News from the AI & ML world
DeeperML - #reasoning
|
Michael Nuñez@AI News | VentureBeat
//
References:
venturebeat.com
, The Algorithmic Bridge
,
Anthropic has been at the forefront of investigating how AI models like Claude process information and make decisions. Their scientists developed interpretability techniques that have unveiled surprising behaviors within these systems. Research indicates that large language models (LLMs) are capable of planning ahead, as demonstrated when writing poetry or solving problems, and that they sometimes work backward from a desired conclusion rather than relying solely on provided facts.
Anthropic researchers also tested the "faithfulness" of CoT models' reasoning by giving them hints in their answers, and see if they will acknowledge it. The study found that reasoning models often avoided mentioning that they used hints in their responses. This raises concerns about the reliability of chains-of-thought (CoT) as a tool for monitoring AI systems for misaligned behaviors, especially as these models become more intelligent and integrated into society. The research emphasizes the need for ongoing efforts to enhance the transparency and trustworthiness of AI reasoning processes. Recommended read:
References :
Ellie Ramirez-Camara@Data Phoenix
//
Google has launched Gemini 2.5 Pro, hailed as its most intelligent "thinking model" to date. This new AI model excels in reasoning and coding benchmarks, featuring an impressive 1M token context window. Gemini 2.5 Pro is currently accessible to Gemini Advanced users, with integration into Vertex AI planned for the near future. The model has already secured the top position on the Chatbot Arena LLM Leaderboard, showcasing its superior performance in areas like math, instruction following, creative writing, and handling challenging prompts.
Gemini 2.5 Pro represents a new category of "thinking models" designed to enhance performance through reasoning before responding. Google reports that it achieved this level of performance by combining an enhanced base model with improved post-training techniques and aims to build these capabilities into all of their models. The model also obtained leading scores in math and science benchmarks, including GPQA and AIME 2025, without using test-time techniques. A significant focus for the Gemini 2.5 development has been coding performance, where Google reports that the new model excels at creating visual. Recommended read:
References :
Matt Marshall@AI News | VentureBeat
//
References:
Microsoft Security Blog
, www.zdnet.com
Microsoft is enhancing its Copilot Studio platform with AI-driven improvements, introducing deep reasoning capabilities that enable agents to tackle intricate problems through methodical thinking and combining AI flexibility with deterministic business process automation. The company has also unveiled specialized deep reasoning agents for Microsoft 365 Copilot, named Researcher and Analyst, to help users achieve tasks more efficiently. These agents are designed to function like personal data scientists, processing diverse data sources and generating insights through code execution and visualization.
Microsoft's focus includes securing AI and using it to bolster security measures, as demonstrated by the upcoming Microsoft Security Copilot agents and new security features. Microsoft aims to provide an AI-first, end-to-end security platform that helps organizations secure their future, one example being the AI agents designed to autonomously assist with phishing, data security, and identity management. The Security Copilot tool will automate routine tasks, allowing IT and security staff to focus on more complex issues, aiding in defense against cyberattacks. Recommended read:
References :
Maximilian Schreiner@THE DECODER
//
Google DeepMind has announced Gemini 2.5 Pro, its latest and most advanced AI model to date. This new model boasts enhanced reasoning capabilities and improved accuracy, marking a significant step forward in AI development. Gemini 2.5 Pro is designed with built-in 'thinking' capabilities, enabling it to break down complex tasks into multiple steps and analyze information more effectively before generating a response. This allows the AI to deduce logical conclusions, incorporate contextual nuances, and make informed decisions with unprecedented accuracy, according to Google.
The Gemini 2.5 Pro has already secured the top position on the LMArena leaderboard, surpassing other AI models in head-to-head comparisons. This achievement highlights its superior performance and high-quality style in handling intricate tasks. The model also leads in math and science benchmarks, demonstrating its advanced reasoning capabilities across various domains. This new model is available as Gemini 2.5 Pro (experimental) on Google’s AI Studio and for Gemini Advanced users on the Gemini chat interface. Recommended read:
References :
Matt Marshall@AI News | VentureBeat
//
References:
AI News | VentureBeat
, Source Asia
,
Microsoft is introducing two new AI reasoning agents, Researcher and Analyst, to Microsoft 365 Copilot. These "first-of-their-kind" agents are designed to tackle complex problems by utilizing methodical thinking, offering users a more efficient workflow. The Researcher agent combines OpenAI’s deep research model with Microsoft 365 Copilot’s advanced orchestration and search capabilities to deliver insights with greater quality and accuracy, helping users perform complex, multi-step research.
Analyst, built on OpenAI’s o3-mini reasoning model, functions like a skilled data scientist and is optimized for advanced data analysis. It uses chain-of-thought reasoning to iteratively refine its analysis and provide high-quality answers, mirroring human analytical thinking. This agent can run Python to tackle complex data queries and can turn raw data scattered across spreadsheets into visualizations or revenue projections. Researcher and Analyst will be available to customers with a Microsoft 365 Copilot license in April as part of a new “Frontier” program. Recommended read:
References :
Maximilian Schreiner@THE DECODER
//
Google has unveiled Gemini 2.5 Pro, its latest and "most intelligent" AI model to date, showcasing significant advancements in reasoning, coding proficiency, and multimodal functionalities. According to Google, these improvements come from combining a significantly enhanced base model with improved post-training techniques. The model is designed to analyze complex information, incorporate contextual nuances, and draw logical conclusions with unprecedented accuracy. Gemini 2.5 Pro is now available for Gemini Advanced users and on Google's AI Studio.
Google emphasizes the model's "thinking" capabilities, achieved through chain-of-thought reasoning, which allows it to break down complex tasks into multiple steps and reason through them before responding. This new model can handle multimodal input from text, audio, images, videos, and large datasets. Additionally, Gemini 2.5 Pro exhibits strong performance in coding tasks, surpassing Gemini 2.0 in specific benchmarks and excelling at creating visually compelling web apps and agentic code applications. The model also achieved 18.8% on Humanity’s Last Exam, demonstrating its ability to handle complex knowledge-based questions. Recommended read:
References :
Ben Lorica@Gradient Flow
//
References:
Gradient Flow
, AI News | VentureBeat
AI is rapidly evolving, moving beyond simple predictions to tackle complex, structured problems. New AI agents, open models, and a focus on productivity are driving this change, with reasoning and deep research capabilities leading the way. China's Monica.ai has developed Manus, an AI agent that autonomously handles tasks like real estate research and resume analysis, demonstrating the potential of multi-agent systems.
OpenAI is also advancing agent-building tools, including the Responses API and Computer Use Tool, enabling developers to create sophisticated applications. Additionally, Google's Gemma 3 offers an open-weight language model family, providing developers with greater flexibility. These advancements signify a shift towards artificial general intelligence (AGI), where AI can genuinely reason and solve novel problems. AI's impact is also evident in content creation, with platforms like n8n enabling users to build content creator agents without coding. These agents can automate tasks and streamline workflows, boosting productivity for individuals and businesses. This reflects a broader trend of AI being used to enhance productivity across various sectors, from research to code generation and content creation. Recommended read:
References :
Matthias Bastian@THE DECODER
//
Baidu has launched two new AI models, ERNIE 4.5 and ERNIE X1, designed to compete with DeepSeek's R1 model. The company is making these models freely accessible to individual users through the ERNIE Bot platform, ahead of the initially planned schedule. ERNIE 4.5 is a multimodal foundation model, integrating text, images, audio, and video to enhance understanding and content generation across various data types. This model demonstrates significant improvements in language understanding, reasoning, and coding abilities.
ERNIE X1 is Baidu's first model specifically designed for complex reasoning tasks, excelling in logical inference, problem-solving, and structured decision-making suitable for applications in finance, law, and data analysis. Baidu claims that ERNIE X1 matches DeepSeek R1’s performance at half the cost. ERNIE 4.5 has shown performance on par with models like DeepSeek-R1, but at approximately half the deployment cost. Recommended read:
References :
Matthew S.@IEEE Spectrum
//
References:
IEEE Spectrum
, Sebastian Raschka, PhD
Recent research indicates that AI models, particularly large language models (LLMs), can struggle with overthinking and analysis paralysis, impacting their efficiency and success rates. A study has found that reasoning LLMs sometimes overthink problems, which leads to increased computational costs and a reduction in their overall performance. This issue is being addressed through various optimization techniques, including scaling inference-time compute, reinforcement learning, and supervised fine-tuning, to ensure models use only the necessary amount of reasoning for tasks.
The size and training methods of these models play a crucial role in their reasoning abilities. For instance, Alibaba's Qwen team introduced QwQ-32B, a 32-billion-parameter model that outperforms much larger rivals in key problem-solving tasks. QwQ-32B achieves superior performance in math, coding, and scientific reasoning using multi-stage reinforcement learning, despite being significantly smaller than DeepSeek-R1. This advancement highlights the potential of reinforcement learning to unlock reasoning capabilities in smaller models, rivaling the performance of giant models while requiring less computational power. Recommended read:
References :
@bdtechtalks.com
//
Alibaba has recently launched Qwen-32B, a new reasoning model, which demonstrates performance levels on par with DeepMind's R1 model. This development signifies a notable achievement in the field of AI, particularly for smaller models. The Qwen team showcased that reinforcement learning on a strong base model can unlock reasoning capabilities for smaller models that enhances their performance to be on par with giant models.
Qwen-32B not only matches but also surpasses models like DeepSeek-R1 and OpenAI's o1-mini across key industry benchmarks, including AIME24, LiveBench, and BFCL. This is significant because Qwen-32B achieves this level of performance with only approximately 5% of the parameters used by DeepSeek-R1, resulting in lower inference costs without compromising on quality or capability. Groq is offering developers the ability to build FAST with Qwen QwQ 32B on GroqCloud™, running the 32B parameter model at ~400 T/s. This model is proving to be very competitive in reasoning benchmarks and is one of the top open source models being used. The Qwen-32B model was explicitly designed for tool use and adapting its reasoning based on environmental feedback, which is a huge win for AI agents that need to reason, plan, and adapt based on context (outperforms R1 and o1-mini on the Berkeley Function Calling Leaderboard). With these capabilities, Qwen-32B shows that RL on a strong base model can unlock reasoning capabilities for smaller models that enhances their performance to be on par with giant models. Recommended read:
References :
@bdtechtalks.com
//
Alibaba's Qwen team has unveiled QwQ-32B, a 32-billion-parameter reasoning model that rivals much larger AI models in problem-solving capabilities. This development highlights the potential of reinforcement learning (RL) in enhancing AI performance. QwQ-32B excels in mathematics, coding, and scientific reasoning tasks, outperforming models like DeepSeek-R1 (671B parameters) and OpenAI's o1-mini, despite its significantly smaller size. Its effectiveness lies in a multi-stage RL training approach, demonstrating the ability of smaller models with scaled reinforcement learning to match or surpass the performance of giant models.
The QwQ-32B is not only competitive in performance but also offers practical advantages. It is available as open-weight under an Apache 2.0 license, allowing businesses to customize and deploy it without restrictions. Additionally, QwQ-32B requires significantly less computational power, running on a single high-end GPU compared to the multi-GPU setups needed for larger models like DeepSeek-R1. This combination of performance, accessibility, and efficiency positions QwQ-32B as a valuable resource for the AI community and enterprises seeking to leverage advanced reasoning capabilities. Recommended read:
References :
Ryan Daws@AI News
//
Alibaba's Qwen team has launched QwQ-32B, a 32-billion parameter AI model, designed to rival the performance of much larger models like DeepSeek-R1, which has 671 billion parameters. This new model highlights the effectiveness of scaling Reinforcement Learning (RL) on robust foundation models. QwQ-32B leverages continuous RL scaling to demonstrate significant improvements in areas like mathematical reasoning and coding proficiency.
The Qwen team successfully integrated agent capabilities into the reasoning model, allowing it to think critically, use tools, and adapt its reasoning based on environmental feedback. The model has been evaluated across a range of benchmarks, including AIME24, LiveCodeBench, LiveBench, IFEval, and BFCL, designed to assess its mathematical reasoning, coding proficiency, and general problem-solving capabilities. QwQ-32B is available as open-weight on Hugging Face and on ModelScope under an Apache 2.0 license, allowing for both commercial and research uses. Recommended read:
References :
Divyesh Vitthal@MarkTechPost
//
References:
LearnAI
, Sebastian Raschka, PhD
Large language models (LLMs) are facing scrutiny regarding their reasoning capabilities and tendency to produce hallucinations, instances where they generate incorrect or fabricated information. Andrej Karpathy, former Senior Director of AI at Tesla, suggests that these hallucinations are emergent cognitive effects arising from the LLM training pipeline. He explains that LLMs predict words based on patterns in their training data, rather than possessing factual knowledge like humans. This leads to situations where models generate plausible-sounding but entirely false information.
Researchers are actively working on improving the reasoning skills of LLMs while minimizing computational costs. One approach involves using distilled reasoners, which allow for faster and more efficient inference. Additionally, Alibaba has unveiled QwQ-32B, a 32 billion parameter AI model that rivals the performance of the much larger DeepSeek-R1. This model achieves comparable results with significantly fewer parameters through reinforcement learning, showcasing the potential for enhancing model performance beyond conventional pretraining and post-training methods. Recommended read:
References :
@techstrong.ai
//
Microsoft has been making significant strides in the realm of artificial intelligence, with developments ranging from new Copilot features to breakthroughs in quantum computing and large language models. The company recently released a native Copilot application for macOS, bringing the AI assistant to Mac users with features similar to the Windows version, including image uploading and text generation. The macOS version includes dark mode and a shortcut command for easy activation. Microsoft also removed limits to Copilot Voice and Think Deeper, allowing Copilot users to have extended conversations with the AI assistant.
Microsoft AI has also released LongRoPE2, a near-lossless method designed to extend the context window of Large Language Models (LLMs) to 128K tokens while retaining a high degree of short-context accuracy. This addresses a key limitation in LLMs, which often struggle to process long-context sequences effectively. In the quantum computing space, Microsoft researchers announced the creation of the first “topological qubits” in a device, representing a potential leap forward in the field. Recommended read:
References :
Esra Kayabali@AWS News Blog
//
Anthropic has launched Claude 3.7 Sonnet, their most advanced AI model to date, designed for practical use in both business and development. The model is described as a hybrid system, offering both quick responses and extended, step-by-step reasoning for complex problem-solving. This versatility eliminates the need for separate models for different tasks. The company emphasized Claude 3.7 Sonnet’s strength in coding tasks. The model's reasoning capabilities allow it to analyze and modify complex codebases more effectively than previous versions and can process up to 128K tokens.
Anthropic also introduced Claude Code, an agentic coding tool, currently in limited research preview. The tool promises to revolutionize coding by automating parts of a developer's job. Claude 3.7 Sonnet is accessible across all Anthropic plans, including Free, Pro, Team, and Enterprise, and via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI. Extended thinking mode is reserved for paid subscribers. Pricing is set at $3 per million input tokens and $15 per million output tokens. Anthropic stated they reduced unnecessary refusals by 45% compared to its predecessor. Recommended read:
References :
@www.analyticsvidhya.com
//
References:
Composio
, Analytics Vidhya
OpenAI recently launched o3-mini, the first model in its o3 family, which includes two specialized variants: o3-mini-high and o3-mini-low. The o3-mini-high variant is designed to spend more time reasoning, providing more in-depth answers, while the o3-mini-low prioritizes speed for quicker responses. Benchmarks indicate that o3-mini displays comparable performance to OpenAI's o1 model, but at a significantly lower cost.
Researchers have noted that o3-mini is roughly 15 times cheaper and about five times faster than o1. Despite being cheaper than GPT-4o, o3-mini has a stricter usage limit of 150 messages per hour, raising questions about OpenAI's subsidization strategy. The model has demonstrated superior performance to o1 on benchmarks such as FrontierMath, Codeforces, and GPQA. Additionally, o3-mini is the first reasoning model from OpenAI to feature official function-calling support, making it particularly useful for AI agents. Recommended read:
References :
@the-decoder.com
//
Perplexity AI has launched Deep Research, an AI-powered research tool aimed at competing with OpenAI and Google Gemini. Using DeepSeek-R1, Perplexity is offering comprehensive research reports at a much lower cost than OpenAI, with 500 queries per day for $20 per month compared to OpenAI's $200 per month for only 100 queries. The new service automatically conducts dozens of searches and analyzes hundreds of sources to produce detailed reports in one to two minutes.
Perplexity claims Deep Research performs 8 searches and consults 42 sources to generate a 1,300-word report in under 3 minutes. The company says that Deep Research tool works particularly well for finance, marketing, and technology research. The service is launching first on web browsers, with iOS, Android, and Mac versions planned for later release. Perplexity CEO Aravind Srinivas stated he wants to keep making it faster and cheaper for the interest of humanity. Recommended read:
References :
Ben Lorica@Gradient Flow
//
Recent advancements in AI reasoning models are demonstrating step-by-step reasoning, self-correction, and multi-step decision-making, opening up application areas that require logical inference and strategic planning. These new models are rapidly evolving, driven by decreasing training costs which enables faster iteration and higher performance. This is also facilitated by the advent of techniques such as model compression, quantization, and distillation, making it possible to run sophisticated models on less powerful hardware.
The competitive landscape is becoming more global, and teams from countries like China are rapidly closing the performance gap, offering diverse approaches and fostering healthy competition. Open AI's deep research feature enables models to generate detailed reports after prolonged inference periods, which directly competes with Google's Gemini 2.0. It appears that one can spend large amounts of money and get continuous and predictable gains with AI models. Recommended read:
References :
@www.trendforce.com
//
References:
Last Week in AI
, bdtechtalks.com
,
OpenAI is enhancing its AI models with improved reasoning and logical capabilities, focusing on chain-of-thought techniques and multimodal approaches. The company's o3-mini model now offers users a more detailed view of its reasoning process, a move towards greater transparency. This change aims to bridge the gap with models like DeepSeek-R1, which already reveals its complete chain-of-thought.
OpenAI also plans to launch the o3 Deep Research agent to both free and ChatGPT Plus users. This advancement will enable the model to generate comprehensive reports following extended inference periods, directly competing with Google's Gemini 2.0 reasoning models. These improvements highlight OpenAI's ongoing efforts to refine AI reasoning and provide users with more insights into the decision-making processes of its models. Recommended read:
References :
Emily Forlini@PCMag Middle East ai
//
References:
www.analyticsvidhya.com
, hackernoon.com
,
DeepSeek is emerging as a notable contender in the AI landscape, challenging established players with its DeepSeek-R1 model. Recent analysis highlights DeepSeek-R1's reasoning capabilities, positioning it as a potential alternative to models like OpenAI's GPT. The company's focus on AI infrastructure and model development, combined with its competitive pricing strategy, is attracting attention and driving its expansion.
The affordability of DeepSeek's models, reportedly up to 9 times cheaper than competitors, is a significant factor in its growing popularity. However, some reports suggest that this lower cost may come with trade-offs in terms of latency and potential server resource constraints, impacting the speed of responses. While DeepSeek is expanding, the Center for Security and Emerging Technology has weighed in on the US and China's race to AI dominance. The DeepSeek-R1 model is built with a Mixture-of-Experts framework that only uses a subset of its parameters per input for high efficiency and scalability. Recommended read:
References :
|
BenchmarksBlogsResearch Tools |