LLMs Process Lists With General Filter Heads

TL;DR

This paper reveals that large language models encode a general, interpretable filtering mechanism similar to functional programming, which can be extracted, reused, and understood through causal analysis.

Contribution

It uncovers a causal, interpretable filter head mechanism in LLMs that generalizes across tasks, formats, and languages, advancing understanding of model internals.

Findings

Filter heads encode a portable filtering predicate.

Models can reapply filtering predicates across different contexts.

Alternative strategies like eager evaluation are also identified.

Abstract

We investigate the mechanisms underlying a range of list-processing tasks in LLMs, and we find that LLMs have learned to encode a compact, causal representation of a general filtering operation that mirrors the generic "filter" function of functional programming. Using causal mediation analysis on a diverse set of list-processing tasks, we find that a small number of attention heads, which we dub filter heads, encode a compact representation of the filtering predicate in their query states at certain tokens. We demonstrate that this predicate representation is general and portable: it can be extracted and reapplied to execute the same filtering operation on different collections, presented in different formats, languages, or even in tasks. However, we also identify situations where transformer LMs can exploit a different strategy for filtering: eagerly evaluating if an item satisfies the predicate and storing this intermediate result as a flag directly in the item representations. Our results reveal that transformer LMs can develop human-interpretable implementations of abstract computational operations that generalize in ways that are surprisingly similar to strategies used in traditional functional programming patterns.