Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
An audit of sixteen open RLVR math corpora covering 1,673,514 training items against six benchmarks with 1,949 questions found that at a cosine threshold of 0.80, 52 of 59 sampled removals (88%) were confirmed false remo…
A Teacher-Student weak supervision framework using a large language model as a Teacher to label unlabeled cybersecurity tweets lets domain-specific pre-trained language models reach F1 scores above 0.92, within 3 percent…
Block Influence (BI) depth-pruning rankings derived from English calibration systematically miss late-layer activity in Indian-language text, with late-block BI 20.7-55.5% higher for Hindi, Marathi and Tamil than for Eng…
On-Policy Reminder Distillation (OPRD), a training method for long-horizon coding agents, improves the benchmark-averaged task resolution rate by 11.3 points over the open-source Qwen3.5-4B base model on SWE-bench Verifi…
Researchers introduced DPA-MoE, a runtime precision control framework that encodes Mixture-of-Experts feed-forward weights as prefix-decodable chunks using SharedScale AnyBCQ and allocates chunks at inference time based …
A new arXiv paper (2610.21436v1) presents HyCo, a hybrid speculative decoding method that assigns draft-tree depth to a server model's native multi-token prediction head running on a mobile NPU while alternative branches…
A paper submitted to arXiv on 9 Oct 2026 reports that much of the reasoning gain from reinforcement learning in NO-THINK mode on hybrid reasoning models comes from "thinking leakage" — partial reactivation of the origina…
A developer built Ranchi Dukaan Pro, a Python and Gradio prototype that uses the TabPFN-TS time-series model to forecast seven-day product demand for a small kirana shop and convert those estimates into reorder recommend…
Alibaba's QwenGyre paper reports that reinforcement-learning rollouts on its 122B model averaged around 2 hours, with many runs on the 2.4T flagship exceeding four hours, as agentic post-training increasingly diverges fr…
Prism ML released Ternary Bonsai 2 9B, a ternary-quantized GGUF build of Qwen3.5-9B that shrinks the vision-language model from 17.9GB to roughly 2.4GB by storing most weights as -1, 0, or +1 with one scale factor per 12…
PrismML released Bonsai 2 in 2B, 4B, and 9B parameter sizes, publishing the models on Hugging Face as GGUF and MLX 2-bit quantized builds under the prism-ml organization. The Ternary-Bonsai-2-9B-gguf build drew 16.4k dow…
Foerster, an MIT-licensed probabilistic programming library for Clojure written by Christian Weilbach, lets programmers write a model once and choose the inference method later, as demonstrated by a two-dice example and …
The New York Times published a visualization built on Epoch AI's estimates of training compute and capability scores, showing that training compute for frontier language models has grown roughly 5x per year since 2020, a…
A developer argues that many LLM calls are effectively multiple-choice or extraction tasks and should be handled by pointer-style models that score spans or labels already present in the input rather than generating text…
Munim, Inc. published a vendor-neutral, fully-cited coverage matrix mapping 16 model-file attack vectors against five open-source scanners (picklescan, fickling, ModelScan, ModelAudit and others), with each cell traceabl…
MithrilMan released Teacher in the Loop, a set of 8M-parameter single-voice text-to-speech models distilled from Kokoro-82M in English and Italian, packaged as 9 MB int8 ONNX files that run entirely in the browser via ON…
Goodfire researcher Nick Cammarata published an essay claiming that roughly 90% of neurons in modern models including Qwen3-32B perform a single interpretable job, and that the canonical "cat-and-car neuron" in Inception…
A new arXiv benchmark, Into the Unknown, tasked nine frontier AI agents with writing 15 follow-up papers each on unpublished research directions contributed by scientists at US Department of Energy national laboratories,…
Running the same Qwen2.5-1.5B-Instruct weights on an Apple M1 MacBook in fp16 and an NVIDIA Jetson Orin Nano Super in fp16 produced scores within half a percentage point, but the generated answers differed on 38 of 250 G…
A paper submitted to arXiv on 9 Oct 2026 presents Minions, a dynamic and distributed virtual GPU cloud that aggregates idle Apple Silicon GPUs across macOS devices to run large-scale machine learning inference workloads.…