NVIDIA releases Llama-Nemotron series of reasoning models, Zero to One: Detailed explanation of AI Agent design patterns

Today's Directory

1. NVIDIA releases Llama-Nemotron series of inference models, with performance exceeding DeepSeek-R1

2. Zero to One: Detailed explanation of AI Agent design pattern

3. Open source project: Build your second brain AI assistant, combining LLM, Agent and RAG technologies

4. RM-R1: An innovative approach to reward modeling as a reasoning process

5. Brain-computer interface breakthrough: Large brain language model for silent speech decoding

6. Ming-Lite-Uni: Advances in a unified architecture for natural multimodal interaction

1. NVIDIA releases Llama-Nemotron series of inference models, with performance exceeding DeepSeek-R1

NVIDIA has officially launched the Llama-Nemotron series of inference models , an open source family of heterogeneous inference models that provides excellent inference capabilities and efficiency, and is open to enterprise-level licenses.

The series includes three specifications:

(1) LN-Nano (8B)

(2) LN-Super (49B)

(3) LN-Ultra (253B)

Notably, LN-Ultra surpasses DeepSeek-R1 in performance, with higher inference throughput and better memory efficiency , and is able to run on a single 8xH100 node.

The innovation of Llama-Nemotron models is that they support dynamic reasoning switching , allowing users to switch between standard chat mode and reasoning mode through a simple system prompt "detailed thinking on/off" during the reasoning process.

The training of these models consists of five stages:

(1) Using Neural Architecture Search to Optimize Inference Efficiency

(2) Knowledge distillation and recovery training of continuous pre-training

(3) Supervised fine-tuning with a mixture of standard instruction data and inference traces from a strong teacher model

(4) Large-scale reinforcement learning on complex mathematics and STEM datasets

(5) Focus on the short-term alignment phase between instruction following and human preferences

Paper title: Llama-Nemotron: Efficient Reasoning Models

Paper link: https://arxiv.org/abs/2505.00949

2. Zero to One: Detailed explanation of AI Agent design pattern

This is a guide to common workflow and agent design patterns, including code snippets for the Gemini model.

The guide details seven core AI agent design patterns:

(1) Prompt Chaining : Chaining LLM calls sequentially, with the output of one serving as the input of the next

(2) Routing : Use LLM to classify input and direct it to the most appropriate specialized task, LLM, or tool

(3) ⏸️ Parallelization : Run multiple independent subtasks simultaneously and aggregate the results to increase speed or enhance quality

(4) Reflection : To achieve self-correction, the agent evaluates its own output according to the standard and iterates based on the feedback

(5) Tool Use : Enables LLM to interact with the outside world by calling external functions or APIs to obtain data or perform operations

(6) Planning : Let the central LLM dynamically decompose complex goals into multi-step plans and delegate execution to work agents

(7) Multi-Agent : Using multiple different agents with specific roles or expertise to achieve a common goal through coordinators or handoffs

Top tip: Start simple! Use workflows for well-defined tasks. If you need to accommodate dynamic problems, choose a proxy, but be aware of the cost/latency and implement robust tracking and error handling.

Article link: https://www.philschmid.de/agentic-pattern

3. Open source project: Build your second brain AI assistant, combining LLM, Agent and RAG technologies

An open source project called "Building Your Second Brain AI Assistant" is launched on GitHub. The project teaches you how to build a personal second brain AI assistant, combining LLM, agent, RAG, fine-tuning and LLMOps technologies.

The main functions of the project include:

(1) Build an agent-based RAG system that interacts with a personal knowledge base (provide Notion example)

(2) Learn production-ready LLM system architecture design and LLMOps best practices

(3) Implement data ETL pipelines for processing custom data, web crawling, and quality scoring using LLM/heuristics

(4) Generate high-quality instruction datasets for fine-tuning through distillation

(5) Fine-tune the Llama model using Unsloth and track the experiment using Comet

(6) Deploy the fine-tuned LLM as a serverless endpoint on Hugging Face

(7) Apply advanced RAG techniques , including context/parent retrieval and vector search

(8) Building agents using smolagents

（9）Use pipeline orchestration (ZenML) and RAG assessment tools (Opik)

Article link: https://github.com/decodingml/second-brain-ai-assistant-course

4. RM-R1: An innovative approach to reward modeling as a reasoning process

The new generative reward model "RM-R1" treats reward modeling as an inference task , significantly improving interpretability and performance.

Reward modeling is crucial for aligning large language models with human preferences via reinforcement learning with human feedback (RLHF). This work introduces a new class of generative reward models, Reasoning Reward Models (REASRMS) , which treat reward modeling as a reasoning task.

The researchers proposed an inference-oriented training process consisting of two key stages:

(1) Distillation of high-quality reasoning chains

(2) Reinforcement Learning with Verifiable Rewards

RM-R1 improves LLM output by self-generating reasoning traces or chat-specific scoring criteria and evaluating candidate responses based on these criteria.

Experiments show that the model achieves state-of-the-art or near-state-of-the-art performance on multiple comprehensive reward model benchmarks, surpassing larger open-weight models (e.g., Llama3.1-405B) and proprietary models (e.g., GPT-4o) by up to 13.8% in accuracy .

Paper title: RM-R1: Reward Modeling as Reasoning

Paper link: https://arxiv.org/abs/2505.02387

5. Brain-computer interface breakthrough: Large brain language model for silent speech decoding

Researchers propose a large brain language model (LBLM) for silent speech decoding , which could enable more natural and flexible communication for active brain-computer interface (BCI) systems.

The research team collected a new silent speech dataset containing more than 120 hours of electroencephalogram (EEG) recordings from 12 subjects, capturing 24 common English words for language model pre-training and decoding.

The study proposed a future spectrotemporal prediction (FSTP) pre-training paradigm to learn effective representations from unlabeled EEG data. Different from existing EEG pre-training methods that mainly follow the mask reconstruction paradigm, the FSTP method adopts autoregressive modeling in both time and frequency domains to capture the temporal and spectral dependencies of EEG signals.

Extensive experiments show that LBLM achieves significant performance improvements over fully supervised and pre-trained baseline models. For example, in a difficult cross-session setting, the proposed model achieves 47.0% accuracy on semantic-level classification and 39.6% on word-level classification, outperforming the baseline methods by 5.4% and 7.3% respectively.

Paper title: Pretraining Large Brain Language Model for Active BCI: Silent Speech

Paper link: https://arxiv.org/abs/2504.21214

6. Ming-Lite-Uni: Advances in a unified architecture for natural multimodal interaction

Ant Group launched Ming-Lite-Uni, an open source multimodal framework with a newly designed unified vision generator and a native multimodal autoregressive model tailored for unified vision and language.

This project provides an open source implementation of the integrated MetaQueries and M2-omni frameworks, while introducing novel multi-scale learnable tokens and multi-scale representation alignment strategies .

By leveraging fixed MLLMs and learnable diffusion models, Ming-Lite-Uni enables native multimodal AR models to perform text-to-image generation and instruction-based image editing tasks, extending their capabilities beyond pure visual understanding.

The experimental results demonstrate the powerful performance of Ming-Lite-Uni and illustrate the smoothness of its interactive process. All codes and model weights have been open sourced to promote further exploration within the community.

Notably, this work aligns with concurrent multimodal AI milestones such as ChatGPT-4o with native image generation capabilities, updated on March 25, 2025, highlighting the broad relevance of unified models like Ming-Lite-Uni on the path to artificial general intelligence (AGI).