Proactive Interference (PI) — Cog4LLM

Minimal read, optional deep‑dive. Interference‑limited retrieval in LLMs.

ICML 2025 Workshop — Accepted Ongoing

What does our test reveal?

Classic retrieval tests (Needle‑in‑a‑Haystack; MRCR) add many similar needles. We show the “haystack” isn’t necessary: by isolating and controlling the number of similar needles, we directly measure how interference limits retrieval. Retrieval declines log‑linearly with interference across major transformers.

In short: a core working‑memory bottleneck from interference, not just context length; a clean way to compare models’ PI susceptibility.

Why cognitive angle (short): After earning a physics degree, I spent two years self‑studying philosophy of science (admitted but not enrolled at University of Bristol) and took a two‑year gap at UChicago Divinity School (Eastern philosophy of religion). These experiences motivated the PI‑for‑LLMs project (accepted to ICML 2025 workshop).
Read more (abstract)

Information retrieval in LLMs intertwines with generation. We adapt proactive interference (PI) from cognitive science: sequentially stream semantically related key‑value updates and query the last value. Accuracy declines log‑linearly toward zero as interference accumulates; prompt‑based mitigation is limited. Findings suggest a working‑memory bottleneck beyond context access.

Read more (cognitive science foundation)

PI is a foundational paradigm for human working memory. Repeating cues updated with new values create outdated associations that must be suppressed to report only the latest value. Humans are resilient; LLMs show interference‑driven limits. Porting this paradigm to LLMs makes interference measurable and comparable.

Links