Research map / Probabilistic and scientific modeling / Uncertainty and Gaussian processes
Experts Don't Cheat: Learning What You Don't Know By Predicting Pairs
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.
Trains models to predict pairs of independent responses (letting them "cheat" by seeing one) so that the cheating amount yields second-order calibrated epistemic variance, distribution-free confidence intervals, and hallucination-controlled decoding.
Problem
Probabilistic predictions from generative models conflate aleatoric noise with epistemic ignorance about the true conditional, so existing epistemic uncertainty methods are overconfident or underconfident when the model underfits, and it is hard to tell how far the model is from the true distributio
Approach
Train a model on pairs of i.i.d. responses per input with cross-entropy; a calibrated pair model must cheat (condition on Y1 to predict Y2) exactly when its marginal differs from the truth. Second-order calibration is shown equivalent to first-order calibration on pairs, so the cheat-corrected variance and confidence give per-input error estimates, confidence intervals and statistical-hallucinatio
Key results
- On CIFAR-10H, Cheat SNGP achieves ECE-2 of 0.009 versus 0.048 for SNGP Cov. and 0.377 for Evidential DL, with top-1 accuracy of 94.9%.
- On the harder scrambled, extra-class CIFAR-10 variant, Cheat SNGP keeps ECE-2 at 0.011 while NN Ensemble degrades to 0.134 and Evidential DL to 0.387.
- On digits-of-pi, filtering by cheat-corrected confidence yields a lower hallucination rate than total log-probability, average token log-probability, and clustering baselines.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Train a SNGP model to predict paired annotator labels instead of using an NN ensemble | CIFAR-10H image classification · CIFAR-10H | ECE-2: 0.039 -> 0.009 | Helped |
| Train a SNGP model to predict paired annotator labels instead of using an NN ensemble | CIFAR-10H image classification · CIFAR-10H w/ Extra Classes, Scrambled | ECE-2: 0.134 -> 0.011 | Helped |
| Use paired-label training with Cheat SNGP instead of an NN ensemble | CIFAR-10H image classification · CIFAR-10H | top-1 accuracy: 94.9 -> 94.9 | No clear effect |
Surprising findings
- Cheat-corrected confidence sometimes exceeds 1 on the digits-of-pi task, due to miscalibration, with outliers around 10^4 for malformed samples.
- Naive NN and ensembles can be badly second-order miscalibrated; ensembles look similar in ECE-1 across tasks yet differ strongly in ECE-2 on the harder variant.
Limitations
- Requires paired responses (two samples from the true conditional per input) in the training or calibration data, which may not be available.
- Pair-calibrated models may overfit or be miscalibrated, producing negative or out-of-range variance estimates that need regularization.
- Confidence intervals become conservative when the epsilon floor is set too small, and the scheme is only worked out for binary Y in the distribution-free result.
Experimental setup
Datasets: CIFAR-10H CIFAR-10N CIFAR-10 synthetic digits-of-pi descriptions Frozen Lake expert demonstrations
Benchmarks: CIFAR-10H digits-of-pi language modeling Frozen Lake
Metrics: ECE-2 ECE-1 top-1 accuracy KL divergence hallucination rate
Baselines: Naive NN NN Ensemble Evidential DL SNGP Cov. Epinet Total LP Avg. Token LP Clustered
Builds on: SNGP Epinet Evidential DL Wide ResNet Transformer
Closest papers by meaning
- Rethinking Aleatoric and Epistemic Uncertainty (ICML 2025 · similarity 0.86)
- Provable Uncertainty Decomposition via Higher-Order Calibration (ICLR 2025 · similarity 0.85)
- Epistemic Uncertainty for Generated Image Detection (NeurIPS 2025 · similarity 0.85)
- Conformal Prediction Beyond the Seen: A Missing Mass Perspective for Uncertainty Quantification in Generative Models (NeurIPS 2025 · similarity 0.85)
- I Don't Know: Explicit Modeling of Uncertainty with an [IDK] Token (NeurIPS 2024 · similarity 0.85)
- Faithful Model Explanations through Energy-Constrained Conformal Counterfactuals (AAAI 2024 · 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.