Persona vectors: Monitoring and controlling character traits in language models
Persona vectors are patterns of neural activity that help monitor and control the character traits of language models, enabling developers to understand and mitigate undesirable personality shifts during training and deployment.
Qwen-Image: Crafting with native text rendering
Qwen-Image is a 20B MMDiT image foundation model that excels in complex text rendering and precise image editing, outperforming existing models in both areas, particularly in rendering Chinese text with high fidelity.
Tokens are getting more expensive
Token costs are rising, contradicting the expectation that cheaper models would improve consumer AI margins, as evidenced by companies like Windsurf and Claude Code facing severe financial strain despite lower LLM inference costs.
New quantum state of matter found at interface of exotic materials
A new quantum state called quantum liquid crystal has been discovered at the interface of a Weyl semimetal and spin ice, revealing unique electronic properties that could lead to advanced technologies.
My Ideal Array Language
An ideal array language must incorporateuser-extensible rank polymorphism to allow for flexible kernel writing, enhancing the language's capability beyond existing frameworks like Numpy and JAX.
Photonic SNN chip reaches 1 GHz and supports on-chip learning
The GHz spiking neuromorphic photonic chip achieves a breakthrough in brain-inspired computing, integrating in situ learning and event-driven dynamics on a silicon platform, addressing critical gaps in emulating brain functions. Link to article
Content-Aware Spaced Repetition
Content-aware memory models enhance spaced repetition systems (SRS) by integrating the semantic meaning of flashcards, allowing for improved retention through contextual relationships between related cards, rather than treating each card in isolation.
Software needs an "independent auditor"
AI-generated code necessitates an independent auditor to ensure quality, as reliance on the same entity for both generation and review can lead to compromised scrutiny and oversight.
Fine-tuned small LLMs can beat large ones with programmatic data curation
Fine-tuning small LLMs on programmatically curated data can outperform large models, achieving up to 30x cost reduction and 4x faster response times across various applications, including data extraction and navigation tasks.
[R] Integrative approach for early detection of Parkinson’s disease and atypical Parkinsonian syndromes leveraging hemodynamic parameters, motion data & advanced AI models
A multimodal hardware-based wearable integrated with a novel machine learning framework shows promise for early and accurate diagnosis of Parkinson’s disease, utilizing time-series data from low-cost sensors processed on microcontrollers.
Rethinking how we measure AI intelligence
Kaggle Game Arena is an innovative, open-source platform that enables head-to-head evaluation of AI models through strategic games, providing a dynamic measure of their capabilities.
CoRGI: Verified Chain-of-Thought Reasoning with Visual Grounding
CoRGI (Chain of Reasoning with Grounded Insights) enhances vision-language models (VLMs) by integrating a visual verification mechanism, addressing the issue of hallucinations in reasoning outputs.
Accurate and Consistent Graph Model Generation from Text with Large Language Models
LLM-based graph model generation often yields models with syntax violations, constraint inconsistencies, and inaccuracies, which can hinder their practical application in software engineering.