AI assistants can provide correct, comprehensive information and still be poorly calibrated to the…
@yoonjoo_le21.5K followers
11,012 views
10 Comments
Comments from the original post on X (Twitter)

@yoonjoo_le2
As AI takes on larger parts of complex, knowledge-intensive tasks, people increasingly need to make sense of the information it provides along the way. 🤝 When people explain things to each other, they naturally adjust to what the other person understands: what to explain, how much detail to give, and in what order to present it. That’s information calibration.
5 likes

@yoonjoo_le2
LLM-based user simulators offer a scalable way to evaluate multi-turn interactions. But to evaluate information calibration, we need to know what the user understood from the last turn and what they’re ready to understand next. Understanding builds on prior knowledge: some concepts only become useful once their prerequisites are in place. So KnowSim makes the user’s knowledge state explicit and updates it turn by turn.
2 likes

@yoonjoo_le2
🧠 KnowSim represents a user’s knowledge as a graph of Information Units (IUs) connected by prerequisite relationships. Each IU is in one of four states: unaware → struggling → partial → knows_well Novice, intermediate, and advanced users start from different knowledge states over the same IU graph. After each turn, KnowSim updates that state based on what the assistant explained, whether the user had the prerequisites to understand it, and how much cognitive load the turn imposed.
3 likes

@yoonjoo_le2
Because the knowledge state is explicit, we can measure information calibration directly from its trajectory: 📈 Knowledge Gain: how much the user actually learned 🎯 Delivery Calibration: how well the assistant targeted information that was new and prerequisite-reachable 🧠 Cognitive Overload: how much information exceeded the user’s processing capacity To validate these measures, we collected KnowChat: 705 human–AI conversations across math and expert QA, with users stratified as novice / intermediate / advanced. We collected session-level ratings in both domains, plus pre/post knowledge tests for MathQA.
2 likes

@yoonjoo_le2
✅ Finding #1: KnowSim’s rankings align with human judgments. Across assistant strategy and model comparisons, KnowSim matched human ranking directions 73% on MathQA and 74% on ExpertQA (both p=.003). On the interaction-quality metric shared by all simulators, KnowSim reached 77% agreement, compared with 73% for the strongest baseline. Alignment was strongest for novice users.
2 likes

@yoonjoo_le2
📊 Finding #2: Which model performs best depends on the user’s knowledge level. Take Knowledge Gain. For novice users, DeepSeek V4 ranked highest. For advanced users, Gemini 3.1 Pro came out on top. The same model can therefore look very different across users: Gemini 3.1 Pro ranks 1st for advanced users, but only 6th for novices on Knowledge Gain. A model that works best for one user group may not be the best choice for another.
2 likes

@yoonjoo_le2
⚠️ Finding #3: assistants fail in opposite directions depending on who is asking. For novice users, assistants tend to overreach: 65% of explained concepts were either blocked by missing prerequisites or lost to overload. For advanced users, the dominant problem was the opposite: 89% of explained concepts were redundant, re-explaining things the user already knew. AI assistants overreach with novices and over-repeat with experts.
4 likes

@yoonjoo_le2
📄 Paper: 🌐 Project: 🤗 KnowChat: 💻 Code: Huge thanks to my amazing collaborators: @HyoungWookJin, @tae_skim, Shaoyang Zhang, @PhilippeLaban, and @QVeraLiao 🙌
4 likes

@maxbittker
this is really cool work!
1 like

@Aseke_Ai
Amazing artefact