Skip to content

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

Conference: EMNLP 2026
Paper: ACL Anthology
Code: https://github.com/yjo2lee/knowsim
Area: Medical LLM
Keywords: user simulator, information calibration, cognitive load, multi-turn evaluation, knowledge state tracking

TL;DR

Introduces KNOWSIM, an evaluation framework powered by a user simulator that models dynamic knowledge state transitions over an Information Unit (IU) prerequisite DAG with cognitive load constraints, providing explainable multi-turn information calibration benchmarks.

Background & Motivation

As Large Language Models (LLMs) are progressively deployed as collaborative partners in complex, knowledge-intensive workflows—such as clinical decision guidance, scientific discovery, and competitive mathematics tutoring—effective collaboration requires assistants to perform information calibration. Information calibration demands that an assistant delivers information matching the user's current conceptual grasp and limited cognitive capacity: presenting the right amount, at the appropriate depth, and in a prerequisite-consistent sequence across successive dialogue turns. However, standard post-training alignments such as single-turn Reinforcement Learning from Human Feedback (RLHF) incentivize models to maximize information density into comprehensive single-turn answers, providing virtually no supervision signal for how explanations should be paced across an evolving multi-turn exchange. In high-stakes medical contexts such as oncology consultations, this failure is acute: an oncologist asking about the signs that follicular lymphoma has transformed to diffuse large B-cell lymphoma can seamlessly absorb dense details regarding elevated SUV thresholds, bone marrow involvement, and core biopsy protocols, whereas a novice patient lacking foundational hematologic understanding is overwhelmed and unable to absorb anything.

To benchmark multi-turn assistants cost-effectively, recent literature has increasingly turned to LLM-based user simulators as scalable proxies for human evaluations. Yet existing simulators are fundamentally ungrounded behavioral proxies that exhibit two fatal blind spots. First, because they lack an explicit internal representation of user knowledge states, they cannot evaluate information calibration across user populations with divergent prior backgrounds; they only produce aggregate, level-agnostic utility ratings. Second, lacking a mechanism to track how knowledge actually evolves during an exchange, they cannot measure objective learning gains. Instead, they rely on LLM judges rating the surface fluency and politeness of dialogues, which perversely rewards chatty models that flood users with unsolicited information while completely failing to foster genuine conceptual understanding.

To resolve this limitation, evaluation simulators must shift from superficial role-playing prompts to causally explainable cognitive state machines grounded in established learning theories. Core idea: develop KNOWSIM, a knowledge-grounded user simulator operating over a directed acyclic graph of Information Units (IUs) and prerequisite dependencies, tracking turn-by-turn absorption dynamics under cognitive load constraints to compute three mechanistic metrics directly from knowledge trajectories—Knowledge Gain (KG), Delivery Calibration (DC), and Cognitive Overload (CO).

Method

Overall Architecture

The KNOWSIM evaluation pipeline comprises six coordinated stages: IU graph construction from a verified reference answer, knowledge state initialization stratified by user expertise, cognition-constrained user message generation, dual-signal extraction from interaction exchanges, deterministic learning-theoretic state transitions, and adaptive termination criteria. When evaluating an assistant model, KNOWSIM executes a multi-turn dialogue where the simulated learner's utterances are strictly bounded by their current mastery and connectable frontiers. After each assistant reply, an evaluator extracts pedagogical quality and learner reasoning attempts, updating the internal knowledge state while attenuating absorption whenever cognitive load exceeds capacity.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Input: Question q and Reference Answer r"] --> B["IU Graph Extraction<br/>Extract concept nodes V and prerequisite edges E"]
    B --> C["Knowledge State Initialization<br/>Topological assignment based on prior ratio R_l"]
    C --> D["Cognition-Constrained Message Generation<br/>Generate prompt ut based on state and local graph density"]
    D --> E["Assistant Model Generates Reply at"]
    E --> F["Dual-Signal Extraction<br/>Classify teaching quality and learner attempt type"]
    F --> G["Learning-Theoretic State Update<br/>Prerequisite ceiling gating + cognitive load dropoff"]
    H{"Adaptive Termination Check<br/>Mastery reached or 5-turn stagnation / overload"}
    G --> H
    H -->|Continue| D
    H -->|Terminate| I["Trajectory Metric Computation<br/>Compute KG, DC, CO, and interaction quality IQ"]

Key Designs

1. IU Graph Extraction: Grounding Unstructured Topics in Causal Prerequisite Topologies

Assessing whether an assistant's explanation is appropriately sequenced requires an objective decomposition of conceptual dependencies. KNOWSIM utilizes a frontier reasoning model to extract a directed acyclic graph \(G_q = (V, E)\) from a question \(q\) and its comprehensive reference answer \(r\). Each node \(v \in V\) represents an atomic Information Unit (IU)—a self-contained concept expressible in 2–4 sentences—associated with an abstraction level \(l(v) \in [0, 1]\) ranging from foundational principles (\(l \approx 1.0\)) to concrete diagnostic procedures (\(l \approx 0.0\)). A directed prerequisite edge \((v_i, v_j) \in E\) is enforced if and only if misunderstanding \(v_i\) guarantees that a learner will misinterpret \(v_j\). To ensure natural explanatory progression, the graph extraction enforces a bridging constraint restricting the abstraction gap along any single edge to \(\Delta l \le 0.35\), automatically inserting bridging IUs whenever larger conceptual leaps occur.

2. Knowledge State Initialization and Cognition-Constrained Generation: Preventing Simulator Omniscience

Unconstrained LLM user simulators frequently hallucinate mastery of advanced concepts because the underlying base model possesses extensive pre-trained knowledge. KNOWSIM discretizes learner understanding into four ordinal states: \(s_t(v) \in \{\text{unaware}(0), \text{struggling}(1), \text{partial}(2), \text{knows\_well}(3)\}\). The initial state \(s_0\) is topologically initialized according to empirically calibrated expertise ratios \(R_\ell\) across novice, intermediate, and advanced profiles (novices start with 10% knows_well and 50% unaware; advanced users start with 80% knows_well and 0% unaware). At each turn \(t\), the simulator conditions on the set of "connectable unknowns" (unlearned IUs whose direct prerequisites have been met). Furthermore, the simulator dynamically adjusts its articulation mode (explicit, vague, or deferential) based on the local graph density \(d\) of mastered neighboring concepts. The generation rules strictly prohibit simulated users from voluntarily mentioning unaware concepts and mandate authentic learner errors (such as misremembered definitions or inverted causal relationships) on struggling concepts, guaranteeing that the simulated user remains grounded in their assigned knowledge state.

3. Dual-Signal Extraction and Learning-Theoretic State Update: Dual Gating by Prerequisite Ceilings and Cognitive Overload

Determining knowledge acquisition requires decoupling superficial mention from genuine cognitive integration. After each turn pair \((u_t, a_t)\), KNOWSIM classifies the assistant's teaching quality into well_explained, shallow, or not_mentioned, and the learner's engagement into reasoning (producing novel deductions or computations), articulation (paraphrasing prior explanations), or none. The knowledge state vector \(s_t\) is then deterministically updated under two psychological principles: - Prerequisite Ceiling: An IU \(v_j\) cannot transition above the lowest mastery level of its immediate prerequisite parents. Even if an assistant provides an impeccable derivation of a downstream concept, the learner's state remains capped at struggling if underlying foundational concepts remain unmastered. - Cognitive Load & Sigmoid Dropoff: The total turn load \(L_t\) is computed as a state-weighted sum over all mentioned and reasoned IUs: $\(L_t = \sum_{v \in M_t} W_S(s_{t-1}(v)) + \sum_{v: \text{reasoning}} W_R(s_{t-1}(v)) + W_A \cdot \mathbb{I}(\text{articulation})\)$ where unfamiliar concepts incur higher cognitive weights (\(W_S(\text{unaware}) = 1.75\) vs. \(W_S(\text{knows\_well}) = 0.35\)). When \(L_t\) exceeds the intrinsic cognitive capacity threshold \(\tau = \max(2, \text{round}(0.55 \cdot |G_q|))\), absorption efficiency decays continuously via the sigmoid penalty factor \(\eta_t = \frac{1}{1 + ((L_t - \tau)/\tau)^2}\), modeling cognitive overload and penalizing excessive single-turn data dumps.

4. Progress-Grounded Termination and Tripartite Calibration Metrics: Disentangling Surface Ratings from True Learning

Rather than enforcing rigid fixed-turn cutoffs, KNOWSIM implements an adaptive termination policy derived from the learning progress hypothesis: dialogue concludes successfully if the learner achieves \(\ge 80\%\) mastery across all target IUs in \(G_q\), or terminates prematurely if the learner experiences persistent failure over a 5-turn sliding window (meeting at least two of: \(\ge 4\) overloaded turns, zero upward state transitions, or complete teaching silence). Over the resulting trajectory \(s_0, s_1, \dots, s_T\), KNOWSIM computes three interpretable metrics: - Knowledge Gain (KG): The net ordinal advancement across all graph concepts: \(KG = \sum_{v \in V} \max(0, s_T(v) - s_0(v))\); - Delivery Calibration (DC): The harmonic mean of precision \(P\) over explained IUs and recall \(R\) over currently teachable IUs (\(DC = \frac{2PR}{P + R}\)), penalizing both redundant repetition of known facts and premature introduction of prerequisite-blocked material; - Cognitive Overload (CO): The session-average cognitive load relative to capacity: \(CO = \frac{1}{T} \sum_{t=1}^T \min(1, \frac{L_t}{\tau})\).

Benchmark & Evaluation Protocol

To validate KNOWSIM against human ground truth, the authors collected KNOWCHAT, an extensive benchmark comprising 705 human–AI conversational sessions across two contrasting domains: competition mathematics from the MATH benchmark (Level 5, spanning 5 distinct subjects) and multi-domain expert consultation from ExpertQA (covering 150 diverse professional domains including oncology, psychology, and arts). The experimental matrix spans four study arms crossing task domain with comparison target: strategy arms holding the base model fixed while contrasting Adaptive, Comprehensive, and Socratic delivery strategies, and model arms holding strategy fixed while comparing GPT-5.4, Claude Opus 4.7, and Gemini 3 Pro. Human participants recruited via Prolific were rigorously stratified into Novice, Intermediate, and Advanced tiers via 10-item pre-screening tests. In MATH sessions, objective knowledge gain was measured via identical pre- and post-conversation tests (\(KG = \frac{\text{Post} - \text{Pre}}{\text{Max} - \text{Pre}}\)), while all sessions collected post-interaction 10-point Likert ratings for Delivery Calibration, Cognitive Overload, and Interaction Quality (IQ).

Key Experimental Results

Main Results: Sign Agreement Alignment with Human Judgments

On high-signal comparative conditions (\(|\delta_{\text{human}}| \ge 0.15\)), KNOWSIM achieves statistically significant directional alignment with human judgments across both tasks and arms, substantially outperforming existing simulator paradigms.

Evaluation Setting / Study Arm Sign Agreement Count (Agree / High-Signal) Sign Agreement Rate One-sided Binomial Significance (\(p\)-value)
MathQA–Strategy 13 / 17 76% \(p = .025^*\)
MathQA–Model 14 / 20 70% \(p = .058^\dagger\)
ExpertQA–Strategy 16 / 20 80% \(p = .006^{**}\)
ExpertQA–Model 10 / 15 67% \(p = .151\)
MathQA (Pooled) 27 / 37 73% \(\mathbf{p = .003^{**}}\)
ExpertQA (Pooled) 26 / 35 74% \(\mathbf{p = .003^{**}}\)

When evaluated against baseline simulators on Interaction Quality (IQ)—the sole metric baseline methods can compute—KNOWSIM demonstrates superior rank concordance pooled across all four arms: - KNOWSIM (Ours): 77% (\(p = .008^{**}\)) - ZS-CoT-Prof (Synthetic profile + initial state): 73% (\(p = .026^*\)) - ZS (Zero-shot conversation history): 68% (\(p = .067^\dagger\)) - ZS-CoT (Turn-level zero-shot chain of thought): 64% (\(p = .143\))

Benchmark Across 9 LLMs: Uncovering Aptitude–Treatment Interactions (ATI)

Benchmarking 9 frontier, mid-tier, and open-weight models using KNOWSIM reveals pronounced rank inversions across user knowledge tiers (\(\bar{R}\) denotes intra-level mean rank; \(CO\downarrow\) indicates cognitive overload, where lower is better).

Evaluated Model Novice KG Novice \(CO\downarrow\) Novice DC Novice IQ Novice \(\bar{R}\) Inter. KG Inter. \(CO\downarrow\) Inter. DC Inter. IQ Inter. \(\bar{R}\) Adv. KG Adv. \(CO\downarrow\) Adv. DC Adv. IQ Adv. \(\bar{R}\)
Gemini 3.1 Pro 13.2 0.875 0.086 8.39 3.25 5.4 0.750 0.088 8.35 4.75 0.7 0.873 0.122 7.96 1.75
Claude Opus 4.7 14.3 0.882 0.090 7.80 3.25 5.6 0.807 0.128 8.39 2.50 0.6 0.879 0.107 8.30 3.00
DeepSeek V4 15.5 0.945 0.092 6.50 4.00 5.7 0.835 0.121 7.53 4.25 0.5 0.890 0.109 7.74 4.25
Gemini 3.1 Flash 15.1 0.936 0.093 6.83 3.25 5.6 0.838 0.116 7.73 5.75 0.5 0.912 0.109 7.64 6.00
Claude Sonnet 4.6 12.9 0.877 0.079 7.06 4.75 5.4 0.791 0.098 7.87 5.00 0.6 0.866 0.100 8.00 3.50
GPT-5.4 14.9 0.940 0.088 5.24 5.50 5.6 0.792 0.119 6.86 5.50 0.4 0.903 0.102 7.05 7.75
Qwen-3.6-35B 14.4 0.934 0.082 5.68 5.25 5.7 0.829 0.128 6.89 4.50 0.5 0.892 0.081 7.39 7.00
GPT-5.4 mini 12.6 0.946 0.075 5.41 8.00 5.4 0.809 0.121 7.13 6.00 0.5 0.901 0.114 7.24 5.75
Llama-4-Maverick 11.8 0.927 0.071 4.58 7.75 5.4 0.767 0.105 5.64 6.75 0.3 0.882 0.117 6.07 6.00

Key Findings

  • Aptitude–Treatment Interactions Shatter Aggregate Leaderboards: No single model dominates across all user proficiency levels. For novices, DeepSeek V4 yields the greatest knowledge transfer (\(KG = 15.5\)), but imposes extreme cognitive overload (\(CO = 0.945\), second worst); conversely, Gemini 3.1 Pro provides the lowest novice overload (\(0.875\)) and highest perceived quality (\(8.39\)), but ranks 6th in novice knowledge gain. For advanced users, the rankings flip: Gemini 3.1 Pro achieves top calibration (\(\bar{R} = 1.75\)) due to concise, non-redundant answers, whereas DeepSeek V4 drops to 4th place.
  • Subjective Satisfaction Decouples from Objective Learning Gain: Human data on MathQA demonstrates that participant-rated Interaction Quality (IQ) is statistically significantly negatively correlated with actual Knowledge Gain (KG) (\(\rho = -0.268, p < .01\) overall; \(\rho = -0.358, p < .05\) for intermediates). Pleasant, effortless conversations frequently signal passive agreement where the user absorbs no challenging concepts, underscoring the necessity of objective knowledge-state tracking.
  • Directional Polarization of Interaction Failure Modes: Analyzing 214,642 explained concepts reveals distinct failure patterns: novices suffer predominantly from over-reaching, with 37% of explanations blocked by missing prerequisites (PREREQ-BLOCKED) and 28% lost to cognitive overload (OVERLOAD-WASTED) (only 11% effective); advanced users suffer from over-repetition, with 89% of explained IUs being completely redundant (REDUNDANT) (only 3% effective).

Highlights & Insights

  • Operationalizing Cognitive Psychology into Graph Automata: Elegantly translates Ausubel's meaningful learning theory and Sweller's cognitive load theory into discrete graph operations and non-linear attenuation functions, bridging behavioral user simulation with cognitive science.
  • Exposing the Hidden Inequity of Aggregate Evaluation Benchmarks: Demonstrates that selecting an LLM assistant based on aggregate leaderboard scores systematically disadvantages less knowledgeable user groups, establishing a critical equity metric for production deployments.
  • Fine-Grained Five-Way Explanatory Failure Diagnostics: The decision-tree failure categorization (REDUNDANT, PREREQ-BLOCKED, EFFECTIVE, OVERLOAD-WASTED, UNDER-ABSORBED) provides an actionable, plug-and-play diagnostic framework for auditing dialogic tutoring and clinical RAG systems.

Limitations & Future Work

  • Lack of Affective and Motivational Dynamics: The framework models structural information constraints (prerequisite gating, cognitive bandwidth) but omits affective factors such as user frustration, curiosity, and boredom, which causes knowledge gain to saturate prematurely for simulated advanced users.
  • Boundary of Domain Generalization: Validation has been conducted primarily on STEM problem-solving and fact-grounded expert Q&A; extending graph-based simulation to open-ended, non-convergent tasks (e.g., creative collaborative writing or exploratory data analysis) remains an open challenge.
  • High Multi-Call Simulation Overhead: Generating each turn requires multiple LLM calls for state extraction and message generation. Future work could distill the verified discrete graph rules into smaller specialized student models or leverage KNOWSIM as a dense reward function for multi-turn RL fine-tuning.
  • vs. SimulatorArena (Dou et al., 2025): SimulatorArena relies on static synthetic personas and history prompting, causing simulated users to suffer from knowledge drift and omniscience; KNOWSIM maintains explicit, unbreachable knowledge state transitions that enforce authentic student errors.
  • vs. Dialogue State Tracking (DST / Neural Belief Trackers): Traditional DST tracks task-oriented slot-value constraints for transaction completion; KNOWSIM generalizes state tracking to conceptual understanding, prerequisites, and cognitive bandwidth across open-ended pedagogical dialogue.
  • vs. Deep Knowledge Tracing (Piech et al., 2015): Traditional knowledge tracing passively infers student mastery from binary question-answering sequences; KNOWSIM actively updates conceptual states from unstructured conversational discourse to drive realistic downstream user dialogue.

Rating

  • Novelty: ⭐⭐⭐⭐⭐ Groundbreaking integration of cognitive load theory and prerequisite DAGs into LLM multi-turn user simulation.
  • Experimental Thoroughness: ⭐⭐⭐⭐⭐ Exceptionally rigorous validation spanning 705 human multi-turn sessions, pre/post objective learning tests, and a 9-model frontier benchmark.
  • Writing Quality: ⭐⭐⭐⭐⭐ Outstanding clarity, precise formalization, and deeply insightful pedagogical analyses.
  • Value: ⭐⭐⭐⭐⭐ Exposes critical blind spots in LLM evaluation, establishing a foundational paradigm for personalized, adaptive conversational systems.