Research map / Learning theory and algorithms / Prediction and decision losses
Conformal Prediction Sets Improve Human Decision Making
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.
Pre-registered randomized controlled trial (600 participants, three classification tasks) showing that conformal prediction sets improve human accuracy over fixed-size top-k sets at matched empirical coverage, though response time effects are inconsistent.
Problem
Whether providing calibrated conformal prediction sets to human decision makers actually improves their accuracy and speed compared with no assistance or fixed-size top-k sets, a claim that had been assumed but not tested with humans.
Approach
Holds empirical coverage fixed across conditions by calibrating RAPS conformal sets to the top-k set's empirical risk, so that any accuracy gain can be attributed to adaptive set sizes and the uncertainty signal they carry, tested in a pre-registered human-subject RCT.
Key results
- On all three tasks, conformal sets beat top-k sets in human accuracy with statistical significance (e.g., ObjectNet p=0.01, d=0.5; GoEmotions p=0.02, d=0.4; Few-NERD p=8e-4, d=0.7).
- Conformal sets beat the no-assistance control on all tasks (p from 5e-5 to 3e-8, d 0.8 to 1.0), while top-k vs control was significant only on GoEmotions and Few-NERD.
- Response time effects were inconsistent: most comparisons were not significant, except conformal vs top-k and conformal vs control on Few-NERD.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| provide conformal sets instead of top-k sets | Human decision makers · ObjectNet | accuracy: p=0.01; d=0.5 | Helped |
| provide conformal sets instead of top-k sets | Human decision makers · GoEmotions | accuracy: p=0.02; d=0.4 | Helped |
| provide conformal sets instead of top-k sets | Human decision makers · Few-NERD | accuracy: p=8e-4; d=0.7 | Helped |
| provide conformal sets instead of no assistance | Human decision makers · ObjectNet, GoEmotions, and Few-NERD | accuracy: p=5e-5, 5e-9, and 3e-8 | Helped |
| provide top-k sets instead of no assistance | Human decision makers · ObjectNet | accuracy: p=0.1; d=0.3 | No clear effect |
| provide top-k sets instead of no assistance | Human decision makers · GoEmotions and Few-NERD | accuracy: p=5e-5 and 0.003 | Helped |
| provide conformal sets instead of top-k sets | Human decision makers · ObjectNet and GoEmotions | response time: p=0.2 and 0.5 | No clear effect |
| provide conformal sets instead of top-k sets | Human decision makers · Few-NERD | response time: p=4e-5; d=0.9 | Helped |
| increase average conformal set size from 2.49 to 2.77 (ablation) | RAPS on GoEmotions · GoEmotions | human accuracy: 59.1±1.2% -> 56.9±1.3% | Hurt |
| replace CLIP ViT-L/14 with the weaker CLIP ViT-B/32 (ablation) | Human-model assistance · ObjectNet | human accuracy | Hurt |
| use a stronger underlying model rather than a weaker one (ablation) | Top-k assistance · ObjectNet | human accuracy | Helped |
Surprising findings
- Top-k sets did not significantly improve accuracy over control on ObjectNet (p=0.1).
- Conformal sets did not consistently speed up decision making; parsing set information may add cognitive load.
Limitations
- Only three classification tasks with general-population participants and a small number of classes (10-20), so generalization to experts or regression is untested.
- Human performance on ObjectNet was below model top-1 accuracy, and poorly performing model classes could drag down human-model teams.
- Some analyses (response time, ablations, per-class results) were not pre-registered and lack statistical testing; the planned Broad Twitter Corpus dataset was replaced with Few-NERD.
Experimental setup
Datasets: ObjectNet GoEmotions Few-NERD
Benchmarks: ObjectNet GoEmotions Few-NERD
Metrics: accuracy response time empirical coverage average set size adoption rate Cohen's d p-value (Welch's t-test)
Baselines: no assistance (control) top-k prediction sets
Builds on: conformal prediction RAPS Welch's t-test
Closest papers by meaning
- Towards Human-AI Complementarity with Prediction Sets (NeurIPS 2024 · similarity 0.91)
- Designing Decision Support Systems using Counterfactual Prediction Sets (ICML 2024 · similarity 0.87)
- Learning to Decide with AI Assistance under Human-Alignment (NeurIPS 2026 · similarity 0.86)
- Conformal Validity Guarantees Exist for Any Data Distribution (and How to Find Them) (ICML 2024 · similarity 0.86)
- Using AI Uncertainty Quantification to Improve Human Decision-Making (ICML 2024 · similarity 0.85)
- Conformal Prediction Sets Can Cause Disparate Impact (ICLR 2025 · similarity 0.85)
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.