Today’s frontier report
Research Pool → Multi-Scout → Editor → Edition
3qk7 · AI LOCAL LABS TECHNICAL SCOUTING

AI FRONTIER

Signals worth understanding before they become obvious.

2026-11-13
01
testdeveloperopen source unlock

LLMON: An LLM-Native Markup Language, now in Granite 4.1 3B and 8B

IBM released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base models and trained to consume and generate LLMON-marked text, which separates instructions from data to counter indirect prompt injection.

  • On the OPI benchmark, LLMON models avoided 47% more attacks than the best frontier model tested and 95% more than the best open-source model.
  • LLMON defines six delimiters installed in the tokenizer as reserved single tokens.
  • Training used five sequential GRPO stages followed by a final DPO stage on a thousand on-policy preference pairs.
  • The models are an experimental preview, ready for exploration and evaluation but not for production.

Builders of RAG or tool-calling agents can swap in the LLMON previews for comparable general performance to Granite 4.1 Instruct with stronger prompt-injection resistance.

Hugging Face Blog
02
testcreativeincremental

Fine-Tuning Guide for TRELLIS.2 3D Generation Models: LoRA Adapters and Containers

Hugging Face released fine-tuning scripts and containers for TRELLIS.2 that train LoRA adapters only on the sparse-structure flow DiT and shape flow DiT, targeting sole geometry accuracy on 188 GSO footwear objects.

  • The experiment uses 188 footwear objects from GSO, which contains 1,030 objects in total.
  • Approximately 9.3 million adapter parameters are trained across both targeted DiTs: rank 8 in the attention and MLP layers of the sparse-structure flow DiT, and rank 16 in the attention layers of the shape flow DiT.
  • The preprocessing container renders 16 views per object with Blender, and the full run for 188 objects took approximately 1 hour on an a100-large (80 GB) job.
  • The training container offers ss and shape configs; examples set max_steps=3000 and i_save=500, and adapter checkpoints can be uploaded to the Hub.

For small 3D datasets, freezing the texture flow DiT and adding LoRA only to shape-related DiTs lets builders correct geometry defects with limited compute.

Hugging Face Blog
03
watchresearchcapability unlock

Google Research introduces wearable-driven CKM risk profiling and personalized coaching framework

Google Research presents an investigational machine learning framework that turns passive wearable signals into a modifiable coachable health age, uses an action-conditioned world model to simulate personalized 12-week "what-if" habit trajectories, and expands to routine blood biomarkers to track cardiovascular-kidney-...

  • In a blinded head-to-head study of 49 participant profiles and 490 paired reviews, lifestyle coaches chose the counterfactual-grounded plan as the best overall plan in 71.4% of evaluations (p < 0.001).
  • Physicians rated 99.2% of counterfactual plans within safe exercise intensity limits and 97.6% accurate on physiological claims.
  • The multimodal CKM health age model was evaluated in 819 WEAR-ME participants, combining wearable signals with cholesterol-to-HDL ratio, eGFR, and HOMA-IR.
  • The authors state that evaluations were conducted offline using retrospective research cohorts and that the models are investigational research prototypes, not clinical diagnostic tools.

For AI builders, this shows a closed-loop pattern of using a world model to generate counterfactual trajectories and a language model to convert them into actionable coaching, validated through blinded expert review and clinical safety metr...

Google Research
04
watchagentscapability unlock

Google Antigravity adds Remote Agents that run AI agents in the cloud

Google Antigravity adds a Remote Agents feature in which AI agents run in the cloud and keep working outside working hours.

  • The feature is called Remote Agents, and the AI agents run in the cloud.
  • According to Publickey, the AI agents keep working outside working hours.

For builders, cloud-hosted agents extend execution beyond a single session, so handoff and monitoring must be designed for unattended runs.

Publickey
05
watchresearchbenchmark jump

Study: AI cuts vulnerability search from nine months to 90 hours

According to a study reported by All About SECURITY, AI reduced vulnerability search time from nine months to 90 hours.

  • The report states vulnerability search time dropped from nine months to 90 hours.
  • Source: All About SECURITY.

For AI builders, this reported reduction indicates AI-assisted vulnerability search can substantially compress security research timelines.

All About SECURITY