Noise, Adaptation, and Strategy: Assessing LLM Fidelity in Decision-Making

Yuanjun Feng, Vivek Choudhary, Yash Raj ShresthaEMNLP 2025

TL;DR

This paper introduces a novel framework to evaluate how well Large Language Models (LLMs) mimic human decision-making in dynamic economic tasks, moving beyond simple outcome optimization. They found that while LLMs default to rational, consistent strategies, targeted interventions like risk framing and in-context learning with human data can nudge them towards more human-like variability, though full fidelity remains elusive.

What Problem Does This Paper Address?

As LLMs are increasingly used to simulate human behavior in social science research, a critical gap exists in evaluating their process-level fidelity. Current evaluations often focus on whether LLMs achieve optimal outcomes (e.g., maximizing profit) rather than how they arrive at those decisions. Human decision-making, however, is characterized by noise, history-dependence, and bounded rationality – factors that are crucial for realistic simulations but often overlooked. This paper aims to bridge this gap by assessing whether LLMs can replicate the variability and adaptive strategies that define human behavior in dynamic decision-making contexts.

Methods

The authors propose a "process-oriented evaluation framework" with three progressive interventions:

  • Intrinsicality: LLMs operate without any explicit guidance, revealing their default decision-making patterns.
  • Instruction: LLMs receive targeted prompts, such as instructions related to risk preference (risk-seeking or risk-averse), to see if they can be modulated.
  • Imitation: LLMs are provided with historical human decision data (in-context learning) and tasked with continuing the observed patterns.

This framework was applied to two classic behavioral economics tasks: a second-price auction and a newsvendor problem. The LLMs used were GPT-40, Claude 3.5 Sonnet, and Claude 3.7 Sonnet, each instantiated with 40 agents assigned unique demographic profiles to mirror human participants.

Datasets

The study did not introduce new datasets but rather used existing human data from prior behavioral economics experiments for comparison. The tasks themselves (second-price auction and newsvendor problem) serve as the experimental environments.

Key Insights and Ideas

  • LLMs Default to Rationality: Without intervention, LLMs consistently exhibit stable, conservative, and profit-maximizing strategies that diverge significantly from the noisy, variable patterns observed in human decision-making.
  • Interventions Matter, But Don't Fully Solve It:
    • Instruction: Risk-framed instructions can push LLMs towards more human-like behavior (e.g., higher reserve prices when risk-seeking), but this often introduces irregularities and doesn't fully capture human variability.
  • Imitation: Providing human data through in-context learning significantly improves LLM alignment, leading to more human-like distributions of decisions and increased behavioral variability. This suggests LLMs can learn underlying behavioral patterns from data.
  • Process Over Outcome: The study strongly advocates for evaluating LLMs based on their decision-making process (variability, adaptation) rather than just the final outcome, as this reveals fundamental differences in how LLMs and humans operate.
  • Human Data is Key for Fidelity: The imitation condition highlights the power of human data in shaping LLM behavior, demonstrating that LLMs can learn and reproduce complex human strategies when exposed to them.

Key Results

  • Auction Task:
    • Intrinsicality: LLMs showed significantly lower variability in reserve prices compared to humans (Entropy of 1.43-2.96 for LLMs vs. 5.08 for humans). Their strategies were largely independent of the number of bidders, unlike humans who increased reserve prices with more bidders.
  • Instruction: Risk-seeking prompts led to drastically higher reserve prices in LLMs (e.g., GPT-40 mean of 81.63), while risk-averse prompts resulted in lower prices closer to the default. Claude models showed a positive correlation between reserve price and bidder count under instruction, a human-like trait.
  • Imitation: KS distances dropped significantly (near 0.05), indicating much closer distributional alignment with human reserve price trajectories. Variability (Entropy) increased to ~4.6-4.7.
  • Newsvendor Task:
    • Intrinsicality: LLMs produced tightly clustered orders near the optimal quantity, lacking the substantial round-to-round variability seen in humans.
  • Instruction: LLMs adjusted orders directionally based on risk framing, but still showed less variability than humans.
  • Imitation: LLMs recovered more human-like variability, though not as closely as in the auction task, attributed to the higher inherent variability in human newsvendor decisions due to fluctuating demand.

Limitations and Future Work

  • Static Interventions: The study used static text-based inputs for interventions. Future work could explore more dynamic approaches like multi-turn interactions, memory-based adaptation, or reinforcement learning fine-tuning.
  • Single-Player Focus: The experiments were primarily single-player profit-driven environments. Extending to multiplayer tasks (e.g., negotiation) would offer a richer test of behavioral fidelity.
  • Participant Pool: The human data came from a specific participant pool, potentially limiting generalizability. Replicating with more diverse populations is suggested.
  • Ethical Considerations: The authors acknowledge the risk of over-interpreting LLM outputs as reflective of human cognition and emphasize responsible reporting. They also note that injecting human data risks importing biases.

Summary updated

BibTeX
@inproceedings{feng2025noise,
  author = {Yuanjun Feng and Vivek Choudhary and Yash Raj Shrestha},
  title = {{Noise, Adaptation, and Strategy: Assessing LLM Fidelity in Decision-Making}},
  booktitle = {Conference on Empirical Methods in Natural Language Processing},
  year = {2025},
}
Figure