Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Independent testing found Jev's calibrated-probability claims fail in practice, with Valeriy M reporting calibration failures on 7 of 8 datasets across 16,500 predictions and the author's own coin-toss, two-dice and thre…
A developer built a Simplified Chinese iPhone app, 他爱我吗, that runs a Core ML conversion of the laya-multilingual model entirely on-device to score candidate answers to relationship questions such as affection, no affecti…
Nvidia released Nemotron 3 Diarization, a roughly 100-million-parameter AI model that identifies which speaker is talking at any moment and can distinguish up to eight speakers, with weights freely available on Hugging F…
A study found that indirect aggression is common in Indian conversations on X about sports and political events, and explored whether machine learning could help identify different forms of aggression in Indian X posts a…
Developer mchinen released Tape Heads, an open-source sonification system that plays back the attention patterns of audio transformer encoders including wav2vec 2.0, MuQ, and HuBERT, converting which heads attend to whic…
A developer built an independent benchmark comparing TypeSafe's Jev model against GPT-4, Claude, and Gemini on classification tasks including spam detection, sentiment analysis, and topic classification, with reproducibl…
Google Research's MSEB (Massive Sound Embedding Benchmark), version 0.1.0, is structured in three layers — types (Sound, SoundEmbedding, Score, TaskMetadata), the MultiModalEncoder contract that user models implement, an…
A developer built "Blog vs Bytecode," a Kaggle benchmark of 28 data-science code snippets paired with blog-style claims, half of which hide real methodological errors such as scaler leakage or shuffled time-series cross-…
A developer describes a paired dual-sample / feature-dropout training pattern for LLM rerankers that injects behavioral signals like CTR, QSS and exposure sequences into prompts while forcing the model to also learn sema…
A developer implemented the FlashAttention-2 forward pass twice — first as a deliberately slow PyTorch reference with a Python double loop over tiles, then as a Triton kernel — to teach kernel optimization. The Triton ve…
A developer built SonicSentinel AI, a dual-model sound-event detection system that pairs a locally trained Python XGBoost classifier with a Google Teachable Machine model, using their disagreements as a signal in the ale…
Xiaomi's MiMo-V2.6 models use groupwise agentic grading to replace binary pass/fail rewards in reinforcement learning, ranking multiple rollouts per task against each other to preserve gradient signal after models pass 8…
Researchers behind the paper "Evolutionary Architecture Search through Grammar-Based Sequence Alignment" (arXiv:2512.04992) use sequence alignment to identify the edits connecting two neural architectures and then recomb…
QLoRA fine-tuning on 150 synthetic examples raised a base Qwen3.5-4B model's adversarial-prompt refusal rate from 79.5% (31 of 39) to 89.7% (35 of 39), while GPTQ and AWQ 4-bit quantization added roughly +0.15 and +0.08 …
Sarvam AI released Saaras V4, a speech-to-text model covering all 22 scheduled Indian languages plus global English accents, available today through Sarvam's API as model="saaras:v4". Sarvam reports state-of-the-art accu…
Weco AI CEO Zhengyao Jiang said in a Machine Learning Street Talk feature that the company's AID system, which applies recursive self-improvement to agentic scaffolding rather than neural weights, produced seven accepted…
A developer built an order risk detector on Laya, an open-source non-autoregressive decision model based on ModernBERT, to test whether a small specialized model can replace large generative LLMs for e-commerce fraud det…
Supersonic Labs, a small AI lab from Brazil, released Julia 1, a 144.3M-parameter open decision model that runs on a plain CPU and is published on Hugging Face under Apache 2.0. Julia 1 starts from JHU CLSP's 140M-parame…
A developer ran an exploratory study testing whether modern LLMs contain internal self-models of other LLMs, using GPT2-medium and Qwen3-4B on 12 post-cutoff news headlines. By having Qwen continue GPT2-generated prefixe…
Dhruva Karkada's paper on arXiv (2602.15029) showed that LLM embeddings of the 12 months project onto a circle with an approximately circulant Gram matrix, a result the author of the post bet a beer he could reproduce wi…