AI summaryⓘ
The authors studied how sparse attention methods, which skip some parts of information to make models faster, actually change what parts of the input influence the model's answers. They showed that removing blocks of data can significantly change model outputs and used special tests (with 'Gold,' 'Poison,' and 'Benign' inputs) to isolate these effects. They found two main effects: the model focuses more on key blocks when sparsity increases but loses the ability to combine information across blocks. Their tests across different models and settings reveal that these changes happen even when overall accuracy stays the same. They also provide tools to measure these effects on any model that uses block-based attention.
Sparse attentionBlock Sparse Flash AttentionCounterfactual auditProbe cardsSignal concentrationIntegration lossCompression ratioKV-cache evictionBlock identitiesModel interpretability
Authors
Xingyu Ren, Youran Sun, Chugang Yi, Haizhao Yang
Abstract
Sparse attention is widely deployed in long-context serving stacks, yet no framework audits how discarding blocks changes the influence of specific content on model output. We first establish that the phenomenon is real and causal: Block Sparse Flash Attention (BSFA) route replay across four architectures changes output decisions in 13 of 16 cells, with zero identity-replay label flips. We then introduce a dense-calibrated counterfactual audit using matched probe cards---Gold (carrying the correct answer label), Poison (carrying a target wrong label), and Benign (filler only)---under six-layout position symmetry, isolating the sparsification-specific effect. Two patterns compete. Signal concentration: the selector preserves Gold and Poison blocks far above filler-matched Benign blocks (G$\approx$P$\gg$B across all model--task pairs). Integration loss: discarding blocks severs cross-block attention---confirmed by an ablation where isolating the probe block collapses its influence from 4.48 logits to zero. Compression ratio governs the balance: a full sweep from mild ($c=0.25$) to aggressive ($c=0.75$) compression across four model--task pairs reveals that three of four cells move toward stronger sparse amplification at higher compression, with two exhibiting sign reversals. Three independent arms---BSFA route replay, controlled block-top-$k$, and KV-cache eviction---converge: sparsification changes content influence in ways aggregate accuracy cannot detect. We provide an open measurement framework deployable on any model exposing block identities.