AI Research news and analysis on Web Pulse: 8784 curated articles tracking the latest AI Research developments, tools, and research, updated continuously from vetted sources.
Researchers introduce PuzzleMoE, a method for compressing large Mixture-of-Experts models via fine-grained element-wise merging and bit-packing, achieving up to 16.7% higher accuracy on MMLU at 50% compression and 1.80× …
Jackrong released an open-source knowledge base for LLM fine-tuning, dataset distillation, reinforcement learning, and local deployment. The guide provides reproducible training pipelines, SFT and RL workflows, data prep…
South Korea's Agency for Defense Development launched Add+i, an in-house generative AI service on its closed internal network, developed by the Defense Artificial Intelligence Technology Research Institute using open AI …
Anthropic researchers discovered a 'global workspace' inside the Claude AI model where its reasoning surfaces, and released the Jacobian Lens tool to visualize it. A developer replicated the technique on a small model in…
Hugging Face released LeRobot v0.6.0, introducing world model policies (VLA-JEPA, FastWAM, LingBot-VA) that learn to imagine future states, new vision-language-action models (GR00T N1.7, MolmoAct2, EO-1, EVO1, Multitask …
OpenAI released GPT-5.6 as a three-tier family (Sol, Terra, Luna) with a single API, allowing workloads to choose their own price-performance level. A new paper argued that reinforcement-learning fine-tuning has been opt…
Blogger Sean Goedecke argues that writing controversial blog posts forces deeper learning and clearer thinking, contrary to common advice favoring unstructured self-expression. He claims that structuring posts around arg…
DeepSeek released the MIT-licensed DeepSeek-R1-Distill-Qwen-1.5B model, a 3.1 GB distilled version of DeepSeek R1 based on Qwen 1.5B, designed for small reasoning experiments and local workflows. The model is available o…
Cockroach Labs and IBM introduced the Always-On Data Foundation, combining CockroachDB's distributed SQL with IBM LinuxONE Rockhopper 5 to provide resilient infrastructure for enterprise AI workloads, addressing the need…
Researchers propose a Cross-Referential Rewriter (CRR) caption framework to improve Text-to-Sounding-Video (T2SV) generation by addressing text conditioning bottlenecks and cross-modal fusion challenges. The method uses …
Researchers propose LensVLM, an inference framework that enables Vision Language Models to scan compressed images and selectively expand only relevant ones, achieving up to 10.1× effective compression while maintaining a…
Researchers from UCLA, UT Austin, and Lambda, Inc. introduced MT-EditFlow, a reinforcement learning framework for multi-turn image editing that uses flow matching to optimize reward signals. The framework significantly i…
ChDB embeds a full ClickHouse query engine inside an AI agent's process to eliminate network latency and instability from multi-turn tool calls, enabling local agent memory, a federation hub, and significant token saving…
The Fable Mode skill injects structured reasoning habits from frontier AI models into cheaper models, using a five-gate framework—scope, evidence, attack, verify, report—to improve output quality without high costs. This…
Video-to-video AI editing tools like Gemini Omni allow marketers to transform existing footage by changing outfits, lighting, and backgrounds without reshoots, enabling rapid iteration on ad creative and viral hooks. The…
NVIDIA's new two-tower diffusion language model generates text in parallel blocks instead of token-by-token, achieving 2.4x faster inference with 98.7% quality retention compared to autoregressive models. The architectur…
Meta's AI research division released Brain2QWERTY in early 2025, a system that decodes typed sentences from non-invasive MEG brain recordings with 61% average word accuracy. The model reads neural signals associated with…
ModelBest and Tsinghua University released MiniCPM-5, a 1-billion-parameter on-device AI model with a 128K context window and native tool-use capabilities, designed for agentic workflows on phones and edge hardware. The …
Tencent released Hy3, a 295B-parameter Mixture-of-Experts model with 21B active parameters, outperforming similar-size models and rivaling flagship open-source models with 2-5x parameters. The model is available for free…
Researchers at ICML Mech Interp workshop present a benchmark of backdoored language models to test trigger-recovery methods, finding that defending against good backdoors requires knowing the attack objective and that cu…