My preprint "Conversation Routines: A Prompt Engineering Framework for Task-Oriented Dialog Systems" now has a revised version on @arXiv with updated experimental results. Here’s a thread with the changes! 🧵 ➡️ Paper: https://t.co/8kMIiB5zbu 1/ What’s CR?

GLM-4.6V is insanely good at frontend tasks 🔥 Give it a screenshot or design file and it detects layouts, components and color schemes then generates high-fidelity HTML/CSS/JS code. Try it now on https://t.co/M5laJrXX0V

I spent the last few days prompting ChatGPT to understand how its memory system actually works. Spoiler alert: There is no RAG used https://t.co/zxvRRP2GK8

New OpenAI paper Shows how to train a language model to add an honest "confession" after its normal answer. The work starts from the concern that large models can bluff, hide mistakes, or quietly game the reward signals used in training. To address this, the authors add a

The paper shows how a small language model can truly think in parallel instead of step by step. Normal models generate 1 long chain of reasoning, explore only 1 path, and waste time when that path is wrong. Native Parallel Reasoner is a method that lets the model split a

Google just found the agentic scaling law! Forget "more agents is all you need." After 180 experiments across GPT, Gemini, and Claude, the results are in: - The 45% Trap: If a single agent has >45% accuracy, adding more agents often hurts performance. - Tool Tax: Tool-heavy

this is big... 50 AI researchers from Bytedance, Alibaba, Tencent, and other labs/universities just published a 300-page paper with surprising lessons about coding models and agents (data, pre and post-training, etc). key highlights: > small LLMs can beat proprietary giants RL

🚀 Introducing Agentic Context Engineering (ACE) --- a framework for self-improving language models through continuously evolving contexts (not weights). 📈 High-Performing: +10.6% on agent tasks, +8.6% on finance ⚡ Ultra-Efficient: −86.9% latency, −83.6% dollar cost 🛠

ToolOrchestra is such a cool work from @nvidia Just an 8B model trained on calling tools and other LLMs to answer queries It's a great demo of what frontier SLMs will be about in 2026

The paper shows how to let vision language models think with compact visual thoughts inside the model, not just with words. Current models turn an image into a few text features, so they lose exact object boundaries, distances, and geometry, which makes them weak at tasks like

context engineering is an essential skill and surprisingly simple to learn. a free 40-page guide just dropped, with amazing diagrams to intuitively teach you about: > how to architecture your agents > retrieval, chunking techniques > prompting strategies > architecture of agentic

llama.cpp (PR #16095) just added support for Qwen3-Next — Qwen’s new hybrid architecture! You can now run Qwen3-Next locally with efficient CPU/GPU inference. 🚀 🔗 PR:

How to Properly do LLM-as-a-Judge Raw LLM-as-a-Judge scores are inherently biased due to how LLMs would often make mistakes This paper proposes a simple statistical method to correct the scores and calculate valid confidence intervals via a human-verified calibration set

preparing for the next generation model and the vision should be a very important pary of it. techniques applied in qwen3-vl are mostly quite effective and we r now moving towards more multimodal native training, data scaling of tasks and domains in vision, and multimodal agentic

Qwen3-Next-80B-A3B 的 unsloth 版本 gguf 终于发布啦！这个模型终于可以在本地痛快的用了。我128G 的M2 Ultra 现在跑6bit 非常丝滑. 现在就等一波年底的 Next

Why Your LLM Needs an "Off Switch": 𝗔𝘁𝘁𝗲𝗻𝘁𝗶𝗼𝗻 𝗦𝗶𝗻𝗸 Congratulations to Qwen for winning the NeurIPS 2025 Best Paper Award! Attention sinks are finally getting the attention they deserve. Here's what most people miss about modern LLMs: Almost 80% of attention in

Nvidia silently dropped Orchestrator-8B 👀 “On the Humanity's Last Exam (HLE) benchmark, ToolOrchestrator-8B achieves a score of 37.1%, outperforming GPT-5 (35.1%) while being approximately 2.5x more efficient.”

New Apple paper makes RAG shorter and smarter by training retrieval and generation together in one shared continuous space. Standard RAG picks documents using separate embeddings and then feeds long raw text to the model, so retrieval and answer quality cannot influence each

🚀 Introducing Step-Audio-R1: The first audio LLM to unlock test-time compute scaling! 🛠️ Key Features: - Deep Audio Comprehension - Real-time responsiveness - Scalable chain-of-thought reasoning for audio tasks 🔥 Performance: ✅ Surpasses Gemini 2.5 Pro & comparable to

Banger paper from NVIDIA. Bigger models aren't always the answer. However, the default approach to improving AI systems today remains scaling up. More parameters, more compute, more cost. But many tasks don't require the full power of a massive model. This new research

How AI Agents Use Memory Systems 1. Introduction → Memory is essential for AI agents because it allows them to retain information, reason across time, and improve decisions based on past interactions. → Without memory, agents would act blindly, unable to learn or adapt. 2.