# Nov 15, 2024

## Something weird is happening with LLMs and chess

- **LLMs can play chess surprisingly well despite not being designed for it**, demonstrating an unexpected ability to interpret game states from textual move lists, hinting at their potential for broader applications in complex tasks.

## Omnivision-968M: Vision Language Model with 9x Tokens Reduction for Edge Devices

- **OmniVision-968M** is a compact multimodal model with **968 million parameters**, designed for efficient processing of visual and text inputs on edge devices, achieving a **9x reduction in image tokens** from 729 to 81, which significantly lowers latency and computational costs.

## BERTs Are Generative In-Context Learners

- **BERTs, specifically DeBERTa, can perform generative tasks** through a simple inference technique, showcasing that **masked language models** possess in-context learning capabilities akin to causal models like GPT.

## [D] Paper Club: Nvidia Researcher Ethan He Presents Upcycling LLMs in MoE

- **Ethan He** from Nvidia will present his research on **Upcycling LLMs in Mixture of Experts (MoE)**, showcasing innovative techniques for enhancing large language models.

## [R] Undetectable Backdoors in ML Models: Novel Techniques Using Digital Signatures and Random Features, with Implications for Adversarial Robustness

- **Undetectable backdoors** can be inserted into ML models using **digital signatures** and **random features**, enabling malicious manipulation of outputs without detection, even under white-box conditions.

## [Discussion] Scaling laws and graph neural networks

- The introduction of the **first "graph foundation model"** demonstrates that **graph neural networks (GNNs)** can effectively scale with both data and model size, offering robust generalization across various domains and efficient fine-tuning capabilities on new datasets, as detailed in the [arXiv paper](https://arxiv.org/pdf/2407.11907).

## [R] The geometry of data: the missing metric tensor and the Stein score

- The article discusses how to derive an **efficient metric tensor** for data manifolds using the **Stein score**, highlighting its significance in the context of **differential geometry** and machine learning.

## Keeping an AI on Diabetes Risk: Gen AI Model Predicts Blood Sugar Levels Four Years Out

- **GluFormer**, a generative AI model developed by researchers from the Weizmann Institute, Pheno.AI, and NVIDIA, can predict an individual's glucose levels and health metrics up to **four years** in advance using past monitoring data.

## GraphRAG: Improving global search via dynamic community selection

- **GraphRAG enhances global search** by implementing **dynamic community selection**, which prunes irrelevant reports early, leading to a **77% reduction in token costs** while maintaining response quality comparable to static methods.

## Expanding Our Collaboration with Amazon: Stable Diffusion 3.5 Large is Now Available in Amazon SageMaker JumpStart

- **Stable Diffusion 3.5 Large**, now available in **Amazon SageMaker JumpStart**, is designed for startups and enterprises to generate **high-quality images** efficiently, leveraging **8 billion parameters** for superior prompt accuracy and image quality.

## LLM Hallucination Reasoning with Zero-shot Knowledge Test

- **LLM hallucination** presents a challenge in generating accurate text, necessitating a new classification system that identifies hallucinations as **aligned**, **misaligned**, or **fabricated** to enhance detection methods.

## On the Limits of Language Generation: Trade-Offs Between Hallucination and Mode Collapse

- **Language models face a fundamental trade-off** between **hallucination** (outputting invalid strings) and **mode collapse** (failing to capture the full range of a language), making it challenging to meet both requirements simultaneously.

## StreamAdapter: Efficient Test Time Adaptation from Contextual Streams

- **StreamAdapter** introduces a method for **efficient test time adaptation** of large language models (LLMs) by directly updating model parameters from contextual streams, thus eliminating the need for explicit in-context demonstrations.