The Anxiety of Influence: Bloom Filters in Transformer Attention Heads

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.

💡 Research Summary

The paper investigates whether individual attention heads in transformer language models implement a form of probabilistic set membership testing akin to Bloom filters. The authors focus on early‑layer heads of GPT‑2 (small, medium, large) and Pythia‑160M, probing each head’s behavior with carefully constructed stimuli that contain exact token repeats, non‑repeats, and semantically similar “near‑miss” tokens. They define three quantitative metrics: selectivity (ratio of attention to a repeated token versus a baseline non‑repeated token), miss rate (fraction of repeated tokens that receive negligible attention to their first occurrence), and a continuous false‑positive (FP) ratio (attention to synonyms relative to true repeats). A head is deemed a Bloom‑filter candidate if selectivity exceeds three times baseline, miss rate is below 10 %, and hit attention is above 0.05.

Applying these criteria to GPT‑2 small (12 × 12 heads) yields four candidate heads: L0H1, L0H5, L1H11, and L3H0. The first three exhibit near‑zero miss rates, massive selectivity (51‑146 × baseline), and low FP ratios (0.01‑0.29). L3H0 initially appears similar but later fails a confound‑controlled capacity test and is re‑classified as a generic prefix‑attention head.

To test the Bloom‑filter hypothesis, the authors vary the number of unique tokens (n) in a fixed‑length (200‑token) context, measuring binary FP rates (fraction of probe tokens that attract attention above 0.1 despite never having appeared). L1H11’s FP curve follows the classic Bloom‑filter formula p ≈ (1 − e^{−kn/m})^k with an R² of 1.0, yielding an estimated bit capacity m ≈ 5 and effective hash count k ≈ 0.86. This head therefore behaves like a low‑capacity, quickly saturating filter. In contrast, L0H1 and L0H5 maintain FP rates below 4 % even when n = 180, indicating a high‑capacity, low‑error filter that far exceeds the nominal 64‑bit capacity of a head’s key dimension. L3H0, when tested under the same controlled conditions, shows a flat FP = 100 % across all n, confirming it does not implement a Bloom filter.

The authors further demonstrate that these three genuine membership‑testing heads are taxonomically independent from previously identified induction heads (which copy patterns A…B…A → B) and previous‑token heads (which attend to the immediate predecessor). No overlap is observed across all heads, and the Bloom‑filter heads are confined to layers 0‑1, suggesting a processing pipeline where membership resolution precedes pattern completion.

Generalization is verified on natural Wikipedia text (WikiText‑103). Across 761 passages, the same heads achieve 15‑54 × selectivity and <1 % miss rates, confirming that the phenomenon is not an artifact of synthetic stimuli. A large‑scale similarity sweep over 1,284 controlled probe tokens shows that FP rates decay monotonically with cosine distance in embedding space, aligning with the notion of distance‑sensitive hashing.

Ablation experiments (mean removal of each head) reveal that disabling these heads degrades performance on both repeated‑token and novel‑token predictions, indicating that while they specialize in membership testing, they also contribute to broader language modeling computations.

The paper’s contributions are threefold: (1) empirical identification of a new functional class of attention heads that act as Bloom‑filter‑like membership testers; (2) demonstration that at least one head follows the theoretical Bloom‑filter capacity curve, while others achieve remarkably low false‑positive rates even at high loads; (3) evidence that these heads form a multi‑resolution system (high‑capacity ultra‑precise vs. low‑capacity broader filters) and are distinct from known head categories.

Limitations include the lack of circuit‑level analysis linking Q‑K weight matrices to explicit hash functions, and uncertainty about whether similar heads persist in larger models (e.g., GPT‑3, LLaMA) or under different training regimes. Future work could probe the learning dynamics that give rise to these heads, explore scaling behavior, and harness them for practical applications such as efficient context management or memory‑constrained inference.

Overall, the study provides the first systematic evidence that transformer architectures internally instantiate probabilistic set‑membership structures, enriching our understanding of how large language models manage and reuse contextual information.