Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty
summary
The gist
Real-world decision-making often involves uncertainty expressed in linguistic rather than numerical terms, and Prospect Theory (PT) provides a classic framework for modeling human behavior under such
In short
This study tested if Large Language Models (LLMs) follow Prospect Theory, a model for human decision-making under uncertainty. The research found that LLM decision parameters are unstable when exposed to linguistic uncertainty, meaning they change significantly based on how uncertainty is expressed in words rather than numbers. This suggests PT frameworks are unreliable for real-world applications involving ambiguity.
Key concepts
- Prospect Theory (PT)
- A behavioral economics framework used to model how people make choices when facing risks and uncertainties. It describes how individuals value gains and losses differently, particularly emphasizing loss aversion, which means the pain of a loss is greater than the pleasure of an equivalent gain.
- Epistemic Uncertainty
- Uncertainty that arises from a lack of knowledge or information about the true state of affairs. In this study, it refers to situations where uncertainty is expressed through vague language or 'epistemic markers' instead of precise numerical probabilities, challenging how models interpret this ambiguity.
- Parameter Instability
- The finding that the specific parameters within Prospect Theory (like risk preference or loss aversion) change significantly when the input uncertainty is switched from numerical values to linguistic markers. This demonstrates that LLMs do not maintain consistent decision-making rules when faced with ambiguous language.
- Scale Dependency
- The observation that Prospect Theory-like behavior in LLMs only emerges reliably once the model reaches a sufficiently large parameter scale. Smaller models do not exhibit these behaviors consistently, indicating that complexity is necessary for PT adherence.
Terminology used across episodes
This episode discusses
- Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty · Paper Radio
- Perceptions of Linguistic Uncertainty by Language Models and Humans
- Mathematical Proof as a Litmus Test: Revealing Failure Modes of Advanced Large Reasoning Models
- DeFine: Decision-Making with Analogical Reasoning over Factor Profiles
- Decision-Making Behavior Evaluation Framework for LLMs under Uncertain Context
- Are LLM-Judges Robust to Expressions of Uncertainty? Investigating the effect of Epistemic Markers on LLM-based Evaluation
- Learning to Ask Like a Physician
- A Conceptual Framework for AI-based Decision Systems in Critical Infrastructures
- Evaluating and Aligning Human Economic Risk Preferences in LLMs
- CostBench: Evaluating Multi-Turn Cost-Optimal Planning and Adaptation in Dynamic Environments for LLM Tool-Use Agents
- NOVA: NOise-aware Verbal Confidence CAlibration for Robust Large Language Models in RAG Systems · Paper Radio
- Revisiting Epistemic Markers in Confidence Estimation: Can Markers Accurately Reflect Large Language Models' Uncertainty?
- An analysis of AI Decision under Risk: Prospect theory emerges in Large Language Models
- Qwen2.5 Technical Report
- Risk Profiling and Modulation for LLMs
- Diversity-Enhanced Reasoning for Subjective Questions
- Calibrating Behavioral Parameters with Large Language Models
- From Automation to Autonomy: A Survey on Large Language Models in Scientific Discovery
- Relying on the Unreliable: The Impact of Language Models' Reluctance to Express Uncertainty
- Navigating the Grey Area: How Expressions of Uncertainty and Overconfidence Affect Language Models
- CritiCal: Can Critique Help LLM Uncertainty or Confidence Calibration?
The paper
Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty · Read on arXiv
Hong Kong University of Science and Technology · Huazhong University of Science and Technology · University of Illinois Urbana-Champaign
Transcript
Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.
Tom: Today's paper: "Rethinking Prospect Theory for LLMs".
Jane: Real-world decision-making often involves uncertainty expressed in linguistic rather than numerical terms, and Prospect Theory (PT) provides a classic framework for modeling human behavior under such uncertainty.
Tom: First, who's behind it and why it matters.
Title and authors: Tom: We started by looking at the title and the authors of "Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty." It tells us immediately that they aren't just applying existing theory; they are actively questioning if Prospect Theory is even a good lens for Large Language Model decision-making.
Jane: The authors are quite a team from institutions like Hong Kong University of Science and Technology, Huazhong University of Science and Technology, and the University of Illinois Urbana-Champaign, which tells us this work has a strong academic foundation.
Lu: I see the core idea is that existing research has tried to fit PT parameters to LLMs, but this paper zeroes in on whether PT actually describes the behavior when uncertainty comes in the form of vague language.
Meng: So, if you look at what they're trying to achieve, it seems their main goal is to show that these traditional PT frameworks can break down when the input is not a clean numerical probability.
Lalam: It really highlights that LLMs aren't just pattern matchers; they are making decisions based on an interpretation of language, and that interpretation isn't always stable.
The paper's summary: Tom: So, let’s talk about what the paper actually summarizes. They developed a three-stage workflow starting by estimating PT parameters from binary choices with precise probabilities to see how well the model captures behavior initially.
Jane: Then they moved into a second stage where they derived probability mappings for epistemic markers and injected those markers into prompts to test if the PT parameters stay stable under that linguistic uncertainty.
Lu: That mapping process, using their "No. Epistemic Marker Probability Mapping by Human," is crucial because it helps define how humans themselves map verbal expressions onto numerical probabilities before testing the model against that human baseline.
Meng: So they are essentially creating a controlled experiment where they test the limits of PT parameters when the input language gets messy, which makes sense for understanding deployment risks.
Lalam: It’s about seeing if those internal psychological structures of risk preference and loss aversion hold up when the uncertainty isn't a simple number but a word like "maybe."
The paper's improvements: Tom: Now, regarding the improvements they suggest for this work, they are clearly advocating for moving beyond just observing instability to actually developing ways to handle it better in real systems.
Jane: They emphasize that the current situation reveals a "representational non-invariance," meaning models alter their revealed preferences when equivalent uncertainty is expressed differently through linguistic markers.
Lu: The paper points out that while risk preference might stay relatively stable, the loss aversion and probability weighting shift more substantially under these conditions, which gives us a specific area to focus on for future work.
Meng: From a practical standpoint, this suggests that we need better ways to calibrate these models dynamically so they don't have such drastic shifts when facing natural language prompts in production.
Lalam: I think the key improvement they are pointing toward is needing a mechanism that can map linguistic markers directly into standardized numerical probabilities before applying any PT logic.
Conclusion: Tom: So, to wrap up, the core implication of "Rethinking Prospect Theory for LLMs: Revealing the Instability of Decision-Making under Epistemic Uncertainty" is that we can't just assume a model’s decision structure will remain consistent when it encounters ambiguous language.
Jane: The paper shows that for Prospect Theory to be a dependable tool, its validity has to hold up even when uncertainty is expressed through epistemic markers, which means models aren't universally adopting PT-like behavior.
Lu: I agree; the finding that large models exhibit more PT-like behavior but still have unstable loss aversion under linguistic uncertainty shows us exactly where their current limitations lie in interpreting human risk communication.
Meng: For implementation, this means we should be very careful about deploying these frameworks in safety-critical areas unless we can guarantee that the input is strictly translated into numeric values first.
Lalam: Ultimately, the paper reminds us that for safety-critical applications where epistemic ambiguity is everywhere, we have to stop treating verbal uncertainty as a direct substitute for numerical certainty and start building systems that handle it differently.
More episodes
- 2610.10613-Temporal transformer CAN encoder with federated lightweight heads for anomaly detection
- 2610.10616-When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry
- 2610.10655-Nullify: Null-Space Activation Steering for Training-Free LLM Unlearning
- 2610.11031-Language Modeling is Monotone Compression
- 2610.01253-Context-Aware Error Mitigation Orchestration for Hybrid Quantum Reinforcement Learning on NISQ Systems
- 2604.24201-CMGL: Confidence-guided Multi-omics Graph Learning for Cancer Subtype Classification
- 2609.34069-Towards Certificate-Driven Software Porting: A Self-Improving Agentic Harness for Scientific Program Optimization
- 2312.01221-Enabling Quantum Natural Language Processing for Hindi Language
- 2508.08833-An Investigation of Robustness of LLMs in Mathematical Reasoning: Benchmarking with Mathematically-Equivalent Transformation of Advanced Mathematical Problems
- 2405.04118-Policy Learning with a Language Bottleneck