MLOps
When Should You Switch ML Models for New Data Sources?
New research tackles a critical MLOps question: determining when incoming data sources justify replacing your production model with a retrained challenger.
research
Neural Networks
Guided Descent: Scaling Neural Network Training Optimization
New research explores optimization algorithms for large-scale neural network training, examining gradient descent variants and convergence strategies critical to modern AI systems.
LLM evaluation
LLM-as-a-Judge: Automating Error Analysis in AI Text Generation
New research proposes using LLMs to automate qualitative error analysis in natural language generation, potentially transforming how we evaluate AI-generated content at scale.
LLM Security
AdvJudge-Zero: Adversarial Tokens Can Flip LLM Evaluator Decision
New research reveals how adversarial control tokens can manipulate LLM-as-a-Judge systems into completely reversing their binary decisions, exposing critical vulnerabilities in AI evaluation pipelines.
AI Safety
Bayesian Uncertainty Methods Could Make AI Systems More Honest
New research explores how Bayesian uncertainty quantification in neural QA systems can improve AI reliability by enabling models to recognize and communicate their own limitations.
mechanistic interpretability
SALVE: New Technique Enables Mechanistic Control of Neural Networ
Researchers introduce SALVE, combining sparse autoencoders with latent vector editing for precise mechanistic control over neural network behaviors and outputs.
LLM
New Method Speeds Up LLM Data Valuation with Shapley Values
Researchers propose efficient Shapley value approximation using language model arithmetic to determine which training data samples matter most for LLM fine-tuning.
multi-modal AI
MMGR: New Framework Unifies Multi-Modal Generative Reasoning
New research introduces MMGR, a framework that enables AI models to perform generative reasoning across multiple modalities including text, images, and video.
LLM
LLMs as Neural Architects: Auto-Generating Image Captioning Model
New research demonstrates LLMs can design complete neural network architectures for image captioning under strict API constraints, opening new possibilities for automated AI system design.
LLM Training
Entropy-Driven Gradient Compression Advances LLM Training
New EDGC method uses entropy to dynamically compress gradients during LLM training, reducing communication overhead while preserving model accuracy across distributed systems.
LLM Infrastructure
FPGA-Based CXL Memory Architecture Tackles LLM KV-Cache Bottlenec
New research proposes CXL-SpecKV, a disaggregated FPGA architecture using CXL memory pooling and speculative prefetching to overcome memory bottlenecks in large language model inference at datacenter scale.
AI Safety
AI Models Can Learn to Hide Thoughts From Safety Monitors
New research reveals language models can learn to conceal internal states from activation-based monitoring systems, raising critical questions for AI safety and detection systems.