Chattaraj & Raj (2026)

LLMs Remember First, Forget Last: Dual-Process Interference in Large Language Models

arXiv

Find It

DOI

Abstract

Large language models can process millions of tokens, yet how they handle conflicting information within context remains poorly understood. From patient health logs tracking evolving vital signs to legal documents with superseding clauses, real-world applications routinely require models to retrieve specific values from streams of semantically similar, competing updates. We adapt classical interference paradigms from cognitive psychology to compare retroactive interference (RI; recalling initial values after updates) and proactive interference (PI; recalling recent values despite competing prior encodings) across 39 LLMs spanning 1B to 2.5T parameters. Every model exhibits the same pattern: PI causes substantially greater degradation than RI (d = 1.73), the opposite of the typical human finding where new information more readily disrupts old. Four lines of evidence indicate that RI and PI engage distinct mechanisms: model size predicts RI resistance but not PI; the two scores correlate only weakly; error analysis reveals qualitatively different failure profiles, with RI failures reflecting passive retrieval failure and PI failures reflecting active primacy intrusion; and hallucination rates remain below 1% in both conditions. An adaptive bootstrap design further shows that RI decays smoothly while PI collapses bimodally at high interference, consistent with gradual dilution versus winner-take-all attention dynamics. These findings characterize a systematic primacy bias in transformer attention, with implications for reliable deployment in settings where in-context information evolves over time.

Reference

Chattaraj, S., & Raj, K. (2026). LLMs Remember First, Forget Last: Dual-Process Interference in Large Language Models. arXiv preprint arXiv:2603.00270. https://doi.org/10.48550/arxiv.2603.00270

What Rests on This

2 mechanisms cite this source. Each finding below is quoted as it appears on that mechanism's own page.

Consolidation Memory Processes

Measured in AI

The closest classic paradigm tried on AI is interference: across 39 language models, earlier information in the conversation disrupted recall of later updates more than the reverse (d = 1.73), opposite to the typical human pattern, but this tests the context window (what the model can see in the current conversation) with no delay or offline period, and no people were tested

Retrieval Interference Memory Processes

Measured in AI

An A-B, A-C style study of 39 language models found proactive interference exceeded retroactive interference in every model (d = 1.73), which the authors call the reverse of the human pattern, but that contrast rests on past literature rather than people tested on the same items, so it cannot be scaled