Research map / Large language models / AI and society
Confounding-Robust Deferral Policy Learning
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes ConfHAI, a confounding-robust deferral policy learning method that combines the marginal sensitivity model with learning-to-defer for offline human-AI collaboration, plus a personalized variant for heterogeneous human experts.
Problem
Offline policy learning for human-AI collaboration when human decision-makers use unobserved confounders, so observational data yields biased treatment-effect and policy estimates.
Approach
Combines learning-to-defer routing with the marginal sensitivity model: a worst-case (pessimistic) self-normalized IPW risk over an uncertainty set of true propensity weights is optimized jointly over the policy and router, with a line-search solver for the inner linear-fractional maximization; a personalized variant routes to specific human experts with expert-specific confounding bounds.
Key results
- On synthetic data, ConfHAI and ConfHAIPerson achieve negative policy regret (outperforming both human-only and algorithm-only teams) for specified log(Gamma) near the true confounding level.
- On HELOC and IST, ConfHAI and ConfHAIPerson achieve the best performance when Gamma is correctly specified, with the personalized variant performing best overall.
- Theory: a finite-sample bound shows the worst-case empirical regret upper-bounds population regret, and Theorem 3 gives a closed-form routing rule to humans when they can exploit the unobserved confounder.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use ConfHAI or ConfHAIPerson instead of the baseline methods | Confounding-robust human-AI collaboration · Synthetic data, homogeneous humans | policy regret | Helped |
| Route instances to specific human experts using personalization | ConfHAIPerson · Synthetic data, heterogeneous humans | policy regret | Helped |
| Use ConfHAI or ConfHAIPerson instead of the baselines | Confounding-robust human-AI collaboration · HELOC | policy regret | Helped |
| Use ConfHAIPerson instead of the other methods | Confounding-robust human-AI collaboration · International Stroke Trial (IST) | policy regret | Helped |
| Use ConfHAI or ConfHAIPerson instead of the baselines | Confounding-robust human-AI collaboration · International Stroke Trial (IST) | policy regret | Helped |
| Use ConfHAI or ConfHAIPerson instead of the other methods | Confounding-robust human-AI collaboration · FOCUS real human responses | policy regret | Helped |
| Increase the specified log(Γ) toward the underlying confounding level (ablation) | ConfHAI and ConfHAIPerson · Synthetic data, homogeneous humans | policy regret | Helped |
| Use personalized routing when human experts have heterogeneous confounding levels (ablation) | ConfHAIPerson · Synthetic data, heterogeneous humans | policy regret | Helped |
| Under-specify or over-specify Γ rather than correctly specify it (ablation) | ConfHAI and ConfHAIPerson · HELOC and IST | policy regret | Mixed |
| Increase human cost (ablation) | Human-only policy · Synthetic data | policy regret | Hurt |
| Increase human cost (ablation) | HAI, ConfHAI, and ConfHAIPerson · Synthetic data | policy regret | Hurt |
| Increase human cost (ablation) | ConfHAI and ConfHAIPerson · Synthetic data | policy regret | No clear effect |
Surprising findings
- The confounding-robust algorithm-only policy (ConfAO) is almost never better than human performance in the synthetic setting, while human-AI complementarity can beat humans for a range of Gamma.
Limitations
- Relies on strict constraints of the marginal sensitivity model, which may be conservative.
- Gamma must be specified via domain knowledge or estimated; over-specification reduces performance gains.
- Real-world experiments use simulated human behavior policies and simulated outcomes (e.g., IST potential outcomes via random forests).
Experimental setup
Datasets: synthetic data following Kallus and Zhou 2021 HELOC International Stroke Trial (IST) FOCUS scientific annotation dataset
Benchmarks: synthetic deferral setup HELOC IST FOCUS
Metrics: policy regret robust regret bound
Baselines: Human Only Algorithm Only (IPW) Confounding-Robust Algorithm Only (ConfAO) HAI (Gao et al. 2021b)
Builds on: Gao et al. 2021b deferral collaboration marginal sensitivity model (Tan 2006) Kallus and Zhou 2021 confounding-robust policy learning self-normalized Hájek estimator
Closest papers by meaning
- When to Act and When to Ask: Policy Learning With Deferral Under Hidden Confounding (NeurIPS 2024 · similarity 0.89)
- Human-AI Collaborative Uncertainty Quantification (ICML 2026 · similarity 0.87)
- Confounding-Robust Policy Improvement (NeurIPS 2018 · similarity 0.87)
- Beyond Single Stationary Policies: Meta-Task Players as Naturally Superior Collaborators (NeurIPS 2024 · similarity 0.85)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.84)
- A No Free Lunch Theorem for Human-AI Collaboration (AAAI 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.