SIGIR 2026 showcases three approaches to a single model aiming to replace an entire cascade

The conference in Melbourne again highlighted the gap between the lab and the production line: researchers continue to push generative retrieval as a full replacement, while industry still relies on classic rankers developed a decade ago. ByteDance unveiled HyFormer, a CTR architecture that fuses user history and hand-crafted features through Global Tokens that undergo two stages in each block—Query Decoding with cross-attention to layer-wise keys/values of the history, followed by Query Boosting with a lightweight MLP-Mixer that incorporates profile, context and candidate. In contrast to LONGER + RankMixer, which joins the streams only after compression, or MTGR and OneTrans, which employ a shared and expensive self-attention, HyFormer preserves a serial separation and uses a small number of global tokens throughout the model depth.

JD.com submitted two separate papers. GenRec tackles the app home page with a decoder-only model that generates Semantic IDs directly from the full catalog, and replaces point-wise NTP with page-wise NTP that predicts orders, clicks and impressions for an entire page in a single step, solving the multiplicity of targets for the same context and producing a denser training signal. Token Merger consolidates the Semantic ID tokens into a single vector and reduces the history length by half, while GRPO adds NLL regularization that prevents the generator from drifting away from real interactions.

OxygenREC extends the same idea to four scenarios—home page, product pages, cart and checkout—using Fast-Slow Thinking: a near-line model creates Contextual Reasoning Instructions that are stored in memory, and a Query-to-Item loss pulls the instruction representations toward target items during training. At inference time the instructions filter relevant events from a long history, and an MCTS per scenario explores sequences of Semantic IDs separately before Joint Pareto Optimization balances the shared policy across competing objectives. The three papers appeared in the same session without any of them reporting a common offline metric that would enable direct comparison.