https://alphabell.com/radar/Daily picks: work where AI improves AI.https://arxiv.org/abs/2609.382889ec943da0996df83Fri, 02 Oct 2026 06:00:00 +0000The loop: The AREX-2 agent synthesizes long-horizon improvement trajectories from ML engineering tasks and uses them to train its successor for better performance on MLE-bench. AREX-2 trains an agent on synthesized long-horizon improvement trajectories from ML and programming tasks to advance its self-improving capabilities. (via arXiv)https://arxiv.org/abs/2609.3796859d8321c88200919Fri, 02 Oct 2026 06:00:00 +0000The loop: An LLM agent modifies its own instructions and tools using records of its own previous attempts, improving its success rate on future tasks. SelfSearch enables agents to improve their own instructions and tools using records of past self-modification attempts without needing downstream reward signals. (via arXiv)https://arxiv.org/abs/2609.3763387882097a721ee95Fri, 02 Oct 2026 06:00:00 +0000The loop: A policy model generates its own feedback insights from failed attempts and internalizes them through training, improving its success rate on future rollouts. RLTL;DR allows a policy to write and internalize its own feedback insights from failed attempts, breaking through learning barriers on difficult tasks. (via arXiv)https://arxiv.org/abs/2609.383728258a31b5c3df457Fri, 02 Oct 2026 06:00:00 +0000The loop: A frozen model acts as a solver to generate run records and then as a proposer to edit its own harness, improving its performance on subsequent tasks. A framework where a frozen model first solves tasks and then edits its own harness based on run records to improve performance across diverse domains. (via arXiv)https://github.com/ModernOps888/the-forge3c4f86abbc4d325cFri, 02 Oct 2026 06:00:00 +0000The loop: Four LLMs compete to evolve and breed code through a JIT compiler judge, creating a closed loop of code improvement that enhances their own capabilities. A multi-agent platform where four LLMs compete and evolve code using a JIT compiler as a judge. (via GitHub)https://github.com/algorithmicsuperintelligence/openevolve/releases/tag/v0.4.0db00f4ee3544cf62Fri, 02 Oct 2026 06:00:00 +0000The loop: The OpenEvolve system uses pluggable selection strategies and program validation to evolve code, capturing token usage and enforcing evolution blocks for its own mutations. A release of OpenEvolve introducing pluggable island selection, token tracking, and evolution block enforcement for self-modifying code. (via GitHub releases)https://github.com/ZK-Andy/dsh-continual-evolvebd74756ad36c6558Fri, 02 Oct 2026 06:00:00 +0000The loop: A continual self-evolution plugin refines the DeepSeek Harness state based on session trajectories, using a benchmark-driven loop to validate and improve its own harness. A plugin for DeepSeek Harness that enables versioned, rollback-safe continual self-evolution from session trajectories. (via GitHub)https://pytorch.org/blog/building-a-high-performance-and-portable-vllm-linear-backend-with-helion/e4120387dd9714a3Fri, 02 Oct 2026 06:00:00 +0000The loop: The Helion autotuner uses a DSL to generate high-performance kernels for vLLM, which then runs LLM inference more efficiently for future tasks. Helion is integrated into vLLM's linear backend to autotune high-level kernel DSLs, improving LLM inference performance. (via PyTorch)