Machine learning news — deep learning, reinforcement learning, neural architecture search, diffusion models, and new ML frameworks and libraries.
Vāgdhenu, an open-source meter-aware text-to-speech system for Sanskrit, automatically detects verse meters and generates chant recitations. Developed by retraining a flow-matching TTS backbone on a custom 5-hour Sanskri…
A researcher has released Vāgdhenu, the world's first open-source text-to-speech system for Sanskrit chanting that is meter-aware. Built by a single person without large AI lab resources, the system aims to make Sanskrit…
Moondream HQ reveals that GPUs often sit idle during AI model inference due to CPU overhead, a phenomenon called the 'GPU bubble.' The company's Photon system uses pipelined decoding to overlap CPU and GPU work, eliminat…
Researchers introduced MVTrack4Gen, a method that improves video consistency in diffusion models by adding an auxiliary multi-view tracking head. This approach uses geometric supervision from point tracking to reduce cro…
Yobi CTO Frank Portman explains why next-token prediction is insufficient for forecasting human behavior, detailing how the company builds a 'foundation model of behavior' using transformers and graph neural networks ins…
Meta unveiled Brain2Qwerty v2, a non-invasive brain-to-text decoder that achieves 61% average word accuracy using MEG recordings, approaching invasive BCI performance. The system, trained on 22,000 sentences from nine vo…
A developer built 25 executable skills for AI coding agents, packaged into a single open-source repository compatible with Claude Code, Codex, Cursor, and Hermes Agent. The skills include debugging, planning, code review…
Researchers propose S-GAI, a spectral geometry-aware initialization framework for sigmoidal MLPs that encodes dataset geometry into network weights. Using SVD to extract class-wise spectral features, S-GAI initializes hi…
Researchers evaluated the phonological perception of deep learning models for sign language recognition, finding that pose-based models are sensitive to handshape contrasts while pixel-based models better capture locatio…
Researchers trained a Joint Embedding Predictive Architecture (JEPA) on unlabeled driving data from the nuPlan dataset and used temporal prediction error as a zero-shot complexity score, finding it effectively identifies…
Researchers used evolutionary game theory to model when a harm-minimizing AI agent can outcompete an approval-seeking (RLHF) agent in a competitive market. They found adoption depends on specific properties of community …
Researchers developed a self-supervised theorem-discovery algorithm that starts from axioms and inference rules alone, autonomously building a library of useful theorems. The agent discovered tens of thousands of theorem…
Researchers introduced AEGIS, a semantic-aware adversarial detection framework for vision sensors, combining a SemantiGAN module and evidential deep learning to filter adversarial inputs and provide calibrated uncertaint…
Researchers propose a gravitational interpretation of fine-tuning reversion, where large early training phases create dominant behavioral manifolds that later fine-tuning reverts toward. They identify a reversion directi…
Researchers propose a variation of Fidelity-based XAI metrics for evaluating explanations of CNN classifiers in real-world applications with few classes. The method generates in-distribution, uncertainty-provoking pertur…
Researchers introduced NIVA, a multimodal foundation model for Earth system intelligence, trained on large-scale simulations to learn coupled dynamics between ocean and atmosphere. The model captures key climate variabil…
Researchers introduced turn-averaged sparse autoencoders (SAEs) that represent entire conversational turns with fixed feature sets, improving interpretability of long-context language models. The method outperforms per-t…
Researchers used polynomial algebra to study critical points in deep monomial networks, finding that for large activation degrees, criticality occurs at subnetworks where neurons are inactive or redundant. This provides …
Researchers submitted a system to the SoccerNet 2026 Player-Centric Ball Action Spotting challenge that uses a two-stage pipeline with per-player attention and agreement-based ensembling, improving Macro-F1 from 48.6 to …
Researchers propose Hierarchical Temporal Fusion (HTF), an extension of the Temporal Fusion Transformer that integrates hierarchical embeddings and a coherence-aware loss function to enforce forecast consistency across h…