PERSONALIZATION · 2026-11-09

Personalization — Paper Daily, 2026-11-09

Today's personalization papers follow three threads. The first is how to describe users and items. Hotel reviews put criteria ratings and free text to work together, and products with random payoffs get a risk-preference curve learned for each user. The second is training systems. SparseLift rebuilds embedding lookups and cross-GPU communication into subgraphs a compiler can read, then optimizes the whole graph. The third is borrowed tools. One paper describes configurations by their past scores to pick algorithms, and another tunes a short sequence of prompt vectors so a small model judge matches human ratings.

3.65

1. SparseLift: Automatically Optimizing Opaque Sparse Paths in Industrial-Scale Recommendation Model Training

In recommendation training, the embedding lookups and cross-GPU communication (the paper calls this the sparse path) appear in the computation graph as opaque nodes. A compiler cannot see inside them, so it cannot optimize them together with the computation around them. SparseLift first uses the model structure and runtime information to rebuild each opaque node as an analyzable subgraph. A cost model then checks, against the current execution plan, whether each candidate optimization actually pays off, and the chosen ones are assembled into a plan that later steady-state steps reuse. On a video-platform recommendation task with 100B sparse parameters, steady-state throughput on 8 GPUs rises by 16.5%. Switching optimizations off one at a time shows that running some sparse work early, so downstream steps wait less, matters most: removing it adds 12.7% to step time.

3.60

2. IncVNM: An Interactive Recommender System with Incremental von Neumann-Morgenstern Utility Learning

For products with random payoffs, such as stocks or insurance, whether a user likes one depends on the whole payoff distribution and on how much risk the user can tolerate. A recommender that learns one score per product goes stale when the distribution shifts. IncVNM learns a utility function for each user, which describes how that person trades risk against return. After the user clicks within a displayed list, the click order is turned into pairwise comparisons, where a product clicked earlier is preferred to one clicked later or not clicked, and the model changes its previous estimate as little as possible to satisfy them. The next list is chosen by scoring products with the current utility estimate and then sampling with entropy regularization. High scorers are more likely to be picked, but the top scorers are not always the only ones shown, so users choose among more varied products and give the model new comparisons. The authors simulate a risk-averse user on 463 S&P 500 stocks over 29 rolling windows. Under every click-noise level tested, IncVNM beats the expected-return, mean-variance, discounted UCB and BPR-variant baselines.

3.60

3. Multicriteria Recommendation System by Leveraging Predefined, Implicit, and Undefined Criteria

A hotel's rating form lists only a few criteria such as location, service and value, but users also write in their reviews about things the form leaves out, like breakfast and noise. This paper splits each review into sentences and computes how similar each sentence is to every predefined criterion with sentence embeddings. A sentence goes to the criterion it most resembles, and sentiment analysis turns it into an implicit rating for that criterion. A sentence that is not similar enough to any predefined criterion (similarity below 0.3) is treated as an undefined aspect and also gets its own rating. A neural network then takes the predefined, implicit and undefined ratings together to predict the overall rating. On TripAdvisor the MSE is 3.08% lower than the strongest baseline, MRRRec, and on Beer Advocate it is 3.65% lower. The gain is modest, but dropping any of the three kinds of rating raises the error.

3.40

4. A Common Language for Dynamic Graphs: Interaction Tokenization Enables Foundation Models

A common task on dynamic graphs is predicting the next interaction, for example which item a user will visit next, and today each dataset usually gets its own separately trained model. This paper tries to build one pretrained model that works across datasets. For each query, the history before it is turned into one row of tabular features: who has contacted whom repeatedly, which intermediate nodes connect them indirectly, and which destinations were visited in what order. An existing tabular foundation model can read such rows directly, so no dedicated graph encoder has to be trained. The authors then keep training that model on a large set of synthetic dynamic graphs. On a real graph it uses labeled samples from the training period as context and updates no parameters. Across 13 datasets in TGB and TGB-Seq, CL-DGFM has the highest MRR on 10, beating the strongest baseline by 15.24 points on Flickr, and it is slightly behind TPNet on Wiki.

3.30

5. Multi-criteria Rating and Review based Recommendation Model

A user rates a hotel on a few criteria such as value, service and cleanliness and also writes a comment. How should the two be combined to predict the overall rating? Earlier models used either the criteria ratings or the aspect ratings extracted from the text. MRRRec uses both. It estimates an implicit rating for each aspect from the comment, fills in criteria the user left unrated by predicting them from the ones that were filled in, and merges everything with attention. On TripAdvisor, against baselines such as ANR and DeepCoNN, MSE is on average 19% lower. In the ablation, dropping the criteria ratings and using only the review raises MSE by 7.7%, while dropping the review and using only the criteria ratings raises it by just 0.6%. So most of the gain comes from the criteria ratings, and the review text adds a smaller part. The authors also admit the model cannot yet tell which aspect in a review corresponds to which criteria rating.

3.15

6. Simple Meta-Learning for HPO and CASH Using Performance Representations

To pick an algorithm and tune its hyperparameters for a new tabular task, people usually use Bayesian optimization. But when the hyperparameters of every algorithm are concatenated into one input, the dimension is high and, for any single configuration, most dimensions are inactive, which is hard to learn. This paper describes a configuration differently. It does not use the hyperparameter values. It uses the vector of scores that configuration got on earlier tasks, much like describing an item by the ratings users gave it in a recommender. Paired with plain Bayesian ridge regression, on the TabArena task of choosing an algorithm and tuning it, it reaches in 73 evaluations the level that ABLR needs 149 evaluations to reach, close to half the trials. Fitting is also far faster, under 3 seconds per task, against about two hours for GP-Hyper. When tuning only a single XGBoost model, ABLR is still best and this method ranks second.

3.05

7. Adapting LLM Judges to Observed Human Annotation Policy through Soft Prompt Tuning

An open-source small model used as a judge for chatbot replies often scores differently from human annotators. This paper freezes the model's parameters and trains only a short sequence of learnable vectors (a soft prompt) placed before the textual rubric, so the judge picks up the annotators' scoring habits. Training and testing use FairBotBench, a benchmark of e-commerce customer-service chats where humans rate each reply for bias, toxicity and personalization. Measured by agreement with human ratings (Krippendorff's alpha), Qwen 3.5 0.8B on toxicity starts out negatively correlated with human ratings and reaches 0.735 after tuning. After tuning, Qwen 3.5 9B beats the two hosted reference judges on all three dimensions. The authors caution that the size of the gain, and whether to keep the textual rubric, vary with the model and the task, so it should be re-validated on your own judge model.

3.05

8. TabAbax: A Tabular Foundation Model Technical Report

To make a tabular foundation model predict better, deepening the in-context-learning Transformer helps very little, and the room lies upstream of it. TabAbax changes two things on top of TabICLv2. One is the synthetic tables used for pretraining. The number of categories in categorical features matches the preset value exactly. Category proportions are drawn from several distributions such as Dirichlet. Regression targets are no longer passed through a monotone stretching of their values after generation, so the target keeps the smooth relation to the features that the generation produced. The other is the encoder. It lets the rows within each column attend to one another and then the columns within each row, and that pair of steps now runs twice instead of once. On the regression benchmark Elo is 336 above TabICLv2, and swapping in only the new synthetic data adds 228, so most of the gain comes from the quality of the pretraining data.