Neural intel Pod

Updates from the PyTorch blog

This research introduces a novel, deep-learning-enabled approach for designing complex multi-port radio-frequency and sub-terahertz circuits and electromagnetic structures.  The methodology overcomes limitations of traditional design methods by using deep learning models to achieve rapid synthesis of arbitrary-shaped structures with desired radiative and scattering properties.  The key innovation...

This research addresses the challenge of improving monocular depth estimation (MDE) using unlabeled data through a novel distillation framework.   The core innovation is Cross-Context Distillation, which combines local and global depth cues to enhance pseudo-label quality and model accuracy.   A multi-teacher distillation approach further leverages complementary strengths of different depth estima...

The research introduces MHA2MLA, a novel fine-tuning framework designed to adapt existing MHA-based language models to the more efficient MLA architecture.   MLA achieves economical inference by compressing the key-value (KV) cache.   MHA2MLA employs partial RoPE and low-rank approximation techniques to minimize performance degradation during the adaptation.   Experiments demonstrate that MHA2MLA,...

This paper introduces SWE-RL, a reinforcement learning (RL) method to improve large language models (LLMs) for software engineering tasks using software evolution data and rule-based rewards.  The approach trains LLMs to autonomously learn from open-source software's lifecycle, including code snapshots, changes, and events.  The resulting model, Llama3-SWE-RL-70B, achieves state-of-the-art per...

Amazon Science highlights research and career opportunities across various fields, with a focus on quantum computing.  It showcases a specific project where SAT solving is used to optimize quantum circuit mapping, leading to significant speed improvements. The method aims to reduce the number of swap gates in quantum circuits, a crucial factor in quantum computation efficiency.  The company is act...

LM Studio has released software development kits (SDKs) for Python and TypeScript, enabling developers to integrate LM Studio's AI capabilities into their own applications.  These MIT-licensed SDKs, lmstudio-python and lmstudio-js, offer core APIs for chat, text completion, embeddings, and a novel agentic tool use API called .act().  The .act() API facilitates agent-oriented programming, allowing...

AI Agents, DeFi, and Decentralized Virtual Ecosystems

LLMs in The Chameleon Game_ Strategic Information Dynamics

These documents detail the AN/UYK-1, a compact and adaptable shipboard computer notable for its "Stored Logic" architecture.  This innovative design allows the computer's logical organization to be defined by software rather than fixed hardware, enabling versatility across various Navy applications. The AN/UYK-1 is engineered for ruggedness, NTDS compatibility, and ease of use, even by non-experts...

Self Debugging code generation models?

This research investigates the strategic capabilities of large language models (LLMs) in scenarios requiring information control.  It introduces a game called "The Chameleon," where LLMs must conceal, reveal, and infer information to succeed as either a chameleon or a non-chameleon player.  The study combines theoretical analysis with empirical results from LLMs like GPT-4 and Gemini 1.5...

This research explores enhancing GameFi platforms by integrating advanced AI agents and decentralized finance (DeFi) mechanisms.  It addresses limitations in current GameFi applications, such as simplistic AI interactions and insufficient community or monetization features.  The authors propose using large language models like GPT-4 to create embodied AI agents that proactively interact with playe...

Amazon Science is focused on improving code generation through debugging.  They use large language models (LLMs) to both generate and debug code, leveraging techniques like supervised fine-tuning and reinforcement learning to train the models.  A key element involves creating synthetic debugging data to overcome the scarcity of real-world examples.  This approach shows significant improvement in c...

The PyTorch blog post focuses on accelerating generative AI models, specifically Segment Anything 2 (SAM2), using native PyTorch.  It details techniques like torch.compile and torch.export for optimized, low-latency inference.  The authors achieved significant performance improvements (up to 13x) by employing ahead-of-time compilation, reduced precision, batched prompts, and GPU preprocessing.  Th...

Visuo-Haptic 6D Object Pose Tracking

Force-Attending Curriculum Training for Contact-Rich Policy Learning

LLMs learn to lie, cheat and KILL in Among Us

Sharp Monocular Metric Depth Estimation

Benchmarking Chain-of-Thought in Large Multimodal Models

Efficient GRPO for Long-Context Reasoning Models

Tunable Length Control for Chain-of-Thought Reasoning

Implementing a Transformer model from scratch is a great way to understand the architecture and its components in depth.

Reflection and Refraction are two fundamental phenomena of light that describe how light interacts with surfaces and different media.

MixGCN: Scalable Graph Convolutional Network Training by Mixture of Parallelism and Mixture of Accelerators is a novel framework designed to address the challenges of training Graph Convolutional Networks (GCNs) on large-scale graphs. GCNs are widely used for graph-based learning tasks, but their scalability is often hindered by memory limitations, communication bottlenecks, and the computational...