Task-level outcomes and representative code changes from AutoRSE experiments.
Loading paper reference results…
These examples have different experiment and evaluation scopes. Each result is reported within its own scope, separately from the task-level means above.
Standard · Claude Code × Claude Opus 4.8
Compare two independent variants of the same HSTU control: cosine learning-rate decay with warmup, and a learnable projection of the supplied frozen 768-dimensional item features into the ID embedding space.
Validation NDCG@10: control 0.06770; training schedule 0.06997; content fusion 0.07216.
Paper case study ↗Standard · Claude Code × DeepSeek V4 Flash · Trial 1
Combine SASRec, HSTU and TIGER predictions with item-similarity scores derived from the latest interaction.
Final submission NDCG@10: 0.0631.
A single archived run; this is not the three-trial mean or an isolated component effect.
Archived-run evidence ↗Idea-injected · Claude Code × DeepSeek V4 Flash
Expand training candidates with in-batch positive items, retain cross-entropy, and mask target duplicates and observed-history negatives.
Models trained with this augmentation were retained in the final ensemble. Its isolated performance contribution was not measured.
Paper appendix case study ↗