A policy agent specifies local intent per action, and a critic evaluates outcomes and failure modes against that intent
What is new.
What it means for your work.
Discover recent technology and evidence. Browse older material in All notes.
Changes to start with
Newly added notes across different industries
AgentRoboCritic · Action-Level Analysis for Evaluating Robot Control Capabilities
TideCache: Transparent Cache Management and Layer-wise Streaming for Hybrid LLM Serving
Virus-like particles enable efficient gene editing in pigs and poultry models
New technology notes
Originals from the past 14 days, ordered by when added to Blackdeck.
A transparent cache pool and cache streamer efficiently manage GPU memory allocation and cache residency across layers.
An automated framework that revises games based on agent play experience to enhance mastery
A novel method efficiently computes feasible stochastic optimal solutions with bounded constraint violation
High-temperature distribution alignment leverages imperfect predictions to expand training data
Virus-like particles (VLPs) successfully deliver gene editing tools in porcine and avian cells and embryos
Integrated analysis of 27 genomic clusters and 50 hub genes reveals coordinated regulatory networks among heat shock-responsive genes
A virtual twin reproduces process variation, and machine learning evaluates multiple failure mechanisms simultaneously.
AI agents can explore multiple design options, integrate them, and iteratively improve designs without fixed design blocks.
A new methodology direction is emerging where AI agents integrate into early design stages of chip development.
By early stopping and appending high-confidence base-only categories, object coverage loss in preference-optimized models is restored without training
A retrieval-augmented refiner uses a shared knowledge base to correct segmentation errors across anatomies without per-anatomy training.
A clock mesh is placed on backside metal layers shared with the power delivery network, enabling efficient co-location and reduced skew.
A defect- and topology-aware chiplet matching framework maximizes functional SoC yield under real-world assembly constraints.
Proposes a method to learn visual task spaces based on visual state changes, especially for cleaning tasks