A new study published in *Thinking & Reasoning* reveals that the ability to use logical intuition—the "smart intuitor" profile where high intelligence leads to accurate gut instincts—is a developmental milestone that matures throughout adolescence. By testing middle and high school students with probability puzzles, researchers found that while older teenagers can use deliberate thought to correct stereotypical biases, younger students lack the underlying mental strategies to override these instincts even with extra time. This suggests that true seamless logic is not immediate but rather an optimized skill built through years of academic practice and cognitive development.
Meta’s new “semi-formal reasoning” technique boosts LLM accuracy for code tasks (review, bug detection, patching) by having the AI reason through code instead of running it. This involves stating assumptions, tracing steps, and drawing conclusions – a structured process that improves results (up to 93% accuracy) and lowers computing costs.
Microsoft's Phi-4-Reasoning-Vision-15B model challenges the trend of ever-larger AI models by demonstrating strong reasoning capabilities with a comparatively compact size. Trained on curated reasoning data, it aims to achieve performance without the massive compute costs associated with frontier models. The model supports multimodal tasks, combining text and image understanding, and offers flexible reasoning modes for different workloads. This research highlights the importance of data quality and training strategy, suggesting that smarter training techniques can be as impactful as simply increasing model size, particularly for AI agents and practical deployments.
Sarvam AI is releasing Sarvam 30B and Sarvam 105B as open-source models, trained from scratch on large-scale, high-quality datasets. These models demonstrate strong reasoning, programming, and agentic capabilities, with optimizations for efficient deployment across various hardware. Sarvam 30B powers Samvaad, while Sarvam 105B powers Indus. The release includes details on the model architecture, training process, benchmark results, and inference optimizations. The models are available on AI Kosh and Hugging Face, and the article details their performance across benchmarks and in real-world applications like webpage generation, JEE problem solving, and conversational agents.
In this tutorial, we build a hierarchical planner agent using an open-source instruct model. We design a structured multi-agent architecture comprising a planner agent, an executor agent, and an aggregator agent, where each component plays a specialized role in solving complex tasks. We use the planner agent to decompose high-level goals into actionable steps, the executor agent to execute those steps using reasoning or Python tool execution, and the aggregator agent to synthesize results into a coherent final response. By integrating tool usage, structured planning, and iterative execution, we create a fully autonomous agent system that demonstrates how modern AI agents reason, plan, and act in a scalable and modular manner.
Qwen3.5-27B is a powerful, multimodal language model designed for versatility and efficiency. It excels in tasks requiring reasoning, coding, and visual understanding thanks to its unified vision-language foundation and efficient architecture utilizing Gated Delta Networks and sparse Mixture-of-Experts. The model supports 201 languages and boasts a native 262,144 token context window, expandable to 1,010,000.
**Key Specs:**
* **Model Type:** Causal Language Model with Vision Encoder, 27 Billion Parameters
* **Architecture:** 64 Layers, 5120 Hidden Dimension
* **Training:** Scalable Reinforcement Learning for real-world adaptability.
**Performance Highlights:** Qwen3.5-27B demonstrates strong performance across a broad spectrum of benchmarks, including: **Knowledge & Reasoning** (MMLU, C-Eval, HLE, GPQA), **Instruction Following & General Agent Capabilities** (IFEval, IFBench, BFCL-V4, TAU2-Bench), **Coding** (SWE-bench, CodeForces), **Long Context Handling** (AA-LCR, LongBench v2), **Vision-Language Understanding** (MMMU, RealWorldQA), and **Multilingual Abilities** (MMMLU, WMT24++).
**Usage & Deployment:**
The model can be served and utilized through several frameworks: **SGLang & vLLM** (for fast, high-throughput inference with features like Multi-Token Prediction), **KTransformers & Hugging Face Transformers** (offering flexibility and lightweight testing options), and a **Chat Completions API** (with OpenAI SDK examples for various input types).
**Key Considerations:**
* Operates in "thinking mode" by default (intermediate thought processes), which can be disabled.
* Well-suited for agent applications, particularly with the Qwen-Agent framework.
* Documentation provides details on API configuration and recommended sampling parameters.
**Key Specs:**
* **Model Type:** Causal Language Model with Vision Encoder, 27 Billion Parameters
* **Architecture:** 64 Layers, 5120 Hidden Dimension
* **Training:** Scalable Reinforcement Learning for real-world adaptability.
**Performance Highlights:** Qwen3.5-27B demonstrates strong performance across a broad spectrum of benchmarks, including: **Knowledge & Reasoning** (MMLU, C-Eval, HLE, GPQA), **Instruction Following & General Agent Capabilities** (IFEval, IFBench, BFCL-V4, TAU2-Bench), **Coding** (SWE-bench, CodeForces), **Long Context Handling** (AA-LCR, LongBench v2), **Vision-Language Understanding** (MMMU, RealWorldQA), and **Multilingual Abilities** (MMMLU, WMT24++).
**Usage & Deployment:**
The model can be served and utilized through several frameworks: **SGLang & vLLM** (for fast, high-throughput inference with features like Multi-Token Prediction), **KTransformers & Hugging Face Transformers** (offering flexibility and lightweight testing options), and a **Chat Completions API** (with OpenAI SDK examples for various input types).
**Key Considerations:**
* Operates in "thinking mode" by default (intermediate thought processes), which can be disabled.
* Well-suited for agent applications, particularly with the Qwen-Agent framework.
* Documentation provides details on API configuration and recommended sampling parameters.
New research introduces Tri-System Theory to explain how we think with AI. It builds on the idea that we have two main thinking styles: System 1 for fast, intuitive thinking. and System 2 for slow, deliberate thinking.
This new theory adds a System 3: thinking with AI. The study found people often "surrender" to AI, meaning they accept AI's answers without much questioning – even if those answers are wrong. This can sometimes improve performance, but often leads to mistakes.
People who trust AI more, and who don't enjoy deep thinking, are more likely to rely on it. In short, we're increasingly letting AI do some of our thinking, and this has both benefits and risks.
This new theory adds a System 3: thinking with AI. The study found people often "surrender" to AI, meaning they accept AI's answers without much questioning – even if those answers are wrong. This can sometimes improve performance, but often leads to mistakes.
People who trust AI more, and who don't enjoy deep thinking, are more likely to rely on it. In short, we're increasingly letting AI do some of our thinking, and this has both benefits and risks.
Google introduces Gemini 3, its most intelligent AI model, enhancing reasoning and multimodal capabilities. It outperforms previous models in benchmarks and is available across Google products like the Gemini app, AI Studio, and Vertex AI.
OpenAI's release of GPT-OSS marks their first major open source LLM since GPT-2, featuring improvements in reasoning, tool usage, and problem-solving capabilities. The article explores its architecture, message formatting, reasoning modes, and tokenizer details.
Trail of Bits announces the open-sourcing of Buttercup, their AI-driven Cyber Reasoning System (CRS) developed for DARPA’s AI Cyber Challenge (AIxCC). The article details how Buttercup works, including its four main components (Orchestration/UI, Vulnerability discovery, Contextual analysis, and Patch generation), provides instructions for getting started, and outlines future development plans.
