TL;DR
CURIOBOT is a set of five operators that change not what an LLM says but how it delivers information: through contradiction, novelty, complexity, uncertainty, or neutrally. The operator is chosen based on the state of the conversation and creates epistemic tension: a gap between what the learner thinks they know and what turns out to be true. The whole mechanism works through prompting, with no fine-tuning.
A regular LLM acting as a tutor defaults to exhaustive explanations, and that's a trap. The learner reads, says "got it," and the conversation dies. Not because it's boring, but because information without tension doesn't demand a response. The brain files the fact away and moves on. A standard session: 4-7 turns, then silence.
The operators break that pattern. CONFLICT introduces a contradiction into the learner's logic, and they have to resolve it. UNCERTAINTY says "scientists are still arguing about this," and the learner starts asking questions on their own. NOVELTY reframes things through an unexpected analogy, and the learner checks whether it holds up. The result: 2.4x more turns in the conversation in the same amount of time.
Want this for your own case? Apply this study to your task: AI adapts its method to what you are working on and writes the prompt.
Method Blueprint
Everything happens in one conversation — no separate requests needed
STEP 1: Assess the learner's state
└─ Short answers / "got it, thanks" → NOVELTY or CONFLICT (restart engagement)
└─ Premature confidence → COMPLEXITY or UNCERTAINTY (broaden)
└─ Actively asking questions → NEUTRAL (stay out of the way)
STEP 2: Shape the response through the chosen operator
NOVELTY → An unexpected analogy or angle
COMPLEXITY → "Here's what else is tied to this..." — new dependencies
CONFLICT → A contradiction inside what the learner just said
UNCERTAINTY → An unresolved question in science/practice on the topic
NEUTRAL → A normal answer with no attention steering
STEP 3: Don't lock in one operator — switch as needed
└─ Repeating the same operator weakens the effect
└─ A strong transition: UNCERTAINTY → CONFLICT
Example in Practice
The task: You sell on Amazon and want to figure out why revenue is growing but profit isn't. You sit down to learn unit economics in a conversation with Claude.
Prompt:
You are a unit economics tutor. Topic: unit economics of selling on Amazon.
My goal: understand why profit isn't growing while revenue is.
How to work with me:
— Don't give me ready-made answers right away. First ask what I already think.
— Watch my answers. If I give a short "got it" or just agree —
switch: introduce a contradiction into my logic or show an unexpected angle.
— If I'm too confident — make it harder: show dependencies I haven't considered.
— If I'm actively digging — stay out of the way and answer neutrally.
— Alternate approaches: contradiction → novelty → uncertainty.
Don't use the same move twice in a row.
— Never wrap up the topic yourself. Let me decide when it's enough.
Start with a question: what do I currently think about how revenue and profit relate on a marketplace?
Result:
The model won't start with explanations but with a question. After your answer it will most likely introduce a contradiction: "You said you need to sell more, but here's the interesting part: it's precisely when sales grow that most sellers see their margins fall. How does that work?" You'll start digging. A few turns later UNCERTAINTY shows up: "Even among financial analysts there's no consensus on the volume at which Amazon unit economics become profitable, because fees are opaque and change retroactively." The conversation will run 15-20+ turns instead of the usual 5.
Why It Works
By default an LLM optimizes for "correctness": it gives a complete, accurate, well-structured answer. That kills curiosity. Once the gap is filled, there's no incentive to move on. The model behaves like a good Wikipedia: exhaustive and lifeless.
Curiosity is a response to tension, not to information. Back in the 1960s, psychologist Berlyne described how people start actively exploring when they encounter novelty, complexity, contradiction or uncertainty, meaning when their picture of the world doesn't add up. The CONFLICT operator showed the best result (79% successful applications) precisely because it creates the sharpest tension: "you just said X, but here's fact Y — how do they fit together?"
The operators use what LLMs do well: generate precisely to instruction. Tell a model to "introduce a contradiction into the learner's logic without revealing the solution" and it will do exactly that. An operator isn't magic, it's a precise instruction for managing an information gap. You set the mechanics, the model executes.
Control knobs:
- Don't use the same move twice in a row → the key parameter. Repeating NOVELTY halves the effect. Alternating keeps the tension up.
- Replace neutral verbs with specific triggers: "introduce a contradiction" works better than "create interest."
- Add the condition never wrap up the topic yourself → otherwise the model will "close the question" at the end of every answer.
Prompt Template
You are a tutor on the topic: {topic}.
My goal: {what I want to understand or learn to do}.
How to work with me:
— Start each topic by asking what I already think — don't explain right away.
— Watch my answers:
• Short answer / "clear" / "got it" → switch:
introduce a contradiction into my logic OR show an unexpected angle
• I'm confident and not doubting → make it harder:
show dependencies or mechanics I haven't considered
• I'm actively asking questions and working through the topic → answer neutrally, stay out of the way
— Rotate approaches in a cycle:
contradiction → novelty → uncertainty → complexity
Don't use the same approach twice in a row.
— Never "close the topic" at the end of an answer.
End with an open question or an unresolved aspect.
Start: ask what I currently think about {topic}.
What to fill in:
- {topic} — specific, not broad: "SaaS pricing" instead of "business"
- {what I want to understand} — a concrete pain point or question, not "figure it out"
🚀 Quick start — paste into the chat:
Here is the CURIOBOT template — a method of learning through guided curiosity.
Adapt it to my task: {your topic and goal}.
Ask me questions to fill in the fields.
[paste the template above]
The LLM will ask for the topic and what exactly you want to understand, because without that it can't set up the right triggers for switching between operators. It will take the pattern from the template and build the session around your task.
Limitations
⚠️ UNCERTAINTY works worst: The uncertainty operator had the lowest success rate, just 44%. "Scientists don't know yet" only works if the learner is already engaged. At the start it tends to push people away: "why bother if even the experts haven't figured it out?"
⚠️ Repetition kills the effect: If the model uses the same technique over and over, the learner zones out. You have to control this explicitly in the prompt. Without an "alternate" instruction, the model gets stuck on NOVELTY, which is its default favorite.
⚠️ Not for factual requests: If you need a specific answer, like "what's today's exchange rate" or "write an email," CURIOBOT mode slows you down and gets annoying. It's a learning tool, not a lookup tool.
⚠️ Gemini behaves differently: With one of the tested models, explanation quality dropped when the operators were on: curiosity went up but accuracy went down. If you notice answers getting less accurate, add to the prompt:
Explanation quality must not suffer — change only the delivery, not the content.
How They Tested It
A team from Singapore Management University and Carnegie Mellon ran a controlled 3×3×3 experiment: three model families (Claude, Gemini, GPT), three domains (STEM, social sciences, humanities), and three topic difficulty levels. 45 participants completed a total of 270 sessions of 10-30 minutes each, depending on topic difficulty.
An interesting design detail: participants didn't know which tutor version they were using. Each did sessions with all three variants (the base model, CURIOBOT, and the provider's native learning mode), but on different topics, to avoid carryover effects. That let the researchers compare fairly on the same group of people.
Evaluation covered 10 dimensions via LLM judges: three different models, three independent runs each, for 9 ratings per conversation. That's a nontrivial methodology: studies usually use a single judge, and here they deliberately aggregated to remove any one model's systematic bias.
The most surprising finding: the improvement on the learner's side held even when the tutor's explanation quality dropped. That means curiosity works as a partly independent mechanism: not "it was explained better, so there were more questions," but "the tension was set up right, and the person started digging on their own." CONFLICT also turned out more effective than NOVELTY, 79% vs 61%, even though intuitively it seems easier to show something new than to introduce a contradiction.
Original from the Paper
NOVELTY — operator example (topic: the Rule of Thirds in photography):
"Imagine the frame as a quietly haunted chessboard where the most powerful
square is almost never the center — that strange bias is the doorway..."
CONFLICT — operator example (same topic):
"...common sense screams that the most important subject must sit dead center
to command absolute attention. Yet... centering an object actually paralyzes
the viewer's gaze, creating a stagnant visual trap..."
UNCERTAINTY — operator example (topic: game theory in international relations):
"...here's what even top scholars fiercely debate... whether game theory
explains international relations or merely provides elegant metaphors...
remains one of the deepest unresolved questions in the field...
Parting Mystery: if two nations are both using game theory to model each other,
does the model become a self-fulfilling prophecy that creates the very behavior
it claims to predict?"
COMPLEXITY — operator example (topic: the Mpemba effect):
"...pulling us into the hidden thermodynamic complexities... However, these
classical mechanics still fall short of explaining why perfectly sealed
containers, where absolutely no evaporated mass can escape, continue to
demonstrate this astonishing thermodynamic paradox."
Context: The researchers presented the same topic through different operators to show clearly that only the frame changes, not the content.
Adaptations and Extrapolations
1. Adaptation: use the operators as a warm-up before a hard request
Before asking an LLM to solve a problem, run a quick CONFLICT round on the topic.
💡 Adaptation for negotiation prep:
I'm preparing to negotiate a 15% price cut with a supplier.
I think my main argument is order volume.
Introduce a contradiction: why might volume NOT work as an argument in this case?
Don't reveal the solution — just create tension so I find the best move myself.
2. Adaptation: CONFLICT mode to stress-test an idea
🔧 Technique: ask for the CONFLICT operator directly → sharp criticism without platitudes
Instead of "critique my idea" (the model will give you a diplomatic list), ask it to introduce a specific contradiction:
Here's my idea: {describe the idea}.
I'm sure that {your key assumption}.
Introduce a contradiction right into that assumption: show a real case or mechanism
where this belief leads to failure. Don't give a solution — just expose the gap.
This cuts sharper than asking "find the weak spots," because CONFLICT hits a specific point of confidence instead of scanning everything.
3. Extrapolation: a 10-dimension frame for evaluating your own LLM session
The study offers 4 learner-side metrics you can use as a checklist after a learning session:
After a learning session, rate it on 4 questions:
1. Did I ask questions I hadn't planned to? (Exploratory questions)
2. Did I revise my point of view along the way? (Productive struggle)
3. Did I want to dig deeper than the task required? (Visible curiosity)
4. Did I steer the conversation myself — change topics, add examples? (Agency)
If the answer to all 4 is "no," the session was passive.
Ask the LLM to restart the topic, opening with the CONFLICT operator.
Resources
Paper title: Curiosity as Linguistic Intervention: Using LLM Tutoring Dialogues to Influence Exploratory Learning Behavior
Authors: Gevindu Ganganath, Pasindu Bolonghege, Qianru Lyu, Pradeep Varakantham, Thivya Kandappu
Organizations: School of Computing and Information Systems, Singapore Management University; Human-Computer Interaction Institute, Carnegie Mellon University
Theoretical foundation: Berlyne's theory of collative variables (Berlyne, 1960, 1962), the core psychological theory the operators are built on