Equitable Evaluation via Elicitation

1. Problem Statement

The paper addresses the issue of endogenous bias in skill evaluation, specifically arising from differences in individuals' self-presentation styles (e.g., modesty vs. self-promotion).

• The Core Issue: Individuals with identical qualifications may present themselves differently due to cultural or personality factors (e.g., gender, collectivist vs. individualist cultures). Modest individuals often omit crucial information, while self-promoters may overstate it.
• Limitations of Existing Solutions:
  • Post-hoc Fairness: Techniques like multicalibration or learning fair representations fail because the necessary data (the "missing" information from modest users) is not present in the initial self-reports.
  • Tone Homogenization: Simply rewriting text to a standard style (e.g., "white cisgender male") cannot recover information the user never supplied.
• Goal: Develop an interactive system that accurately determines a candidate's skills while allowing them to speak in their natural voice, thereby mitigating bias by actively eliciting missing information rather than passively analyzing static text.

2. Methodology

The authors propose an interactive AI elicitation system that functions as a dynamic interview. The system consists of two main components trained jointly using a Reinforcement Learning (RL) framework.

A. System Architecture

1. Question Selector ($q_\psi$): An "Actor" that selects the next question from a fixed, human-curated bank of questions ($Q$) based on the current conversation transcript. It aims to minimize prediction uncertainty.
2. Skill Predictor ($f_\phi$): A "Critic" that maps the conversation transcript to a probability score ($\hat{y} \in [0, 1]$) indicating the likelihood of the user possessing a target skill.
3. Synthetic Human Simulator ($s()$): A fine-tuned Large Language Model (LLM) trained on real professional profiles to act as a "ground truth" simulator. It generates responses to questions, allowing the system to be trained without risking harm to real humans or requiring massive labeled datasets.

B. Training Framework (Actor-Critic)

The system uses a Rollout-based Actor-Critic approach:

• Reward Signal (Certainty Maximization): Instead of waiting for the final answer to calculate error, the Question Selector is rewarded for selecting questions that maximize immediate information gain.
  • The system simulates $M$ future steps (rollouts) using the simulator $s()$.
  • It calculates a certainty measure $Z = 2|f_\phi(x_{i+M}) - 0.5|$.
  • The policy is updated via REINFORCE to maximize this certainty, encouraging the selection of questions that resolve ambiguity quickly.
• Supervised Learning: The Skill Predictor ($f_\phi$) is trained using Mean Squared Error (MSE) against the ground-truth labels provided by the synthetic simulator.

C. Fairness Enforcement: Multi-Accuracy

To address systematic model bias, the authors enforce Multi-Accuracy with respect to a set of auxiliary functions $\mathcal{C}$ (representing personality traits, specifically the Big 5: Openness, Conscientiousness, Extroversion, Agreeableness, Neuroticism).

• Definition: A predictor is $(\mathcal{C}, \epsilon)$-multi-accurate if the covariance between the prediction error and any function $c \in \mathcal{C}$ is small ($\le \epsilon$).
• Implementation:
  • The raw predictor $f_\phi$ is corrected by adding a linear combination of the auxiliary functions: $f^*_\phi(x) = \sigma(\sum l_c c(x) + \sigma^{-1}(f_\phi(x)))$.
  • Calibration weights $\{l_c\}$ are solved via a convex optimization problem minimizing a specific loss function (Equation 2).
  • Interleaved Training: Fairness calibration is not a one-time post-processing step. It is interleaved with the RL training loop. After every few epochs, the system recalibrates the predictor to ensure the multi-accuracy constraint holds, creating a "virtuous cycle" where fairness constraints sharpen the reward signal for the actor.

3. Key Contributions

1. Active Elicitation for Fairness: This is the first work to actively elicit additional information to address feature insufficiency in algorithmic fairness, rather than relying solely on available data.
2. Synthetic Humans for Training: A novel application of LLMs in computational social science, where a fine-tuned LLM generates synthetic human profiles to bootstrap training data, avoiding privacy risks and data scarcity.
3. Dynamic Fairness Constraints: The system ensures that every epochal predictor in a dynamic training sequence satisfies fairness constraints, diverging from static post-hoc correction methods.
4. Preservation of Voice: The system allows users to respond in their natural style while the AI adapts the questioning strategy to uncover the truth, avoiding the burden of forcing users to conform to a specific communication style.

4. Results

The prototype was tested on a simulated "leadership" skill evaluation task with 20% of synthetic users possessing the skill.

• Fairness Loss Reduction: The multi-accuracy fairness loss (covariance between error and personality traits) dropped from 0.048 (before correction) to 0.002 (after correction), well below the target threshold of 0.01.
• Accuracy Improvement: Contrary to the common trade-off between fairness and accuracy, enforcing multi-accuracy increased test set accuracy from 0.937 to 1.0.
• Training Dynamics:
  • Applying fairness corrections at every epoch yielded the best results.
  • Even a single correction at the end of training significantly improved fairness, though repeated corrections provided the most robust stability.
  • The system converged within 100 epochs.

5. Significance and Future Work

• Theoretical Impact: The paper provides an "existence proof" that fairness can be improved in interactive social settings by moving beyond passive data analysis to active information gathering. It formalizes equitable elicitation as a problem where ground truth is discoverable through interaction.
• Practical Application: The framework is applicable to professional networking platforms (e.g., LinkedIn onboarding), corporate reorganizations, and other scenarios where matching skills to needs is critical.
• Future Directions:
  • Extending the model to estimate multiple skills simultaneously.
  • Applying the framework to domains like medical or financial recommendations.
  • Formalizing equitable elicitation for scenarios involving inherent aleatoric uncertainty (predicting future events) rather than just ground-truth discovery.

In summary, the paper demonstrates that by combining interactive elicitation, synthetic data generation, and interleaved multi-accuracy constraints, it is possible to build evaluation systems that are both highly accurate and robust against biases inherent in human self-presentation.