# Jun 16, 2025

### Nanonets-OCR-s
- **Nanonets-OCR-s** is a cutting-edge **image-to-markdown OCR model** that intelligently transforms documents into structured markdown, enhancing compatibility with **Large Language Models (LLMs)** for advanced processing.

### Meta's Llama 3.1
- **Meta's Llama 3.1 can reproduce 42% of _Harry Potter and the Sorcerer's Stone_, indicating significant memorization capabilities that raise concerns about copyright infringement.** This finding suggests that the model's training may have included extensive exposure to popular texts, leading to high rates of verbatim recall.

### First 2D, non-silicon computer
- **Penn State researchers have developed the world's first 2D, non-silicon computer**, utilizing atom-thin materials like molybdenum disulfide and tungsten diselenide to create a complementary metal-oxide semiconductor (CMOS) capable of simple operations, marking a significant shift in semiconductor technology.

### Salesforce study finds LLM agents flunk CRM and confidentiality tests
- **LLM agents** achieve only a **58% success rate** on single-step CRM tasks, dropping to **35%** for multi-step tasks, indicating significant limitations in their operational capabilities.

### Chemical knowledge and reasoning of large language models vs. chemist expertise
- **ChemBench** is a newly developed framework that evaluates the chemical knowledge and reasoning abilities of large language models (LLMs), revealing that top models can outperform expert chemists in specific tasks while still struggling with basic concepts and overconfident predictions.

### [R] Vision Transformers Don't Need Trained Registers
- **Vision Transformers** can effectively operate without **trained registers**, as demonstrated in recent research that explores artifacts in attention and feature maps, paralleling findings in **large language models**.

### Why Claude's Comment Paper Is a Poor Rebuttal
- **Apple's paper** reveals that **Large Reasoning Models (LRMs)** exhibit significant limitations in computation and reasoning, challenging their viability as a foundation for AGI, a sentiment echoed by experts like Subbaro Kambhampati and Yann LeCun.

### [P] Research Scientists + Engineers for Generative AI at NVIDIA
- NVIDIA is seeking **senior and principal research scientists** to advance **generative AI**, focusing on **training and deploying frontier-scale models** and optimizing architectures for enhanced AI performance.

### Breaking Quadratic Barriers: A Non-Attention LLM for Ultra-Long Context Horizons
- The proposed **non-attention architecture** for large language models (LLMs) can efficiently manage **context windows** of hundreds of thousands to millions of tokens, circumventing the **quadratic memory** issues of traditional Transformers.

### Quantum mechanics provide truly random numbers on demand
- **Quantum mechanics enables the generation of truly random numbers** through a novel service called CURBy, which utilizes a Bell test to ensure randomness that is certifiable and traceable.

### The Illusion of Thinking: A Reality Check on AI Reasoning
- **AI reasoning models exhibit a sudden collapse in performance** when faced with complex tasks, revealing a critical threshold beyond which they fail to execute knowledge reliably, even with ample computational resources.

### 🤗Groq on Hugging Face Inference Providers 🔥
- **Groq** is now an official **Inference Provider** on Hugging Face, enhancing serverless inference capabilities for a variety of models, including **Meta's LLama 4** and **Qwen's QWQ-32B**.

### EMLoC: Emulator-based Memory-efficient Fine-tuning with LoRA Correction
- **EMLoC** introduces a **memory-efficient fine-tuning framework** that allows users to fine-tune large models within the same memory budget as inference, utilizing a lightweight emulator and **LoRA Correction** for improved performance.

### Mitigating Hallucination Through Theory-Consistent Symmetric Multimodal Preference Optimization
- **Symmetric Multimodal Preference Optimization (SymMPO)** enhances **Direct Preference Optimization (DPO)** by implementing rigorous theoretical alignment and direct preference supervision, effectively reducing hallucination in Multimodal Large Language Models (MLLMs).

### LoRA-Gen: Specializing Large Language Model via Online LoRA Generation
- **LoRA-Gen** enhances **edge-side models** by generating **LoRA parameters** from a large cloud-side model, enabling **flexible specialization** without the need for specialized training.