HugAgent: Evaluating LLMs in Simulating Human-Like Individual Reasoning on Open-Ended Tasks

AI summary

Overview

Research area: Human simulation with large language models — specifically individual-level (intra-agent) reasoning fidelity, benchmark design, and belief dynamics under counterfactual evidence.

Technical level: Intermediate. The benchmark tasks and metrics are described in accessible terms, but readers will benefit from familiarity with LLM evaluation protocols (accuracy, MAE, directional accuracy), retrieval-augmented generation, and basic Bayesian/causal-model vocabulary.

One-sentence scope: The paper introduces HugAgent, a human-grounded benchmark built from questionnaires and think-aloud interviews with 54 retained participants across healthcare, surveillance, and zoning, to test whether LLMs can predict a specific person's belief state and belief updates rather than reproducing population-level consensus.

What This Paper Is About

LLMs are increasingly used to simulate people — role-playing individuals, building digital twins, generating "silicon" samples — but because they are pretrained on population-level corpora, they tend to collapse into an average voice that erases individual reasoning styles and belief trajectories. The paper asks whether models can move from simulating the average to simulating the individual: given partial evidence of a person's prior views, can a model predict that specific person's behavioral responses and reasoning dynamics in out-of-distribution scenarios? HugAgent is the authors' benchmark for making this "average-to-individual reasoning adaptation" a measurable task.

Key Contributions

1. A formalization of average-to-individual reasoning adaptation. The authors define an individual's belief state as a distribution over d factors conditioned on context, and belief evolution as an update operator applied under an intervention, yielding tasks of Belief State Inference and Belief Dynamics Update. They anchor this in Bayesian conditioning / Probabilistic Language of Thought and Structural Causal Models as normative baselines rather than assumptions.

2. A human-grounded benchmark with two tasks. HugAgent provides 356 Belief State Inference items and 1,386 Belief Dynamics Update items (1,742 total) from 54 retained participants, spanning healthcare, surveillance, and zoning, plus a human test–retest reliability ceiling.

3. Baselines and a diagnosed asymmetry. Experiments with GPT, Gemini, LLaMA, Qwen, Claude, DeepSeek, and memory-augmented baselines (RAG, RAG with Full Context, Generative Agents reproduced with Qwen2.5-32B-instruct) show models recover a person's belief state reasonably well but struggle to predict belief updates under intervention.

4. Open-source infrastructure. The complete data collection pipeline and a companion semi-structured interview chatbot (TraceYourThinking) are released, nominally under the MIT license for code and CC BY-NC 4.0 for de-identified participant data.

Main Findings

• Models approach but do not match human belief-state inference. Human Belief State Inference accuracy is 84.84% (SD = 8.90, 95% CI: [80.00, 89.68]). Best-performing LLMs trail by 7–9 points. Qwen-max reaches 77.40% ±0.27, RAG-FC 77.56% ±0.38, Qwen2.5-7B-instr. 77.18% ±0.73, and GPT-4o 74.66% ±0.24.

• Belief-update prediction lags further behind humans. Human Belief Dynamics Update accuracy is 85.66% (SD = 7.66, 95% CI: [80.91, 90.40]) with MAE 0.68 (SD = 0.20) and directional accuracy 88.92% (SD = 9.89). GPT-4o reaches 63.11% ±0.14 accuracy, MAE 1.29, directional accuracy 82.27%; Gemini 2.0 Flash reaches 60.55% ±0.08 accuracy, MAE 1.35, directional accuracy 83.31%.

• Aggregate baselines trail sharply. Global Majority scores 65.77% on Belief State Inference and 58.18% on Belief Dynamics Update with MAE 2.54 and 4.44 ATI; Random Guess scores 51.89% ±3.96 and 43.12% ±0.78 with 0.00 ±5.80 ATI. Human ATI is 100.00.

• Cross-domain transfer degrades performance substantially. For GPT-4o, Belief State Inference drops from 74.66% to 58.56% and Belief Dynamics Update from 63.11% to 44.64% when context comes from a different domain. The degradation is most pronounced in belief updating.

• For Qwen2.5-32B-instr., transferring Surveillance context to Zoning drops BDU accuracy from 62.25% in-domain to 24.84%, with MAE rising from 1.33 to 1.94.

• More context helps belief-state inference but not belief updating. Belief State Inference improves by +5.2 points at full context (intervals wide; no individual comparison significant after correcting for five comparisons), while every Belief Dynamics Update estimate stays within 1.0 point of baseline with intervals contained within ±2.7 points. Accuracy at the 5-QA baseline is 71.1% for BSI and 66.2% for BDU. For 43% of the 162 participant–domain cells, accuracy is identical across all context lengths; where context length matters, shorter context usually performs best, and full context is uniquely best for only 2 of 54 participants in each domain.

• Individual context beats population priors. For GPT-4o, moving from No-Context (demographics only) to Full-Context raises Belief State Inference accuracy from 58.49% to 74.66% and Belief Dynamics Update accuracy from 39.83% to 63.11%.

• Cross-person context destroys the gain. Using one participant's QA context to predict another participant's responses drops GPT-4o belief-dynamics-update accuracy to 39.30% with MAE 1.93, indicating the full-context gains come from identity-specific patterns rather than richer detail alone.

• Errors are dominated by change detection, not direction inference. In Healthcare, GPT-4o attains 49.36% change-detection accuracy versus 88.89% direction-inference accuracy (overall Dir. Acc. 77.03%); Qwen2.5-32B-Instruct records 33.19% versus 84.29% (Dir. Acc. 68.96%). Models appear change-averse, defaulting to "no change."

• Low error magnitude does not imply correct direction. o3-mini has an MAE of 1.22 yet lower directional accuracy (71.29% ±2.87) than GPT-4o, confirming magnitude alignment alone cannot ensure correct update directionality.

• Human reliability ceiling based on limited data. Of 54 participants contributing across all sessions, 18 completed the 14-day retest and 13 were retained after a demographic consistency check. The paper notes the human track and test–retest subset are modest relative to conventional large-scale datasets.

Methodology in Plain English

The authors built a two-stage pipeline. First, a questionnaire collected demographics, baseline stances on a 1–10 scale, reason weights on a 1–5 scale, and counterfactual interventions, which supply the gold labels for belief updates and anchor free text to decision factors. Second, a semi-structured chatbot interview elicited 8–20 question–answer pairs per participant, mixing open-ended elaboration (Context QAs) with concise polarity judgments (GT QAs). Each transcript supports both benchmark tasks, and survey-provided updates were never revealed in dialogue to prevent leakage.

The chatbot maintains a causal belief network of factors to identify the most critical nodes and edges for follow-up, and a question generator derives targeted, context-specific probes from that network. Participants were recruited through Prolific with informed consent and fair compensation (compensation exceeded USD 8,350, covered in part by the City Science group at the MIT Media Lab and in part by the authors). From over 120 participants, 54 were retained after predefined quality-control filtering.

Evaluation covers Belief State Inference (accuracy) and Belief Dynamics Update (accuracy within a tolerance band of ±1 for 5-point and ±2 for 10-point scales, MAE normalized to a 5-point scale, directional accuracy, and an average-to-individual score derived via hierarchical aggregation following SuperGLUE). Leakage control included masking attribution targets, drawing interventions from external surveys, and presenting each item independently with minimal-overlap prompts. The human ceiling came from a 14-day test–retest study reported with Intraclass Correlation and quadratic-weighted kappa with 95% confidence intervals.

Why This Matters

Impact on research. HugAgent reframes human simulation evaluation from inter-agent to intra-agent fidelity, from static behavioral outcomes to reasoning trajectories, and from researcher-authored vignettes to open-ended elicited data. It also provides a diagnostic frame — cross-domain swap tests, context-length sweeps, population-prior versus full-context comparisons, cross-person controls, and directional error decomposition — that surfaces structured biases standard accuracy and MAE metrics obscure.

Real-world applications:

• Policy simulation, where practitioners currently use LLM "silicon samples" to approximate public opinion on contested issues such as healthcare, surveillance, and zoning.
• Digital twins of individuals for personalized advisory or research purposes, where identity consistency across contexts matters.
• Survey and interview methodology, through the open-sourced semi-structured interview chatbot that elicits out-loud reasoning on arbitrary topics.
• Evaluation of memory-augmented and retrieval-based agent architectures, since RAG (k=5), RAG with Full Context, and Generative Agents were all tested.

Industry relevance. Any product that personalizes to a user — recommender systems, conversational assistants, health or civic engagement tools — depends on modeling a specific person rather than an average user. The paper's finding that retrieval-augmented and memory-based architectures do not close the gap suggests that adding context or memory alone is insufficient, and that calibrated change detection may be a more productive design target.

Future Directions

• Identity-conditioned representations with cross-domain consistency constraints. The authors suggest that because full-context gains disappear under cross-person and cross-domain settings, future models may need explicit identity conditioning rather than associative matching.

• Calibrated change detection. The directional error decomposition shows models preserve prior beliefs even when cues signal change; the authors propose continuous output representations or confidence-aware designs to better calibrate when to update.

• Context prioritization rather than context expansion. Since longer context did not reliably improve belief updating and sometimes increased confidence on wrong updates, the paper calls for prioritizing high-signal evidence over larger context windows.

• Expanding the participant pool, domains, and controversy levels. The limitations section notes that individual reasoning may behave differently in lower-stakes topics, interpersonal scenarios, scientific reasoning, consumer choice, or culturally specific contexts, and asks that future benchmarks report within-person, cross-domain generalization alongside standard accuracy. Topic selection is framed as an explicit experimental variable rather than a hidden confound.

Target Audience

Researchers and practitioners working on LLM-based human simulation, digital twins, and pluralistic alignment; benchmark designers interested in individualized rather than aggregate evaluation; computational social scientists and cognitive scientists studying belief change and reasoning traces; policy researchers who use synthetic samples; and applied ML engineers building personalization, memory, or retrieval systems who need to know where those architectures break down on belief updating.

Authors’ abstract

Simulating human reasoning in open-ended tasks has been a long-standing aspiration in AI and cognitive science. While large language models now approximate human responses at scale, they remain tuned to population-level consensus, often erasing the individuality of reasoning styles and belief trajectories. To advance the vision of more human-like reasoning in machines, we introduce HugAgent (Human-Grounded Agent Benchmark), a benchmark for average-to-individual reasoning adaptation. The task is to predict how a specific person would reason and update their beliefs in novel scenarios, given partial evidence of their past views. HugAgent adopts a dual-track design: a synthetic track for scale and systematic stress tests, and a human track for ecologically valid, "out-loud" reasoning data. This design enables scalable, reproducible evaluation of intra-agent fidelity: whether models can capture not just what people believe, but how their reasoning evolves. Experiments with state-of-the-art LLMs reveal persistent adaptation gaps, positioning HugAgent as the first extensible benchmark for aligning machine reasoning with the individuality of human thought. Our benchmark and chatbot are open-sourced as HugAgent (https://anonymous.4open.science/r/HugAgent) and TraceYourThinking (https://anonymous.4open.science/r/trace-your-thinking).