User Modeling
Overview
User modeling builds representations of people from interaction histories, queries, clicks, conversations, sensor streams, or learning outcomes so systems can predict needs and adapt recommendations, tutoring, and agents. Common approaches include sequential and graph encoders for behavioral data, latent-state and hierarchical persona models, retrieval-based or compact memory for language models, and synthetic-user simulation. Systems are evaluated on recommendation and next-action tasks, personalization benchmarks, and educational or conversational settings, with growing attention to temporal change, uncertainty, and evidence for inferred traits. A central open tension is whether these models genuinely capture individual users: synthetic personas and automatic judges can miss real preferences, while richer behavioral data raises privacy and over-inference risks.
Latest papers
163 papers14 papers in the last four weeks, down 26% on the four weeks before. 0.1% of all new papers.
Probe What-If Worlds Before You Ship: Persona Simulation for Real-World Decision Making
SeqFold: Timestamp-Addressable Compressed States for User Action Sequence Backbones
SeqFold, a compressed-state interface that combines learned multi-channel fusion, channel-specific temporal decay, and Lookup-Free Quantization (LFQ) to produce compact discrete codes from frozen backbone states. Task-specific decoders then map these shared codes into flexible embedding spaces for downstream models. To make timestamp-addressable export practical at scale, we further develop a fused streaming implementation that updates the fusion recurrence after every action while materializing states only at target timestamps. On large-scale proprietary datasets, SeqFold reduces per-state representation storage by while preserving most of the downstream ranking performance of the uncompressed model, showing that SeqFold is an effective interface for reusing long-history sequence backbones in downstream models.