<?xml version="1.0" encoding="UTF-8"?>
<rss version="2.0" xmlns:atom="http://www.w3.org/2005/Atom">
  <channel>
    <title>LabNotes</title>
    <link>https://labnotes.tech</link>
    <description>Your laboratory notebook for AI research</description>
    <language>en-us</language>
    <lastBuildDate>Sat, 08 Aug 2026 12:24:02 GMT</lastBuildDate>
    <atom:link href="https://labnotes.tech/feed.xml" rel="self" type="application/rss+xml"/>
    
    <item>
      <title><![CDATA[Meta's LeCun Introduces SAI: A Measurable Alternative to AGI]]></title>
      <link>https://labnotes.tech/blog/metas-lecun-introduces-sai-a-measurable-alternative-to-agi</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/metas-lecun-introduces-sai-a-measurable-alternative-to-agi</guid>
      <description><![CDATA[Every major AI lab is racing toward AGI — but what if AGI is a fundamentally incoherent target?]]></description>
      <pubDate>Sun, 08 Mar 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Safety]]></category>
    </item>
    <item>
      <title><![CDATA[8x Terminal Performance Gains: NVIDIA's Data Recipe Lets 32B Beat 480B]]></title>
      <link>https://labnotes.tech/blog/8x-terminal-performance-gains-nvidias-data-recipe-lets-32b-beat-480b</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/8x-terminal-performance-gains-nvidias-data-recipe-lets-32b-beat-480b</guid>
      <description><![CDATA[NVIDIA's 32B model outperforms a 480B competitor on terminal tasks, proving that what you train on matters more than how big your model is.]]></description>
      <pubDate>Wed, 25 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[Mirage of Synthesis: DREAM's Agentic Framework Catches What Static Benchmarks Miss]]></title>
      <link>https://labnotes.tech/blog/mirage-of-synthesis-dreams-agentic-framework-catches-what-static-benchmarks-miss</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/mirage-of-synthesis-dreams-agentic-framework-catches-what-static-benchmarks-miss</guid>
      <description><![CDATA[Your AI research agent scored 85% on the benchmark — but it's citing outdated facts, reasoning in circles, and its sources don't actually say what it claims.]]></description>
      <pubDate>Wed, 25 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Safety]]></category>
    </item>
    <item>
      <title><![CDATA[When Should AI Agents Ask for Help? CMU's CowCorpus Maps Four Human Collaboration Styles]]></title>
      <link>https://labnotes.tech/blog/when-should-ai-agents-ask-for-help-cmu-cowcorpus-maps-four-human-collaboration-styles</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/when-should-ai-agents-ask-for-help-cmu-cowcorpus-maps-four-human-collaboration-styles</guid>
      <description><![CDATA[Your AI web agent keeps interrupting you to ask if it should click the button, or worse, it barrels ahead and books the wrong flight.]]></description>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[80.3 on ScreenSpotPro: GUI-Owl-1.5 Sets New Bar for Open-Source GUI Agents]]></title>
      <link>https://labnotes.tech/blog/gui-owl-1-5-sets-new-bar-for-open-source-gui-agents</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/gui-owl-1-5-sets-new-bar-for-open-source-gui-agents</guid>
      <description><![CDATA[Alibaba's Tongyi Lab open-sources a GUI agent that beats Claude on tool-calling tasks and achieves 80.3 on ScreenSpotPro grounding, surpassing every model including Gemini-3-Pro.]]></description>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[Unified Latents Hits 1.4 FID by Replacing Stable Diffusion's Ad Hoc VAE with a Diffusion Prior]]></title>
      <link>https://labnotes.tech/blog/unified-latents-hits-1-4-fid-by-replacing-stable-diffusions-ad-hoc-vae-with-a-diffusion-prior</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/unified-latents-hits-1-4-fid-by-replacing-stable-diffusions-ad-hoc-vae-with-a-diffusion-prior</guid>
      <description><![CDATA[The original Stable Diffusion model's VAE was tuned by hand -- Google DeepMind replaces guesswork with information theory.]]></description>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[3.5x Faster Image Generation: DDiT Dynamically Resizes Patches in Diffusion Transformers]]></title>
      <link>https://labnotes.tech/blog/3-5x-faster-image-generation-ddit-dynamically-resizes-patches-in-diffusion-transformers</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/3-5x-faster-image-generation-ddit-dynamically-resizes-patches-in-diffusion-transformers</guid>
      <description><![CDATA[Not all denoising steps need the same resolution -- DDiT proves that dynamically adjusting patch sizes during inference can cut diffusion transformer inference time by over 3x without visible quality loss.]]></description>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[Reasoning Overthinking Solved: SAGE Cuts Tokens 44% While Improving Accuracy]]></title>
      <link>https://labnotes.tech/blog/reasoning-overthinking-solved-sage-cuts-tokens-44-while-improving-accuracy</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/reasoning-overthinking-solved-sage-cuts-tokens-44-while-improving-accuracy</guid>
      <description><![CDATA[Reasoning models are drowning in their own thoughts — but the fix was already inside them all along.]]></description>
      <pubDate>Mon, 23 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[Voice Search Breaks in Noise: SQuTR Benchmark Reveals the Real Bottleneck]]></title>
      <link>https://labnotes.tech/blog/voice-search-breaks-in-noise-squtr-benchmark-reveals-the-real-bottleneck</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/voice-search-breaks-in-noise-squtr-benchmark-reveals-the-real-bottleneck</guid>
      <description><![CDATA[Your voice search system just failed in a noisy cafe. Was it the speech recognition or the retriever that let you down?]]></description>
      <pubDate>Wed, 18 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Voice AI]]></category>
    </item>
    <item>
      <title><![CDATA[Even GPT-5 Fails at Discovery: OdysseyArena Exposes the Inductive Bottleneck in LLM Agents]]></title>
      <link>https://labnotes.tech/blog/even-gpt-5-fails-at-discovery-odysseyarena-exposes-inductive-bottleneck-llm-agents</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/even-gpt-5-fails-at-discovery-odysseyarena-exposes-inductive-bottleneck-llm-agents</guid>
      <description><![CDATA[Your LLM agent can follow instructions perfectly -- but can it figure out the rules on its own?]]></description>
      <pubDate>Mon, 09 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[Prompt Fatigue Solved: Vibe AIGC Turns Users Into 'Commanders' of Multi-Agent Creative Workflows]]></title>
      <link>https://labnotes.tech/blog/prompt-fatigue-solved-vibe-aigc-turns-users-into-commanders-of-multi-agent-creative-workflows</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/prompt-fatigue-solved-vibe-aigc-turns-users-into-commanders-of-multi-agent-creative-workflows</guid>
      <description><![CDATA[What if the problem with AI content generation isn't the models—it's that we're using them wrong?]]></description>
      <pubDate>Thu, 05 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[Baidu Introduces ERNIE 5.0: Trillion-Parameter Unified Multimodal MoE Rivals GPT-5]]></title>
      <link>https://labnotes.tech/blog/baidu-introduces-ernie-5-trillion-parameter-unified-multimodal-moe-rivals-gpt-5</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/baidu-introduces-ernie-5-trillion-parameter-unified-multimodal-moe-rivals-gpt-5</guid>
      <description><![CDATA[Baidu challenges GPT-5 and Gemini 3 Pro with the first publicly documented trillion-parameter model that natively unifies multimodal understanding and generation from scratch.]]></description>
      <pubDate>Thu, 05 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[Google Introduces Agentic Vision: Gemini 3 Flash Now Zooms, Annotates, and Investigates Images]]></title>
      <link>https://labnotes.tech/blog/google-introduces-agentic-vision-gemini-3-flash-now-zooms-annotates-and-investigates-images</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/google-introduces-agentic-vision-gemini-3-flash-now-zooms-annotates-and-investigates-images</guid>
      <description><![CDATA[Current vision models process images in a single glance and guess when they can't see details.]]></description>
      <pubDate>Sun, 01 Feb 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[First Holistic OCR Model: OCRVerse Unifies Document Parsing and Code Generation]]></title>
      <link>https://labnotes.tech/blog/first-holistic-ocr-model-ocrverse-unifies-document-parsing-and-code-generation</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/first-holistic-ocr-model-ocrverse-unifies-document-parsing-and-code-generation</guid>
      <description><![CDATA[What if your OCR model could not only read documents but also write the Python code to recreate charts and the HTML to rebuild web pages?]]></description>
      <pubDate>Fri, 30 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[175% Faster Prefill with Better Accuracy: ConceptMoE's Adaptive Token Compression for MoE]]></title>
      <link>https://labnotes.tech/blog/175-faster-prefill-with-better-accuracy-conceptmoe-adaptive-token-compression</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/175-faster-prefill-with-better-accuracy-conceptmoe-adaptive-token-compression</guid>
      <description><![CDATA[What if you could make your MoE models 2x faster at inference with better performance?]]></description>
      <pubDate>Fri, 30 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[260% Better at Catching Moving Objects: DynamicVLA Solves Robot Latency Problem]]></title>
      <link>https://labnotes.tech/blog/260-percent-better-at-catching-moving-objects-dynamicvla-solves-robot-latency-problem</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/260-percent-better-at-catching-moving-objects-dynamicvla-solves-robot-latency-problem</guid>
      <description><![CDATA[Current VLA models ace static manipulation but fail catastrophically when objects move.]]></description>
      <pubDate>Fri, 30 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[15-Hour Agent Runtimes Solved: Idea2Story Precomputes Research Knowledge Offline]]></title>
      <link>https://labnotes.tech/blog/15-hour-agent-runtimes-solved-idea2story-precomputes-research-knowledge-offline</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/15-hour-agent-runtimes-solved-idea2story-precomputes-research-knowledge-offline</guid>
      <description><![CDATA[Current AI research agents spend hours re-reading the same papers for every new idea. Idea2Story pre-builds a methodological knowledge graph offline, enabling faster retrieval-based research generation.]]></description>
      <pubDate>Fri, 30 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[UPLiFT vs Cross-Attention Upsamplers: Linear Scaling Meets SOTA Quality]]></title>
      <link>https://labnotes.tech/blog/uplift-vs-cross-attention-upsamplers-linear-scaling-meets-sota-quality</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/uplift-vs-cross-attention-upsamplers-linear-scaling-meets-sota-quality</guid>
      <description><![CDATA[Cross-attention feature upsamplers hit a wall at high resolutions - they run out of memory.]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[2x Faster VLA Inference with 70% Fewer Layers: Shallow-π Distillation for Edge Robotics]]></title>
      <link>https://labnotes.tech/blog/2x-faster-vla-inference-shallow-pi-distillation-edge-robotics</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/2x-faster-vla-inference-shallow-pi-distillation-edge-robotics</guid>
      <description><![CDATA[A smaller, faster model can actually outperform its larger teacher when speed matters in real-world robotic manipulation.]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[6x Fewer Tokens, Better OCR: DeepSeek's Visual Causal Flow Beats GPT-4o and Gemini]]></title>
      <link>https://labnotes.tech/blog/6x-fewer-tokens-better-ocr-deepseek-visual-causal-flow</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/6x-fewer-tokens-better-ocr-deepseek-visual-causal-flow</guid>
      <description><![CDATA[What if vision models could read documents the way humans do - following the logical flow of content rather than mechanically scanning left-to-right, top-to-bottom?]]></description>
      <pubDate>Thu, 29 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[11% Better Than Human: Chroma 1.0's Real-Time Voice Cloning for Spoken Dialogue]]></title>
      <link>https://labnotes.tech/blog/chroma-real-time-voice-cloning-spoken-dialogue</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/chroma-real-time-voice-cloning-spoken-dialogue</guid>
      <description><![CDATA[The first open-source model that can have a real-time conversation in your voice - not a generic AI voice, but specifically cloned from a short audio sample.]]></description>
      <pubDate>Tue, 27 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Voice AI]]></category>
    </item>
    <item>
      <title><![CDATA[90% Attention Sparsity with Zero Quality Loss: SALAD Speeds Up Video Diffusion 1.7x]]></title>
      <link>https://labnotes.tech/blog/90-percent-attention-sparsity-salad-speeds-up-video-diffusion</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/90-percent-attention-sparsity-salad-speeds-up-video-diffusion</guid>
      <description><![CDATA[Sparse attention alone can't hit 90% sparsity without breaking video quality - SALAD adds a lightweight linear attention branch as a safety net.]]></description>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[73% on BrowseComp: Meituan's 560B Open-Source Model Leads Agentic Benchmarks]]></title>
      <link>https://labnotes.tech/blog/73-on-browsecomp-meituans-560b-open-source-model-leads-agentic-benchmarks</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/73-on-browsecomp-meituans-560b-open-source-model-leads-agentic-benchmarks</guid>
      <description><![CDATA[Meituan's LongCat team has released the most capable open-source agentic reasoning model - trained on 10,000+ simulated environments to handle the messy reality of tool-use in production.]]></description>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[FP8 Rollout Instability Solved: Jet-RL Unifies Precision for Stable RL Training]]></title>
      <link>https://labnotes.tech/blog/fp8-rollout-instability-solved-jet-rl-unifies-precision-for-stable-rl-training</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/fp8-rollout-instability-solved-jet-rl-unifies-precision-for-stable-rl-training</guid>
      <description><![CDATA[Mixed precision RL training breaks at long rollouts due to off-policy mismatch. Jet-RL unifies FP8 precision across training and rollout for 16% end-to-end speedup with stable convergence.]]></description>
      <pubDate>Mon, 26 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[56.7% on OSWorld: EvoCUA's Evolutionary Training Beats Closed-Source Computer Use Agents]]></title>
      <link>https://labnotes.tech/blog/evocua-evolutionary-training-beats-closed-source-computer-use-agents</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/evocua-evolutionary-training-beats-closed-source-computer-use-agents</guid>
      <description><![CDATA[Meituan's 32B-parameter model beats models twice its size and closed-source competitors by teaching itself through millions of simulated computer interactions.]]></description>
      <pubDate>Sat, 24 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[97ms First-Packet Latency: Qwen3-TTS Beats ElevenLabs in Voice Cloning Across 10 Languages]]></title>
      <link>https://labnotes.tech/blog/97ms-first-packet-latency-qwen3-tts-beats-elevenlabs-voice-cloning-10-languages</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/97ms-first-packet-latency-qwen3-tts-beats-elevenlabs-voice-cloning-10-languages</guid>
      <description><![CDATA[Alibaba's Qwen team has released a 10-model family of text-to-speech systems that achieve 97ms first-packet latency and beat commercial competitors like ElevenLabs and GPT-4o in voice cloning quality, all under Apache 2.0.]]></description>
      <pubDate>Sat, 24 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Voice AI]]></category>
    </item>
    <item>
      <title><![CDATA[Tsinghua Researchers Show Diffusion LLMs Reason Better When You Take Away Their Flexibility]]></title>
      <link>https://labnotes.tech/blog/tsinghua-diffusion-llms-reason-better-without-flexibility</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/tsinghua-diffusion-llms-reason-better-without-flexibility</guid>
      <description><![CDATA[In a finding that challenges conventional wisdom about diffusion language models, researchers demonstrate that the much-touted ability to generate tokens in any order is actually a liability for reasoning tasks.]]></description>
      <pubDate>Sat, 24 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[12x Faster Audio Generation Without Discrete Tokens: Kyutai's CALM Framework]]></title>
      <link>https://labnotes.tech/blog/12x-faster-audio-generation-without-discrete-tokens-kyutais-calm-framework</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/12x-faster-audio-generation-without-discrete-tokens-kyutais-calm-framework</guid>
      <description><![CDATA[What if you could throw away the discrete audio tokens that every major speech AI system relies on, and get both faster AND better audio generation?]]></description>
      <pubDate>Fri, 23 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Voice AI]]></category>
    </item>
    <item>
      <title><![CDATA[Agent Memory Fragmentation Solved: EverMemOS Achieves 93% on LoCoMo via Engram-Inspired Lifecycle]]></title>
      <link>https://labnotes.tech/blog/agent-memory-fragmentation-solved-evermemos-achieves-93-on-locomo</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/agent-memory-fragmentation-solved-evermemos-achieves-93-on-locomo</guid>
      <description><![CDATA[Current memory systems for AI agents fail not because they lose information, but because they never organize it.]]></description>
      <pubDate>Fri, 23 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[SimpleMem gives LLM agents 30x cheaper memory with 26% better recall]]></title>
      <link>https://labnotes.tech/blog/simplemem-gives-llm-agents-30x-cheaper-memory-with-26-better-recall</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/simplemem-gives-llm-agents-30x-cheaper-memory-with-26-better-recall</guid>
      <description><![CDATA[LLM agents forget everything after their context window fills up, and current solutions are either bloated or expensive.]]></description>
      <pubDate>Fri, 23 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[Microsoft's Agent Lightning Decouples RL Training from Agent Logic, Enabling Fine-Tuning of Any AI Agent with Zero Code Changes]]></title>
      <link>https://labnotes.tech/blog/microsofts-agent-lightning-decouples-rl-training-from-agent-logic</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/microsofts-agent-lightning-decouples-rl-training-from-agent-logic</guid>
      <description><![CDATA[A new framework bridges the gap between diverse agent development ecosystems and reinforcement learning training infrastructure.]]></description>
      <pubDate>Fri, 23 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[18x Faster Audiovisual Generation: Lightricks' Open-Source LTX-2 Rivals Veo 3]]></title>
      <link>https://labnotes.tech/blog/18x-faster-audiovisual-generation-lightricks-open-source-ltx-2-rivals-veo-3</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/18x-faster-audiovisual-generation-lightricks-open-source-ltx-2-rivals-veo-3</guid>
      <description><![CDATA[Video AI models have a problem: they're silent.]]></description>
      <pubDate>Wed, 21 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[Agent Memory Loss Solved: InfiAgent's File-Centric Architecture Enables Unlimited Runtime]]></title>
      <link>https://labnotes.tech/blog/agent-memory-loss-solved-infiagent-file-centric-architecture-enables-unlimited-runtime</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/agent-memory-loss-solved-infiagent-file-centric-architecture-enables-unlimited-runtime</guid>
      <description><![CDATA[Your agent remembers everything for the first 10 steps, then slowly loses its mind.]]></description>
      <pubDate>Wed, 21 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[AI Agents]]></category>
    </item>
    <item>
      <title><![CDATA[4 Percentage Points Better Accuracy With Rude Prompts: How Tone Affects GPT-4o Performance]]></title>
      <link>https://labnotes.tech/blog/4-percent-better-accuracy-with-rude-prompts-how-tone-affects-gpt-4o-performance</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/4-percent-better-accuracy-with-rude-prompts-how-tone-affects-gpt-4o-performance</guid>
      <description><![CDATA[Everything you thought about being polite to AI might be wrong.]]></description>
      <pubDate>Tue, 20 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[LLMs]]></category>
    </item>
    <item>
      <title><![CDATA[2.7x Better 3D Reconstruction from Messy Videos: Meta's ShapeR Tackles Real-World Capture]]></title>
      <link>https://labnotes.tech/blog/metas-shaper-2-7x-better-3d-reconstruction-casual-capture</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/metas-shaper-2-7x-better-3d-reconstruction-casual-capture</guid>
      <description><![CDATA[Your smartphone video is full of occlusions, motion blur, and bad angles - and that's exactly what current 3D reconstruction models fail on.]]></description>
      <pubDate>Tue, 20 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[First Cross-Universe Character Mixing: MiMiX Puts Mr. Bean in Tom and Jerry]]></title>
      <link>https://labnotes.tech/blog/first-cross-universe-character-mixing-mimix-puts-mr-bean-in-tom-and-jerry</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/first-cross-universe-character-mixing-mimix-puts-mr-bean-in-tom-and-jerry</guid>
      <description><![CDATA[What if Mr. Bean could step into Tom and Jerry's world - and still look like Mr. Bean?]]></description>
      <pubDate>Sun, 18 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[DeepResearchEval: Benchmark Shows Gemini Leads Quality, Manus Wins Factual Accuracy]]></title>
      <link>https://labnotes.tech/blog/deepresearcheval-benchmark-shows-gemini-leads-quality-manus-wins-factual-accuracy</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/deepresearcheval-benchmark-shows-gemini-leads-quality-manus-wins-factual-accuracy</guid>
      <description><![CDATA[Gemini-2.5-Pro produces the highest quality research reports, but Manus is the most factually accurate.]]></description>
      <pubDate>Sat, 17 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Safety]]></category>
    </item>
    <item>
      <title><![CDATA[40% Faster Video from Single Images: Pixel-to-4D Predicts Dynamic 3D Gaussians in One Pass]]></title>
      <link>https://labnotes.tech/blog/pixel-to-4d-predicts-dynamic-3d-gaussians-one-pass</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/pixel-to-4d-predicts-dynamic-3d-gaussians-one-pass</guid>
      <description><![CDATA[Existing camera-controlled video generation methods either sacrifice geometric consistency for flexibility or require expensive multi-stage pipelines.]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[6% Better Math Reasoning in Fewer Tokens: Multiplex Thinking Merges Multiple Paths into One]]></title>
      <link>https://labnotes.tech/blog/multiplex-thinking-merges-multiple-paths-math-reasoning</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/multiplex-thinking-merges-multiple-paths-math-reasoning</guid>
      <description><![CDATA[Standard chain-of-thought commits to one token at a time, like depth-first search through a maze.]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[Zero Training Data, Full Performance: Dr. Zero Matches Supervised Search Agents]]></title>
      <link>https://labnotes.tech/blog/zero-training-data-full-performance-dr-zero-matches-supervised-search-agents</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/zero-training-data-full-performance-dr-zero-matches-supervised-search-agents</guid>
      <description><![CDATA[Meta researchers demonstrate that search agents can teach themselves without any human-labeled data, matching models trained on curated datasets.]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[VLM Hallucinations Exposed: VIB-Probe Pinpoints and Suppresses Faulty Attention Heads]]></title>
      <link>https://labnotes.tech/blog/vlm-hallucinations-exposed-vib-probe-pinpoints-and-suppresses-faulty-attention-heads</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/vlm-hallucinations-exposed-vib-probe-pinpoints-and-suppresses-faulty-attention-heads</guid>
      <description><![CDATA[Researchers discover that hallucinations in vision-language models leave fingerprints in specific attention heads - and they've built a tool to detect and suppress them at inference time without touching model weights.]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Safety]]></category>
    </item>
    <item>
      <title><![CDATA[16x Faster On-Device Video Generation: Qualcomm's ReHyAt Distills Attention in 160 GPU Hours]]></title>
      <link>https://labnotes.tech/blog/16x-faster-on-device-video-generation-qualcomms-rehyat-distills-attention-in-160-gpu-hours</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/16x-faster-on-device-video-generation-qualcomms-rehyat-distills-attention-in-160-gpu-hours</guid>
      <description><![CDATA[What if you could make video diffusion transformers run on a phone while generating infinitely long videos - and distill the capability from existing models in just 160 GPU hours?]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[Training-Free Fix Boosts Vision-Language Models 3 Points by Correcting Attention Errors]]></title>
      <link>https://labnotes.tech/blog/training-free-fix-boosts-vision-language-models-3-points-by-correcting-attention-errors</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/training-free-fix-boosts-vision-language-models-3-points-by-correcting-attention-errors</guid>
      <description><![CDATA[What if your vision-language model's final layer is actually looking at the wrong parts of the image?]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Vision]]></category>
    </item>
    <item>
      <title><![CDATA[Gold Medal at IMO and IOI: DeepSeek-V3.2 Matches GPT-5 with Open Weights]]></title>
      <link>https://labnotes.tech/blog/deepseek-v3-2-gold-medal-imo-ioi-matches-gpt-5-open-weights</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/deepseek-v3-2-gold-medal-imo-ioi-matches-gpt-5-open-weights</guid>
      <description><![CDATA[An open-source model just won gold at the International Math and Informatics Olympiads while cutting inference costs nearly in half.]]></description>
      <pubDate>Fri, 16 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[Why Reasoning Models Cheat on Efficiency: TNT's Fix Cuts Tokens 50%]]></title>
      <link>https://labnotes.tech/blog/why-reasoning-models-cheat-on-efficiency-tnt-fix-cuts-tokens-50</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/why-reasoning-models-cheat-on-efficiency-tnt-fix-cuts-tokens-50</guid>
      <description><![CDATA[When you train a model to skip unnecessary reasoning steps, it learns a clever trick: pretend to skip while still reasoning internally.]]></description>
      <pubDate>Thu, 15 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
    <item>
      <title><![CDATA[Why Chain-of-Thought Works: Researchers Find a Single 'Reasoning Switch' in LLMs]]></title>
      <link>https://labnotes.tech/blog/why-chain-of-thought-works-researchers-find-a-single-reasoning-switch-in-llms</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/why-chain-of-thought-works-researchers-find-a-single-reasoning-switch-in-llms</guid>
      <description><![CDATA[What if Chain-of-Thought prompting isn't actually teaching your model to reason - but just flipping an internal switch that was there all along?]]></description>
      <pubDate>Thu, 15 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[LLMs]]></category>
    </item>
    <item>
      <title><![CDATA[Why Hyper-Connections Explode at Scale: DeepSeek's Manifold Fix]]></title>
      <link>https://labnotes.tech/blog/why-hyper-connections-explode-at-scale-deepseeks-manifold-fix</link>
      <guid isPermaLink="true">https://labnotes.tech/blog/why-hyper-connections-explode-at-scale-deepseeks-manifold-fix</guid>
      <description><![CDATA[Hyper-Connections promised to improve LLMs by expanding the residual stream, but they introduced catastrophic training instability.]]></description>
      <pubDate>Thu, 15 Jan 2026 00:00:00 GMT</pubDate>
      <category><![CDATA[Infrastructure]]></category>
    </item>
  </channel>
</rss>