Reproduce “KUAISHOU Explorer LLM-Rec Challenge 2026: Reasoning Generative Recommendation” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jiangxia Cao, Hao Peng, Wenlong Xu, Jiaxin Deng, Zhixin Ling, Xingmei Wang, Kun Shang, Can Tang and 92 more
OneReason evaluates a single model on item understanding, user interests, multi-domain recommendation, and commonsense questions. Recommendation merges 32 thinking-mode and 32 direct-mode candidates, counting a hit when any matches a target item. The leading aggregate score was 1.44 versus 1.42 for second place. Its ablation found masking final-code losses improved performance over full supervision and down-weighting within this combined-mode setup.
Key insights
- Reasoning quality and recommendation performance do not necessarily move together; training should be assessed by recommendation results, not the presence or fluency of explanations alone.
- The winning team reported that masking target-answer loss on self-sampled reasoning trajectories improved recommendation and candidate complementarity, while supervising both reasoning and repeated answers reduced diversity.
- Rewarding partially correct SID prefixes can improve early-code accuracy without reliably improving complete item-code prediction, making credit assignment across code levels important.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.