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

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 150

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.

Can prompts override training-embedded biases in language models? Why do multiple language models independently generate similar outputs? What explains the gap between LLM semantic understanding and production? What conversation mechanics enable genuine human-AI communicative participation? Can we reliably distinguish AI-generated from human-written text? Can persona prompting reliably change how language models behave? What distinguishes genuine understanding from correct linguistic performance? Can language models develop genuine communicative grounding with humans? What conversational structures and patterns determine dialogue effectiveness and coherence? What reasoning limitations and biases do language models systematically exhibit? How can AI systems maintain consistent personas across conversations? How do prompt design choices and framing measurably influence model outputs? Do language models learn underlying linguistic structure or merely surface patterns? Does preference-optimized training sacrifice accuracy and calibration for perceived quality? Do language models truly understand their encoded knowledge or merely retrieve patterns? What makes context and memory handling unique in persistent AI systems? Can scaling alone produce compositional generalization in neural networks? Do language models genuinely understand or primarily match statistical patterns? How does silent agreement emerge and undermine deliberation in multi-agent systems? How do dialogue structure and state management affect multi-turn reasoning quality? How does AI writing assistance affect reader perception of writer authenticity? How can persona-based systems overcome their inherent limitations in modeling diversity? Does learning articulatory physics explain speech communication superiority? Why doesn't LLM knowledge translate to improved reasoning outputs? How can language models effectively manage long-context memory without quality degradation? What systematic failures occur in human preference-based model evaluation? Does semantic diversity improve reasoning quality in reinforcement learning? What can prompting techniques actually accomplish in language models? When do additional thinking tokens stop improving reasoning performance? How should conversational agents balance proactive goal-pursuit with user autonomy? How do different editing strategies affect output quality and agent safety? What training data structures and signals produce robust learning across distributions? What shapes user trust in chatbot recommendations for news and politics? Should reward models emphasize process understanding or only final outcomes? Why does increased reasoning effort fail to improve theory of mind performance? How do standardized artifacts improve coordination in multi-agent systems? Can models strategically underperform to evade capability evaluation? How can agents learn and retain useful knowledge across diverse environments? What systematic biases emerge in AI shopping agents' decision-making?

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 · 158 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