A new arXiv paper argues the way social platforms generate recommendations has a quiet math problem: the search method itself can block correct answers from ever showing up.
Today's generative recommenders, in the style of the TIGER model, decode discrete Semantic IDs using finite-width beam search. The paper shows this creates a reachability constraint: a held-out target item can be perfectly scoreable yet still never enter the returned list, because the beam search never reaches it. The researchers propose two fixes. CQG-Single generates one continuous retrieval query in a single pass. CQG-AR autoregressively builds continuous residual updates step by step. Both drop the discrete token vocabulary and code lookup entirely, so every encoded item becomes directly scoreable through dense or approximate-nearest-neighbor retrieval. Across seven runs, CQG-Single beat the baseline every time while using 55.3% as many trainable parameters, and CQG-AR matched or beat it on the ML-1M, Beauty, and Sports datasets.
This matters because generative recommenders borrowed their decoding approach from large language models without much scrutiny of what that borrowing costs. Treating beam search's blind spots as a tuning problem, rather than a structural one, may have left platforms optimizing the wrong layer. A smaller model beating a bigger one on its own turf also suggests real savings in training and serving cost, assuming the result holds outside these benchmark datasets.
There's a catch: the paper's own depth ablation found that adding more continuous refinement steps doesn't reliably improve accuracy. So this isn't a clean win, it's a trade of one unsolved tuning problem for another.