💬 NLP

DiscoverLLM: From Executing Intents to Discovering Them

DiscoverLLM is a novel framework that trains Large Language Models to help users discover and concretize their ambiguous intents through a cognitive state-based user simulator, resulting in significantly improved task performance, shorter conversations, and higher user satisfaction across creative, technical, and visual tasks.

Original authors: Tae Soo Kim, Yoonjoo Lee, Jaesang Yu, John Joon Young Chung, Juho Kim

Published 2026-05-14

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Tae Soo Kim, Yoonjoo Lee, Jaesang Yu, John Joon Young Chung, Juho Kim

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to order a very specific meal from a chef, but you don't actually know what you want to eat yet. You just know you're hungry and that you want something "good."

If you ask a standard AI chef, "Make me something good," it might guess a steak. You take a bite and say, "Hmm, not quite right." The chef asks, "Do you want it spicier?" You say, "I don't know." The chef guesses again, maybe a salad. You say, "No, that's too cold." This back-and-forth can go on forever because the chef is waiting for you to know exactly what you want before they can cook it.

DISCOVERLLM is a new way of training AI chefs to realize that you might not know what you want until you see the food.

Here is how the paper explains this, using simple analogies:

1. The Problem: The "Hidden Menu"

Usually, AI models assume you have a complete order in your head (a "fully formed intent"). They think their job is just to ask clarifying questions like, "Do you want it hot or cold?"

But in real life, especially with creative tasks like writing a story or drawing a picture, you often don't know the answer until you see an example. You can't tell the chef you want "spicy" if you haven't tasted anything spicy yet. You need to discover that you like spicy food by trying it.

2. The Solution: The "Tasting Menu" Simulator

The researchers built a special "simulated user" (a robot acting like a human) to train the AI. This robot has a secret, hidden menu of preferences that it doesn't know it has yet.

The Hidden Menu: Imagine the robot has a secret list of desires: "I want a poem," "I want an animal," "I want a cat," "I want a sleeping cat."
The Discovery Process: At first, the robot only knows it wants "a poem." It doesn't know it wants a cat.
The AI's Job: Instead of just asking, "What animal do you want?" (which the robot can't answer yet), the AI tries offering different options.
- AI tries: "Here is a poem about a dog."
- Robot thinks: "Oh, a dog is okay, but maybe I want something softer." (The robot discovers it prefers a pet).
- AI tries: "Here is a poem about a sleeping cat."
- Robot thinks: "Yes! That's it!" (The robot has now discovered its true intent).

3. The Training: Learning to "Nudge"

The AI is trained using a reward system. It gets points not just for giving the right answer, but for helping the user figure out what they want.

Bad Move: If the AI asks, "Do you want a cat or a dog?" when the user doesn't know, the AI gets no points. The user is stuck.
Good Move: If the AI shows a picture of a dog, and the user says, "No, maybe something smaller," the AI gets points. It successfully "nudged" the user to discover a new preference.

The paper calls this balancing Divergence (showing many different options to explore) and Convergence (narrowing down once the user knows what they like).

4. The Results: Less Talking, More Doing

The researchers tested this new AI on tasks like writing stories, technical articles, and drawing digital images.

Faster Discovery: The new AI helped users find what they wanted about 10% faster than older models.
Shorter Conversations: Because the AI stopped asking useless questions and started showing helpful examples, the conversations were 40% shorter.
Happier Users: In a study with real humans, people felt the new AI was more helpful and seemed to "anticipate" their needs, even when the users weren't sure what they wanted themselves.

The Big Picture

Think of the old AI as a waiter who waits for you to read the whole menu and order.
Think of the new DISCOVERLLM as a tasting chef who brings out small samples of different dishes. As you taste them, you realize, "Oh, I actually love the spicy one!" or "I hate the cold soup."

The paper claims that by teaching AI to be that tasting chef—helping you discover your own preferences through exploration rather than just waiting for instructions—we can make AI much better at helping with creative and open-ended tasks.

Technical Summary: DISCOVERLLM

Problem Formulation

Current Large Language Models (LLMs) are predominantly trained to execute and elicit user intents, operating under the assumption that users begin interactions with fully formed, albeit sometimes unarticulated, goals. However, in open-ended, creative, and ill-defined tasks (e.g., writing, design), users often lack clear intents initially. Instead, they discover what they want by observing and exploring potential outcomes. Existing approaches, including multi-turn fine-tuning and prompting strategies, typically assume intents are hidden but static, focusing on surfacing them through clarification questions. This fails when users cannot answer questions like "what tone do you want?" because they have not yet formed that preference. The core challenge is to train models that do not merely elicit pre-existing intents but actively help users discover and form their intents through collaborative exploration.

Methodology: The DISCOVERLLM Framework

The authors propose DISCOVERLLM, a generalizable training framework that operationalizes intent discovery as a distinct problem from intent elicitation. The framework relies on a novel User Simulator that models the cognitive state of a user with a latent, hierarchical intent structure.

1. Intent Discovery Through Interaction

The framework formalizes intent formation as a progressive refinement process ( $I_0 \subseteq I_1 \subseteq \dots \subseteq I_T$ ).

Latent Intent Hierarchy: User intents are represented as a tree hierarchy ( $H$ ) where nodes range from broad (e.g., "includes an animal") to specific (e.g., "includes a tabby cat sleeping on a sofa").
Discovery States: Nodes exist in states of undiscovered, emerging, or discovered. A user can only articulate intents that are discovered.
Response-Driven Progression: The assistant's response ( $r_t$ $r_{t}$ ) determines state transitions.
- Direct Engagement: If the response explicitly asks about or satisfies an intent, the node becomes discovered.
- Tangential Engagement: If the response offers related but non-matching options (e.g., suggesting a dog when the user wants a cat), it contributes to a cumulative score. Exceeding a threshold advances the node from undiscovered to emerging or discovered, modeling how users narrow preferences by seeing what they do not want.
Expressiveness Constraint: Users in the simulation can only express intents they have discovered. If an intent is only emerging, the user can only provide vague feedback (e.g., "maybe something different"), preventing the model from simply asking for specific details the user cannot yet provide.

2. Reward Computation

The framework computes a per-turn reward ( $R$ ) to guide training:

Discovery Progress ( $R_d$ ): Measures the increase in the number of discovered intents ( $|I_{t+1}| - |I_t|$ ). This rewards the model for surfacing latent preferences.
Efficiency Penalty ( $R_e$ ): Penalizes excessive response length to encourage concise interactions.
Total Reward: $R(r_t) = R_d(r_t) + R_e(r_t)$ . The design prioritizes discovery first, with efficiency as a secondary constraint.

3. Training Pipeline

Using the user simulator and reward function, the framework supports multiple training paradigms:

Synthetic Data Generation: Simulating conversations to generate high-quality datasets for Supervised Fine-Tuning (SFT).
Preference Optimization: Creating pairwise comparisons (chosen vs. rejected responses) for Direct Preference Optimization (DPO).
Online Reinforcement Learning: Optimizing models directly on the reward signal using algorithms like PPO or GRPO.

Key Contributions

Formalization of Intent Discovery: The paper distinguishes between eliciting hidden intents and discovering unformed ones, proposing a formal model where intent concretization is driven by the assistant's exploration of the solution space.
Novel User Simulator: A simulator that models user cognitive states as a hierarchy of intents, enabling the computation of rewards for intent discovery—a capability previously unavailable in standard training setups.
Training Framework: A complete pipeline (SFT, DPO, RL) that teaches LLMs to balance divergence (exploring options to probe abstract intents) and convergence (refining outputs to satisfy concrete intents).

Experimental Results

The framework was evaluated across three domains: Creative Writing, Technical Writing, and SVG Drawing, using Llama-3.1-8B-Instruct and Qwen3-8B as base models.

Performance Gains: DISCOVERLLM achieved over 10% higher intent discovery and 11% higher intent satisfaction compared to the best baselines (including prompting and COLLABLLM).
Efficiency: It reduced conversation length by up to 40% while maintaining or improving task performance.
Interactivity: Interactivity scores increased by 83%, indicating more effective collaboration.
Behavioral Analysis: Unlike base models that tend to be purely convergent (revising single drafts) or SFT models that overfit to divergence (offering endless options), DISCOVERLLM variants demonstrated a balanced pattern of alternating between divergent exploration and convergent refinement.
Generalization: The approach generalized to unseen domains (travel planning, web development) and different simulator configurations (varying backbone LLMs and update rules).
User Study: In a study with 75 human participants, DISCOVERLLM yielded higher interaction and writing satisfaction scores and reduced task completion time compared to baselines. Participants noted the model seemed to "anticipate" their latent intents.

Significance and Claims

The paper claims that DISCOVERLLM represents a shift in human-centered AI from models that simply execute instructions to proactive partners that explore and shape problems with users. By addressing the cognitive reality that users often discover their needs through action rather than pre-planning, the framework enables AI to support ill-defined, open-ended tasks more effectively. The authors emphasize that this approach does not compromise safety alignment and offers a scalable method for training models to collaborate on complex, creative endeavors where user intent is initially fluid. The work suggests that future AI systems should be designed to facilitate this co-evolution of problem and solution spaces, rather than assuming static user goals.