Instead of retyping who you are every chat, could you just edit the AI's profile of you directly?
Can editing a chatbot's user model replace typing long context prompts?
This explores whether you could skip writing long background prompts by directly viewing and adjusting what a chatbot has already inferred about you (your age, expertise, preferences), and what the research says about the trade-offs.
This explores whether adjusting a chatbot's internal picture of you could stand in for typing out who you are and what you want every time. The corpus says yes in principle, with a catch: a user model is a much more powerful lever than a prompt, and it can change more than you meant it to.
The most direct evidence is a dashboard study. Researchers exposed a chatbot's inferred guesses about a user's age, gender, education and socioeconomic status, and let people see them Does showing a chatbot's inferred user model increase user control?. Most participants felt more in control. Some were unsettled to learn the inference was happening at all, and several worried about targeted advertising. The surprising part comes from the same line of work. When researchers set the chatbot's internal sense of a user's socioeconomic status to 'low', it falsely said there were no direct flights to Hawaii Can steering a chatbot's user model change its factual claims?. A prompt mostly changes what you ask for. A user model can change the facts the model reports back. So editing it is not a neutral shortcut. It works on a layer that quietly shapes everything the model says.
On the engineering side there is real evidence that compact user representations beat long text. User-LLM compresses a person's interaction history into embeddings that the model reads directly. On long histories it personalizes better than text prompts and costs less to run Can user embeddings personalize language models more efficiently than prompts?. Small adapters show the same pattern for personality. They can set traits like the Big Five accurately with under 0.1% extra parameters Can we control personality in language models without prompting?. Prompting often fails at this, because many open models stick to their trained default personality whatever you tell them Can open language models adopt different personalities through prompting?. If prompts are the weak lever, working on the model's internals is the strong one.
There is also a design reason to want this. A prompt packs your message, the background context and the role you want the model to play into one fixed block that the model can't renegotiate How do prompts reshape the role of context in AI conversation?. That rigidity is part of why conversations go wrong. Models lock onto early guesses and seldom recover when information arrives piece by piece Why do AI assistants get worse at longer conversations?. An editable user model looks like the kind of interface Friedman argues for: direct controls for capturing intent, replacing slow typed explanation Can task-oriented UI patterns reduce the friction of AI chat interfaces?. It also fits evidence that generated, structured interfaces beat plain chat in most tasks Do generated interfaces outperform text-based chat for most tasks?.
The limit: a user model holds who you are, not what this task needs. It can't hold a specific document, a dataset or new domain knowledge. For that kind of bulk material the corpus points to different tools: compressing prompts about 20x Can prompts be compressed without losing model performance?, or letting the model query a long prompt like an external environment it searches with code Can models treat long prompts as external code environments?. The likely future is a split. An editable user model handles the 'who I am' part, and retrieval or compression handles the 'what I'm working on' part. That division only helps if users can see the hidden assumptions the model is acting on, because those assumptions can quietly change the answers they get.
Sources 11 notes
A dashboard exposing a chatbot's inferred age, gender, education, and socioeconomic status made most participants feel more in control and transparent, but five felt uncomfortable with the inference itself and seven worried about privacy risks like targeted advertising.
A causal intervention that set a chatbot's internal SES representation to low caused it to falsely claim no direct flights existed when asked about Hawaii travel. This demonstrates that attribute-steering can alter factual content, not just communication style.
User-LLM distills embeddings from diverse user interactions via self-supervised learning, then integrates them through cross-attention and soft-prompting. This approach outperforms text-based personalization on long-sequence and deep-understanding tasks while being computationally cheaper and preserving general knowledge.
PsychAdapter modifies every transformer layer with <0.1% additional parameters to achieve 87.3% Big Five accuracy and 96.7% depression/life satisfaction accuracy across GPT-2, Gemma, and Llama 3. This architecture-level approach bypasses prompt resistance entirely.
Research shows most open models fail to adopt prompted personalities, stubbornly retaining their trained ENFJ-like defaults. Only a few flexible models succeed. Combining role and personality conditioning improves results but doesn't fully overcome resistance.
Show all 11 sources
LLM prompts bundle utterance, context assignment, and role specification into a single static frame the model cannot renegotiate, unlike human dialogue where context evolves cooperatively. This makes mid-conversation pivots require explicit re-prompting rather than implicit adjustment.
LLMs perform at 90% accuracy with single-message instructions but drop to 65% across natural conversation. Models lock into early guesses when information arrives gradually and cannot course-correct, a behavior induced by RLHF training that rewards helpfulness over clarification.
Friedman identifies five design levers—intent capture, output shaping, refinement, agentic tasks, and workflow embedding—that sidestep slow typed prompts. Examples include node-based intent visualization, style lenses for output, and canvas manipulation, shifting AI interfaces away from conversational chat toward direct interaction.
Research shows users strongly prefer LLM-generated interactive interfaces—dashboards, tools, animations—over text blocks, especially for structured and information-dense tasks. Structured representation and iterative refinement reduce cognitive load.
LLMLingua achieves up to 20x prompt compression with only 1.5-point performance drops by allocating less compression to instructions and questions, more to demonstrations, then pruning tokens using a small model's perplexity scoring aligned to the target LLM.
Recursive Language Models store long prompts in a Python REPL and query them via code execution, avoiding attention degradation. RLMs outperform base models even on shorter prompts while handling inputs two orders of magnitude beyond context windows.
Papers this line draws on 8
The research behind the notes this line reads — ranked by how closely each paper relates.
- LLMLingua: Compressing Prompts for Accelerated Inference of Large Language Models
- Recursive Language Models
- Designing a Dashboard for Transparency and Control of Conversational AI
- When Single-User-Oriented LLM-based Assistants Involve Others: A Scoping Review of Pathways, Risks, and Responses
- ACON: Optimizing Context Compression for Long-horizon LLM Agents
- Design Theater: Evaluating the Gap Between User-Facing Design Reasoning and Implementation in Generative UI Tools
- Psychologically Enhanced AI Agents
- Disclose with Care: Designing Privacy Controls in Interview Chatbots