Research map / Architectures and efficiency / RNNs and normalization
Preliminary Evaluation of the Test-Time Training Layers in Recommendation System (Student Abstract)
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Applies TTT-Linear test-time-training layers as the feature extractor of a sequential recommender (TTT4Rec) and compares it with DIN, GRU4Rec, SASRec and ComiRec variants on Amazon Beauty, Electronics and ML-1M.
Problem
Whether test-time training (TTT) layers, an efficient sequence-modeling architecture for short contexts, are effective for sequential recommendation compared with standard recommenders.
Approach
Replaces the usual sequence encoder in a sequential recommender with a TTT-Linear layer (self-supervised weight updates per token) followed by RMSNorm, and dot-products the last-click hidden state with a target-item MLP representation.
Key results
- On Electronics, TTT4Rec reports the best NDCG@5 (0.3486), NDCG@10 (0.3892), HR@5 (0.4712) and HR@10 (0.5968) among compared models.
- On ML-1M, TTT4Rec achieves NDCG@5 0.5239 and NDCG@10 0.5577, slightly above SASRec (0.5217/0.5574), but lower HR@10 (0.7877 vs 0.7919).
- On Beauty, TTT4Rec is comparable to baselines (NDCG@10 0.2890 vs ComiDR 0.2925), with ComiRec reported as overall stronger but requiring about 50 epochs to match TTT4Rec's single-epoch performance.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use TTT4Rec instead of ComiSA | Recommendation on Beauty · Beauty | NDCG@5: 0.2594 -> 0.2571 | Hurt |
| Use TTT4Rec instead of ComiDR | Recommendation on Beauty · Beauty | NDCG@10: 0.2925 -> 0.2890 | Hurt |
| Use TTT4Rec instead of SASRec | Recommendation on Beauty · Beauty | HR@5: 0.3854 -> 0.3866 | Helped |
| Use TTT4Rec instead of ComiDR | Recommendation on Beauty · Beauty | HR@10: 0.4902 -> 0.4848 | Hurt |
| Use TTT4Rec instead of GRU4Rec | Recommendation on Electronic · Electronic | NDCG@5: 0.3450 -> 0.3486 | Helped |
| Use TTT4Rec instead of GRU4Rec | Recommendation on Electronic · Electronic | NDCG@10: 0.3844 -> 0.3892 | Helped |
| Use TTT4Rec instead of GRU4Rec | Recommendation on Electronic · Electronic | HR@5: 0.4666 -> 0.4712 | Helped |
| Use TTT4Rec instead of GRU4Rec | Recommendation on Electronic · Electronic | HR@10: 0.5882 -> 0.5968 | Helped |
| Use TTT4Rec instead of SASRec | Recommendation on ML-1m · ML-1m | NDCG@5: 0.5217 -> 0.5239 | Helped |
| Use TTT4Rec instead of SASRec | Recommendation on ML-1m · ML-1m | NDCG@10: 0.5574 -> 0.5577 | Helped |
| Use TTT4Rec instead of SASRec | Recommendation on ML-1m · ML-1m | HR@5: 0.6790 -> 0.6834 | Helped |
| Use TTT4Rec instead of SASRec | Recommendation on ML-1m · ML-1m | HR@10: 0.7919 -> 0.7877 | Hurt |
Surprising findings
- ComiRec-style models underperform the simpler baselines on Electronics and ML-1M in the reported table, despite being described as advanced.
- Mini-batch size of 1 or 10 gave the best TTT-Linear performance, with larger mini-batches linking unrelated tokens and hurting results.
Limitations
- Preliminary student abstract with a small set of baselines and no variance or significance reporting.
- Results are inconsistent across datasets; TTT4Rec does not uniformly outperform baselines (e.g., Beauty NDCG@10 and HR@10).
Experimental setup
Datasets: Amazon Beauty Amazon Electronics MovieLens-1M
Benchmarks: Amazon Beauty Amazon Electronics MovieLens-1M
Metrics: NDCG@5 NDCG@10 HR@5 HR@10
Baselines: DIN GRU4Rec SASRec ComiRec-SA ComiRec-DR
Builds on: test-time training layers SASRec ComiRec
Closest papers by meaning
- Test-Time Training Provably Improves Transformers as In-context Learners (ICML 2025 · similarity 0.84)
- Actions Speak Louder than Words: Trillion-Parameter Sequential Transducers for Generative Recommendations (ICML 2024 · similarity 0.83)
- Would You Like Your Data to Be Trained? A User Controllable Recommendation Framework (AAAI 2024 · similarity 0.83)
- In-Place Test-Time Training (ICLR 2026 · similarity 0.82)
- A Layer Selection Approach to Test Time Adaptation (AAAI 2025 · similarity 0.82)
- Language Ranker: A Lightweight Ranking framework for LLM Decoding (NeurIPS 2025 · similarity 0.82)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.