Research map / Multi-agent systems and games / Human-AI and goal-conditioned
Does Your AI Agent Get You? A Personalizable Framework for Approximating Human Models from Argumentation-based Dialogue Traces
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes Persona, a framework that updates a probabilistic human model over argumentation dialogue via a Bayesian update combined with a prospect-theory probability weighting function, personalized per user and validated on a 184-participant human study.
Problem
AI agents in argumentation-based XAI typically use static, predetermined human models, failing to capture evolving and uncertain human beliefs during dialogue.
Approach
Represent the human model as a probability distribution over propositional models, updated by Bayesian conditioning on exchanged arguments whose probabilities are transformed from human-reported confidence via a prospect-theory weighting function with personalizable parameters (s, r) learned from user model rankings.
Key results
- Persona achieves significantly higher Spearman correlation than HM1 and HM2 in model approximation across rounds 2-5 (e.g., Round 4 p=3.76e-5 vs HM1), and vs HM2 in rounds 2, 3, 5.
- In argument belief estimation, Persona significantly outperforms HA (p=2.1e-11) but is not statistically different from the SBU ablation (p=0.569).
- Persona takes about 0.6 seconds per participant for hyperparameter pair computation, versus 9 s for HM1 and 41 s for HM2.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| learn participant-specific prospect-theory parameters instead of using shared parameters (ablation) | Persona · Human model approximation, Rounds 2–4 | Spearman rank correlation distribution in [0.75, 1]: Round 2: 0.353 -> 0.375; Round 3: 0.359 -> 0.391; Round 4: 0.337 -> 0.408 | Helped |
| add the probability weighting function to Bayesian updating (ablation) | Persona · Human model approximation, Rounds 2–5 | Spearman rank correlation distribution in [0.75, 1]: Round 2: 0.326 -> 0.375; Round 3: 0.353 -> 0.391; Round 4: 0.380 -> 0.408; Round 5: 0.451 -> 0.467 | Helped |
| use three prior rounds rather than one or two to learn personalized parameters | Persona · Human model approximation, Round 4 | Spearman rank correlation: p = 4.43 × 10^{-4}; p = 0.015 | Helped |
| use two prior rounds rather than one to learn personalized parameters | Persona · Human model approximation, Round 4 | Spearman rank correlation: p = 0.925 | No clear effect |
| add participant-specific parameter learning and probability weighting | Persona · Argument belief estimation | Spearman rank correlation: p = 0.569 | Mixed |
| use participant-specific rather than shared parameters (ablation) | Persona · Argument belief estimation | Spearman rank correlation: p = 0.002 | Helped |
| use Persona's personalized weighting and belief updates rather than HA | Persona · Argument belief estimation | Spearman rank correlation: p = 2.105 × 10^{-11} | Helped |
| use Persona rather than HM1 | Persona · Human model approximation, Rounds 2–5 | Spearman rank correlation: Round 2: p = 2.408 × 10^{-6}; Round 3: p = 3.254 × 10^{-4}; Round 4: p = 3.760 × 10^{-5}; Round 5: p = 0.006 | Helped |
| use Persona rather than HM2 | Persona · Human model approximation, Round 4 | Spearman rank correlation: p = 0.006 | No clear effect |
Surprising findings
- The simple Bayesian update (SBU) ablation performs statistically on par with full Persona for argument belief estimation, with SBU winning in non-high correlation ranges.
- Hunter's HM2 performs close to Persona in round 4 due to randomness in its ranking procedure.
Limitations
- Natural language to logic translation is not addressed; additional time would be needed for it.
- Evaluation is limited to a single fictional venue-selection scenario with propositional logic.
- Differences from the SBU ablation for argument belief estimation are not significant, and Persona underperforms in medium correlation ranges.
Experimental setup
Datasets: human-subject study dialogue traces (Luminara Gardens scenario)
Benchmarks: human-subject study dialogue traces (184 participants)
Metrics: Spearman's rank correlation paired Student's t-test p-values runtime
Baselines: Generic SBU HM1 HM2 HA
Builds on: prospect theory structured deductive argumentation Hunter's probabilistic human model updating
Closest papers by meaning
- Enhancing Personalized Multi-Turn Dialogue with Curiosity Reward (NeurIPS 2025 · similarity 0.86)
- Theory of Mind and Persuasion Beyond Conversation: Assessing the Capacity of LLMs to Induce Belief States via Planning and Action (NeurIPS 2026 · similarity 0.85)
- Consistently Simulating Human Personas with Multi-Turn Reinforcement Learning (NeurIPS 2025 · similarity 0.84)
- Towards Trustworthy Autonomous Systems via Conversations and Explanations (AAAI 2024 · similarity 0.84)
- Utilizing Human Behavior Modeling to Manipulate Explanations in AI-Assisted Decision Making: The Good, the Bad, and the Scary (NeurIPS 2024 · similarity 0.84)
- Talk, Evaluate, Diagnose: User-aware Agent Evaluation with Automated Error Analysis (ICLR 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.