Machine Learning News

A new analysis from the Journal of Machine Learning Research explains that Mixture-of-Experts (MoE) routing can cause large language models to produce different outputs across runs due to discrete expert selection amplif…

China deployed three homegrown AI weather models—Shanghai AI Laboratory's Fengwu, Huawei's Pangu, and Fudan University's Fuxi—to track Typhoon Dolphin alongside traditional numerical systems, marking a milestone for mach…

A study published in Nature Genetics identified 766 genes associated with schizophrenia, including 641 not found in previous transcriptomic analyses, using AI-based computational models to reconstruct gene activity in th…

ONNX Runtime, an open-source inference engine, enables interoperability of machine learning models across frameworks and hardware platforms by leveraging the ONNX standard. It simplifies deployment by allowing models tra…

Researchers propose AXPO (Agent eXplorative Policy Optimization), a reinforcement learning method that addresses the Thinking-Acting Gap in multimodal agentic reasoning by resampling tool calls for all-wrong subgroups. A…

SpatialClaw, a training-free framework that uses code as the action interface for spatial reasoning, achieves 59.9% average accuracy across 20 benchmarks, outperforming the recent spatial agent by +11.2 points. The frame…

Researchers introduced a novel think-answer distillation framework that masks salient reasoning prefixes to force compact vision-language models to anchor their thinking on visual evidence, reducing visual forgetting. Th…

Researchers introduced Zone of Proximal Policy Optimization (ZPPO), a method that keeps a teacher model inside prompts rather than policy gradients, improving knowledge distillation for small student models. On the Qwen3…

Researchers introduced GenRecal, a general-purpose distillation framework that transfers knowledge from large vision-language models (VLMs) to smaller ones by aligning feature representations across heterogeneous archite…

A new study identifies Relevant Visual Information Shift (RVIS) during decoding as the primary cause of performance degradation in visual token pruning methods for Multimodal Large Language Models (MLLMs) on complex reas…

Researchers propose Masters (Masking Teacher and Reinforcing Student), a mask-progressive reinforcement learning distillation framework that enables compact vision-language models to learn from large teachers by first ma…

Researchers propose Recursive Think-Answer Process (R-TAP), a method enabling large language models (LLMs) and vision-language models (VLMs) to iteratively refine their answers, outperforming single-pass approaches. R-TA…

Researchers introduced Unified Reinforcement and Imitation Learning (RIL), a training algorithm that combines reinforcement learning with adversarial imitation learning to create lightweight vision-language models (VLMs)…

WebAI released TwIL-LM, a two-model family of formal-logic reasoners at 1.7B and 3B parameters, with the 3B model TwIL-LM3 outperforming gpt-oss-120b on four of five formal-reasoning lanes while running locally on an iPh…

Researchers propose VLsI (Verbalized Layers-to-Interactions), a new family of vision-language models in 2B and 7B sizes that uses layer-wise distillation with intermediate verbalizers to align smaller models with larger …

An AI engineering intern shared five lessons learned during the first seven weeks of an internship, including the importance of monitoring both training and validation loss to catch overfitting, and the finding that simp…

A new study by researchers including Ziqian Zhong finds that when teacher models imitate other models, students fine-tuned on their answers inherit the imitated model's detectable writing signature but not its identity c…

VectorWare has made Rust's portable SIMD (core::simd) work natively on GPUs, allowing the same SIMD code to run on x86, ARM, and NVIDIA GPUs without modification. This direct mapping of SIMD vectors to GPU warps eliminat…

Researchers introduced CARV, a compute-aware variance-accounting framework that improves Monte Carlo estimation in diffusion-model pipelines, achieving 2-3x effective compute multipliers in text-to-3D distillation and at…

Loka, in collaboration with Arcee AI and AWS, post-trained an open model to perform scientific research tasks, achieving a score increase from 70.8% to 81.2% on the held-out Drug Tool evaluation and a combined score of 0…