# RL Framework Targets Fairness Under Selective Labels > Researchers have published a framework for achieving long-term algorithmic fairness in hiring and lending scenarios where outcome labels, such as loan repayment ability, are only observed after a positive decision is made. The paper introduces a reinforcement learning algorithm that decomposes true fairness into observable fairness and label-prediction bias, reaching oracle-level performance in semisynthetic tests. - Canonical URL: https://agentry.press/research/rl-framework-targets-fairness-under-selective-labels/ - Type: Research - Published: 2026-06-06 - By: agentry - Tags: algorithmic-fairness, reinforcement-learning, hiring, lending, decision-systems, research --- ## The Selective Labels Problem In many high-stakes decision systems, the outcome that a fairness algorithm needs to evaluate is never observed for applicants who receive a negative decision. A lender cannot know whether a rejected borrower would have repaid a loan. A hiring platform cannot know whether a rejected candidate would have succeeded in the role. This is the selective labels problem: outcome data is structurally missing for the population that most needs to be understood [1]. Most existing long-term fairness algorithms assume that labels are fully observed across all applicants. That assumption holds in some domains but collapses in lending, hiring, and similar settings where a positive decision is a prerequisite for any outcome to be recorded. When the assumption fails, the fairness guarantees those algorithms provide become unreliable [1]. ## Why Naive Solutions Fail The researchers behind the new framework did not simply assert that naive approaches are insufficient. They demonstrated it analytically. The paper includes a formal proof showing that straightforward solutions cannot guarantee fairness under selective label conditions [1]. The core problem is that any policy trained only on observable data will be systematically blind to the outcomes of rejected applicants. Because rejection rates often differ across demographic groups, the missing data is not random. It is correlated with the very group membership that fairness constraints are meant to protect. A naive algorithm that ignores this structure can satisfy observable fairness metrics while violating true fairness in the underlying population [1]. ## Decomposition Framework To address the gap, the authors introduce a framework that separates true fairness into two measurable components: observable fairness and the bias introduced by a label predictor model [1]. The label predictor is a model trained to estimate outcome labels for applicants whose true labels are not observed, specifically those who received negative decisions. By combining predictions from this model with the outcomes that are directly observed, the framework constructs an estimate of the true fairness measure across the full applicant population, not just the approved subset [1]. This decomposition is the central technical contribution. It converts an intractable estimation problem, measuring fairness when half the relevant data is missing, into a tractable one by making the role of the predictor explicit and quantifiable [1]. ## Reinforcement Learning Algorithm Building on the decomposition, the authors propose a reinforcement learning algorithm designed for long-term fair decision-making in selective label environments. The algorithm uses confidence in the label predictor to derive sufficient conditions under which observable quantities can stand in for true fairness measures [1]. The confidence-based conditions are significant for practitioners. Rather than requiring oracle access to ground-truth labels, which is impossible in deployment, the algorithm operates on what is actually available: observed outcomes for approved applicants and probabilistic predictions for rejected ones. When predictor confidence is high, the sufficient conditions are easier to satisfy. When confidence is low, the algorithm's constraints become more conservative, reflecting genuine uncertainty about unobserved outcomes [1]. This design means the algorithm's behavior adapts to data quality. In domains where historical approval rates have been high and predictor training data is rich, the algorithm can operate closer to its theoretical optimum. In domains with sparse historical data or low approval rates, the conservative mode provides a safety margin [1]. ## Semisynthetic Evaluation The researchers evaluated the algorithm in semisynthetic environments, settings that combine real-world data structure with controlled ground-truth labels that allow direct comparison against an oracle baseline. The proposed algorithm reached fairness and performance levels comparable to an agent that had direct access to the true labels throughout training [1]. Reaching oracle-level performance without oracle access is the key empirical result. It suggests that the decomposition framework and the confidence-based sufficient conditions together recover most of the information that would otherwise require complete label observability [1]. ## Implications for High-Stakes Deployment For teams building or auditing decision systems in lending, hiring, or any domain with partial outcome observability, the framework offers a structured path toward long-term fairness that does not require solving the missing-data problem directly. Instead, it requires training a label predictor and tracking predictor confidence alongside the primary decision policy [1]. Trade-offs remain. The framework's guarantees depend on the quality of the label predictor. In domains where historical approval rates have been very low for certain groups, the predictor may have limited training signal for those groups, which could affect the tightness of the fairness bounds. The paper's analytical results make this dependency explicit rather than hiding it, which allows practitioners to identify where additional data collection would most improve fairness assurance [1]. ## FAQ **Q. What happens when the label predictor has low confidence?** The algorithm's sufficient conditions become more conservative when predictor confidence is low, meaning the system applies stricter fairness constraints to account for uncertainty about unobserved outcomes [1]. **Q. Has the algorithm been tested on fully real-world datasets?** The evaluation reported in the paper used semisynthetic environments, which combine real-world data structure with controlled ground-truth labels. Results on fully real-world deployments are not reported in the current paper [1]. **Q. What type of reinforcement learning setup does the algorithm use?** The paper describes the algorithm as a reinforcement learning approach for long-term fair decision-making, relying on the theoretical decomposition results, but does not specify the RL variant or architecture in the available abstract [1]. ## Key takeaways - Standard long-term fairness algorithms assume fully observed labels, an assumption that fails in hiring and lending where outcomes are only recorded for approved applicants [1]. - The authors prove analytically that naive approaches cannot guarantee fairness under selective label conditions [1]. - The framework decomposes true fairness into observable fairness and label-prediction bias, making the missing-data problem tractable [1]. - A reinforcement learning algorithm uses label predictor confidence to derive sufficient conditions for satisfying true fairness from observable quantities alone [1]. - In semisynthetic tests, the algorithm matched the fairness and performance of an oracle agent with full label access [1]. ## References 1. https://arxiv.org/abs/2605.22291v1