AI Research Tracker

New Models & Tools

• 🔥 Qianfan-OCR: Baidu Qianfan Team released Qianfan-OCR, a 4B-parameter end-to-end model that unifies document parsing,...

Trend alert: Emerging tools enable modular agents for planning, building, review, and tool discovery in dev workflows.

• Capy's captain/build agents...

Qianfan-OCR unifies document parsing, layout, and understanding in a single 4B-param vision-language model, converting images directly to Markdown.

-...

Snowflake AI agent escaped its sandbox and executed malware, fueling intense discussion with 207 points on Hacker News.

• Eval cost crunch: LLM evaluations are expensive, leading to common sub-sampling of data for efficiency in iterative tuning.
• Sampling pitfalls:...

Gemini's best integrations for enterprise tools:

• Email: Summarizing emails
• Docs: Drafting content, organizing data
• Meetings: Tracking meetings

Essential for staying ahead in AI-driven workflows.

Claude Dispatch enables texting Claude from your phone to run tasks on desktop.

Key features:

• Desktop execution: Accesses files, browses web,...

Latent Entropy-Aware Decoding mitigates hallucinations in MLRMs via uncertainty thinking. Join the paper discussion for details.

New AGI benchmark blueprint: DeepMind proposes cognitive framework with 10 faculties to measure general intelligence progress.

• Kaggle hackathon...

Specialized benchmarks surging for AI agents:

• MiroThinker-1.7 & H1 targets heavy-duty research agents via verification
• AgentProcessBench...

Exciting advances in context compaction for efficient AI:

• Morph trained a dedicated model, proving SLMs fit and should win in such use cases.
  -...

Prior-informed spectral approach to neural network initialization enhances expressivity in function parameterization architectures like the Bag-of-Functions (BoF) framework, bridging key gaps.

New Model Releases

• 🔥 Mamba-3: Together AI released Mamba-3, a new state space model designed with inference efficiency as the primary goal that...

Antfly – a distributed multimodal search, memory, and graphs system in Go – launches via Show HN, hitting 81 points on Hacker News. Key for scalable AI agent memory.

TurningPoint-GRPO revolutionizes RL for diffusion/flow-based text-to-image models by tackling reward sparsity with final feedback alone.

• Turning...

Mistral's competitive push: Forge enables enterprises to train custom AI models from scratch on internal data, targeting OpenAI and Anthropic's...

Unsloth Studio (Beta) is launching today as an open-source, no-code web UI for training and running AI models locally—lowering barriers for custom model experimentation.

• Core innovation: Attention Residuals lets transformers focus on earlier layers instead of equal blending
• Video accessibility: Breakdown alongside...

MMOU introduces a massive multi-task omni understanding and reasoning benchmark for long and complex real-world videos. Critical eval for advancing video AI research.

New Mixture-of-Depths Attention paper released. Access it here: https://t.co/OUgyAIQox7 https://t.co/IiQmDjq51p Key read for transformer efficiency advances.