Hacker News with Generative AI: Reasoning

Reflections on Neuralese (greaterwrong.com)
With the recent breakthroughs taking advantage of extensive Chain of Thought (CoT) reasoning in LLMs, there have been many attempts to modify the technique to be even more powerful.
Beyond Semantics: Unreasonable Effectiveness of Reasonless Intermediate Tokens (arxiv.org)
Recent impressive results from large reasoning models have been interpreted as a triumph of Chain of Thought (CoT), and especially of the process of training on CoTs sampled from base LLMs in order to help find new reasoning patterns.
Claude 4 (anthropic.com)
Today, we’re introducing the next generation of Claude models: Claude Opus 4 and Claude Sonnet 4, setting new standards for coding, advanced reasoning, and AI agents.
The Unreasonable Effectiveness of Reasonless Intermediate Tokens (arxiv.org)
Recent impressive results from large reasoning models have been interpreted as a triumph of Chain of Thought (CoT), and especially of the process of training on CoTs sampled from base LLMs in order to help find new reasoning patterns.
Analyzing, Predicting, and Controlling How a Reasoning Model Will Think (arxiv.org)
Long chain-of-thought (CoT) is an essential ingredient in effective usage of modern large language models, but our understanding of the reasoning strategies underlying these capabilities remains limited.
Absolute Zero: Reinforced Self-Play Reasoning with Zero Data (arxiv.org)
To address these concerns, we propose a new RLVR paradigm called Absolute Zero, in which a single model learns to propose tasks that maximize its own learning progress and improves reasoning by solving them, without relying on any external data.
Llama-Nemotron: Efficient Reasoning Models (arxiv.org)
We introduce the Llama-Nemotron series of models, an open family of heterogeneous reasoning models that deliver exceptional reasoning capabilities, inference efficiency, and an open license for enterprise use.
Step-by-step reasoning verifiers that think (arxiv.org)
Step-by-step verifiers -- also known as process reward models (PRMs) -- are a key ingredient for test-time scaling.
Reasoning Models Can Be Effective Without Thinking (arxiv.org)
Recent LLMs have significantly improved reasoning capabilities, primarily by including an explicit, lengthy Thinking process as part of generation. In this paper, we question whether this explicit thinking is necessary.
M1: Towards Scalable Test-Time Compute with Mamba Reasoning Models (arxiv.org)
Effective reasoning is crucial to solving complex mathematical problems. Recent large language models (LLMs) have boosted performance by scaling test-time computation through long chain-of-thought reasoning.
RAG Without Vectors – Reasoning-Based RAG using PageIndex (ycombinator.com)
Traditional vector-based RAG often struggles with retrieval accuracy because it optimizes for similarity, not relevance. But what we truly need in retrieval is relevance, which requires reasoning. When working with professional documents that require domain expertise and multi-step reasoning, vector-based RAG and similarity search often fall short.
Search-R1: Training LLMs to Reason and Leverage Search Engines with RL (arxiv.org)
Efficiently acquiring external knowledge and up-to-date information is essential for effective reasoning and text generation in large language models (LLMs).
Sketch-of-Thought: Efficient LLM Reasoning (arxiv.org)
Recent advances in large language models have demonstrated remarkable reasoning capabilities through Chain of Thought (CoT) prompting, but often at the cost of excessive verbosity in their intermediate outputs, which increases computational overhead.
Engineering Reasoning LLMs: Notes and Observations (thelis.org)
Reasoning models are LLMs that are designed to tackle complex tasks that require multiple steps of reasoning. OpenAI’s o1 was the first mainstream reasoning model, and showed remarkable results on highly complex questions. A torrent of papers were published subsequent to o1, each exploring how to build these types of systems. Most recently, DeepSeek-R1 has captured attention as a state-of-the-art reasoning model with a published methodology.
Cognitive Behaviors That Enable Self-Improving Reasoners (arxiv.org)
Test-time inference has emerged as a powerful paradigm for enabling language models to ``think'' longer and more carefully about complex challenges, much like skilled human experts.
Order Doesn’t Matter, But Reasoning Does (arxiv.org)
Claude 3.7 Sonnet and Claude Code (anthropic.com)
Today, we’re announcing Claude 3.7 Sonnet1, our most intelligent model to date and the first hybrid reasoning model on the market.
General Reasoning: Free, open resource for building large reasoning models (gr.inc)
1,568,725 questions and 268,241 chain-of-thought traces to train open models
Grok 3 Beta – The Age of Reasoning Agents (x.ai)
We are thrilled to unveil an early preview of Grok 3, our most advanced model yet, blending superior reasoning with extensive pretraining knowledge.
Reasoning models are just LLMs (antirez.com)
It’s not new, but it’s accelerating. People that used to say that LLMs were a fundamentally flawed way to reach any useful reasoning and, in general, to develop any useful tool with some degree of generality, are starting to shuffle the deck, in the hope to look less wrong. They say: “the progresses we are seeing are due to the fact that models like OpenAI o1 or DeepSeek R1 are not just LLMs”.
LIMO: Less Is More for Reasoning (arxiv.org)
We present a fundamental discovery that challenges our understanding of how complex reasoning emerges in large language models.
Demystifying Long Chain-of-Thought Reasoning in LLMs (arxiv.org)
Scaling inference compute enhances reasoning in large language models (LLMs), with long chains-of-thought (CoTs) enabling strategies like backtracking and error correction.
Understanding Reasoning LLMs (sebastianraschka.com)
This article describes the four main approaches to building reasoning models, or how we can enhance LLMs with reasoning capabilities. I hope this provides valuable insights and helps you navigate the rapidly evolving literature and hype surrounding this topic.
Microsoft Phi 4 with R1 Reasoning (huggingface.co)
These LoRA adapters were trained using diverse reasoning datasets that incorporate structured Thought and Solution responses to enhance logical inference.
A Visual Guide to Reasoning LLMs (maartengrootendorst.com)
DeepSeek-R1, OpenAI o3-mini, and Google Gemini 2.0 Flash Thinking are prime examples of how LLMs can be scaled to new heights through “reasoning“ frameworks.
Efficient Reasoning with Hidden Thinking (arxiv.org)
Chain-of-Thought (CoT) reasoning has become a powerful framework for improving complex problem-solving capabilities in Multimodal Large Language Models (MLLMs). However, the verbose nature of textual reasoning introduces significant inefficiencies.
OpenAI launches o3-mini, its latest 'reasoning' model (techcrunch.com)
OpenAI on Friday launched a new AI “reasoning” model, o3-mini, the newest in the company’s o family of reasoning models.
DeepSeek-R1 at 3,872 tokens / second on a single Nvidia HGX H200 (nvidia.com)
DeepSeek-R1 is an open model with state-of-the-art reasoning capabilities. Instead of offering direct responses, reasoning models like DeepSeek-R1 perform multiple inference passes over a query, conducting chain-of-thought, consensus and search methods to generate the best answer.
Open-R1: an open reproduction of DeepSeek-R1 (huggingface.co)
OpenAI’s o1 model showed that when LLMs are trained to do the same—by using more compute during inference—they get significantly better at solving reasoning tasks like mathematics, coding, and logic.
Bespoke-Stratos: The unreasonable effectiveness of reasoning distillation (bespokelabs.ai)
We trained Bespoke-Stratos-32B, our reasoning model distilled from DeepSeek-R1 using Berkeley NovaSky’s Sky-T1 data pipeline.