🤖 AI

Beyond Rational Illusion: Behaviorally Realistic Strategic Classification

This paper introduces the Prospect-Guided Strategic Framework (Pro-SF), a novel approach to strategic classification that moves beyond idealized rationality by integrating prospect theory to model and learn from agents' psychologically biased, behaviorally realistic strategic manipulations.

Original authors: Xinpeng Lv, Yunxin Mao, Renzhe Xu, Chunyuan Zheng, Yikai Chen, Haoxuan Li, Yang Shi, Jinxuan Yang, Zhouchen Lin, Yuanlong Chen, Yuanxing Zhang, Shaowu Yang, Wenjing Yang, Haotian Wang

Published 2026-06-09

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Xinpeng Lv, Yunxin Mao, Renzhe Xu, Chunyuan Zheng, Yikai Chen, Haoxuan Li, Yang Shi, Jinxuan Yang, Zhouchen Lin, Yuanlong Chen, Yuanxing Zhang, Shaowu Yang, Wenjing Yang, Haotian Wang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are a teacher grading exams, and you announce the passing score is 80. In the old way of thinking about how students behave, we assume every student is a perfectly logical robot. If a robot scores 79, it calculates: "I need 1 more point. The cost of studying is low. I will study and get an 81." If it scores 60, it thinks: "I need 20 points. That's too much work. I give up."

This paper argues that real humans are not robots. They are messy, emotional, and influenced by psychological tricks that make them act differently than the "perfect robot" predicts.

Here is the breakdown of the paper's ideas using simple analogies:

1. The Problem: The "Robot" Assumption is Wrong

Current AI systems that make decisions (like loan approvals or hiring) are built on the idea that people will only try to "game the system" if the math says it's worth it.

The Reality: Humans often stop trying even when the math says they should succeed, or they try way too hard when they shouldn't.
The Paper's Claim: If we build AI assuming people are perfect robots, the AI will fail in the real world. It will either be too strict (rejecting good people) or too loose (letting bad actors slip through).

2. The Three "Human Glitches"

The authors identify three specific psychological "glitches" that cause humans to deviate from the robot logic:

Glitch A: Loss Aversion (The "Pain of Effort")
- Analogy: Imagine you are offered a chance to win $100, but you have to pay $80 to play. A robot sees this as a $20 profit and plays. A human, however, feels the pain of losing $80 much more strongly than the joy of gaining $100. They might decide, "It's not worth the stress," and walk away, even though they could have won.
- Result: People give up on manipulating their features (like fixing their credit score) even when they could succeed.
Glitch B: Reference Bias (The "Rough Estimate")
- Analogy: A robot knows the passing grade is exactly 79.5. A human looks at their test and thinks, "I'm somewhere in the 70s." They don't see the exact line; they see a fuzzy zone. If they think they are "close enough" to the line, they might try a little bit. If they think they are "far away," they might not try at all, even if they are actually closer than they think.
- Result: People make decisions based on their own fuzzy, subjective feelings of where they stand, not the actual math.
Glitch C: Probability Distortion (The "Lottery Ticket" Mindset)
- Analogy: If a robot sees a 1% chance of getting a loan, it ignores it. A human, however, often thinks, "But it could happen! I'm the lucky one!" They overestimate tiny chances and underestimate sure things.
- Result: People might try to manipulate their data for a "long shot" chance that a robot would never bother with, or they might ignore a very likely rejection.

3. The Two Ways AI Fails

Because current AI assumes people are robots, it makes two specific mistakes when real humans show up:

Mistake 1: Over-Defense (The "Over-Protective Parent")
- The AI thinks, "Everyone will try to cheat to get over the line!" So, it moves the goalposts way back to make it harder to pass.
- The Reality: Because of "Loss Aversion," many people actually didn't try to cheat because the effort felt too painful.
- The Result: The AI rejects honest people who didn't even try to cheat, just because it was scared they might have.
Mistake 2: Under-Defense (The "Gullible Guard")
- The AI thinks, "People will only cheat just enough to barely pass." So, it sets a defense that stops small cheats.
- The Reality: Because of "Probability Distortion," some people think they have a huge chance of success and go "all in," cheating way more than the AI expected.
- The Result: The AI's defense is too weak, and these aggressive cheaters slip right through.

4. The Solution: Pro-SF (The "Psychologist AI")

The authors propose a new framework called Pro-SF (Prospect-Guided Strategic Framework). Instead of assuming people are robots, this AI assumes people are humans with the three "glitches" mentioned above.

How it works: The AI simulates the decision-making process using Prospect Theory (a famous psychological theory). It asks: "If a human feels the pain of effort, has a fuzzy idea of where they stand, and overestimates their luck, how will they actually behave?"
The Analogy: Instead of building a wall for a robot army, the AI builds a fence designed for a crowd of tired, optimistic, and slightly confused humans.

5. The Results

The authors tested this new "Psychologist AI" on real-world data (like credit scores and spam filters) and fake data.

Finding: When people acted like real humans (with biases), the old "Robot AI" failed badly. The new "Psychologist AI" (Pro-SF) performed much better.
Bonus: Even when people did act like perfect robots, the new AI didn't do worse; it just stayed competitive.

Summary

This paper says: Stop treating humans like math equations. If you want your AI to work in the real world, you have to account for the fact that humans are emotional, bad at math, and sometimes give up too soon or try too hard. By building AI that understands these human quirks, we can make systems that are fairer and more accurate.

Technical Summary: Beyond Rational Illusion: Behaviorally Realistic Strategic Classification

1. Problem Definition: The Limitation of Rationality in Strategic Classification

Strategic Classification (SC) addresses the interaction between decision-makers and agents who manipulate their observable features to secure favorable outcomes. Existing SC frameworks rely on the rational-agent assumption, positing that agents modify features if and only if the expected benefit strictly outweighs the cost (a best-response function).

The authors argue this assumption is a "rational illusion" that fails to capture real-world decision-making, which is systematically shaped by cognitive biases. The paper identifies a gap where standard SC models cannot account for agents who deviate from rational utility maximization due to psychological factors. This mismatch leads to two specific failure modes in deployed classifiers:

Over-defense: The classifier shifts its decision boundary to counter manipulations that agents, due to biases like loss aversion or reference bias, never actually perform. This results in false negatives (rejecting genuine positive instances).
Under-defense: The classifier fails to anticipate manipulations that agents perform more aggressively than rational models predict (e.g., due to probability distortion or overestimation of success), leading to false positives.

The paper formalizes a new problem setting: Behaviorally Realistic Strategic Classification (BR-SC), where agents' manipulations are driven by psychological biases rather than pure rational optimization.

2. Methodology: The Prospect-Guided Strategic Framework (Pro-SF)

To address BR-SC, the authors propose Pro-SF, a framework grounded in Prospect Theory (Kahneman & Tversky, 2013). Instead of a standard utility function, Pro-SF models agent behavior using a prospect-based utility function that integrates three key psychological mechanisms:

Loss Aversion: Agents perceive losses (effort/cost) more heavily than equivalent gains. The framework uses an asymmetric value function where the perceived weight of cost is amplified ( $\kappa > 1$ ), potentially causing agents to abandon beneficial manipulations they would otherwise undertake under a rational model.
Reference Bias: Agents evaluate outcomes relative to a subjective reference point ( $r$ ) rather than an absolute threshold. This is modeled via a quantization operator, acknowledging that agents often have coarse-grained self-assessments (e.g., estimating a score range rather than a precise value).
Probability Distortion: Agents distort objective probabilities, typically overweighting small probabilities of success and underweighting large ones. This is modeled using a probability weighting function ( $w(p)$ ) with an inverse-S shape, leading agents to "gamble" on unlikely opportunities or misjudge the effort required to cross a threshold.

The Framework Mechanics:

Agent Model: The optimal manipulated feature $x'$ is determined by maximizing the prospect-based utility $U_P$ , which combines the distorted probability of acceptance, the asymmetric value of gains/losses relative to a reference point, and the cost of manipulation.
Classifier Learning: The decision-maker trains a classifier to minimize expected loss against the behaviorally realistic manipulation function $b_P(x)$ , rather than the rational best-response $b_R(x)$ .
Parameter Learning: The behavioral parameters ( $\alpha, \beta, \kappa, \gamma$ ) are inferred from observed data pairs of original and manipulated features using maximum likelihood estimation.

The framework is analyzed as a Stackelberg game, where the authors prove the existence, local stability, and local uniqueness of the equilibrium under standard compactness conditions.

3. Key Contributions

Problem Formalization: The paper introduces and defines BR-SC, formally characterizing the limitations of the rational-agent assumption and identifying over-defense and under-defense as the primary failure modes resulting from behavioral mismatch.
Framework Proposal: It proposes Pro-SF, the first principled framework to integrate loss aversion, reference bias, and probability distortion into the strategic classification Stackelberg game.
Empirical Validation: Extensive experiments on synthetic and real-world datasets (Credit, Adult, Diabetes, German, Spam) demonstrate that Pro-SF achieves robust performance across diverse behavioral regimes (fully rational, non-rational, and mixed populations), outperforming rational-based baselines in non-rational settings while maintaining competitiveness in rational ones.

4. Experimental Results

Performance: In non-rational and mixed behavioral settings, Pro-SF consistently outperforms rational-based models across all datasets. For instance, on the Credit dataset, Pro-SF achieved 82.41% accuracy under non-rational manipulation compared to 69.54% for the rational-based model.
Ablation Studies: Removing any single behavioral component (loss aversion, reference bias, or probability distortion) degrades performance. Reference bias was found to be the most critical for preventing over-defense, while probability distortion and loss aversion were crucial for mitigating under-defense.
Sensitivity: The framework demonstrates robustness to variations in hyperparameters (e.g., loss aversion coefficient $\kappa$ , probability distortion $\gamma$ ), indicating it does not rely on fine-tuned settings to function effectively.

5. Significance and Claims

The paper claims that the rational-agent assumption, while mathematically convenient, often "collapses in real-world settings." By bridging machine learning and behavioral economics, Pro-SF offers a more reliable approach for deploying strategic classifiers in domains such as hiring, credit scoring, and college admissions.

The authors position their work as a necessary refinement of SC to handle the "non-rational behaviors" that characterize real-world agents. They assert that Pro-SF provides a "behaviorally grounded approach" that ensures robustness not just against strategic gaming, but against the specific psychological biases that drive such gaming. The paper concludes by noting that future work will extend this framework to richer and more diverse deployment settings, but does not currently propose specific new applications beyond the existing domains of SC.