Researchers posted the preprint “Cut Binary Cross Entropy: Efficient Large-Vocabulary Loss and Gradient Kernels for Sequential Recommendation”1, which describes CutBCE as an exact BCE loss and gradient operator for large-vocabulary sequential recommendation workloads.
The work matters because the paper targets a training bottleneck that the authors say appears in industrial recommender systems with very large item catalogs, where standard binary cross-entropy can exhaust accelerator memory. According to the arXiv preprint1, those systems can operate over catalogs of “10^5--10^7 items,” and standard BCE materializes a dense logits tensor in High Bandwidth Memory, creating what the authors describe as prohibitive O(BNV) memory use and out-of-memory failures.
The preprint remains an arXiv posting rather than a peer-reviewed paper. It says CutBCE is implemented in JAX and Pallas and combines an “exact fused reformulation” for dense background loss and sparse target corrections with a custom vector-Jacobian product and a TPU backward kernel that computes logit tiles on-chip so logits and gradients do not reside in HBM. The abstract also says the method includes dynamic VMEM budgeting, sharding-aware collective hoisting for distributed meshes, and count-based training metrics at what the authors describe as zero overhead.
The performance figures currently available are limited to the authors’ own reported results in the preprint. On single-chip TPU v5e and v6e mini-benchmarks, the abstract1 says CutBCE eliminates out-of-memory errors with up to 91.9% speedup. On 8-chip TPU slice training for multi-label SASRec with 876k items on Yambda-50M, it says CutBCE reduces peak HBM by 65.7%, saving more than 14 GiB per chip, and increases training speed by 225.9% with comparable accuracy.
Those numbers establish what the authors claim for the tested configuration, not how the method performs across other model families, hardware targets or independent reproductions. The arXiv page does not provide peer-review status beyond the preprint posting, and the abstract-level record does not include methodology detail such as variance or repeated-run reporting.
The paper is listed on arXiv as arXiv:2610.055591 under machine learning, with additional subject tags for artificial intelligence, hardware architecture and information retrieval. The abstract also says CutBCE is open-sourced at “this https URL,” but the visible arXiv abstract page does not show the repository destination.