AI makes AI better.

7 picked from 104 candidates · 31 sources read · 57 loops indexed

1. 01
   Loop of the dayPaperarXiv2 Oct 2026Training data

   Recursive Self-Improvement in Unified Multimodal Models

   The loop: A unified multimodal model generates images and writes programs to evaluate them, using the verified results to train its own visual understanding and generation.

   Cross-capability self-improvement allows a multimodal model to use one modality to verify and generate training data for another, preventing error accumulation.

   loop fit 10/10Huijuan Wang, Chufan Shi et al.via arXiv
   
2. 02
   PaperarXiv2 Oct 2026Tools and harness

   VERSE: Verified Self-Evolving Optimizer for Agent Harnesses

   The loop: The VERSE optimizer tests draft edits and replays failures to revise an agent harness alongside its own prompts and tools, improving its ability to diagnose and fix future errors.

   loop fit 9/10Zekai Wang, Yingqiang Ge et al.via arXiv
   
3. 03
   Forum postLessWrong4 Oct 2026Training data

   Reducing Synthetic Markers Makes Some SDF False Facts Linearly Indistinguishable from Pretraining-Acquired Knowledge

   The loop: A synthetic document generator produces training data with reduced markers, which finetunes a model to internalize false facts that evade middle-layer linear probes.

   loop fit 6/10Jason Zengvia LessWrong
   
4. 04
   Blog postPyTorch1 Oct 2026Kernels and compilers

   Optimizing Jagged Flash Attention with TLX: The Road Toward SOTA FA4 on Blackwell

   The loop: The Jagged Flash Attention kernel improves inference efficiency on Blackwell, which makes the Generative Ads Model faster.

   loop fit 6/10Han Xu, Jacky Zhou, Jackie (Jiaqi) Xu, Hongtao Yu, Peng Chen (Dev Infra),...via PyTorch
   
5. 05
   PaperarXiv2 Oct 2026Training data

   Recursive Harness Self-Improvement for Frontier Reasoning Data Synthesis

   The loop: A data synthesis system recursively improves its own generation harness by converting intermediate solver failures into reusable skills, producing progressively harder reasoning tasks.

   loop fit 9/10Wenlong Zhang, Zhengbo Jiao et al.via arXiv
   
6. 06
   Blog postAi21 Oct 2026Tools and harness

   Introducing Olmo-core 3: Open, scalable training infrastructure for large MoEs

   The loop: Olmo-core 3 provides an open training stack that improves the efficiency of training MoE models, which are then used to build better versions of the stack.

   loop fit 6/10via Ai2
   
7. 07
   PaperarXiv29 Sep 2026Self-reward and self-play

   AdviSD: Learning to Advise Frontier LLMs via Targeted Multi-Turn Self-Distillation

   The loop: An advisor model reflects on completed interactions to propose corrections, then uses targeted self-distillation to improve the advice it issues to a frozen language-model executor.

   loop fit 9/10Rishabh Agrawal, Hejie Cui et al.via Hugging Face Papers
   
8. 08
   PaperarXiv27 Sep 2026Self-modifying agents

   Rep2Skill: Representation-Guided Skill Self-Evolution for LLM Agents

   The loop: The Rep2Skill agent models its own internal representation trajectories to localize execution errors, generating textual feedback to evolve its external skills.

   loop fit 9/10Euntae Choi, Su-Min Song et al.via Semantic Scholar