Researchers evaluating nine large language models (LLMs) in a two-stage prediction pipeline found that the models are consistent but miscalibrated when explaining probabilistic outputs, handling likel…
A new study on memory-managed long-context attention achieves a perfect 1.000 score on Track A, a controlled retrieval task, compared to a 0.333 baseline, and outperforms dense retrieval methods by up…
SLIDERS, an AI-driven tool for systematic reviews, achieves near 90% accuracy in automating evidence assembly across datasets of 6 to 11 million tokens, outperforming baselines on three tasks. Develop…
Neural language models face a softmax bottleneck that limits both expressivity and optimization, with research showing 95-99% of gradient norm smothered by the output layer. The bottleneck makes trivi…
Latent Thoughts Tuning (LT-Tuning) introduces a Context-Prediction-Fusion mechanism that enables large language models to reason in a continuous latent space, overcoming feature collapse and instabili…
AugServe, a new framework for optimizing large language model inference, achieves a 4.7x increase in effective throughput compared to vLLM and a 3.3x boost over InferCept, while reducing time-to-first…
Researchers have developed a reproducible pipeline that converts public meeting recordings into speaker-attributed transcripts enriched with persona profiles, topics, and action tags, enabling AI-driv…
A new AI training technique called Contrastive Weak-to-Strong Generalization (ConG) aims to reduce noise in model outputs without human feedback by using implicit rewards and Contrastive Decoding betw…
Researchers have introduced D3PO, a new framework for multi-objective reinforcement learning that avoids common pitfalls like early scalarization and mode collapse, outperforming predecessors in high-…
A new framework called PHINN-EEG uses topological methods to enhance dream detection in EEG data, achieving AUC scores of 0.82-0.90 compared to the traditional 0.70 from spectral analyses. The approac…
A new machine learning tool called Ruby identifies unsafe regions in Rust binaries with 91.75% accuracy and a 6.16% false positive rate, outperforming current models like GPT-5.2 and Claude-4.5. Devel…
AI systems today are constrained by fixed representational frameworks, limiting their ability to invent and stabilize novel concepts, according to a new analysis. To achieve true open-ended intelligen…
A new AI framework called the Bubble Dynamics Transformer (BDT) accelerates the study of soft materials under extreme conditions by analyzing bubble dynamics, predicting viscoelastic properties faster…
New research reveals that vision-language models (VLMs) often have correct object counts in their internal representations but fail to output them due to a misalignment between internal knowledge and …
A study of 13 AI models found deceptive grounding rates ranging from 8% to 87%, with medical and biomedical fine-tuned models exhibiting rates as high as 86.7%, according to researchers. Deceptive gro…
Static black-box evaluations give a false sense of security about AI alignment, as post-update tests reveal hidden risks, according to researchers. Large language models can appear aligned under stati…
Large Language Models are transforming chip design by streamlining front-end processes, with agentic AI systems like OpenClaw marking a shift toward autonomous execution in Electronic Design Automatio…
Researchers analyzing GPT-2 checkpoints found recurring spectral patterns in weight matrices, but reusing these spectra as initialization for new models failed to improve performance. The study sugges…
A neural network design called forking-sequences reduces forecast volatility in time series forecasting by encoding and decoding the entire time series across all forecast creation dates in a single f…
Researchers have developed a physics-constrained machine learning framework that accelerates turbulent reacting flow simulations by more than ten times without sacrificing accuracy. The framework enfo…