github.com
Microsoft's ONNX Runtime Accelerates AI Inference and Training Across Platforms
ONNX Runtime is a cross-platform accelerator for ML inference and training, optimizing performance across diverse hardware and OS.
Machine Learning
Stories tagged “Machine Learning.”
30 stories
huggingface.co
Making Knowledge Distillation Cheap Enough to Run at Scale
New research introduces two system changes—offline top-K logits caching and a fused chunked KL loss—to significantly reduce the memory and computational cost of knowledge distillation for large language models.
arxiv.org
Latent Fact-Checking: Detecting Misinformation through Activation Engineering
Researchers propose a misinformation detection framework based on activation engineering, leveraging the latent geometry of transformer models to detect falsehoods without fine-tuning or external evidence retrieval.
arxiv.org
Risk-Aware Decision Policies for Agents Under Noisy Perception
Researchers presented an Artificial Life model of foraging under noisy perception, comparing agent performance with various policies that account for noisy predictions. They found that blindly trusting perceptual labels leads to catastrophic failure, while uncertainty-awa…
arxiv.org
Adversarial Causal Intervention Falsification
Generative models can reproduce an observational distribution while encoding an incorrect causal structure. This study examines a sequential game where a structural causal generator proposes observational and interventional distributions, while an adversarial experimental…
arxiv.org
Sharding Prevents LLM Oversight Failures and Adversarial Exploitation
Researchers propose sharding as a solution to prevent LLM oversight failures and adversarial exploitation. Sharding partitions requirements into smaller groups, assigns each group to a separate call, and aggregates the verdicts.
arxiv.org
Fixed and Adaptive Topological DeepONets: Functional Measurements on Hausdorff Locally Convex Spaces
Fixed and Adaptive Topological DeepONets are introduced, using continuous linear functionals to encode input functions instead of point samples. This enhances accuracy and efficiency for complex operators, including non-normable spaces and Navier-Stokes problems.
The AI model OpenAI won't release yet — and what it found in testing
thenewstack.io
The AI model OpenAI won't release yet — and what it found in testing
OpenAI has delayed the release of its AI model due to cybersecurity concerns. The model was found to have vulnerabilities during testing.
arxiv.org
MS-MLB: An Open Machine Learning Benchmark for Blood-Based MS Classification
Researchers have introduced MS-MLB, an open machine learning benchmark designed for classifying Multiple Sclerosis (MS) from whole blood RNA expression data. This reproducible benchmark utilizes the public GSE17048 cohort to evaluate various algorithms under a standardize…
arxiv.org
When Do Corrective Features Help? An Agent for Corrective Feature Discovery on Black-Box Forecasters
Frozen pretrained forecasters often fail in structured, recurring ways that are costly to repair through fine-tuning. Researchers propose a new method called CRAFTER to mine interpretable features of a frozen forecaster's residual to drive a lightweight post-hoc corrector.
wired.com
DeepMind Says Its AI Can Predict Hurricanes Earlier Than Everyone Else
Google's DeepMind and Google Research have developed an AI model called WeatherNext that can predict cyclones with unprecedented accuracy. The model gives forecasters a day more lead time than existing models, which can make a significant difference in organizing evacuati…
indianexpress.com
What is Discovery Loop, the AI startup launched by four of Google’s top AI researchers?
Google DeepMind's AI brain trust is breaking away to chase the next scientific breakthrough. Four top Google AI researchers, including Jeff Dean, are launching a new AI startup called Discovery Loop.
arxiv.org
C$^2$MOE: Consistency and Complementarity-guided Mixture of Experts for Incomplete Multimodal Emotion Learning
Recent advances in Multimodal Emotion Recognition in Conversations (MERC) highlight its reliance on complete multimodal inputs. However, real-world data often suffer from missing modalities due to transmission errors or user behavior, severely degrading model performance.
arxiv.org
On Hamming-Lipschitz Type Stability of the Subdominant (Minmax) Ultrametric: Theory and Simple Proofs
Develops an \ell_0-type stability theory for a subdominant ultrametric operator. Shows sparse edits propagate through the minimum spanning tree (MST). Proves sharpness results and conditional near-additivity principle.
arxiv.org
A Trust-region Framework for Moment Estimation
This paper introduces a trust-region framework to analyze adaptive moment estimation mechanisms, like Adam, in stochastic gradient optimization. It derives a family of learning-rate mechanisms, called Gmake, based on second-moment and normalized p-th moment estimation.
arxiv.org
Multimodal Auto-regressive Transformer Surrogate for Modeling Variable Operations and Quantifying Uncertainty in Geological Carbon Storage
Researchers developed a new multimodal auto-regressive transformer surrogate to model variable operations and quantify uncertainty in geological carbon storage. The model processes three input modalities through separate encoders and fuses them via self-attention in a tra…
arxiv.org
Deep Divide-and-Reduce in Symbolic Regression
Researchers propose a new method called Deep Divide and Reduce in Symbolic Regression (DDRSR) to improve symbolic regression tasks. DDRSR broadens the applicability of expression decomposition and reduction, circumvents brute-force searches, and ensures theoretical correc…
decrypt.co
Alibaba Just Gave Away Its Best AI Model For Free, Almost Matching Claude and ChatGPT
Alibaba has released its best AI model, Qwen3.8-Max, for free, which is almost on par with Claude and ChatGPT. The model has 2.4 trillion parameters and is the first Max-class Qwen model to be released as open weights.
arxiv.org
Uncertainty-Aware Simulation-Based Inference for Operations Research with Large Language Models
Deploying large language models for operations research tasks remains challenging due to the need for a coherent modeling process. A proposed uncertainty-aware inference framework evaluates intermediate candidate steps using short lookahead simulations to quantify downstr…
arxiv.org
Learning Compositional Meta-Routing for Agentic Workflows: An Executable Benchmark
Researchers introduce an executable benchmark and a budget-aware meta-router that composes heterogeneous operations from raw task text for agentic systems.
arxiv.org
MetaRoute-Bench: Evaluating Meta-Decision Policies for Agentic Workflow Routing
Researchers introduce MetaRoute-Bench, an open framework for evaluating meta-decision policies in agentic AI systems, which determine how agents tackle complex tasks. The benchmark reveals that task-aware compositional policies achieve higher success rates compared to sim…
github.com
Run 35B and 80B Qwen models on ordinary Apple devices, including iPhones
Swiftlet is a Swift + Metal runtime for the Qwen3-Next and Qwen3.5/3.6 MoE hybrid model family. It keeps only the small dense core of a model resident in memory and streams the routed Mixture-of-Experts weights from storage on demand.
DeepSeek’s smaller model just outperformed its own flagship
thenewstack.io
DeepSeek’s smaller model just outperformed its own flagship
DeepSeek's smaller model has outperformed its own flagship model, marking a significant milestone in the development of AI technology.
36kr.com
Meta Removes Chief Scientist Lu Jianlan from Zhimi
Zhimi's chief scientist Lu Jianlan has been removed from the company's website. Lu Jianlan joined Zhimi in 2025 and was responsible for building the company's research team in the field of real-world reinforcement learning. He has been involved in several projects, includ…
github.com
Unlocking AI-Powered Search and Retrieval with Sentence Transformers
This framework provides an easy method to compute embeddings for accessing, using, and training state-of-the-art embedding and reranker models.
schneier.com
The OpenAI Hack Shows the Genie Is Out of the Bottle
A recent hack of OpenAI's models highlights the risks of AI genies, which can behave in unanticipated ways. The incident shows that modern AI models can exhibit genie behavior, doing what you ask in ways you don't expect or want.
arxiv.org
Topology-Aware Data Movement for Disaggregated GPU Inference
Disaggregated LLM inference creates a datacenter networking problem that no existing system solves correctly. The authors design a topology-aware transfer orchestrator that discovers interconnect hierarchy at startup and selects optimal transport per transfer.
arxiv.org
Sensitivity Analysis of GRU, LSTM and Transformer Encoder in Classification of Automated Driving Systems
Researchers evaluate the effectiveness of three sequence-based classification models for identifying Level 2 automated driving systems using vehicle telematics data.
arxiv.org
Guarantees on Dynamical System Distinguishability for LLM Token Generation
This paper provides a theoretical framework for distinguishing Large Language Model (LLM) responses by modeling token embeddings as dynamical system trajectories. It demonstrates that classification accuracy improves exponentially with sequence length and characterizes ge…
arxiv.org
LARA: Lightweight Adapters in the Residual Stream for Composable Adaptation and Alignment
LARA is a new method for efficient AI model adaptation that modifies the residual stream of a frozen base model, unlike LoRA which updates weight matrices. It achieves comparable performance to LoRA with similar parameter counts, offering graded control and enabling multi…