Key insights from Mark Zuckerberg's podcast with Dwarkesh Patel, where he talks about Meta's AGI plan, covering various topics like: Lllama 4, gaming of benchmarks, AI Code generation within Meta, Bottlenecks for superhuman intelligence, and Monetizing AGI.👉 Benchmark gaming.-.

Reducing test-time compute and scaling LLM Inference at Test-time for stateful AI systems, with Sleep-time compute. Sleep-time compute, allows models to “think” offline about contexts before queries are presented: by anticipating what queries users might ask and pre-computing

s1: simple Test-time Scaling approach to exceed OpenAI’s o1-preview performance. Test-time scaling is a language modeling approach that uses extra test-time compute to improve performance, as also recently been shown by OpenAI’s o1 model. This paper seeks simplest approach to

AgoraBench: a benchmark that systematically evaluates LLMs synthetic data generation capabilities through standardized settings and metrics. This paper proposes AGORABENCH, a benchmark for evaluating LMs’ data generation capabilities across nine settings, combining three domains

Newer Llama-3.3-70B-Instruct model released, with better benchmark scores and newer capabilities. 𝗞𝗲𝘆 𝗵𝗶𝗴𝗵𝗹𝗶𝗴𝗵𝘁𝘀:.- leverages latest advancements in post-training techniques including online preference optimization,.- outperforms Google’s Gemini 1.5 Pro, OpenAI’s

PerfCodeGen : training-free framework to improve performance of LLM Generated Code with Execution Feedback.This paper proposes PERFCODEGEN, a training-free framework that enhances the performance of LLM-generated code by incorporating feedback based on runtime during test case

ALMA: method for aligning LLMs with minimal human annotation by using base model inherent capabilities. LLM alignment typically require millions of human annotations or rely on external aligned models for synthetic data generation. To address it, this paper introduces ALMA:

HunyuanVideo: A Systematic Framework For Large Video Generative Models. In this paper, authors present HunyuanVideo, a novel open-source video foundation model that exhibits performance in video generation that is comparable to, if not superior to, leading closed-source models.

RedStone: data pipeline designed to create specialized large-scale datasets by leveraging the vast and diverse data from Common Crawl. This paper from Microsoft introduce REDSTONE, an innovative and scalable pipeline engineered to extract and process data from Common Crawl,

Building a Visual AI Agent for Video Search and Summarization with NVIDIA AI Blueprint. Building a visual AI agent capable of understanding long-form videos requires a combination of VLMs and LLMs ensembled together with datastores. Nvidia AI's blueprint provides a recipe for

Mixture-of-Experts(MoE) does not relieve the massive memory requirements of networks, which limits their practicality in LLMs. Current approaches remove entire layers of MoE to reduce memory, but the performance degradation is still notable. To address it, this paper propose

KV Cache compression with Inter-Layer Attention Similarity for efficient Long-Context LLM Inference. For longer context LLM Inference, existing methods, including selective token retention and window-based attention, improve efficiency but risk discarding important tokens needed

cDPO: contrastive DPO algorithm to identify critical tokens for enhancing LLM’s reasoning abilities. LLMs tend to produce positive outcomes when forced to decode other tokens instead of critical tokens. Based on this observation, this paper propose a novel approach — cDPO -

Critic-V: framework to enhance feedback quality in visual perception and reasoning processes of Vision-Language Models (VLMs).VLMs often generate inaccurate or irrelevant responses due to issues like hallucinated image understandings or unrefined reasoning paths. To address it,

Star Attention: block-sparse attention mechanism for efficient LLM inference over long sequences.Inference with Transformer-based Large Language Models (LLMs) on long sequences is both costly and slow due to the quadratic complexity of the self-attention mechanism. To address

Template-based Data Generation: approach for generating mathematical datasets using LLM generated meta-templates.This paper introduces Template-based Data Generation (TDG), a novel approach that leverages LLMs (GPT-4) to automatically generate parameterized meta-templates, which

SageAttention2: accurate 4 Bit Attention for Plug-and-play Inference Acceleration. Although quantization for linear layers has been widely used, its application to accelerate the attention process remains limited. To further enhance the efficiency of attention computation done by

Marco-o1: open reasoning models powered by Chain-of-Thought (CoT) fine-tuning, Monte Carlo Tree Search (MCTS), reflection mechanisms. Marco-o1 not only focuses on disciplines with standard answers, such as mathematics, physics, and coding—which are well-suited for reinforcement