SeqLLM: Integrating Behavioral Sequences into LLMs for Risk Control
August 5, 2026
SeqLLM enables LLMs to model long behavioral sequences by using a discrete vocabulary of behavioral tokens and a lightweight projector. This framework allows WeChat Pay to perform high-stakes merchant risk control without catastrophic forgetting of language capabilities.
HOW THIS AFFECTS YOU
●
builderYou can extend LLMs to process non-textual event sequences by using similar projector-based alignment.
●
researcherThis provides a method for grounding discrete behavioral tokens in an LLM's semantic space.