Statistical uncertainty from random-diagonal seeds

Report confidence intervals for perplexity degradation by evaluating TurboAngle across multiple random ±1 diagonal seeds D and ascertain the statistical significance of ΔPPL differences on the order of 0.001 to ensure robust conclusions.

Background

TurboAngle relies on a fixed random diagonal rotation shared across layers, heads, and tokens. The paper evaluates with a single seed and notes small ΔPPL differences that may be within statistical noise.

Computing confidence intervals over multiple seeds would clarify the stability of results and the reliability of very small reported differences, which is important for fair comparisons and ablation studies.

References

Confidence intervals over multiple seeds for the random diagonal $D$ are not reported; $\Delta\mathrm{PPL}$ differences below approximately $0.001$ should be interpreted with appropriate caution.

— TurboAngle: Near-Lossless KV Cache Compression via Uniform Angle Quantization  (2603.27467 - Patel, 29 Mar 2026) in Conclusion — Limitations

Single training run per model; multi-seed replication was not completed and remains future work.

— Per-Aetiology Contrastive Severity Embeddings with Phonological Pseudo-Labelling for Multilingual Dysarthric Speech  (2609.21789 - Muller et al., 18 Sep 2026) in Section 5, Discussion, Limitations (i)