CONFIDENCE-GATED DISTILLATION OF LANGUAGE MODEL SEMANTICS FOR COLD-START RECOMMENDATION

Collaborative filtering fails on items with no interaction history, and a large language model is an inexpensive source of the side information that could fill the gap. The open question is not whether such side information helps but which part of it does, and existing evaluations answer neither: the protocols labelled cold-start are usually leave-one-out splits in which the held-out item is still trained, and the controls that would isolate the language model’s contribution are absent. We present DualDistill, which distils an LLM-derived item geometry into a graph collaborative backbone through four auxiliary objectives and combines the two branches with a dual confidence gate that is made identifiable by a fused ranking loss. We evaluate it under a strict item cold-start protocol in which held-out items have exactly zero training interactions, alongside standard leave-one-out, on three catalogs spanning two orders of magnitude in density. DualDistill reaches 0.1203 nDCG@20 on ML-1M against 0.0540 for the strongest baseline (123%), and under strict cold start it reaches 0.0796 against 0.0619 (29%), with consistent gains on Amazon-Beauty and Book-Crossing. A sixteen-way ablation over 5 seeds separates the components: replacing the semantic neighbourhood with a random graph is the single most damaging change (-0.0144 nDCG@20), and freezing the gates costs -0.0094. Inspecting the trained gate shows it does not degenerate — its output spans 0.25 to 1.00 across the catalog, so it is selecting rather than rescaling. We also report where the method does not win: on Book-Crossing a tuned item-based nearest-neighbour baseline is within 8% of the full model, and substituting raw catalog metadata for the LLM text costs only -0.0035 on ML-1M, which bounds how much of the gain is attributable to world knowledge rather than to text of any kind.