Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
A test by an unnamed team found that a quantized 4B multimodal model running on a single AWS Graviton4 CPU server with 16 vCPUs and 32 GB of RAM completed 360 of 380 enterprise AI tasks (94.7%) across 13 industries, with…
In a blog post, an unnamed author argues that AI performance will not plateau, countering the common belief that progress is driven by brute-force scale with diminishing returns. The author contends that combining genera…
SGLang, the open-source inference project associated with Ying Sheng and Banghua Zhu, published deployment recipes for Alibaba's Qwen3.8-27B model, enabling single-GPU serving with NVFP4 quantization and DFlash2 speculat…
A new study from arXiv researchers finds that language models leak sensitive in-context secrets through benign outputs, with 2-digit secrets reconstructed with near-perfect accuracy and 4-digit secrets at 82% exact match…
A researcher at Alignment of Complex Systems showed that a 32B open-source model, Qwen2.5-Coder-32B, can subtly introspect when external concepts are injected into its activations, despite appearing unable to do so. By a…
A developer found that DSPy's compiled programs re-send up to 20 few-shot demonstrations on every inference call, adding significant token overhead that is invisible in the source code. The default optimizer attaches 4 b…
A developer's final part of a CUDA/ROCm optimization series details implementing Mixture of Experts (MoE) inference, including a custom top-k gating kernel, expert parallelism with All-to-All communication, and CPU offlo…
Moderna and Merck's experimental mRNA cancer vaccine intismeran autogene met its primary endpoint of recurrence-free survival and a key secondary endpoint in a Phase 3 trial of 1,137 patients with resected stage IIB-IV m…
A developer shared on Hacker News that they trained a 125M parameter transformer-based language model to autocomplete piano melodies entirely on-device, with no internet connection required. The model tokenizes MIDI data…
A new CUDA kernel correctness harness compiles kernels with nvcc, loads them into Python via ctypes, and checks outputs against NumPy baselines, with timing via CUDA events. All 23 tests pass on an NVIDIA T4 GPU, coverin…
A developer argues that AI applications are software and should be tested with the same rigor as traditional code, advocating for evaluation datasets, prompt testing, and context testing. The developer highlights that pr…
Lino Avila and the JX Planet X project have released JX N-Body Engine 0.1.0, a Python toolkit for high-precision Newtonian N-body simulations with built-in numerical validation. The software uses a sixth-order Yoshida in…
A paper from Amazon Science presented at AAAI 2026 found that agentic keyword search achieves 94.5% of RAG's faithfulness score without a vector database, suggesting retrieval quality increasingly comes from the agent's …
A new arXiv paper (2608.19202v1) proposes active inference as a framework for AI agents to acquire context efficiently, balancing default assumptions against clarifying questions, retrieval calls, or tool calls under tok…
A study submitted to arXiv on 2 Jun 2026 compares transformer models BERT, RoBERTa, and BART for text summarization, examining their architectures, pretraining strategies, and suitability for extractive and abstractive t…
Researchers propose R2-OPD, a method that filters on-policy distillation rewards by reasoning progress to improve language model post-training. The approach, detailed in arXiv:2608.19408v1, suppresses teacher-derived rew…
A new arXiv preprint (2608.19535v1) proposes telemetry-informed adaptive compression for retrieval-augmented generation (RAG) on edge devices, based on experiments on the NVIDIA Jetson AGX Thor with Llama and Qwen genera…
Researchers introduced QDOS (Quality-Diversity Offline Skill learning), a unified pipeline for offline-to-online reinforcement learning that uses an Advantage-Weighted Quality-Diversity pretraining objective to extract d…
Researchers propose the Correlation-Aware Memory Arbitration (CAMA) framework to address Memory Correlation Bias in long-term multi-agent systems, where memories from different agents may share upstream sources, causing …
DiDi Global Inc. has deployed GenMatch, the first end-to-end generative matching framework for micro-view order-dispatching in ride-hailing, in its production environment. The framework, detailed in a new arXiv paper (26…