Search papers, labs, and topics across Lattice.
This paper investigates the causal impact of sparse attention mechanisms on model output in long-context foundation models, specifically through the lens of Block Sparse Flash Attention (BSFA). By employing a dense-calibrated counterfactual audit with matched probe cards, the authors reveal that sparsification significantly alters the influence of specific content, demonstrating that signal concentration and integration loss compete in determining output decisions. The findings indicate that higher compression ratios can amplify the effects of sparsification, leading to unexpected shifts in content influence that aggregate accuracy measures fail to capture.
Sparsification in attention mechanisms can drastically alter content influence, with higher compression ratios leading to surprising shifts in model output that standard accuracy metrics overlook.
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.
