Daily Research

Paper of the Day

A lightweight reading queue for papers on memory systems, efficient models, agent infrastructure, and trustworthy interfaces.

Paper of the DayDeepSeek AI / GitHub

DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

Xin Cheng, Xingkai Yu, Chenze Shao, Jiashi Li, Yunfan Xiong, Anyi Xu, Huishuai Zhang, Dongyan Zhao, Wenfeng Liang (DeepSeek-AI, Peking University)

Efficient Inference, Speculative Decoding, and On-Device Acceleration

DeepSeek releases DSpark, an open-source speculative decoding framework that accelerates per-user generation for DeepSeek-V4 models by 60–85% over prior methods. It attaches a lightweight draft module (semi-autoregressive) to existing model weights and uses confidence scheduling for verification batches. Maintains identical output quality while dramatically improving throughput and reducing latency on the same hardware. Fully open-sourced with paper, code (DeepSpec), and checkpoints.

Why it matters: Today's paper of the day, directly referenced in the new Recursive Singularity article. A practical example of substrate acceleration: speculative techniques that make small, fast local models (like the LFM2.5 browser agent) even more responsive without retraining. Open research that lowers the barrier for on-device intelligence harnesses. Essential reading for anyone building efficient agent systems.

Read PDFView Archive
Paper of the DayarXiv:1705.08039

Poincaré Embeddings for Learning Hierarchical Representations

Maximilian Nickel, Douwe Kiela

Geometric Memory and Hierarchical Representations

Introduces Poincaré embeddings, a method for learning hierarchical representations in hyperbolic space (the Poincaré ball model). The approach exploits the exponential volume growth of hyperbolic space to embed tree-like and hierarchical data with much lower distortion than Euclidean embeddings, enabling more efficient and accurate modeling of taxonomies, knowledge graphs, and structured memory.

Why it matters: Foundational paper for the geometric memory direction. Directly supports the idea that memory lives naturally on hyperbolic manifolds for hierarchical, branching, long-term coherence. Essential reading for anyone exploring Liquid Memory or manifold-aware state. The 2018 follow-up in the Lorentz model is also worth studying for numerical stability in practice.

Read PDFView Archive
Paper of the DayarXiv:2402.17764

The Era of 1-bit LLMs: All Large Language Models are in 1.58 Bits

Shuming Ma, Hongyu Wang, Lingxiao Ma, Lei Wang, Wenhui Wang, Shaohan Huang, Li Dong, Ruiping Wang, Jilong Xue, Furu Wei

Efficient Models and Systems Infrastructure

Introduces BitNet b1.58, a variant where every parameter of the LLM is ternary, taking values in {-1, 0, 1}. The authors report that from 3B parameters upward it matches full-precision (FP16/BF16) Transformer LLMs of the same size and training budget on perplexity and end-task accuracy, while being substantially cheaper in latency, memory, throughput, and energy.

Why it matters: Pick of the day, and the closest thing to a north star for the blog's efficient-models thread: the same ternary idea that could eventually put a capable assistant directly in the browser. Reported claim is parity with full precision at 3B+; the bigger claim of a new compute-optimal regime still needs broad independent replication and mature kernels to land in everyday tooling. (interpretation)

Read PDFView Archive
Paper of the DayarXiv:2605.12013

L2P: Unlocking Latent Potential for Pixel Generation

Zhennan Chen, Junwei Zhu, Xu Chen, Jiangning Zhang, Jiawei Chen, Zhuoqi Zeng, Wei Zhang, Chengjie Wang, Jian Yang, and Ying Tai

Computer Vision and Artificial Intelligence

L2P studies a latent-to-pixel transfer approach for pixel-space diffusion models. The authors freeze intermediate layers from pretrained latent diffusion models, train shallow layers for latent-to-pixel transfer, and report native high-resolution generation without the usual VAE bottleneck.

Why it matters: A useful paper for thinking about how much prior knowledge can be transferred without retraining an entire generative model from scratch.

Read PDFView Archive
Paper of the DayOpenBMB MiniCPM technical report

BitCPM-CANN: Native 1.58-Bit Large Language Model Training on Ascend NPU

BitCPM Team

Efficient Models and Systems Infrastructure

BitCPM-CANN documents native 1.58-bit large language model training on Ascend NPU infrastructure. It sits in the MiniCPM ecosystem and is useful for readers tracking practical work on ternary and ultra-low-bit model efficiency.

Why it matters: A good companion read for the blog's efficient-models thread: it connects model compression, training systems, and hardware-aware deployment.

Read PDFView Archive

Curated with care • 5 papers in archive