The Anxiety of Influence: Bloom Filters in Transformer Attention Heads

2026-02-19

AI Summary

This study investigates whether attention heads in Transformer models possess Bloom filter–like membership testing capabilities—specifically, the ability to determine whether a token has previously appeared in the context. Through comprehensive analyses of attention patterns, modeling of false positive rates, control of confounding variables, and ablation experiments, the work provides the first evidence of genuine membership-testing heads in early layers that operate at multiple resolutions and are sensitive to token distance. These heads are prevalent across models such as GPT-2 and Pythia, exhibit remarkably low false positive rates (0–4%), and demonstrate 43% higher generalization performance compared to heads that merely respond to repeated tokens, revealing their dual role in processing both recurring and novel tokens.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models
• Computer Vision: Large Vision Models

Application Category

• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Web Mining and Content Analysis: Large pretrained models with web data
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Some transformer attention heads appear to function as membership testers, dedicating themselves to answering the question "has this token appeared before in the context?" We identify these heads across four language models (GPT-2 small, medium, and large; Pythia-160M) and show that they form a spectrum of membership-testing strategies. Two heads (L0H1 and L0H5 in GPT-2 small) function as high-precision membership filters with false positive rates of 0-4\% even at 180 unique context tokens -- well above the $d_\text{head} = 64$ bit capacity of a classical Bloom filter. A third head (L1H11) shows the classic Bloom filter capacity curve: its false positive rate follows the theoretical formula $p \approx (1 - e^{-kn/m})^k$ with $R^2 = 1.0$ and fitted capacity $m \approx 5$ bits, saturating by $n \approx 20$ unique tokens. A fourth head initially identified as a Bloom filter (L3H0) was reclassified as a general prefix-attention head after confound controls revealed its apparent capacity curve was a sequence-length artifact. Together, the three genuine membership-testing heads form a multi-resolution system concentrated in early layers (0-1), taxonomically distinct from induction and previous-token heads, with false positive rates that decay monotonically with embedding distance -- consistent with distance-sensitive Bloom filters. These heads generalize broadly: they respond to any repeated token type, not just repeated names, with 43\% higher generalization than duplicate-token-only heads. Ablation reveals these heads contribute to both repeated and novel token processing, indicating that membership testing coexists with broader computational roles. The reclassification of L3H0 through confound controls strengthens rather than weakens the case: the surviving heads withstand the scrutiny that eliminated a false positive in our own analysis.

Problem

• membership testing
• Bloom filters
• Transformer attention heads
• false positive rate
• confound controls

Innovation

• Bloom filter
• membership testing
• transformer attention heads
• false positive rate
• confound control