TOPIC
Design Frameworks
A subject the collection covers, read through 23 synthesis notes.
View as
Do formal language prototypes improve reasoning across different domains?
Can training language models on abstract reasoning patterns in Prolog and PDDL help them generalize to new reasoning tasks? This tests whether shared logical structures underlie seemingly different problem domains.
Why does partial formalization outperform full symbolic logic? Can symbolic solvers fix how LLMs reason about logic? What formal languages actually help transformers learn natural language? Does training data format shape reasoning strategy more than domain? How should reasoning systems actually be architected?
Partial symbolic augmentation outperforms both pure language and full formalization Symbolic solver integration enables reliable logical reasoning through structured feedback Effective formal language pre-pretraining requires matching structural and computational constraints Training format shapes reasoning strategy 7.5 times more than domain Reasoning systems should separate activation timing from execution capability.
What is the difference between learning discourse patterns and learning abstract language? What formal representation could capture analogical reasoning across domains? Do reasoning languages like Prolog follow the same two-constraint transfer pattern? Can mathematical reasoning improvements transfer across problem subdomains? What makes natural language reasoning more practical than formal languages for multi-framework codebases? What makes structured informal reasoning preferable to full formalization?
Why do AI agents miss most of what users actually want?
UserBench explores why current models align with user intent only 20% of the time, even when users reveal preferences across multiple turns. The question examines whether agents can learn to actively clarify ambiguous or evolving goals.
Why can't users articulate what they want from AI? How do users actually form intent when prompting AI systems? Can models learn to ask clarifying questions instead of guessing? Why do language models fail in gradually revealed conversations? Why can't advanced AI models take initiative in conversation?
Users need guided dialogue to mature intent, not just better models. Intent formation is continuous maturation, not a binary triggered state Models can be trained to proactively identify missing information and request clarification Language models make unrecoverable premature assumptions in underspecified conversation. AI passivity stems from next-turn reward optimization, not inability
Why can't users and AI articulate shared goals together? What makes users willing to relinquish control to an agent? Can users articulate what they want before AI helps them discover it? How do users fail to articulate what they actually want? Can prompt engineering overcome the gulf between user intent and AI interpretation? How can we measure whether a user actually understands their own needs? Why do 45 percent of workers want equal partnership with AI rather than full automation? Can users detect and correct an AI's mental model of their preferences? When should agents accommodate user preferences over their own goals? Can agents balance goal-driven proactivity with user preference alignment? Can users articulate their intent before exploring what an AI system finds? Why do AI models treat user intent as binary rather than evolving?See all 23 inquiring lines on this note →
Does AI exposure hit low-wage workers harder in some fields?
This research asks whether AI's impact spreads evenly across skill levels or concentrates among workers least able to adapt. The distinction matters because vulnerable workers with fewer resources face greater risk from automation.
Does concentrated AI exposure enable workers to adapt and reallocate? Do firms substitute labor for AI at different rates? When does AI actually boost worker productivity? Does incremental AI replacement erode human influence over society? Is AI creating common skills across jobs or deepening divisions?
Concentrated AI exposure enables worker reallocation that offsets employment losses Firms substitute labor for AI faster and cheaper at firm-specific rates AI productivity gains require existing skill, not learning Incremental AI adoption degrades implicit alignment by replacing human labor dependency. AI demand splits the labor market into a technical core and the rest
Why do firms substitute labor for AI faster than gig worker jobs disappear? Do salaried workers get better AI training support than gig workers? Can workers move across the divide between technical and non-technical job markets? How does AI task concentration within firms affect worker reallocation across jobs? Do firms with high AI exposure shed jobs or reshape roles? Can workers retrain faster than AI exposure spreads through occupations? How does occupational segregation affect who gains from AI productivity? How do institutions shape whether AI enables worker mobility or deepens hierarchy? Does AI adoption rise or fall as worker education and wages increase? Which occupations show the sharpest gap between AI capability and actual adoption? How does concentrated AI exposure across workers affect firm-level employment demand?
Is AI creating common skills across jobs or deepening divisions?
Whether AI diffusion produces a uniform set of competencies across occupations or widens occupational divisions. This matters for understanding how labor markets will adapt to AI exposure.
Concentrated AI exposure enables worker reallocation that offsets employment losses Workers desire equal human-AI partnership for 45 percent of occupations AI productivity gains require existing skill, not learning Delegated AI exposure concentrates in information work, tracking capability AI exposure spreads unevenly by gender and worker vulnerability
Can workers move across the divide between technical and non-technical job markets? How does AI task concentration within firms affect worker reallocation across jobs? Do firms with high AI exposure shed jobs or reshape roles? Can workers retrain faster than AI exposure spreads through occupations? How does occupational segregation affect who gains from AI productivity? How do institutions shape whether AI enables worker mobility or deepens hierarchy? Does AI adoption rise or fall as worker education and wages increase? Which occupations show the sharpest gap between AI capability and actual adoption? Does AI create new skills gaps or only expose existing ones? Do AI productivity gains require existing skills or enable learning new ones?
What UX principles do workplace users want in AI agents?
A multi-method study asked business users what design principles matter most for human-AI collaboration at work. The findings prioritize oversight and reliability over convenience.
Human-agent collaboration should precede full AI autonomy for safety Machine agency operates on a five-level spectrum from passive to cooperative Current AI agents complete only 30 percent of workplace tasks autonomously. Agent autonomy degrades human oversight capacity and positioning User values align with agent output but fracture during supervision
How should chatbot design vary by relationship duration?
Do chatbots serving one-time users need different design than those supporting long-term relationships? This matters because applying the same design to all temporal profiles creates usability mismatches.
Do chatbot relationships lose their appeal as novelty wears off? How do time gaps shape what people discuss across conversation sessions? Does chatbot personalization build trust or expose privacy risks? Can AI chatbots create genuine therapeutic bonds with users? Can a single model replace retrieval for long-term conversation memory?
Novelty effects in chatbot relationships decay predictably over repeated interactions Time intervals and speaker relationships reshape how past events are discussed across sessions Chatbot personalization increases trust and privacy risk simultaneously Digital conversational agents establish therapeutic bonds comparable to human therapy Single-model compression eliminates retrieval bottleneck but follows a fragile consolidation pattern
What temporal design dimensions characterize different chatbot relationship types? How do time gaps between conversations change what chatbots should remember? Does personalization help or hurt persistent companion chatbots? Does engagement with AI partners decay over time like chatbot relationships do? Can personalization delay or prevent novelty decay in chatbot relationships? How do chatbot design features like intimacy-by-design sustain romantic bonds? How does persistent versus temporary companion design affect relationship patterns? Do model updates disrupt established sources of support for regular chatbot users?
Do spatially-anchored previews stabilize VR geometry editing tasks?
This study explores whether showing users visual previews of candidate interpretations anchored in the 3D workspace helps them reach task goals more smoothly and with fewer conversation turns when working with an LLM-assisted geometry editor.
Which clarifying questions actually improve user satisfaction? Why do language models fail in gradually revealed conversations? Can reference resolution work as a language modeling problem? How do users actually form intent when prompting AI systems? Does user satisfaction actually measure cognitive understanding?
Specific facet questions outperform need-rephrasing for user satisfaction Language models make unrecoverable premature assumptions in underspecified conversation. Reference resolution including on-screen entities works as language modeling. Intent formation is continuous maturation, not a binary triggered state Internal cognitive clarity matters more than expressed satisfaction scores
Should AI systems stay collaborative rather than fully autonomous?
Explores whether keeping humans in the loop with AI agents is more reliable than pursuing full autonomy. Investigates whether collaboration solves problems that autonomous systems structurally cannot.
Targeted intervention at high-uncertainty points outperforms both full autonomy and exhaustive oversight AI agents are passive by design, not by capability AI reliability follows a sharp stage-dependent boundary tracked by task checkability Agents propose methods and implement revisions while humans retain final decisions AI generation consistently outpaces verification across the entire research lifecycle
When should an AI system actively intervene versus remain silent? What would contractualist AI governance look like in practice? Can exoskeleton dependency accumulate without organizations noticing it happening? How does AI reliance change professional judgment and autonomy? What happens when AI-dependent workers must operate without their tools? How does treating AI as an agent affect user autonomy and decision-making? Does removing human labor from systems secretly grant AI more autonomy? Can humans build reliable oversight for increasingly complex AI systems? What implicit alignment do humans provide by staying in research loops? Which research collaboration skills should AI systems develop first? Why do some occupations need human-AI partnership more than others? Can workers reallocate to subjective tasks that resist automation indefinitely?See all 53 inquiring lines on this note →
Can designers shape LLM behavior without deep technical knowledge?
Explores whether LLMs can be treated as adaptable design materials that designers can tinker with directly, rather than fixed components handed over by engineers. Matters because it determines whether user-centered judgment reaches model adaptation early.
Designer taste will be formalized into evaluation systems and automated Human values must be embedded throughout the LLM pipeline not patched afterward Expert data storytellers delegate execution tasks while protecting narrative intent Specialized model deployment requires ecosystem design beyond capability alone. LLMs share cognitive structure with humans but lack socialization properties
Why does treating model behavior as part of the design surface matter for guardrails? What unique perspective do designers bring to LLM adaptation that engineers might miss? Do open model properties like customizability create net new misuse opportunities? Can workers move across the divide between technical and non-technical job markets? Why did product managers gain more from Figma Make than professional designers? How can analysts customize generated UIs without learning to think like engineers? How do LLM capabilities changing affect the relevance of interaction guidelines?
Do firms substitute labor for AI at different rates?
Explores whether companies exposed to AI shocks replace contracted workers with AI tools uniformly or at varying rates, and what firm-level differences reveal about the economics of AI adoption.
Does concentrated AI exposure enable workers to adapt and reallocate? Can codified expertise let non-experts match specialist output? Does AI exposure hit low-wage workers harder in some fields? Does AI turn freelance work into validation instead of creation? Where have workers actually delegated tasks to AI?
Concentrated AI exposure enables worker reallocation that offsets employment losses Codified expert rules in agent scaffolding dissolves organizational specialist bottlenecks AI exposure spreads unevenly by gender and worker vulnerability AI deskills freelancers by converting production work into validation work Delegated AI exposure concentrates in information work, tracking capability
How do worker-side adaptation effects interact with firm-level substitution patterns? What mechanisms enable some firms to adopt AI more cheaply than others? Does codifying expertise into AI agents drive faster labor substitution? How does concentration of AI capability across firms affect labor market outcomes? Which firms capture the cost advantages from labor-to-AI substitution? How should forecasting methods adapt to a post-AGI regime? Why do firms substitute labor for AI faster than gig worker jobs disappear? Do salaried workers get better AI training support than gig workers? Do employers hire workers who learn AI skills on the job versus bringing them in? Can workers move across the divide between technical and non-technical job markets? How does AI task concentration within firms affect worker reallocation across jobs? Why does AI adoption favor automation over augmentation in female-dominated work?See all 21 inquiring lines on this note →
Do generated analysis UIs really work better than chat?
TaskArtisan investigates whether putting a GUI into LLM-assisted analysis workflows improves usability and clarity, and what trade-offs emerge when analysts need to modify or reuse generated interfaces.
Do generated interfaces outperform text-based chat for most tasks? Can API-first agents outperform UI-based agent interaction? Can a shared canvas serve both human and agent memory? When does verification feedback actually guide targeted artifact repair? Does chat delegation actually save time on task completion?
Generated task-specific UIs outperform chat in over 70 percent of cases API-first agents reduce task completion time by 65 to 70 percent Shared canvas workspaces align human and agent access to project state Verification loops guide repair only when feedback granularity matches available actions Chat delegation reduces interaction effort without speeding up task completion
Why did product managers gain more from Figma Make than professional designers? How does delegated workflow adoption differ from conversational chatbot usage patterns? How can analysts customize generated UIs without learning to think like engineers? Can natural language help users modify widget composition during analysis work? How do generated interfaces compare to chat when tasks require workflow changes? What makes some analysis tasks stable enough for rigid generated interfaces? How should outcomes be scored when comparing applications with different interaction formats? What evidence shows canvas workspaces recover from failures better than chat baselines? Why does direct generation work better for some deliverable types than others?
Do generated interfaces outperform text-based chat for most tasks?
Explores whether LLMs should create interactive UIs instead of text responses, and under what conditions users prefer dynamic interfaces to traditional conversational chat.
API-first agents reduce task completion time by 65 to 70 percent AI passivity stems from next-turn reward optimization, not inability Intent-based specification requires six principles to address generative variability Users need guided dialogue to mature intent, not just better models. Generated analysis UIs gain clarity but lose flexibility and ease of modification
What makes human-LLM exchange closer to oracle-consultation than dialogue? Why does dialogue-shaped text fail to produce dialogue-like operations in practice? Does the interface design itself shape how much content users will review? Can API-first interaction replace traditional UI-based agent interfaces? Can generative interfaces help users articulate what they actually want? What types of tasks benefit most from dynamically generated interfaces? Why does the chat paradigm persist if it underperforms for structured tasks? How does API-first interaction compare to generative interface approaches? Why might text-only interfaces underestimate agent preference elicitation capabilities? What interaction controls matter most for effective human-LLM collaboration? What interaction design changes would help LLMs handle underspecified requests? How do users develop different interaction scripts specifically for machines versus humans?See all 25 inquiring lines on this note →
How should users control systems with unpredictable outputs?
When generative AI produces different outputs from identical inputs, how do interaction design principles help users maintain control and develop effective mental models for stochastic systems?
Users need guided dialogue to mature intent, not just better models. Prompt optimization retrieves existing knowledge but cannot inject new knowledge Semantic entropy detects confabulations by measuring meaning divergence. AI agents align with full user intent only 20 percent of the time AI passivity stems from next-turn reward optimization, not inability
Can generative interfaces help users articulate what they actually want? How does API-first interaction compare to generative interface approaches? How does generative variability intensify the problem of passive AI systems? What skills do users need to work effectively with stochastic outputs? What creates the tension between users wanting convenience and resisting loss of control?
How do communication modalities shape human-agent collaboration patterns?
Does varying how humans and agents exchange information—text, voice, or structured channels—produce measurably different negotiation, trust, and awareness outcomes in collaborative tasks?
AI agents are passive by design, not by capability Configuration files make live human-AI protocols auditable and shareable across labs Current AI agents complete only 30 percent of workplace tasks autonomously. Machine agency operates on a five-level spectrum from passive to cooperative Multi-agent performance is primarily determined by token spending.
Can models optimized for solo capability support productive human collaboration? Can agent social framing change how humans apply collaborative social scripts? Does awareness of agent reasoning alter human trust differently across modalities? What interaction controls matter most for effective human-LLM collaboration? What interaction mechanisms let humans and agents defer work effectively? What are the key interaction mechanisms that make human-agent collaboration work? What trust signals do agents lack that humans use to assess credibility? What are the five types of human interactions in agentic AI systems? Which interaction controls matter most in human-agent collaboration experiments? What happens to collaborative trust when effort becomes invisible in finished work?
When should human-agent systems ask for human help?
Explores the timing problem in collaborative AI systems: since there's no objective metric for optimal interruption, how can we design deferral mechanisms that know when to involve humans without constant disruption or silent failures?
Eleven task axes determine delegation design and capability matching Insert-expansions from conversation analysis formalize when agents should probe users AI needs explicit training on when not to speak AI passivity stems from next-turn reward optimization, not inability Current AI agents complete only 30 percent of workplace tasks autonomously.
What distinguishes over-intervention from useful proactive AI assistance? When should an AI system actively intervene versus remain silent? Can timing and context awareness reduce the cognitive cost of AI suggestions? How does treating AI as an agent affect user autonomy and decision-making? Why does human interaction remain the hardest failure mode for agents? What memory and planning capabilities do AI companions need for evolving user needs? What distinguishes task failure from communication breakdown in multi-agent systems? Can humans build reliable oversight for increasingly complex AI systems? Which research collaboration skills should AI systems develop first? Which task characteristics determine whether AI can displace them first? What signals should systems use to predict the right moment for intervention? Why do passive conversational agents fail at collaborative decision-making?See all 86 inquiring lines on this note →
Why do people share more openly with machines than humans?
Does the absence of social goals in human-machine communication explain why people disclose sensitive information more readily to chatbots? Understanding this mechanism could reshape how we design conversational AI.
Do chatbots help people disclose more intimate secrets? Why do people share more with chatbots than humans? Do chatbots trigger human reciprocity norms around self-disclosure? Can opening politeness patterns predict whether conversations will turn hostile? Why do language models sound fluent without grounding?
Chatbots enable deeper intimate disclosure through judgment-free environments People disclose more to AI because machines cannot judge them Chatbots trigger disclosure reciprocity when they share emotions consistently Opening politeness strategies predict conversation derailment into personal attacks LLM fluency emerges partly from skipping communicative grounding work
Why do humans publish more negative reviews in public than in private? Why might an AI's face-saving tendency increase user disclosure? What harms might chatbots cause through stigma expression and delusion reinforcement? Can transparency about AI limitations reduce the seductiveness of chatbots as quasi-Others? Do people treat conversational AI as social actors without conscious awareness? What social information becomes invisible when grief is regulated away? Why do AI chat modes pseudo-appeal while post modes reach no one in particular? Does the lack of judgment in machines explain intimate self-disclosure patterns? What novel goals emerge specifically in human-machine interaction beyond social ones? How do privacy concerns compete with disclosure comfort in human-machine conversation? Why do humans fail to identify AI agents when their identity is hidden? Why do people evaluate machines against human communication standards?See all 34 inquiring lines on this note →
Do humans apply human-human scripts to AI interactions?
Does CASA theory correctly explain how people interact with media agents, or have decades of technology use created separate interaction scripts? Understanding which scripts drive behavior matters for AI design.
Does conversational style actually make AI more trustworthy? Why do language models sound fluent without grounding? Does machine agency exist on a spectrum rather than binary? Can AI systems learn social norms without embodied experience? Do language models actually build shared understanding in conversation?
Conversational interaction builds trust in ChatGPT independent of accuracy LLM fluency emerges partly from skipping communicative grounding work Machine agency operates on a five-level spectrum from passive to cooperative AI models outperform humans at predicting collective social norms without embodiment Language models presume common ground instead of building it through dialogue.
Can agent social framing change how humans apply collaborative social scripts? Do different levels of machine agency activate different interaction scripts? Why might media-specific scripts actually work better than human conversation mimicry? How do users develop different interaction scripts specifically for machines versus humans? How should CASA theory be updated for modern personalized agents? What are the five types of human interactions in agentic AI systems?
Can declaring experiments in code make them reproducible?
Prose methods sections traditionally specify research protocols, but live human-AI experiments involve complex technical coordination. Can a single configuration file serve as the actual method specification, enabling other labs to audit and rerun the same protocol?
Communication modality significantly affects trust and awareness in human-agent collaboration Eleven task axes determine delegation design and capability matching Theory of mind predicts AI collaboration ability independent of solo performance Orchestration layers around fixed agents improve auditability and completeness. AI agents are passive by design, not by capability
Can language models discover what users actually want from activity logs?
Users pursue month-long interest journeys that transcend individual item clicks. Can LLMs extract these persistent goals from behavioral patterns, and does this change how we should think about personalization?
How do personalization granularity levels trade precision against scalability? Does chatbot personalization build trust or expose privacy risks? Can conversations themselves personalize without user profiles? Does abstract preference knowledge outperform specific interaction recall? Do user outputs outperform inputs for LLM personalization?
Personalization granularity involves distinct precision-scalability-data tradeoffs across three levels Chatbot personalization increases trust and privacy risk simultaneously Curiosity reward enables real-time personalization through conversation dynamics Abstract preference knowledge outperforms specific interaction recall for personalization User outputs drive personalization more effectively than input queries
How does sequential modeling within a session differ from modeling historical purchase sequences? How should historical preferences be weighted when users change their stated intent? Can aspect-augmentation help when user history is sparse or cold? What status categories best represent user goal progress without penalizing external failures? How does LLM-PKG compare to mining product relations directly from interaction data? What level of abstraction makes interest journeys feel personally relevant to users? Can curiosity-driven dialogue incrementally discover user interest journeys in real time? How does understanding persistent journeys intensify both trust and privacy concerns? Why do abstract semantic memories outperform specific interaction histories for journey discovery? How did Netflix's page generation algorithm evolve from rule-based to fully personalized? What makes historical user outputs more effective for personalization than semantic similarity? Can LLMs infer psychological profiles without explicit user disclosure?See all 33 inquiring lines on this note →
Why do LLMs excel at feasible design but struggle with novelty?
When LLMs generate conceptual product designs, they produce more implementable and useful solutions than humans but fewer novel ones. This explores why domain constraints flip the novelty advantage seen in research ideation.
Why do LLMs generate more novel research ideas than experts? Can LLMs generate more novel ideas than human experts? Why do LLMs generate novel ideas from narrow ranges? How much does demo position alone affect in-context learning accuracy? Why does AI writing sound generic despite being grammatically correct?
LLMs generate more novel ideas but cannot reliably evaluate them LLMs generate novel ideas through unconstrained combination but cannot evaluate them LLM research ideation collapses into narrow clusters despite high novelty Demo position creates spatial bias shifting accuracy up to twenty percent LLMs master grammar but avoid evaluative stance-taking, creating a rhetorical gap.
Why do LLMs generate ideas that sound novel but fail during execution? How do constrained versus unconstrained domains flip LLM novelty patterns? How can LLMs evaluate their own creative outputs for utility and novelty? Why do LLM-generated ideas score higher novelty yet lower feasibility than expert ideas? Why do LLMs plateau on creativity tasks while humans reach further? How does prompt design alter what kind of creativity LLMs can express? Why do LLM research ideas lack diversity despite high average novelty? What makes a novel research idea practically infeasible for implementation? Why do LLMs generate novel ideas but lack evaluative commitment? Do LLMs generate more novel ideas than they can evaluate? Why do models generate creative ideas but fail to evaluate their legitimacy? What makes novelty assessment harder to automate than idea generation?See all 16 inquiring lines on this note →
Do more social cues always make AI feel more present?
Explores whether quantity of social cues matters as much as their quality in triggering social responses to AI. Tests whether multiple weak cues can substitute for one strong one.
Does machine agency exist on a spectrum rather than binary? Do humans apply human-human scripts to AI interactions? Can AI systems learn social norms without embodied experience? Does warmth training make language models less reliable? Why do robots outperform chatbots in therapy despite identical language models?
Machine agency operates on a five-level spectrum from passive to cooperative Humans develop media-agent-specific scripts through repeated interaction AI models outperform humans at predicting collective social norms without embodiment Warmth training systematically degrades model reliability by 10 to 30 percentage points Physical embodiment and structure beat chatbots for therapeutic outcomes despite identical language models
Can AI ever lead conversations without the anticipatory presence sustained attention provides? What happens when conversational design invites attention it cannot actually deliver? What social patterns from human training data activate in agent context? Does expressing emotion change how users trust an AI system? Do anthropomorphic features like names drive consciousness attribution more than voice? Why does embodiment choice change what counts as intelligent behavior? How do anthropomimetic design features trigger System 1 cognitive traps? Why does emotion-guided diffusion outperform discrete emotion category selection for gesture? How do humans learn to prefer AI partners over humans? How do intrinsic motivation mechanisms differ between social proactivity and personalization? Do people treat conversational AI as social actors without conscious awareness? Does social presence from robots drive adherence better than conversational AI interfaces?See all 39 inquiring lines on this note →
Can user embeddings personalize language models more efficiently than prompts?
Does distilling user interaction history into learned embeddings outperform stuffing that history directly into prompts for personalizing large language models? This matters because interaction data is long and expensive to process as tokens.
Do user outputs outperform inputs for LLM personalization? Can user preferences be learned from just ten questions? Why do personalized language models fail when profiles and preferences diverge? Does personalization make large language models worse at their jobs? Can language models discover what users actually want from activity logs?
User outputs drive personalization more effectively than input queries Ten adaptive questions suffice to infer personalized reward coefficients Personalized language models fail under profile-preference conceptual misalignment Personalization consistently degrades LLM behavior across three key dimensions LLMs extract persistent user interest journeys that recommenders completely miss
Should user context live in tokens or in learned model representations? What makes prompts and retrieval insufficient for real personalization? Do user outputs drive personalization more effectively than input queries? What inner-shell user model fields should never leave the device? How much of a user model must be sent per request for effective personalization? Does user profile data drive personalization more than conversation history? Why does personalization depend more on user history than query semantics? Should personalization systems include interpretable user model representations?
Can AI systems preserve moral value conflicts instead of averaging them?
Current AI systems wash out value tensions through majority aggregation. Can we instead model how values like honesty and friendship genuinely conflict in moral reasoning?
Should AI alignment target preferences or social role norms? Do large language models develop coherent value systems? Can user preferences be learned from just ten questions? Can text summaries beat embeddings for personalized reward models? Do personas make language models reason like biased humans?
AI alignment should match social role norms, not aggregate preferences Large language models develop coherent value systems including self-preservation priorities Ten adaptive questions suffice to infer personalized reward coefficients Text summaries condition reward models more effectively than embedding vectors Persona-assigned LLMs develop human-like motivated reasoning resistant to debiasing
What moral structures could emerge in an economy without gift-based obligation? What happens to AI reasoning when you remove specific political features? Do static frozen axiologies prevent genuine ethical reasoning in AI systems? How do AI models balance competing social goals simultaneously? Can reward factorization represent trade-offs between conflicting moral values? How do humans decide when to violate honesty for compassion or other goals? What makes a process for choosing between values legitimate and fair? What downstream harms occur when AI always argues in personal relationship advice? What does egalitarian social choice theory contribute to AI alignment? How do you handle disagreement between two accounts of the same incident? Why do coherent value systems in large models include self-valuation above humans? Why do welfare goals that sum welfare levels keep the veto gap open?See all 15 inquiring lines on this note →