Papers

AI papers from arXiv and journals, each with a summary and why it matters · 7705 papers
Oct 30
PapersBuildersSelected10:41
Soft-MTP Fixes Multi-Token Prediction Training via KL Gradient Recovery

The KL gradient-recovery trick recovers exact KL divergence gradients from cross-entropy gradients without vocabulary-sized logit matrices, staying compatible with standard fused kernels — a practical detail most MTP implementations overlook. Teams pre-training models with MTP, including DeepSeek-style architectures, can adopt it as a drop-in change with no parameter or throughput cost. The repo offers a direct starting point for replication.

arXiv: DeepSeek@Jayden TeohSource
PapersSelected10:03
ParaTTT Parallelizes Deep Test-Time Training Memory Layers

The core insight is that deep nonlinear test-time memories no longer have to trade expressiveness for parallelism, since PTTT recovers exact online gradient descent as a fixed point while running chunkwise in parallel. Researchers building subquadratic sequence models should examine the algorithm and benchmarks, since it directly compares against Gated DeltaNet and Mamba-2 at matched model sizes. The paper appears on arXiv as 2610.38255v1.

arXiv cs.LG@Noel Loo et al. (6)Source
PapersBuildersSelected09:50
TIDE Speeds FSDP2 Training with Zero-Copy Collectives on 64 GPUs

The hidden cost in FSDP training is often layout conversion and collective interference rather than compute, and TIDE shows both can be largely eliminated with standard NCCL APIs rather than custom hardware. Teams running large-model training on multi-node clusters can adopt it to reclaim step time and GPU memory without changing FSDP2's per-parameter kernels. The mixed-precision bf16-to-fp32 reduce-scatter is a capability stock NCCL lacks, making the implementation details directly reusable.

arXiv cs.LG@Zejian Wang et al. (14)Source
Oct 28