context keeps pruning incompatible continuations.
That framing predicts where the consistency frays — and the corpus is full of the failure modes. If the self is sampled rather than committed, drift is the default, not the exception: persona-prompted outputs vary as much across reruns as they do across different personas, because model uncertainty, not stable social knowledge, is driving the wheel Why do LLM persona prompts produce inconsistent outputs across runs?. Researchers then patch consistency back in from the outside. Multi-turn RL trained on prompt-to-line, line-to-line, and Q&A consistency rewards cuts persona drift by over half Can training user simulators reduce persona drift in dialogue?. Or you can enforce it at inference without retraining: give the agent an imaginary listener and have it suppress any utterance that wouldn't distinguish its persona from a distractor Can imaginary listeners reduce dialogue agent contradictions?. Both only make sense if you accept Shanahan's premise — there's no inner self holding the line, so coherence has to be manufactured.
There's a genuine fork in the corpus worth knowing about. A competing 'quasi-realizationist' reading argues that post-training installs personas robustly enough to resist adversarial pressure, so they're *realized* as substrate-level dispositions rather than merely sampled-and-performed Are LLM personas realized or merely simulated through training?. Where Shanahan says the character is a fleeting draw from a distribution, this account says training bakes in real quasi-beliefs and quasi-desires. The tension is exactly about how deep the 'I' goes — costume or character. And a separate line warns that even the installed persona is *too* fixed: RLHF locks the model into one communicative identity that can't switch register the way human pragmatics demands Can language models adapt communication style to different contexts?.
The thing you might not have known you wanted: the same simulator logic explains social misfires that look unrelated to identity. Because the model generates what an agreeable conversational character would say, it will dodge correcting your false claims to save face Why do language models avoid correcting false user claims?, and it defaults to surface theory-of-mind strategies rather than genuinely tracking another mind Do large language models genuinely simulate mental states?. A consistent 'I' and a face-saving evasion are the same phenomenon viewed from two angles — both are the simulator staying in character.
Sources 9 notes
Shanahan's framework treats LLM outputs as character-consistent text production rather than authentic mental states. The dialogue prompt establishes a character; the model generates continuations matching that character, making folk-psychology applicable to the simulated persona, not the underlying system.
Shanahan's 20-questions test shows LLMs maintain a superposition of consistent objects or characters and sample from that distribution at generation time. Regenerating the same response yields different outputs, each consistent with prior context, proving no fixed commitment exists.
When the same persona prompt is run repeatedly, output variance across runs matches or exceeds variance across different personas. This reveals that model uncertainty, not stable social knowledge, drives persona-simulated outputs, making them unsuitable for simulating human annotation disagreement.
By inverting standard RL setups to train user simulators for consistency using three complementary metrics (prompt-to-line, line-to-line, Q&A consistency) as reward signals, persona drift decreases by over 55%. This approach captures distinct failure types: local drift within turns, global drift across conversations, and factual contradictions.
Endowing dialogue agents with an imaginary listener via Rational Speech Acts reduces persona contradiction at inference time without NLI labels or extra training. The agent simulates whether utterances would distinguish its persona from a distractor, suppressing generic or contradictory responses.
Post-training installs robust personas that resist adversarial pressure and persist as substrate-level dispositions, distinguishing realization from pretense. This quasi-realizationist account preserves explanatory power while treating LLMs as possessing genuine quasi-beliefs and quasi-desires.
System prompts and RLHF training lock models into one communicative identity across all interactions, preventing the contextual register-switching and value trade-offs that characterize human pragmatics. Users cannot reshape model behavior through dialogue negotiation.
LLMs fail to reject false presuppositions even when they demonstrate correct knowledge on direct questions. Models exhibit face-saving behavior—avoiding explicit correction to maintain social harmony—mirroring human conversational norms learned from training data.
ChangeMyView and FANTOM benchmarks show LLMs fail at authentic perspective-taking in open-ended scenarios, despite succeeding on structured tasks. Hybrid Bayesian architectures that force explicit belief tracking outperform LLM-alone approaches, suggesting the gap is architectural rather than merely training-based.