We introduce UNISON, a unified framework that bridges multiple instance learning and recommendation systems by treating both as instances of learning from scored bag collections of (item, score) pairs. Unlike traditional approaches that address bag-level classification and personalized recommendation as separate problems, UNISON decomposes learning into three interconnected stages: (1) learning universal item representations, (2) modeling bag-specific preference patterns through lightweight personalized parameters, and (3) inferring bag-level characteristics from learned preferences. We justify the architectural choice of linear bag-specific scoring through a first-order analysis: when item embeddings are sufficiently smooth in the score, a linear function approximates the score up to a Taylor remainder controlled by the curvature of the underlying preference function. For cold-start scenarios, we show that the cold-bag fitting problem reduces to ridge-less linear regression with a closed-form solution, and provide convergence guarantees for the iterative variant. Empirically, UNISON matches or outperforms state-of-the-art approaches across diverse domains: entertainment (books, movies), immunology (T cell receptor (TCR) based classification), and microbiome, offering superior computational efficiency and interpretability. A controlled synthetic experiment further confirms that the bag-specific parameters (wj,bj) recover the latent factors that drive item scores even when the projection φ has no access to the original scoring mechanism (AUC > 0.998). UNISON is available at https://github.com/louzounlab/UNISON.
Figures & tables
Figure 1: UNISON operates in three stages (Left plot): (Stage 1) Raw items from a bag (e.g., books) are processed by domain-specific encoders to produce latent representations zi . (Stage 2) Items are partitioned into a support set (known scores) and a query set (unobserved scores). An MLP, shared across bags, projects items into a low-dimensional space. For each bag, linear parameters (wj,bj) are optimized at inference time to predict the known support set scores. These parameters are then used to predict scores for the query set. (Stage 3) (wj,bj) are treated as functional embeddings, serving as input to a final MLP that predicts bag-level characteristics (e.g., user demographics). Comprehensive performance analysis and ablation studies (Right plot). (A) Conceptual Workflow Example: Raw items are transformed into latent embeddings, which are then used to optimize user-specific weights (wj,bj) based on the support set. These functional embeddings enable dual-task inference: predicting ratings for unobserved items (query set) and classifying bag-level traits (e.g., gender) without auxiliary metadata. (B) Predicted vs. ground-truth scores for 10 representative cold-start users, demonstrating the model’s ability to recover individual preference patterns. (C) Comparison of performance with (brown) and without (blue) the shared MLP projection. The left bars represent the Spearman correlation (Stage 2), and the right bars represent the AUC for bag-level classification (Stage 3). (D) Impact of support set size on performance. The green curve (left y-axis) tracks Stage 2 Spearman correlation. The orange curve (right y-axis) tracks Stage 3 AUC. (E) Histogram of items per user across the dataset. (F) Distribution of individual Spearman correlation coefficients across all users. (G) Stage 2 Spearman correlation (left triplet) and Stage 3 AUC (right triplet) comparison of three optimization regimes: sequential (Stage 2 then Stage 3), joint (Stage 2 and 3 simultaneously), and joint training followed by Stage 3 fine-tuning.
Figure 2: Left -plot - Training/Test definitions and recommendation benchmarks on Book-Crossing and MovieLen-1M. (A) Training/Test definitions: Users and books are partitioned into warm (red) and cold (blue) sets. Training is conducted exclusively on the WU-WI quadrant, while inference is evaluated on WU-CI, CU-WI, and CU-CI scenarios. Each user is associated with a support set (observed ratings, always consisting of warm items) and a query set for evaluation. (B) Mean Absolute Error (MAE ↓ ) across cold-start scenarios for MovieLen and Book-Crossing datasets. Each group shows three methods: MeLU (orange), ColdNAS (green), and UNISON (blue). (C) Spearman Correlation (SCC ↑ ) across cold-start scenarios for MovieLen and Book-Crossing datasets. The right hand plot is similar to Figure 1, but for MovieLens (A) Predicted vs. ground-truth scores for 10 representative cold-start users, (B) Stage 2 Ablation, (C) Support Set Sensitivity (D) Bag Statistics.
Figure 3: Effect of support set size on UNISON performance across three domains: TCR (left), and Microbiome (right). For MovieLens and TCR, we report both SCC (green) and AUC (orange). For Microbiome, we compare Stage 2 (solid, light) vs Stage 3 (dashed, dark) AUC across three datasets (M1-M3). Error bars show standard error across 5 random seeds.
Appendix figures & tables10 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 4: Empirical verification of Lipschitz property. Each blue point represents distances before ( x -axis) and after ( y -axis) projection ϕ for an item pair. The orange dashed line ( y=0.014x ) bounds nearly all pairs, confirming L≈0.014 . The red line ( y=0.003x ) shows average scaling. All points remain well below the diagonal ( y=x , black), validating the boundedness and contractive assumptions in Theorem 1.
Dataset
Bags
Items
Ratings
Sparsity
MovieLens
6,040
3,706
1,000,209
95.53%
BookCrossing
278,858
271,379
1,149,780
99.99%
Vaccine TCR
125
24,386
125,000
58.98%
Emerson TCR
600
82,565,945
600,000
99.98%
Microbiome 1
445
448
27,037
86.44%
Microbiome 2
344
175
11,627
80.69%
Appendix
Table 1: Dataset statistics.
Dataset
Bag
Item
Range
MovieLens
Age, Zip
Genre, Dir.
1∼5
Book-Crossing
Age
Auth., Pub.
1∼10
Vaccine TCR
Response
Vgene, cdr3
1.1e−5∼0.17
Emerson TCR
CMV
Vgene, cdr3
1e−5∼0.001
Microbiome 1
Gastric
Sequence
−1∼4.95
Microbiome 2
Parkinson
Sequence
−1∼3.97
Appendix
Table 2: Detailed content attributes and rating ranges for each dataset.
Hyperparam
MovieLens
Book-Crossing
TCR
Microbiome
Stage 2
Enc Dim ( d )
3584
3584
75
8
Latent Dim ( d′ )
48
48
64
8
MLP Arch.
1536-768-384-192-96
1536-768-384-192-96
64-48
32-16-8
MLP Dropout
0.15
0.15
0.03
0.07
Stage 3
Appendix
Table 3: Hyperparameters across all stages.
Dataset
Method
WU-CI
CU-WI
CU-CI
MAE ↓
MAE ↓
MAE ↓
Movie Lens
MeLU
0.9275
0.7854
0.9235
ColdNAS
0.8996
0.7896
0.8836
UNISON
0.8937
0.7579
0.8551
Book Crossing
MeLU
1.4468
1.3924
1.4302
ColdNAS
3.5052
3.3487
3.4908
Appendix
Table 4: Stage 2 performance comparison of recommendation methods across cold scenarios (MAE).
Dataset
Method
WU-CI
CU-WI
CU-CI
MAE ↓
MAE ↓
MAE ↓
Movie Lens
BPR
0.8639
0.8746
0.8563
NGCF
0.8406
0.8522
0.8357
LightGCN
0.8242
0.8410
0.8294
MetaHIN
0.7967
0.8066
0.7993
MeLU
0.9196
0.8104
0.9091
Appendix
Table 5: Stage 2 performance comparison of recommendation methods across cold scenarios (MAE).
Dataset
Method
WU-CI
CU-WI
CU-CI
SCC ↑
SCC ↑
SCC ↑
Movie Lens
MeLU (original enc.)
0.300
0.476
0.291
MeLU (Qwen3)
− 0.103
0.080
− 0.036
MeLU (Qwen3 + PCA)
0.299
0.299
0.286
UNISON (Qwen3)
0.261
0.377
0.323
Book Crossing
MeLU (original enc.)
0.224
0.185
0.111
Appendix
Table 6: Stage 2 performance comparison under different embedding configurations (Spearman correlation).
Scenario
MeLU
UNISON (Opt. CU-CI)
UNISON (Opt. WU-CI)
UNISON (Opt. CU-WI)
WU-CI
0.301
0.260
0.311
0.248
CU-WI
0.476
0.377
0.366
0.441
CU-CI
0.291
0.323
0.318
0.320
Appendix
Table 7: Spearman correlation (SCC ↑ ) on MovieLens under different UNISON optimization targets.
Scenario
MeLU
UNISON (Opt. CU-CI)
UNISON (Opt. WU-CI)
UNISON (Opt. CU-WI)
WU-CI
0.301
0.260
0.311
0.248
CU-WI
0.476
0.377
0.366
0.441
CU-CI
0.291
0.323
0.318
0.320
Appendix
Table 8: Spearman correlation (SCC ↑ ) on MovieLens under different UNISON optimization targets.
Configuration
AUC S2 ↑
AUC S3 ↑
Microbiome #1
Full UNISON (with MLP)
0.905
0.722
w/o MLP ( ϕ )
0.610
0.582
Microbiome #2
Full UNISON (with MLP)
0.885
0.627
w/o MLP ( ϕ )
0.642
0.545
Appendix
Table 9: Ablation study results for Microbiome datasets. The results demonstrate the impact of the MLP projection ϕ on performance across Stage 2 (S2) and Stage 3 (S3).
We introduce RecPFN, a prior-fitted network that brings in-context learning to sequential recommendation. RecPFN is pretrained entirely on synthetic clickstream environments sampled from a broad structural causal prior, enabling it to amortize Bayesian-style inference from a small support set. At inference, a lightweight decoder-only transformer conditions on a handful of domain sequences and produces next-item predictions for queries in a single forward pass, without any weight updates. Across eight public benchmarks, RecPFN achives state-of-the-art zero-shot performance while remaining strongly competitive with supervised methods in low-compute and low-data regimes. It is deployment-efficient and robust to domain shift, outperforming strong zero-shot baselines that rely on large real-interaction corpora. RecPFN provides a practical path toward generalizable, data-efficient recommenders and opens avenues for richer priors, longer-context ICL, and multimodal extensions. Code for training and evaluation is publicly available at https://github.com/SAP-samples/tabular-ai-recpfn/.
Large language models (LLMs) struggle to accumulate evidence across multiple rounds of user interactions, failing to update their beliefs in a manner consistent with Bayesian inference. Existing solutions require fine-tuning on sensitive user interaction data, limiting their applicability in privacy-conscious settings. We propose AdaptFuse, a training-free framework that externalizes probabilistic computation entirely from the LLM, where a symbolic module maintains a Bayesian posterior over a discrete hypothesis set, while a frozen LLM contributes semantic reasoning via multi-sample Dirichlet aggregation. The two signals are combined through entropy-adaptive fusion, which automatically weights each source by its predictive confidence, shifting reliance from the LLM to the symbolic posterior as evidence accumulates. We evaluate across three domains: flight recommendation, hotel recommendation, and web shopping; on Gemma2 9B, Llama3 8B, and Qwen~2.5 7B. AdaptFuse consistently outperforms both prompting baselines and fine-tuned Bayesian Teaching models on all tasks, with accuracy improving monotonically over interaction rounds. These results demonstrate that principled inference-time algorithms can substitute for fine-tuning in personalized recommendation, without storing or training on sensitive user data.
Fangzhou Lin, Zilin Dai, Shuo Xing +8
Texas A&M University · Worcester Polytechnic Institute · Tohoku University +1
In modern recommendation systems, retrieval serves as a primary stage responsible for filtering billions of candidate items down to thousands prior to refined ranking. To make this massive search effective and efficient, the system relies on ranking accuracy and indexing efficiency. However, these two objectives are traditionally misaligned: while the former optimizes for the alignment between ranking predictions and user behavior, the latter optimizes for a structural grouping of item representations which enables fast search among billions of candidates. Thus, despite extensive efforts to scale up interaction modeling for retrieval, they remain fundamentally limited by the structural misalignment between the ranking objectives and the proximity-learned index. In this work, we address this long-standing dichotomy by proposing a new holistic retrieval framework, OneShot. It is an end-to-end, in-model index learning framework that natively aligns index learning with ranking objectives. Using this joint learning as a structural foundation, OneShot pushes the boundaries of retrieval expressiveness by scaling interaction modeling with neural scoring beyond the persistent dot-product bottleneck. OneShot is fully deployed in Instagram's industrial short-video recommendation system, driving significant wins in user daily sessions, engagement, and time-spent. Additionally, OneShot achieves a 20% recall gain at the operational ranking volume and a 10x efficiency improvement at an equivalent recall level.