Russian teams from T-Bank, MIPT, HSE, and Innopolis University presented three papers on recommender systems at the ACM CIKM '26 conference in Rome. The E²L³AE model fine-tunes a text encoder for LLM embeddings jointly with a recommendation model and outperformed the previous solution on all five test datasets during cold start, while the MoE-IGR architecture replaces the shared weights of a graph model with dynamically selected “experts.” The third paper, “Split Decisions,” demonstrates that standard offline evaluation of recommenders on random data splits systematically overestimates model performance on future interactions. Conference proceedings were published on November 6, 2026, and all three papers are fully readable via DOI.

What happened

At the 35th International ACM Conference on Information and Knowledge Management (ACM CIKM '26), held in Rome and publishing its proceedings on November 6, 2026, Russian teams from T-Bank, MIPT, HSE, and Innopolis University presented three papers on recommender systems—systems that suggest products and content to users. The E²L³AE model (Kurdun, Liakhnovich, Pechatov, Lashinin; pages 7388–7393 of the proceedings) trains a previously “frozen” text encoder for LLM embeddings jointly with a linear recommendation autoencoder, rather than separately: on five datasets—cosmetics, toys, groceries, office supplies, and clothing—it outperformed L³AE on all tests, achieving +20.3% recall and +17.2% ranking quality on the office supplies dataset; verified annotations confirmed improvements over both L³AE and other baselines. The MoE-IGR architecture from the paper “Give Me MoE” (Mardanova et al.; pages 7578–7583 of the proceedings) replaces the common update of the inductive graph model PinSAGE with dynamic layer-wise routing between “experts”; routing analysis showed expert specialization on cold and “warm” items, and the model achieved up to +16% Recall@20 and up to +34.4% NDCG@20 on new items. The paper “Split Decisions: How Evaluation Protocols Change Recommender Model Selection” introduces a model-selection regret metric and notes that in 17 out of 18 model comparisons, models selected on random data splits underperformed the choice made on a global-temporal split on future interactions—the gap reached 49% on the movies dataset and 81% on the short video dataset.

Context

To understand the significance of these works, context on industrial recommender systems is needed. Their bottleneck is cold start: for a product just appearing in the catalog, or a user without interaction history, a classic collaborative model trained on data about who viewed and bought what can hardly suggest anything meaningful. One way to fill this gap is using text descriptions of products as embeddings from large language models; however, in previous approaches, the encoder for such embeddings was “frozen”: it built representations trained on general language tasks and not adapted to a specific catalog or ranking task, and it is precisely this gap that E²L³AE addresses. The second path is graph models like PinSAGE, which operate on an interaction graph and, in inductive mode, transfer knowledge to previously unseen nodes, but typically update representations of all objects with shared weights; the idea of MoE (Mixture of Experts) is that instead of one common network, data is dynamically routed through specialized “expert” blocks, and MoE-IGR transfers this principle to graph-based recommendations. The third layer of context is the measurement methodology itself: offline evaluation of recommenders relies on splitting the dataset into training and test parts, and random splitting that ignores time mixes future events into the test set, whereas global-temporal splitting keeps the test strictly after training, so the difference between protocols determines which model appears best.

Why this matters for the industry

For the industry, these works offer three practical layers. First, cold start becomes cheaper: both architectures provide verified annotation-confirmed gains precisely where there is no interaction history, with the fine-tuned E²L³AE encoder itself acting as a competitive semantic recommender, and MoE-IGR expert routing ensuring measurable and interpretable specialization on cold and “warm” items; for founders, this is a signal that the “data bridge” from years of logs is no longer the only barrier to quality recommendations on new items. The second layer is methodological and the most implementable: the choice of data splitting scheme becomes part of the model selection decision, so for chronological forecasting, a model should be selected under a global-temporal scheme or honestly report induced regret, rather than presenting results on random splits as an evaluation of future interaction behavior. The third layer is an architectural trend: the combination of “fine-tuned LLM encoder plus classic collaborative model” in E²L³AE and “mixture of experts in graphs” in MoE-IGR looks like a testable template that product teams and ML platforms can already use to prepare their cold-start pipelines.

Why this matters for users

All three papers are published in open access under the CC BY license and are fully readable via DOIs 10.1145/3799682.3839945, 10.1145/3799682.3839967, and 10.1145/3799682.3839942, so conclusions can be verified against the original sources rather than summaries. If you build recommendations yourself, the practical takeaway from “Split Decisions” is to compare a model selected on a random split versus one selected on a global-temporal split: the discrepancy in metrics will reveal your model-selection regret and may change the final model choice, and such a run requires only repeated offline experiments and costs almost nothing. The second step is to read the papers via DOI and assess the transferability of E²L³AE and MoE-IGR recipes to your data before any deployment. If you follow the industry, this is a concrete example of a trend: hybrids of “LLM semantics plus collaborative model” and “mixture of experts in graphs” are gradually moving from papers into a standard toolkit for cold start. If you are simply a user of marketplaces and streaming services, you will see the benefit later and indirectly: when such methods reach production, new products and catalog novelties will be able to be recommended meaningfully from their first appearance.

What is still unknown / limitations

Several circumstances require caution. The impressive percentages are best points on specific datasets, not average results: “up to +34.4% NDCG@20” and “up to +16% Recall@20” for MoE-IGR are maxima on new items, while +20.3% recall and +17.2% ranking quality for E²L³AE were obtained on the office supplies dataset. There is a minor inconsistency in the materials: the summary cites +34.4% NDCG@20 on new items, while the post cites +34.5% on MovieLens against PinSAGE, so the exact figure and conditions should be verified against the full text of the paper. Code and model weight availability is not confirmed, and there is no data on latency or training cost, so both architectures remain recipes for reproduction rather than ready-to-integrate components. The gains have not yet been re-verified by independent groups, and the timeline for establishing chronologically honest evaluation in industry pipelines is an interpretation, not an established fact.

Sources

Author

Look at AI, editorial team