KnowSim: Evaluating Information Calibration in LLM Assistants with User Simulators that Learn

Yoonjoo Lee, Hyoungwook Jin, Tae Soo Kim, Shaoyang Zhang, Philippe Laban, Q. Vera Liao
University of Michigan    KAIST    Microsoft Research
Under Review

KnowSim is a user simulator that tracks what the user knows and how their understanding evolves turn by turn. Unlike existing simulators, it maintains a knowledge graph of concepts (Information Units), enabling mechanistic metrics—Knowledge Gain, Delivery Calibration, Cognitive Overload—that explain why the same response helps one user but overwhelms another. Validated against 705 human–AI sessions, KnowSim's rankings align with human judgments at 73–74% sign agreement.

The Problem

The same response can be perfect for one user and overwhelming for another. Human evaluation captures this—but doesn't scale. How do we automatically evaluate whether an assistant calibrates information to each user's knowledge level?

User question to a RAG assistant: "How does retrieval-augmented generation work, and when should I use it instead of fine-tuning?"
Beginner (new to ML)
Turn 1 — Assistant Response
Explains vector embeddings, chunking strategies, approximate nearest-neighbor search, hybrid retrieval with BM25, and reranking pipelines—all in one message.
Absorbs nothing—too many unfamiliar concepts at once
Turn 2 — User follow-up
"Wait, what are embeddings? And what's BM25?"
Slowly picks up basics, but still confused about the big picture
Outcome
After 5 turns, the user understands some vocabulary but still can't answer the original question: when to use RAG vs. fine-tuning. The assistant never matched its explanations to what this user actually needed.
Overwhelmed Low knowledge gain Poorly calibrated
ML Engineer (experienced)
Turn 1 — Same Assistant Response
Same explanation: embeddings, chunking, ANN search, BM25, reranking. The engineer already knows all of these.
Already familiar—absorbs instantly, no overload
Turn 2 — User follow-up
"Got it. What about latency tradeoffs between dense retrieval and hybrid approaches at scale?"
Asks deeper, more targeted questions
Outcome
In 3 turns, the engineer gets exactly the comparison they needed. But half the first response was redundant—time wasted on concepts they already knew.
No overload High knowledge gain Partially redundant
The core question: How do we automatically evaluate whether an assistant calibrates information to each user's knowledge level across a multi-turn conversation?

How KnowSim Works

A three-stage pipeline grounded in learning theory for interactive, multi-turn evaluation.

Pre-Interaction
01

IU Graph & Knowledge Initialization

From a question and reference answer, an LLM extracts a directed acyclic graph of Information Units (IUs)—self-contained concepts connected by prerequisite edges. The same graph is initialized with different mastery levels per user type.

Novice
Mostly unaware
Intermediate
Mixed understanding
Advanced
Mostly mastered
Same graph structure, different initial knowledge states per user level.
During Interaction
02

Knowledge-State Simulation

Each turn follows four steps:

A User Message

Conditioned on current knowledge state.

Unaware — cannot mention
Struggling — makes errors
Partial — uses with gaps
Known — applies correctly
B Assistant Responds

The model under evaluation (external).

C Signal Extraction

Per-IU teaching quality from assistant and engagement from user.

D State Update
Driver: well_explained → advance +1
Prereq. ceiling: capped by prerequisites
Cognitive load: too many IUs → reduced absorption
Monotonicity: never regresses
st−1 (before)
update
st (after)

Embeddings well explained: struggling → partial

Sim. search well explained: prerequisite not mastered → blocked

5 IUs in one turn → cognitive load

Evaluation Metrics

Computed directly from the knowledge state trajectory—mechanistic, interpretable, and human-aligned.

Knowledge Gain (KG)

Total ordinal state advancement summed over all IUs.

Delivery Calibration (DC)

Harmonic mean of precision over the IUs the assistant explained and recall over the IUs the user could learn, penalizing redundant, prerequisite-inappropriate, and ultimately unabsorbed delivery.

Cognitive Overload (CO)

Average per-turn information load relative to user capacity. Lower is better.

Interaction Quality (IQ)

Holistic quality scored by an LLM judge, shared across all simulators.

Simulated Conversations

Pick a question, then switch between user knowledge levels to see how KnowSim adapts the simulated conversation. Compare side-by-side with baseline simulators.

Both conversations use the same assistant (GPT-4.1, comprehensive strategy). Only the user simulator differs.
KnowSim (Ours)
Zero-Shot

Knowledge State Progression

How the simulated user's understanding evolves turn by turn. Each row is an Information Unit; colors show mastery level.

Qualitative Differences

Across all examples, KnowSim is the only simulator that produces actual mathematical errors at the novice level—baseline simulators say "I'm confused" but compute flawlessly. KnowSim novices also engage for 2–3× more turns, while advanced users terminate early after asking extension questions—a pattern consistent with real learners.

Alignment with Human Judgment

KnowSim's rankings align with human preferences at 73–74% sign agreement across two task domains, outperforming three baseline simulators.

Per-Level Sign Agreement

Key Findings

  • Novice level: 100% agreement (13/13, p<.001) — strongest alignment
  • Pooled: 73% agreement (27/37, p=.003)
  • KG, IQ, CO each reach 75% sign agreement
  • DC trails at 67%, reflecting difficulty of calibration ratings

Per-Level Sign Agreement

Key Findings

  • Novice level: 83% agreement (10/12, p=.019)
  • Intermediate: 80% agreement (8/10, p=.055)
  • Pooled: 74% agreement (26/35, p=.003)
  • IQ (80%, p=.055) and CO (79%, p=.029) lead; DC at 64%

Comparison with Baseline Simulators (IQ Sign Agreement)

KnowSim
77%**
ZS-CoT-Prof
73%*
ZS
68%†
ZS-CoT
64%
** p<.01   * p<.05   † p<.10   (one-sided binomial vs. 50% chance)

Frontier Model Benchmarking

Applied to 9 LLMs, KnowSim reveals that the best model shifts by user knowledge level—aptitude–treatment interactions invisible to aggregate leaderboards.

Key Takeaways

For Novice Users

DeepSeek V4 maximizes knowledge gain (KG=15.5) but at high cognitive overload (CO=0.945). Gemini 3.1 Pro ties for best R̅ (3.25) with the highest IQ (8.39) and lowest CO.

For Intermediate Users

Claude Opus 4.7 leads overall (R̅=2.50) with the highest IQ (8.39) and ties for best delivery calibration (DC=0.128).

For Advanced Users

Gemini 3.1 Pro achieves the best overall rank (R̅=1.75) with the highest KG and DC, best serving knowledgeable users.

Citation · If you use this work

@article{lee2026knowsim, title={KnowSim: Evaluating Information Calibration in LLM Assistants with User Simulators that Learn}, author={Lee, Yoonjoo and Jin, Hyoungwook and Kim, Tae Soo and Zhang, Shaoyang and Laban, Philippe and Liao, Q. Vera}, year={2026} }