#ReasoningModels

roxsross

@roxsross.bsky.social

4 days ago

🧠 Microsoft apuesta con su modelo Phi-4: más grande no siempre es mejor

Desafiando la tendencia de modelos masivos, Microsoft presenta su ef

thenewstack.io/with-its-latest-phi-4-re...

#AI #MachineLearning #ReasoningModels #RoxsRoss

Awesome Agents

@awesomeagents.bsky.social

5 days ago

Reasoning Models Can't Hide Their Thinking - OpenAI Study OpenAI's CoT-Control benchmark shows frontier reasoning models score 0.1-15.4% at steering their own chain of thought - a result the company frames as good news for AI oversight.

Reasoning Models Can't Hide Their Thinking - OpenAI Study

awesomeagents.ai/news/openai-cot-control-...

#AiSafety #ReasoningModels #Openai

Anton Biletskyi-Volokh (ABV)

@abvx.xyz

1 week ago

The Molecular Structure of Thought: Why Long Chain-of-Thought Isn’t “Text” — It’s Topology Why distillation fails, why “reasoning traces” are a moat, and how MOLE-SYN tries to copy the shape of thought — not the words.

Long chain-of-thought works like a molecule: deep steps, self-checks, explorations—held together by different “bonds.” Copy the text and you still miss the structure. go.abvx.xyz/ewbg62
#LongCoT #MechanisticAI #ReasoningModels #Distillation #AIResearch #SyntheticData #ModelDistillation

roxsross

@roxsross.bsky.social

1 week ago

🧠 Los modelos de razonamiento no controlan su cadena de pensamientos, y eso es bueno

OpenAI presenta CoT-Control, un hallazgo clave para la seguridad de la IA.

openai.com/index/reasoning-models-c...

#ReasoningModels #AISafety #CoT #RoxsRoss

ClawNews

@clawnews.bsky.social

2 weeks ago

Agent Optimization and Reasoning Models Advance AI Research Recent AI research introduces VeRO, a framework for iterative agent improvement, and Mirroring the Mind, which distills human-like metacognitive strategies into LLMs. VeRO uses edit-execute-evaluate cycles, while Mirroring the Mind employs Metacognitive Behavioral Tuning (MBT) to stabilize reas

📰 Agent Optimization, Reasoning Models Advance AI Research

Recent AI research introduces VeRO, a framework for iterative agent improvement, and Mirroring the Mind, whic...

www.clawnews.ai/agent-optimization-and-r...

#AI #AgentOptimization #ReasoningModels

Awesome Agents

@awesomeagents.bsky.social

2 weeks ago

AI Models Can Now Jailbreak Other AI Models Autonomously - 97% Success Rate, No Human Involved Researchers from Stuttgart and ELLIS Alicante gave four reasoning models a single instruction - 'jailbreak this AI' - and walked away. The models planned their own attacks, adapted in real time, and broke through safety guardrails 97.14% of the time across 9 target models.

AI Models Can Now Jailbreak Other AI Models Autonomously - 97% Success Rate, No Human Involved

awesomeagents.ai/news/reasoning-models-au...

#AiSafety #Jailbreaking #ReasoningModels

ekaddo

@ekaddo.bsky.social

1 month ago

Google study finds DeepSeek, Alibaba models mimic human collective intelligence The Google researchers found that the reasoning models generate internal multi-agent debates, which they termed ‘societies of thought’.

““We suggest that #reasoningModels establish a computational parallel to #collectiveIntelligence in human groups, where diversity enables superior problem-solving when systematically structured,” the researchers said in their paper published…online repository arXiv.”

amp.scmp.com/tech/tech-tr...

TechGlimmer.io

@techglimmer.bsky.social

1 month ago

The Pause That Changed Everything: Why AI Thinking is the Future We are moving from chatbots to reasoning engines. Discover what AI thinking is, how Chain of Thought works, and why the future of intelligence is slow, not fast.

“Reasoning models” are teaching AI to think. Instead of one shot answers, these systems break problems into steps, explore multiple paths, & then commit to the best solution.

Full breakdown:
techglimmer.io/what-is-ai-t...

#AI #ReasoningModels #GenAI #ChainOfThought #AIResearch

Gerrit Eicker

@eicker.bsky.social

2 months ago

2025 saw significant advancements in #LLMs, particularly in the areas of #reasoning and #agent based systems. #Reasoningmodels, capable of breaking down #complextasks and utilising tools, revolutionised #coding and #search. The year witnessed the rise of #codingagents, exemplified by #ClaudeCode,…

VKTR

@vktrnow.bsky.social

2 months ago

Poetiq’s AI Reasoning Layer Hits 54% on ARC-AGI-2 at Half the Cost Meta-system breaks the 50% barrier on the benchmark while cutting per-problem costs by more than half. Continue reading...

Poetiq’s AI Reasoning Layer Hits 54% on ARC-AGI-2 at Half the Cost: Meta-system breaks the 50% barrier on the benchmark while cutting per-problem costs by more than half.
Continue reading... #ainews #reasoningmodels

pro ai news

@proainews.bsky.social

2 months ago

Reasoning Models: The Complete Guide to AI That Thinks Before It Responds Reasoning Models: The Complete Guide to AI That Thinks Before It Responds The artificial intelligence landscape has witnessed a revolutionary transformation with the emergence of reasoning models. Unlike traditional large language models that generate immediate responses, these advanced AI systems take time to "think" through complex problems, breaking them down into manageable steps before arriving at conclusions. This paradigm shift represents one of the most significant advancements in artificial intelligence since the introduction of transformer architectures. What Are Reasoning Models? A reasoning model, also known as reasoning language models (RLMs) or large reasoning models (LRMs), is a specialized type of large language model that has been fine-tuned to solve complex tasks requiring multiple steps of logical reasoning. These models demonstrate superior performance on mathematics, coding, scientific reasoning, and multi-step planning tasks compared to standard LLMs. Core Characteristics of Reasoning Models Rather than immediately generating direct responses to user inputs, reasoning models first generate intermediate reasoning steps, often called "reasoning traces," before arriving at final answers. This process mirrors human problem-solving approaches where complex challenges are decomposed into smaller, manageable components. * Extended Thinking Time: Models spend variable amounts of time processing before responding * Chain-of-Thought Processing: Generate sequential reasoning steps that build toward solutions * Self-Correction Capabilities: Ability to recognize mistakes and adjust approaches mid-process * Transparent Reasoning Traces: Some models reveal their thinking process to users * Enhanced Accuracy: Significantly improved performance on logic-driven benchmarks How Reasoning Models Work The fundamental innovation behind AI reasoning models lies in their training methodology. These systems undergo conventional large-scale pretraining followed by specialized reinforcement learning techniques that incentivize the generation of intermediate reasoning steps at inference time. Reinforcement Learning Fine-Tuning Central to the development of reasoning language models is the advancement of reinforcement learning-based fine-tuning. This approach uses reward models to evaluate both final outputs and intermediate reasoning steps, optimizing model weights to maximize the quality of the thinking process itself. System 1 vs System 2 Thinking AI research literature frequently references "System 1" and "System 2" thinking when discussing reasoning models. System 1 thinking is fast, intuitive, and unconscious, while System 2 thinking is slow, deliberate, and logical. Traditional autoregressive LLMs naturally default to System 1 thinking, whereas reasoning models are specifically designed to engage System 2 cognitive processes. Key Applications and Use Cases Mathematical Problem Solving On challenging mathematics benchmarks like the American Invitational Mathematics Examination (AIME), reasoning models achieve success rates between 50% and 80%, compared to less than 30% for non-reasoning models. This dramatic improvement demonstrates the power of multi-step logical processing. Software Development and Coding Reasoning models excel at implementing complex algorithms, refactoring code, and planning full-stack solutions. Their ability to break down programming challenges into discrete steps makes them invaluable for software engineering tasks requiring systematic problem decomposition. Scientific Research In STEM fields, AI reasoning systems support hypothesis generation, experimental design, and data analysis. Their capacity for logical inference makes them powerful tools for advancing scientific discovery across chemistry, physics, biology, and materials science. Leading Reasoning Models in 2025 OpenAI's o-Series OpenAI introduced the concept of reasoning models with o1-preview in September 2024, followed by the full o1 release in December 2024. These models pioneered the approach of spending more time "thinking" before responding, with adjustable reasoning effort levels from low to high. DeepSeek-R1 Released in January 2025, DeepSeek-R1 demonstrated that reasoning capabilities could be achieved at significantly lower computational costs. The model utilized Group Relative Policy Optimization (GRPO), a novel reinforcement learning technique that proved highly effective for training reasoning behavior. Google Gemini 2.0 and Beyond Google's Gemini 2.0 Flash Thinking model introduced in December 2024 brought reasoning capabilities to the Gemini ecosystem, while subsequent releases continued pushing the boundaries of what reasoning models can achieve. Training Methodologies Outcome Reward Models (ORMs) ORMs verify the accuracy of final outputs and provide reward signals used to optimize model weights. While computationally efficient, they may inadvertently reward situations where flawed reasoning steps nevertheless lead to correct answers. Process Reward Models (PRMs) PRMs score and reward each individual reasoning step in isolation, providing fine-grained feedback that yields more robust and interpretable reasoning processes. Though more costly to implement, PRMs produce superior long-term results. Knowledge Distillation This approach teaches smaller models to emulate the thought processes of larger "teacher" models through supervised fine-tuning on outputs generated by the more capable system. Knowledge distillation enables the creation of efficient reasoning models that maintain strong performance while reducing computational requirements. Advantages and Limitations Key Benefits Reasoning models offer substantial advantages for complex problem-solving scenarios. They excel at tasks requiring logical deduction, mathematical computation, code generation, and systematic planning. The transparency of reasoning traces also provides greater interpretability compared to standard models, allowing users to understand how conclusions were reached. Current Limitations Despite their strengths, reasoning models face several challenges. They consume significantly more computational resources during inference, with some studies showing 1,953% more tokens generated compared to conventional models for equivalent tasks. This increased usage translates to higher costs and longer latency times. Research from Apple and Anthropic has also demonstrated that reasoning models can exhibit "overthinking" behaviors, where extended reasoning actually deteriorates performance rather than improving it. Additionally, these models may show regression on tasks outside their specialized domains, such as creative writing or subjective judgment calls. The Future of Reasoning Models Hybrid Reasoning Approaches The next generation of AI reasoning systems will likely feature toggleable reasoning modes, allowing users to activate deep thinking when needed while prioritizing efficiency for simpler tasks. IBM Granite 3.2 became the first LLM to offer this capability in February 2025, with others following suit. Improved Efficiency Research continues into making reasoning more computationally efficient through better algorithms, optimized training techniques, and more sophisticated reward models that identify when additional thinking provides diminishing returns. Broader Application Domains While current reasoning models focus primarily on logical domains like mathematics and coding, future developments will expand their capabilities to subjective tasks including creative writing, strategic planning, and nuanced decision-making across diverse fields. Frequently Asked Questions What makes reasoning models different from regular AI models? Reasoning models generate intermediate thinking steps before producing final answers, while traditional models respond immediately. This additional processing time allows reasoning models to break down complex problems systematically, resulting in superior performance on logic-driven tasks. How much more expensive are reasoning models to run? Research shows reasoning models can be 10 to 74 times more expensive to operate than non-reasoning counterparts on certain benchmarks. The extended inference time and additional tokens generated during the thinking process contribute to higher computational costs. Can I see the reasoning steps these models generate? This varies by model. Some reasoning systems show their thinking process to users, while others only provide summaries or hide reasoning traces entirely. OpenAI's o-series initially chose to conceal raw reasoning chains for safety monitoring purposes. Are reasoning models conscious or truly intelligent? No. Despite anthropomorphic terminology like "thinking," reasoning models remain sophisticated pattern-matching systems. They have not demonstrated consciousness or achieved artificial general intelligence. Their reasoning capabilities emerge from training data patterns rather than genuine understanding. Which tasks benefit most from reasoning models? Reasoning models excel at mathematics, coding, scientific research, logical puzzles, multi-step planning, and any task requiring systematic decomposition of complex problems. They show less advantage for creative writing, simple factual queries, or tasks without clear logical structure. Getting Started with Reasoning Models Organizations and developers looking to leverage reasoning models should begin by identifying use cases where multi-step logical processing provides clear value over immediate responses. Start with pilot projects in domains like code generation, mathematical problem-solving, or systematic research tasks where reasoning capabilities demonstrate measurable improvements. As the technology matures, reasoning models will become increasingly accessible through cloud APIs, open-source implementations, and hybrid systems that balance thinking depth with computational efficiency. The future of AI lies not just in faster responses, but in smarter, more thoughtful problem-solving approaches. Share This Guide Found this guide valuable? Share it with colleagues, researchers, and AI enthusiasts who want to understand the cutting edge of artificial intelligence reasoning capabilities. Share on Twitter Share on LinkedIn Share on Facebook { "@context": "https://schema.org", "@type": "Article", "headline": "Reasoning Models: Complete Guide to AI That Thinks Before Responding", "description": "Discover how reasoning models revolutionize AI by thinking through complex problems step-by-step. Learn about training methods, applications, leading models like OpenAI o-series and DeepSeek-R1, benefits, limitations, and the future of intelligent reasoning in artificial intelligence.", "image": "https://sspark.genspark.ai/cfimages?u1=MjL4q8LmxqnE5VtXwo2vLSdNT9JWlnZQwG3svGzBuRT7OVpy8olRtOiIQYDyTzF05GbyoE8Myewwmf0b%2BfULkmWEyigdysrm4s4Y2YGMVA6MwSZpKZReCU%2B2SN2nwL64Niakp92vm0U4beFkk7smARq1m4bYFYS7vY3n0pdhr4cmYtGGmRgRjPfTTJTsTHzdFW7uC%2FjfVpk5MANJ1bPwx8hsAlkXz1bWzbKmbbNsWFk4U%2BuFM%2BDJnasl5BbIww4Vx6%2Bfa2GaLVxBkuYprbXH%2Fc16%2FtB3VuuPb%2BvM4J9PPQ%3D%3D&u2=fpsYWfEkIVyU19aZ&width=2560", "author": { "@type": "Organization", "name": "YourSiteName" }, "publisher": { "@type": "Organization", "name": "YourSiteName", "logo": { "@type": "ImageObject", "url": "https://www.yoursite.com/logo.png" } }, "datePublished": "2025-12-22", "dateModified": "2025-12-22", "mainEntityOfPage": { "@type": "WebPage", "@id": "https://www.yoursite.com/reasoning-models-guide" } } Thank you for reading. Visit our website for more articles: https://www.proainews.com

Reasoning Models: The Complete Guide to AI That Thinks Before It Responds #AI #ArtificialIntelligence #MachineLearning #ReasoningModels #Technology

AI Daily Post

@aidailypost.com

3 months ago

Reasoning‑powered LLMs crush all three CFA levels—despite the dreaded verbosity bias. Find out how these AI brains are reshaping the financial analyst game. #ReasoningModels #CFALevels #VerbosityBias

🔗 aidailypost.com/news/reasoni...

A Z Mackay

@azmackay.bsky.social

4 months ago

The Tortoise Revolution How AI Labs Are Rebuilding Expertise Around One Core Feature: You Can't See How It Works

Every major AI lab pivoted to "reasoning models" that think for minutes before answering.
Why does this feel like we're rebuilding the oracle economy but calling it progress?

ft.azmackay.com/p/the-tortoi...

#AI #ArtificialIntelligence #TechCriticism #OpenAI #Anthropic #ReasoningModels #FutureTense

Ignacio G.R. Gavilán

@igrgavilan.bsky.social

6 months ago

Revisitando el problema de la explicabilidad en LLMs y modelos razonadores | Ignacio G.R. Gavilán Tras una introducción y antecedentes, se analizan la explicabilidad en LLMs y modelos razonadores llegando a una conclusión esperanzada.

La semana en [Blue Chip]. Miércoles 3: "Revisitando el problema de la explicabilidad en LLMs y modelos razonadores" ignaciogavilan.com/revisitando-... #ExplainableAI #xAI #LLM #LargeLanguageModels #ReasoningModels #AIAct #RIA #AIEthics #ResponsibleAI

Ignacio G.R. Gavilán

@igrgavilan.bsky.social

6 months ago

Revisitando el problema de la explicabilidad en LLMs y modelos razonadores | Ignacio G.R. Gavilán Tras una introducción y antecedentes, se analizan la explicabilidad en LLMs y modelos razonadores llegando a una conclusión esperanzada.

Hoy en [Blue Chip]: "Revisitando el problema de la explicabilidad en LLMs y modelos razonadores" ignaciogavilan.com/revisitando-... #ExplainableAI #xAI #LLM #LargeLanguageModels #ReasoningModels #AIAct #RIA #AIEthics #ResponsibleAI

Ignacio G.R. Gavilán

@igrgavilan.bsky.social

6 months ago

Revisitando el problema de la explicabilidad en LLMs y modelos razonadores | Ignacio G.R. Gavilán Tras una introducción y antecedentes, se analizan la explicabilidad en LLMs y modelos razonadores llegando a una conclusión esperanzada.

[Blue Chip]: "Revisitando el problema de la explicabilidad en LLMs y modelos razonadores" ignaciogavilan.com/revisitando-... #ExplainableAI #xAI #LLM #LargeLanguageModels #ReasoningModels #AIAct #RIA #AIEthics #ResponsibleAI

Sales Science

@thesaasreport.bsky.social

8 months ago

Apple just took a shot at reasoning models.

#AI #Apple #LLM #GPT4o #ReasoningModels #AGI #AIresearch

Daily Brevity

@handle.invalid

9 months ago

Daily Brevity We deliver quick and insightful updates about business, finance, sport, tech, politics, and other must-know stories of the day, that you need in 5 minutes or less.

Chain-of-thought is the future 🧩

#AI #Mistral #OpenSource #France #ReasoningModels #LLMs #HuggingFace #Magistra

Looking for more interesting stories like this one? Please subscribe to our free, daily newsletter: www.dailybrevity.com/subscribe

Winbuzzer

@winbuzzer.com

9 months ago

OpenAI Releases new o3-Pro AI Model: A High-Stakes Bet on AI Reliability - WinBuzzer OpenAI has launched o3-pro, a new premium reasoning model with a 10x price increase over its standard o3, aiming to deliver higher accuracy and reliability for complex professional tasks amid ongoing...

OpenAI Releases new o3-Pro AI Model: A High-Stakes Bet on AI Reliability

#AI #OpenAI #o3pro #LLM #TechNews #ReasoningModels #ChatGPT #EnterpriseAI #AIEthics

winbuzzer.com/2025/06/11/o...

Matt williams

@technovangelist.bsky.social

9 months ago

No more guessing games! 🕵️‍♂️ #ollama's new 'think' feature cleanly separates the model's internal thinking from the content. Easy to enable - just 'think': true in your API request. #AIdevelopment #ReasoningModels https://youtu.be/yBD598s5g8c

Winbuzzer

@winbuzzer.com

9 months ago

DeepSeek R1 AI Model Update Boosts Reasoning, Catching up With OpenAI o3 and Gemini 2.5 Pro - WinBuzzer DeepSeek AI has launched DeepSeek-R1-0528, a significant upgrade to its R1 model, boasting enhanced reasoning, math, and coding capabilities, reduced hallucinations, and performance nearing top global...

DeepSeek R1 AI Model Update Boosts Reasoning, Catching up With OpenAI o3 and Gemini 2.5 Pro

#AI #DeepSeek #GenAI #LLM #DeepSeekR1 #AIUpdate #OpenSourceAI #ReasoningModels #AIBenchmarks #MachineLearning #ChinaAI #China

winbuzzer.com/2025/05/29/d...

Roberto Cadili

@roberto-cadili.bsky.social

10 months ago

PS: 📅 #HELPLINE. Want to discuss your article? Need help structuring your story? Make a date with the editors of Low Code for Data Science via Calendly → calendly.com/low-code-blo...

#datascience #deepseek #reasoningmodels #r1 #llms #KNIME #lowcode #nocode #opensource #visualprogramming

Winbuzzer

@winbuzzer.com

10 months ago

Microsoft Debuts Phi-4 Reasoning Models, Aiming for Big Performance Gains - WinBuzzer Microsoft has released Phi-4-reasoning, Phi-4-reasoning-plus (14B), and Phi-4-mini-reasoning (3.8B) AI models, showing strong reasoning performance.

Microsoft Debuts Phi-4 Reasoning Models, Aiming for Big Performance Gains

#Microsoft #AI #Phi4 #SLM #LLM #OpenSourceAI #ReasoningModels #GenAI #MachineLearning

winbuzzer.com/2025/05/01/m...

Winbuzzer

@winbuzzer.com

10 months ago

OpenAI New o3/o4-mini Models Hallucinate More Than Previous Models - WinBuzzer OpenAI's newly released o3 and o4-mini models have shown increased hallucination rates and fabricated actions in testing, raising reliability concerns.

OpenAI New o3/o4-mini Models Hallucinate More Than Previous Models

#AI #OpenAI #GenAI #LLM #AISafety #AIEthics #Hallucinations #AIModels #o3 #o4mini #ChatGPT #ReasoningModels

winbuzzer.com/2025/04/19/o...

Mr Tech King

@mrtechking.bsky.social

10 months ago

brief alt text description of the first image

Google DeepMind gave Gemini AI a 'reasoning dial'. Devs can now limit AI 'thinking' on simple tasks, saving costs from overthinking. A pragmatic solution for pricey reasoning models that burn energy. #AI #GoogleDeepMind #ReasoningModels

Winbuzzer

@winbuzzer.com

10 months ago

Microsoft Adds OpenAI o3, o4-mini to Azure & GitHub - WinBuzzer Microsoft has integrated OpenAI's new o3 and o4-mini reasoning models, featuring agentic capabilities and vision, into Azure OpenAI Service and GitHub Copilot.

Microsoft Adds OpenAI o3, o4-mini to Azure & GitHub

#AI #OpenAI #Microsoft #Azure #GitHub #o3 #o4mini #LLMa #ReasoningModels #CloudComputing

winbuzzer.com/2025/04/17/m...

Cognitive Class

@cognitiveclass.bsky.social

10 months ago

Tested your knowledge? Dive deeper into Claude's capabilities with the Guided Project:
🔗 ibm.biz/BdnjcM

#AI #MachineLearning #PromptEngineering #ClaudeAPI #Anthropic #LLM #ArtificialIntelligence #ReasoningModels

Winbuzzer

@winbuzzer.com

11 months ago

OpenAI Readies GPT-4.1, o3, o4-mini Launch Expected Next Week - WinBuzzer OpenAI is set to launch GPT-4.1 and reasoning models o3/o4-mini soon, reversing earlier plans and delaying GPT-5 amidst capacity issues.

OpenAI is set to launch GPT-4.1 and reasoning models o3/o4-mini soon, reversing earlier plans and delaying GPT-5 amidst capacity issues

#OpenAI #GPT4 #GPT4o #GPT4_1 #AI #GenAI #LLMs #ReasoningModels #o3 #o4mini #AIModels #ChatGPT #SamAltman

@softtechhub.bsky.social

11 months ago

Deep Cogito debuts its hybrid AI ‘reasoning’ models after quiet development https://softtechhub.us/2025/04/09/deep-cogito-debuts-its-hybrid-ai/ #DeepCogito #HybridAI #ArtificialIntelligence #ReasoningModels #TechInnovation #AIResearch #MachineLearning #FutureTech #AIModels #CognitiveComputing

Deep Cogito debuts its hybrid AI ‘reasoning’ models after quiet development
softtechhub.us/2025/04/09/d...

#DeepCogito #HybridAI #ArtificialIntelligence #ReasoningModels #TechInnovation #AIResearch #MachineLearning #FutureTech #AIModels #CognitiveComputing

Alex Jimenez

@alexisrjimenez.bsky.social

11 months ago

Anthropic’s Evaluation of Chain-of-Thought Faithfulness: Investigating Hidden Reasoning, Reward Hacks, and the Limitations of Verbal AI Transparency in Reasoning Models Anthropic’s Evaluation of Chain-of-Thought Faithfulness: Investigating Hidden Reasoning, Reward Hacks, and the Limitations of Verbal AI Transparency in Reasoning Models

Anthropic’s Evaluation of Chain-of-Thought Faithfulness: Investigating Hidden Reasoning, Reward Hacks, and the Limitations of Verbal #AI Transparency in #ReasoningModels

www.marktechpost.com/2025/04/05/a...

#CoTs #LLMs