Perplexity Ranks Activation Quantisers but Understates What They Remove
Anish Sathyanarayanan, Min Xu
原文摘要
Four-bit activation quantisation of language models is usually tuned and compared by perplexity. We ask what that misses, and find three things. First, perplexity gets the order of quantisers right but the size of their damage wrong. Across four models it ranks 33 of 34 pairs of settings correctly by closed-book QA loss, yet a 6-16% rise in perplexity comes with a 10-23% drop in QA accuracy, and the lost facts are mostly rare ones. Second, the ranking only holds while the quantisation error acts like random noise. Without rotation, rounding small values to zero deletes part of the signal, and this hurts in-context retrieval far more than noise of the same size would. The crest factor of the quantiser's input (its peak-to-average ratio) helps tell the two cases apart. Third, rotating in blocks, the usual workaround when a layer's width has no ready-made Hadamard matrix, can land a model in between and leave outliers behind; 10 of 42 popular open models have such a width. A full-width Hadamard exists for every such width we found, costs about the same, and on Qwen3-4B halves the closed-book TriviaQA loss (0.112 to 0.060, over three rotation seeds). We recommend reporting a knowledge-heavy task alongside perplexity, and rotating at full width wherever blocks leave an input outlier-heavy.
cs.LG