Think Harder Where It Matters: Item-Level Adaptive Recursion for Sequential Recommendation

Think Harder Where It Matters. Item-Level Adaptive Recursion for Sequential Recommendation. By Zhanshuo Cao and four others, from Renmin University of China. Published on the twenty-first of October. Its code repository is public and real: about thirty source files.

Sequential recommenders read a user's history of interactions and predict what comes next. Most of them spend the same fixed computational budget on every item in that history, regardless of how much the item contributes to the final prediction. The authors challenge this de facto design with two diagnostic findings. The first is that computational need is not uniform. Items within the same sequence vary drastically in how much computation they require. Most converge after shallow processing, while a critical minority stabilizes only at full depth.

The second finding is that fixed allocation is suboptimal. Given oracle access to the target item, a non-uniform assignment of depth consistently outranks the fixed one. Uniform computation therefore loses twice over. It wastes capacity on items that have already converged, and it sacrifices overall recommendation quality at the same time. The fixed budget is neither efficient nor the best choice for accuracy.

Those observations motivate a shift, from fixed-depth computation to adaptive reasoning at the level of each individual item. LoopRec is the system built on that shift. In it, depth becomes something each item earns, rather than something every item is handed by default.

LoopRec applies a single weight-shared block recursively, so going deeper means passing through the same block again rather than through new layers. At each recursion boundary, a lightweight gate makes a binary decision, exit or continue, for every item independently. Items whose representations have sufficiently matured exit early. Those that need deeper reasoning carry on into subsequent recursions. The fixed global depth becomes a learned, per-item property, and it is learned without auxiliary depth labels. The design is also readily adaptable to diverse backbone architectures, which is what lets it be tested widely.

Extensive experiments on public benchmarks with four backbone families show LoopRec consistently outperforming fixed-depth counterparts matched to it. Further analysis looks at where the learned allocation chooses to spend its depth. The patterns are interpretable, and line up with the earlier post-hoc oracle analysis: lower-frequency items and more recent interactions naturally receive deeper computation. That agreement matters, because the oracle had access to the target item and the gate does not. A gate that learns, from training alone, to spend depth roughly where the oracle would have spent it is evidence that the per-item decisions are tracking something real rather than noise. The code is publicly available.

The paper's arc runs from diagnosis to design to confirmation. It measured that items need different amounts of thought, built a recommender that lets each item decide for itself, and then found the model's own choices agreeing with what the oracle had suggested. Rarer items and recent interactions get more attention, and the model arrived at that allocation without ever being told to.