qeinstein/adaptive-computation

Investigating representation-guided adaptive inference for low-resource African NLP, with a focus on predicting per-example computational difficulty from multilingual model representations. **Structure, Association, and Decision Value: Representation-Based Difficulty Estimation for Adaptive Inference in African-Language NLI** Code, cached model outputs, and analysis for a study of whether internal representation statistics can estimate the per-example value of escalating an input to a more capable model, in low-resource multilingual NLI. The answer, under the tested setup, is no — and the reasons arise before the representation analysis itself. `results/LEDGER.md` is the source of truth for every number reported in the paper, including the analyses that were superseded during the study. Four findings. First, evaluation validity: AfriXNLI is a translation of XNLI, with 1,047 of its 1,050 English examples appearing verbatim in XNLI evaluation data, and its English development split *is* XNLI validation. Contamination is split-specific, and the development–test accuracy gap fingerprints which split a checkpoint saw. Second, capability ordering: parameter count does not reliably order accuracy across African languages, with the larger checkpoint winning in seven languages and losing in eight, and no significant aggregate difference. Third, language confounding: across three multilingual representation spaces, angular dispersion is consistently the most language-determined statistic (η² ≈ 0.48–0.58) and effective rank the least (0.08–0.15), so pooled correlations inflate one and mask the other; a stronger claim, that η² predicts pooling bias generally, did **not** replicate and is not made. Fourth, target dependence and routing: effective rank predicts the probability gain from escalation but not whether escalation changes the prediction, while confidence does the reverse, and under the tested models, signals, and compute budgets no routing method beats always-expensive inference — though an oracle exceeds it by 11 points at 60% of the compute. To reproduce, create the environment: ```bash python3 -m venv .venv .venv/bin/pip install torch transformers " …