README.md

metadata

license: mit
task_categories:
  - question-answering
language:
  - en
tags:
  - adversarial-robustness
  - certified-radius
  - span-extraction
  - question-answering
  - retrieval-augmented-generation
size_categories:
  - 10K<n<100K

Span Extraction Adversarial Geometry

Reproducibility data for "Why Additive Span Extraction Heads Cannot Be Improved by Coupling: Exact Adversarial Radii, Sign Attacks, and Encoder-Level Certified Training."

Key Results

Method	EM	ε* median	Δε*
Additive baseline	59.8%	2.17	—
Elsayed (global margin)	65.6%	2.96	+36%
CBCT (RoBERTa)	66.0%	3.31	+52%
CBCT (BERT-large)	62.0%	4.71	+113%
CBCT (DistilBERT)	57.8%	2.96	+45%
Biaffine spectral	65.4%	0.93	-57%

Statistical tests: CBCT vs baseline Wilcoxon p = 10⁻⁴², CBCT vs Elsayed p = 10⁻⁶.

Contents

• data/logits/: 14,863 examples × 10 extraction heads × 6 configs
• data/hidden_states/: Hidden state vectors for SQuAD examples
• data/results/: All experiment outputs (baseline comparison, multi-model CBCT, paired tests, theory verification, hyperparameter sensitivity, bucket analysis, cross-dataset)
• verify_all.py: Checks every number in the paper against data files

Full code release on GitHub after acceptance.

Verification

python verify_all.py

Citation

@article{ahamad2026span,
  title={Why Additive Span Extraction Heads Cannot Be Improved by Coupling:
         Exact Adversarial Radii, Sign Attacks, and Encoder-Level Certified Training},
  author={Ahamad, Arif Mohamed Khan Rabi and Douangnouanexay, Laddaphone},
  year={2026}
}