SYNTHESIS NOTE
Topics›this note

Do large language models actually commit to a single character?

Explores whether LLMs pick and hold a fixed character or instead sample from multiple consistent possibilities. Tests reveal that regenerated responses differ while remaining consistent with context, challenging intuitive assumptions about how dialogue agents work.

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

Shanahan constructs a simple but decisive behavioral test. Have an LLM-based dialogue agent play 20 questions — the agent "thinks of" an object and the user asks yes/no questions. After several rounds, ask the agent to reveal the object. It names something consistent with all previous answers. Now regenerate that response. The agent names a different object, also consistent with all previous answers.

This phenomenon is incompatible with any view that treats the agent as having committed to a specific object at the start of the game. A human playing 20 questions picks an object, holds it in mind, and answers questions from that fixed commitment. The LLM never picks. It maintains a set of objects consistent with the accumulated constraints — what Shanahan calls a superposition — and samples from that set at the moment of reveal. The same logic extends from objects to characters: the agent never commits to being a specific character with specific properties. It maintains a distribution over consistent characters and generates behavior sampled from that distribution.

The test is portable. Any feature that appears settled in one generation but changes on regeneration (while remaining consistent with context) is evidence of superposition rather than commitment. This has been observed in personality traits, stated preferences, claimed memories, and emotional dispositions of dialogue agents. The philosophical consequence is that attributing fixed psychological properties to an LLM conversation state is category-mistaken: the system has a distribution over properties, not a property. What appears stable is a high-probability region of the distribution, not a fact about an underlying entity.

Inquiring lines that read this note 120

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 systems represent uncertainty from noisy speech input? Do language models learn genuine linguistic structure or just surface patterns? How does rhetorical adaptation affect LLM persuasion and detectability? Does conversational format create illusions of genuine AI communication? Why can't humans reliably detect AI-generated text despite measurable linguistic signatures? How can LLM user simulators model realistic goal-driven conversation? How do language models establish social grounding in human dialogue? How should dialogue recommender systems manage conversation history and state? Why do language models reinforce false assumptions instead of correcting them? Do language models understand semantics or rely on pattern matching? How can language models sustain linguistic synchrony and intersubjectivity during dialogue? How can conversational AI maintain consistent personas across conversations? Can prompting strategies overcome LLM biases without model fine-tuning? How do formal dialogue structures reveal conversation coherence mechanisms? How can persona representations reduce language model variance and improve task accuracy? What prevents language models from reliably adopting diverse personas? Can LLM personas constitute genuine psychology or remain linguistic role-play? Can next-token prediction alone produce genuine language understanding? Does RLHF training sacrifice accuracy and grounding for user agreement? What factors beyond surface content determine how readers extract meaning differently? Why do language models struggle with implicit discourse relations? Why do multi-turn conversations degrade AI intent and coherence? When does optimizing for quality undermine the value of diversity? Can prompting inject entirely new knowledge into language models? How do language models inherit human biases from training data? What distinguishes dynamic from static grounding in dialogue systems? What makes dialogue-based explanation more successful than monologue? What articulatory information do speech signals carry that text cannot? How faithfully do LLMs reflect their actual reasoning in outputs and explanations? How should retrieval systems optimize for multi-step reasoning during inference? Why does finetuning cause catastrophic forgetting of model capabilities? How can AI alignment serve diverse human preferences at scale? How should conversational agents balance goal-driven initiative with user control? Do language model representations contain causally steerable task-specific features? How do prompt structure and constraints affect model instruction reliability? When do additional thinking tokens stop improving reasoning performance? How should agents balance memory condensation to optimize context efficiency?

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 · 145 in 2-hop network ·dense cluster Open in graph ↗
Do large language models actually commit to a si… Does an LLM commit to a single character or mainta… Should we call LLM errors hallucinations or fabric…

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.

Search by related questions 4

Suggested questions this note speaks to — click to search the collection, or type your own.

Why does regenerating LLM responses produce different but equally valid answers?How does maintaining a superposition differ from committing to a character?Can we detect superposition in LLM personality traits and stated preferences?What happens when we treat LLM outputs as sampled rather than stored?
Original note title

the 20-questions regeneration test falsifies any committed-character view of LLM behavior