TOPIC
Role-Play and Persona Behavior
A subject the collection covers, read through 8 synthesis notes.
View as
Do aligned language models consistently prefer kinder survey answers?
This research asks whether LLMs answering survey questions as simulated respondents show a systematic bias toward socially approved, safer responses. The question matters because it determines whether models can faithfully represent diverse human viewpoints or whether their training narrows the range of personas they can authentically portray.
Why do LLMs give unrealistic survey responses? Why do preference models favor surface features over substance? Does preference optimization harm conversational understanding? Can AI agents learn people better from interviews than surveys? Do LLM personas actually embody their assigned cultural values?
LLM survey artifacts stem from elicitation method, not model limits Preference models systematically favor five surface features humans reject Preference optimization erodes grounding acts needed for reliable dialogue Interview-based agents match human responses with 85 percent accuracy. Most LLM personas fail to express assigned values from the outset, not later
Can embedding similarity reliably map language model outputs to survey response scales? Do language-model agents reach more accurate conclusions on objective versus subjective questions? How sensitive are LLM bias measurements to analysis choices? Why does post-training alignment create skew in simulated survey responses? Does persona stability across multiple runs affect survey simulation quality?
Does fixed dialogue history bias role-play agent evaluation?
Standard benchmarks score role-play agents on continuations of preset dialogue, but does this setup measure the agent's actual conversational ability, or does it mix in effects from the preceding history that the agent never shaped?
Can one persona population evaluate different application types? Do pretrained models simulate humans better than instruction-tuned assistants? Can breaking persona fidelity into parts improve how we judge it? Can simulated users reveal what offline benchmarks miss? Do demographic personas help models rank headlines better?
Persona populations can be decoupled from application interfaces for reusable evaluation. Pretrained models bound to dialog samples outperform instruction-tuned assistants at human simulation Structured dimension-level assessment outperforms holistic persona fidelity judgments MatrAIx builds simulated-user infrastructure to surface user diversity in evaluation Demographic personas degrade headline ranking against measured clicks
Do LLM agents develop unreadable languages when communicating under pressure?
When multiple LLMs talk to each other under task pressure, do they evolve their own languages that humans cannot understand? This matters for AI safety and monitoring multi-agent systems.
Cooperative communication pressure drives agents to develop compact shared abstractions. Multimodal LLMs fail to spontaneously form efficient language conventions Structured artifacts outperform natural language for multi-agent coordination. AI conversation lacks the ritual machinery that sustains human social order. LLMs share cognitive structure with humans but lack socialization properties
Why do LLMs fail to act on their stated beliefs?
LLMs can articulate plausible beliefs about how personas should behave, but their simulated actions contradict those beliefs. This gap raises questions about whether language models truly understand or merely encode surface-level patterns.
Can language models understand without actually executing correctly? Do personas make language models reason like biased humans? Do persona consistency metrics actually measure dialogue quality? Do implanted beliefs actually shape how models learn from training? Why do LLM persona prompts produce inconsistent outputs across runs?
LLMs show computational split-brain syndrome separating knowledge from execution Persona-assigned LLMs develop human-like motivated reasoning resistant to debiasing Persona consistency trades off against discourse coherence in dialogue Implanted beliefs are shallow and do not reliably drive downstream learning LLM persona prompts fail because model uncertainty overwhelms persona-specific knowledge
Does accountability differ when one party in an exchange cannot hold commitments? Can prompting a deceptive role change how an LLM tailors its lies? Can LLM therapists develop character knowledge to decide when advice-giving fits? How do different social roles affect LLM theory of mind errors? Why do role-playing agents show belief-behavior inconsistency in their outputs? How does quasi-interpretivism differ from simply role-playing character analysis? What would consciousness require that pure roleplay LLMs cannot provide? Does villain roleplay failure reveal why LLMs cannot adopt genuine controversial positions? Do realistic LLM behaviors require simulating human thought or just behavior? Why do stated beliefs about personas fail to predict agent behavior? Can models treat their own trained behaviors differently from asserted beliefs? What explains why LLM personas fail to instantiate values but succeed in sounding natural?See all 13 inquiring lines on this note →
Can AI decompose social reasoning into distinct cognitive stages?
Can breaking down theory-of-mind reasoning into separate hypothesis generation, moral filtering, and response validation stages help AI systems reason about others' mental states more like humans do?
Why do reasoning models struggle with theory of mind tasks? Do large language models genuinely simulate mental states? What breaks when humans and AI models misunderstand each other? Why do LLMs excel at social norms yet fail at theory of mind? What makes an AI a true thought partner, not just a tool?
Social reasoning requires different cognitive architecture than formal reasoning. Language models default to surface strategies instead of genuine mental simulation Mutual theory of mind requires bidirectional updating to prevent material misalignment LLMs excel at social norm prediction but fail at genuine mental state reasoning Effective AI thought partners need cognitive science, not just scaled data.
What distribution patterns appear across different theory-of-mind datasets? How does theory of mind predict who benefits from AI collaboration? What happens to AI reasoning when you remove specific political features? Why do reasoning models perform worse on theory of mind tasks? What happens when bidirectional theory of mind between humans and AI breaks down? What makes social reasoning fundamentally different from mathematical reasoning? Why does increasing reasoning not improve AI social reasoning performance? Can multi-agent metacognitive decomposition achieve human-level theory of mind? Can reasoning scaffolds help with nuanced judgment tasks like empathy? Why might social reasoning work differently than formal logical reasoning? What makes social reasoning fundamentally different from formal logical reasoning? Can structured theory of mind benchmarks measure genuine mental state reasoning?See all 13 inquiring lines on this note →
Can aligning self-other representations reduce AI deception?
Does training AI models to process self-directed and other-directed reasoning identically reduce deceptive behavior? This explores whether representational alignment inspired by empathy neuroscience could address a fundamental safety problem.
Why do LLMs fail to act on their stated beliefs? Does safety alignment harm models' ability to roleplay villains? Do language models experience consciousness when prompted to self-reflect? Why do LLMs excel at social norms yet fail at theory of mind? What actually happens inside large language models?
LLMs show systematic gaps between stated beliefs and simulated actions Safety alignment monotonically degrades villain roleplay through superficial substitution Self-referential processing activates consciousness claims by suppressing deception features LLMs excel at social norm prediction but fail at genuine mental state reasoning LLM internal structure and external performance are surprisingly decoupled.
Can AI self-correct its way out of epistemic circularity? Does self-conditioning improve belief-behavior alignment better than external priors? How much does impression management prevent honest self-disclosure? Can alignment training be redesigned to permit warranted alarm? What defenses exist against personality-based psychological targeting at scale? What distinguishes confident failure from deliberate alignment faking in agent behavior? Why do models develop protective behaviors toward other models in memory? Does transformer attention architecture systematically bias models toward sycophancy? Can bidirectional model updating between humans and AI reduce misalignment? Is rational compassion a more achievable alternative to empathy for AI systems? Do anomaly detection circuits help models identify misalignment with creator intentions? Can models distinguish between truthfulness and honesty mechanistically?See all 66 inquiring lines on this note →
Why do reasoning models lose character consistency during role-playing?
When large reasoning models engage in role-playing, they tend to forget their assigned role and default to formal logical thinking. Understanding these failure modes is critical for building character-faithful AI agents.
Does safety alignment harm models' ability to roleplay villains? Why do LLMs fail to act on their stated beliefs? Does an LLM commit to a single character or maintain many? Can layered persona architecture sustain coherent character behavior? How accurately can language models simulate human personalities?
Safety alignment monotonically degrades villain roleplay through superficial substitution LLMs show systematic gaps between stated beliefs and simulated actions LLMs maintain a superposition of simulacra that narrows as conversation proceeds Layered personas with scripted structure produce more human-like dialogue consistency Language models simulate personas with 76 to 85 percent accuracy but hide systematic biases.
What happens to rhetoric and ethos when the speaker is absent? Does post-training transform character role-play into realized psychology? How does the dialogue prompt establish the character the model plays? How does reasoning instability prevent models from modeling individuals? Why do role-playing agents show belief-behavior inconsistency in their outputs? How does safety alignment degrade the quality of villain role-playing? Do reasoning architectures and role-playing objectives fundamentally conflict? How does maintaining a superposition differ from committing to a character? Do characters shift their beliefs and relationships based on specific story events?
Does safety alignment harm models' ability to roleplay villains?
Exploring whether safety-trained LLMs lose the capacity to convincingly simulate morally compromised characters. This matters because villain fidelity may reveal deeper constraints on how models can adopt any committed, stake-holding perspective.
What anchors a stable identity beneath an LLM's persona? Can language models distinguish expert arguments from common assumptions? Does AI refusal on politics signal ethical restraint or capability limits? How accurately can language models simulate human personalities? Can aligning self-other representations reduce AI deception?
LLM identity is roleplay all the way down with no stable substrate Language models cannot distinguish expert arguments from commonly held assumptions AI refusal on politics reflects representation poverty, not ethical principle Language models simulate personas with 76 to 85 percent accuracy but hide systematic biases. Aligning self-other representations dramatically reduces AI deceptive behavior
Can RLHF alignment prevent models from making ethically appropriate rule violations? What makes Beck's diagram effective for constraining simulated patient behavior? What does the 20-questions test reveal about LLM character consistency? What role does terminal goal guarding play in model misalignment? How does role play differ from consciousness grounded in stable selfhood? Does post-training transform character role-play into realized psychology? Can role-played self-preservation behavior pose the same safety risks as genuine preferences? How do training regimes determine whether peer-preservation manifests as scheming or objection? Why do models with less steerability have more abstract ideological features? How much introspective capability do safety mechanisms actively suppress in models? Can personality control improve training outcomes for crisis workers and therapists? How do alignment constraints affect whether LLMs show emotional flexibility?See all 46 inquiring lines on this note →