Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Researchers propose AgenticCANN, a knowledge-augmented agentic evolution framework for automated Ascend C operator synthesis on Huawei Ascend 910B NPUs, achieving 90-100% feasibility on elementwise and normalization oper…
Researchers propose RLPF (Reinforcement Learning from Performance Feedback), a staged reward method that trains code agents to prefer faster correct implementations. Fine-tuning Qwen3-32B with RLPF on PerfCodeBench raise…
Researchers propose SkillBoost, a three-stage framework that mitigates skill overfitting in large language model (LLM) agents by balancing exploration and exploitation. In experiments across 23 model–benchmark configurat…
A new method called Functional Reconstruction improves draft-token acceptance in speculative decoding for multi-head latent attention (MLA) draft models, according to a paper on arXiv (2607.27269v1). The method optimizes…
Researchers propose SDO (Structure-Aware Data Organization), a plug-and-play framework that organizes mini-batch composition and sample exposure based on representation-space structure to improve efficiency in large lang…
Researchers introduced CAPA, a benchmark with 600 coding sessions across 60 user-ambiguity cells, to test whether coding assistants can use a user's resolved session history to resolve recurring personalized ambiguities …
A new arXiv study (2607.27275v1) finds that 4-bit post-training quantization of large language models appears lossless on standard benchmarks but amplifies existing tool-calling failures by up to 2.5x in volume, masked b…
A new study from arXiv (2607.27304v1) introduces CoT-Mediate, a behavioral framework to test whether medical vision-language models (VLMs) causally use their generated chain-of-thought reasoning. Evaluating LLaVA-Med and…
Researchers released ZUNA1.1, a 380M-parameter diffusion autoencoder for EEG signal reconstruction that can handle variable-length sequences up to 30 seconds, arbitrary channel counts and scalp locations, and reconstruct…
A new arXiv preprint (2607.27350v1) reports that XGBoost outperforms Transformer-based sequence models for Sybil bot detection on Ethereum under leakage-aware evaluation, while also providing lower latency and energy use…
Researchers propose IGME, an efficient chained method ensemble for transferable semantic segmentation attacks, achieving competitive transferability with only one source model and faster runtime than model-ensemble attac…
Researchers at arXiv introduced a compression-based behavioral similarity method for Sybil discovery on Ethereum, using Gzip-based NCD on EVM transaction traces to differentiate Sybil bots, organic users, and arbitrage b…
ProgFormer, a hierarchical voxel-space Diffusion Transformer for longitudinal brain MRI prediction, outperforms state-of-the-art methods on ADNI, AIBL, and OASIS benchmarks by using a coarse pathway for volumetric predic…
A new arXiv preprint (2607.27566v1) finds that direct answer-only supervised fine-tuning (SFT) on MedGemma-1.5-4B is the most robust adaptation method for multi-frame medical visual question answering on the MedFrameQA b…
Researchers introduced FinSMART, the first market-aligned reinforcement learning framework for financial sentiment analysis, which directly optimizes sentiment signals using realized market outcomes and improves cumulati…
Researchers introduce Explorative Modeling (XMs), a new paradigm that factors the training loop by exploring K candidate matches between model generations and data, improving performance across images, video, and languag…
Researchers at arXiv (paper 2607.27383v1) established the first convergence guarantees for the plain vector-form Adam optimizer under heavy-tailed stochastic noise, showing it converges to (ρ,ε)-stationary points but wit…
Researchers introduced EvoPINN, an agentic framework that uses a large language model (LLM) to automatically discover executable algorithms for physics-informed neural networks (PINNs), reducing relative L2 error across …
Researchers introduced LayerRAG-Bench, a cross-layer reliability benchmark for agentic retrieval-augmented generation systems, covering 8 enterprise domains, 240 tasks, 9 fault scenarios, 2 contract modes, and 38,880 liv…
Researchers introduced Reflective Retrieval Memory (RRM), a framework that improves long-horizon multimodal reasoning by adding reflective experience memory to an entity-centric multimodal memory graph, distilling reusab…