EPIC Adds Explicit Item-Level Competition to Diffusion Recommenders, Beating Baselines on 4 Amazon Benchmarks

_reachsumit · x · 2026-09-04

EPIC introduces explicit posterior item conditioning into Semantic ID masked-diffusion recommendation, adding item-level competition to denoising.

• Problem: a partial SID can map to multiple feasible items at each denoising step, while prior methods rely on position-wise token predictions
• Approach: build a personalized posterior over feasible candidate items from the generation context and recent user interactions, then project it back to unresolved SID positions to guide token decisions
• The pretrained backbone stays frozen, with no extra decoder forward pass
• Consistent gains over strong baselines on four Amazon benchmarks, driven mainly by personalized transition evidence that preserves promising item hypotheses

Original post →

More from Research

• Large Drug Discovery Model LDDM debuts with X-ray-validated novel hits across 5 targets — IAmSamFin · 2026-09-20
• After Chess and Go: Can Any AI Engine Actually Beat Humans at Scrabble? — zuilserip · 2026-09-20
• Textbook author: 99.9% accuracy can mean zero scientific discoveries — bravo_abad · 2026-09-20
• rasbt: Jev's Secret Sauce Is Data, Not the Algorithm — Laya Rival Falls Far Short — RichmanRonald · 2026-09-20
• Sebastian Raschka open-sources an end-to-end 'AI text detector from scratch' project — rasbt · 2026-09-20
• Sebastian Raschka: restricting LLM outputs to an action space is just a classic encoder classifier — rasbt · 2026-09-20

Research channel →