Conformal Prediction Eval

conformal-prediction-eval

Uncertainty-Aware Transformers: Conformal Prediction for Language Models — Vellore et al. (2026) (arXiv:2604.08885, 2026)

What this evaluates

Evaluates the ability of conformal prediction frameworks to produce statistically valid prediction sets with instance-level uncertainty quantification for encoder-only transformers, measuring both classification accuracy and calibration efficiency across standard NLP benchmarks.

Datasets

• GLUE — total ?; splits: test (-1)
• SuperGLUE — total ?; splits: test (-1)

Metrics

• Test Accuracy (primary) — range: [0, 1]
  • Fraction of test instances where the model's predicted class matches the ground truth label.
• Correct Efficiency — range: [0, 1]
  • Proportion of correctly classified instances that yield a singleton prediction set. Maximizes decisiveness by favoring smaller prediction sets for correct predictions to reduce review cost.

Input / output format

Input: Text inputs for binary or multi-class classification tasks (e.g., single sentences, sentence pairs, or question-answer passages).

Output: A predicted class label and a conformal prediction set (subset of candidate labels) with associated credibility/confidence scores.

Scoring recipe

def compute_metrics(predictions, gold_labels, prediction_sets):
    correct = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    accuracy = correct / len(gold_labels)
    
    correct_singleton = sum(1 for p, g, s in zip(predictions, gold_labels, prediction_sets) if p == g and len(s) == 1)
    correct_total = sum(1 for p, g in zip(predictions, gold_labels) if p == g)
    efficiency = correct_singleton / correct_total if correct_total > 0 else 0.0
    
    return {'test_accuracy': accuracy, 'correct_efficiency': efficiency}

Common pitfalls

• Assuming aggregate (marginal) coverage guarantees equitable performance across all classes; minority or hard classes often suffer severe undercoverage.
• Overlooking resource constraints: full pairwise distance computations on long-context datasets can exceed GPU memory, forcing approximations or limiting model sizes.
• Interpreting neighbor distances as direct causal explanations rather than plausibility/consistency diagnostics in representation space.

Evidence (verbatim from paper)

Across models, CONFIDE improves correct efficiency while preserving accuracy; however, we frequently observe substantial undercoverage on hard or minority classes (e.g., CoLA “unacceptable,” BoolQ “false”), and no approach reaches the nominal $1-\varepsilon$ target when the exchangeability condition is violated.

Citation

@misc{vellore2026uncertainty,
  title={Uncertainty-Aware Transformers: Conformal Prediction for Language Models},
  author={Vellore et al. (2026)},
  year={2026},
  note={arXiv:2604.08885}
}

• arXiv: 2604.08885