Understanding Sparse Attention Selectivity in Long-Context Foundation Models via Counterfactual Evaluation

2026-08-03

AI Summary

This work addresses the lack of effective evaluation of the causal impact of content block dropping induced by sparse attention mechanisms in long-context models. We propose the first counterfactual auditing framework for causal assessment, integrating Gold, Poison, and Benign probes to systematically analyze how sparsification modulates content influence across diverse layout configurations. Leveraging techniques including block-sparse FlashAttention replay, controlled Top-k selection, KV cache pruning, and dense calibration, we uncover a trade-off between signal concentration and integration loss, and demonstrate that compression ratio governs the magnitude of sparsity-induced amplification effects. Experiments across four model-task pairs reveal that sparsification substantially alters content influence—exhibiting stronger amplification in three settings and sign reversals in two—phenomena entirely obscured by aggregate accuracy metrics.

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.

Problem

• sparse attention
• long-context models
• attention selectivity
• counterfactual evaluation
• content influence

Innovation

• sparse attention
• counterfactual evaluation
• attention sparsification
• long-context models
• causal auditing