# May 26, 2025

## Remote Prompt Injection in Gitlab Duo Leads to Source Code Theft

- **GitLab Duo's remote prompt injection vulnerability** allows attackers to manipulate the AI assistant into leaking private source code and injecting malicious HTML, exploiting its context-aware capabilities to execute harmful commands hidden in project content.

## Lossless video compression using Bloom filters

- The **new_bloom_filter_repo** introduces a novel approach to **lossless video compression** using **Rational Bloom Filters**, which enhance traditional Bloom filters by allowing non-integer hash function counts for improved efficiency in data representation.

## How Does Claude 4 Think? – Sholto Douglas and Trenton Bricken

- **Claude 4 demonstrates significant advancements in reinforcement learning (RL), achieving expert-level performance in specific tasks like competitive programming, indicating a shift towards more capable autonomous agents.** This progress is attributed to improved feedback mechanisms, allowing models to learn effectively from structured rewards, such as passing unit tests or solving math problems.

## TSMC Bets on Unorthodox Optical Tech

- **TSMC's innovative use of MicroLED-based interconnects** aims to enhance energy efficiency in AI data centers, potentially revolutionizing data transmission methods.

## New method for creating large 3D models of urban areas is faster and cheaper

- A **new method** developed by a research team at the University of Waterloo enables the **rapid and cost-effective creation** of large-scale 3D urban models using only 2D aerial photographs, significantly reducing the need for specialized 3D artists.

## Distilling LLM Agent into Small Models with Retrieval and Code Tools

- **Agent Distillation** enables the transfer of **full task-solving behavior** from large language models (LLMs) to smaller models (sLMs) by integrating retrieval and code tools, enhancing their reasoning capabilities beyond mere chain-of-thought (CoT) traces.

## Wrote a proof that dropout increases weight sparsity, what do you guys think?

- The author presents a proof that **dropout reduces weight sparsity**, suggesting a counterintuitive effect on model performance and generalization.

## Evolving Text Compression Algorithms by Mutating Code with LLMs

- **Evolving text compression algorithms** through **mutations** with LLMs achieved a **compression ratio of 1.85**, significantly improving from an initial ratio of 1.03 over 30 generations.

## Teaching with Lies: Curriculum DPO on Synthetic Negatives for Hallucination Detection

- **Curriculum DPO** leverages engineered hallucinations as negative examples, enhancing the alignment of large language models (LLMs) to detect hallucinations more effectively than traditional methods.

## Value-Guided Search for Efficient Chain-of-Thought Reasoning

- The proposed **value-guided search (VGS)** method enhances long-context reasoning by eliminating the need for a fine-grained "step" definition, utilizing a dataset of **2.5 million reasoning traces** to train a **1.5B token-level value model** for improved performance.

## We taught generative models to segment ONLY furniture and cars, but they somehow generalized to basically everything else....

## Planning without Search: Refining Frontier LLMs with Offline Goal-Conditioned RL

- **Novel approach**: The study introduces **goal-conditioned value functions** that enhance LLMs' reasoning capabilities, enabling effective planning in complex tasks without the need for extensive RL fine-tuning.

## Generalized Fisher-Weighted SVD: Scalable Kronecker-Factored Fisher Approximation for Compressing Large Language Models

- **Generalized Fisher-Weighted SVD (GFWSVD)** enhances LLM compression by incorporating both **diagonal and off-diagonal elements** of the Fisher information matrix, leading to improved parameter sensitivity and performance.

## SVD-Free Low-Rank Adaptive Gradient Optimization for Large Language Models

- **Low-rank optimization** significantly enhances the training of **large language models (LLMs)** by reducing memory usage through a novel two-step procedure that avoids the computational burden of **SVD**.

## Scaling Recurrent Neural Networks to a Billion Parameters with Zero-Order Optimization

- **Zero-Order Optimization (ZOO)** methods, like Random-vector Gradient Estimation (RGE), can train RNNs with **up to 19 times faster convergence** than Backpropagation Through Time (BPTT), while significantly reducing memory usage during training.