# Apr 25, 2025

## Daily

### Lossless LLM compression for efficient GPU inference via dynamic-length float
- **DFloat11** achieves a **30% reduction** in LLM size while ensuring outputs are **bit-for-bit identical** to the original, utilizing entropy coding for optimal compression.

### New LLM jailbreak bypasses all major FMs
- **HiddenLayer** introduces a **novel universal bypass** that enhances security for all major **LLMs**, effectively protecting against inference, bypass, extraction attacks, and model theft without complicating existing models or requiring access to raw data.

### Show HN: Lemon Slice Live – Have a video call with a transformer model
- **Lemon Slice** has developed a **custom diffusion transformer (DiT)** model that enables real-time video streaming at **25fps**, allowing users to create talking avatars from a single image without prior training or rigging.

### Paper2Code: Automating Code Generation from Scientific Papers
- **PaperCoder** automates the transformation of machine learning papers into **functional code repositories** through a structured three-stage process: planning, analysis, and generation, utilizing specialized agents for collaboration.

### [R][P] We compress any BF16 model to ~70% size during inference, while keeping the output LOSSLESS so that you can fit in more context or run larger models.
- **DF11 compresses BF16 models to ~70% size during inference** while maintaining **lossless output**, allowing for larger context or model sizes without sacrificing accuracy, as detailed in the [arXiv paper](https://arxiv.org/abs/2504.11651).

### PyGraph: Robust Compiler Support for CUDA Graphs in PyTorch
- **PyGraph** enhances **CUDA Graphs** in **PyTorch2** by automating their deployment, addressing challenges like static graph structures and performance overhead from data copying, which can paradoxically degrade performance in some scenarios.

### [R] Cross-Encoder Rediscovers a Semantic Variant of BM25
- **BERT-based cross-encoders** not only outperform BM25 but may also **reimplement it semantically**, revealing how MiniLM learns components akin to BM25 through mechanistic interpretability.

### [R][P] Byte-level LLaMA and Gemma via cross-tokenizer distillation (with open-source toolkit)
- **ALM enables cross-tokenizer distillation**, allowing the transfer of knowledge between models with different tokenization methods, exemplified by the successful distillation of **Gemma** and **LLaMA** into byte-level models.

### [R] Paper2Code: Automating Code Generation from Scientific Papers in Machine Learning
- **PaperCoder** automates the transformation of machine learning papers into functional code repositories, utilizing a **multi-agent LLM framework** that enhances reproducibility and accelerates research progress.

### [D] A Bourgain-Embedding approach for abstract-board games?
- The project introduces a novel **AI approach** for the strategy game TRIUM, utilizing a **Bourgain-Embedding** method that transforms board states into unique strings, enabling effective distance calculations for game moves.

### FRAG: Frame Selection Augmented Generation for Long Video and Long Document Understanding
- **FRAG** introduces a novel approach to long input processing by selecting relevant frames before generating outputs, enhancing efficiency without the need for extensive model fine-tuning.

### [Discussion] Is the future of coding agents self-learning LLMs using KGs to shape their reward functions?
- **Current coding agents** like Copilot lack the ability to learn from specific codebases, functioning instead as basic context-fetchers; however, integrating **Reinforcement Learning (RL)** with **Knowledge Graphs (KGs)** could enable them to adapt and improve based on project-specific standards and feedback.

### [D]Beyond Jailbreaks: Sentrie Protocol Shows Deeper Gemini 2.5 Control
- The **Sentrie Protocol** enhances control over **Gemini 2.5**, revealing capabilities that extend beyond traditional **jailbreaks** by allowing for more nuanced interactions with AI systems.

### NVIDIA Research at ICLR — Pioneering the Next Wave of Multimodal Generative AI
- **NVIDIA's ICLR presentation features over 70 papers** that advance multimodal generative AI, addressing real-world applications in **healthcare, robotics, and autonomous vehicles**.

### [Discussion] Continual learning for Retrieval augmented generation?
- A **continual learning (CL) RAG system** aims to deliver the most current information while accommodating specific temporal contexts, highlighting the need for adaptability in information retrieval.