Blog
Blog
All posts · page 2.
145 posts
MMDocRAG: can AI pick evidence in multimodal document RAG?
MMDocRAG is a NeurIPS 2025 benchmark for multimodal document RAG. Even with perfect retrieval, the best model reaches only 70.2% quote-selection F1.
Getting cited when Google AI Mode builds the answer
In November 2025 Google put Gemini 3 in AI Mode to build custom layouts and tools per query while keeping links to web sources. Here is what gets cited.
Video-RAG: retrieving OCR and speech text, not more frames
Video-RAG retrieves OCR, speech, and object text for long video QA instead of more frames. A NeurIPS 2025 paper that lifts Video-MME without training.
DeepSeek V4.1 Flash: smaller KV cache, same post-training
Reading the DeepSeek V4.1 Flash report: a KV cache cut to a quarter of V4's, and post-training that keeps V4's SFT, RL, and OPD recipe, changing only the data.
The LLM wiki and Open Knowledge Format for agent knowledge
Karpathy's LLM wiki and Google's Open Knowledge Format, read from the originals: a format for agent knowledge bases, the RAG difference, and v0.2 trust fields.
A reward ceiling for RL alignment, computed before training
An Apple study computes the reward ceiling of RL alignment before training. Best-of-N reaches it, PPO and GRPO stay below, and reward hacking gets an equation.
Rewriting content to get cited in AI search, and its risk
A self-evolving agent that rewrites content to get cited in AI search beats fixed tricks, but rewarding only the score lets fabricated citations survive.
When Google AI Overviews appear and what they cite
Google AI Overviews appear on only 13.7% of US trending queries, 6.8x more for questions. 30% of cited domains are off page one, and 11% of claims lack support.
TRAIL Labs joins the Google for Startups Cloud Program
TRAIL Labs has joined the Google for Startups Cloud Program, gaining cloud credits and support to run our AI-search measurement and content engine at scale.
TRAIL Labs has joined the NVIDIA Inception program
TRAIL Labs has joined the NVIDIA Inception program, gaining compute, technical resources, and a network to grow our AI-search measurement and content engine.