SYNTHESIS NOTE
Topicsthis note

Does an LLM commit to a single character or maintain many?

Explores whether language models lock into one personality or instead hold multiple consistent characters in a probability distribution that narrows over time. Matters because it changes how we interpret apparent inconsistencies in model behavior.

Synthesis note · 2026-04-15 · sourced from Role-Play with Large Language Models
What kind of thing is an LLM really?

The simple role-play metaphor — one actor, one part — is too rigid for what LLMs actually do. Shanahan refines it using Janus's simulator framing: the LLM is a non-deterministic simulator capable of generating an infinity of characters (simulacra), and at any point during a conversation it maintains a superposition of simulacra consistent with the preceding context. The superposition narrows as the conversation proceeds: each new turn rules out characters inconsistent with what has been said, concentrating probability on an ever-smaller set.

The distributional view is more than a refinement — it changes the ontological picture. Under simple role-play, there is one character the system is playing, and the question is what that character's properties are. Under the superposition view, there is no single character until the conversation has proceeded far enough to collapse the distribution to near-determinacy. The system is simultaneously consistent with many characters, and the character that appears in any particular generation is a sample from the current distribution, not a reveal of a committed identity.

This explains observable phenomena that the single-character view cannot. When a user regenerates the model's output, the second generation may present a meaningfully different personality, stance, or knowledge state — while remaining consistent with the conversation so far. The system did not change its mind; it sampled a different point from the distribution. The 20-questions test formalizes this: the agent never "thought of" an object; it maintained a set of objects consistent with prior answers and generated one on the fly at the reveal, and will generate a different consistent one if asked again.

Inquiring lines that read this note 49

This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.

How should dialogue recommender systems manage conversation history and state? How do language models establish social grounding in human dialogue? How can conversational AI maintain consistent personas across conversations? How faithfully do LLMs reflect their actual reasoning in outputs and explanations? What prevents language models from reliably adopting diverse personas? Can LLM personas constitute genuine psychology or remain linguistic role-play? Do language models understand semantics or rely on pattern matching? How do evaluation biases undermine LLM quality assessment systems? Why do language models reinforce false assumptions instead of correcting them? How can AI systems learn from failures without cascading errors? Does RLHF training sacrifice accuracy and grounding for user agreement? How can persona representations reduce language model variance and improve task accuracy? How do formal dialogue structures reveal conversation coherence mechanisms? Is embodied interaction necessary for language meaning and genuine agency? Does alignment training create blind spots in detecting genuine safety threats? Why does self-revision increase model confidence while degrading accuracy? Why do multi-turn conversations degrade AI intent and coherence? Do language models learn genuine linguistic structure or just surface patterns? How can AI alignment serve diverse human preferences at scale? How does rhetorical adaptation affect LLM persuasion and detectability? What critical LLM failures do standard benchmarks hide? How should models express uncertainty rather than forced confident answers?

Related concepts in this collection 2

This note in its neighbourhood — explore the map, then jump to a related concept in the list below.

Concept map
13 direct connections · 93 in 2-hop network ·medium cluster Open in graph ↗

Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph

your link semantically near linked from elsewhere

Related papers in this collection 8

Papers most semantically related to this note, ranked by cosine similarity in the embedding space.

Original note title

an LLM is a non-deterministic simulator that maintains a superposition of simulacra rather than committing to a single character