Alphabell.
Updated daily

Daily radar

New work where AI improves AI, picked each day from public sources by our editor pipeline. Every item states its loop and links to the original.

What the radar picked in the last 14 editions, by part of the AI stack
Agents and harness 39Compute stack 14Data and training signal 20Search 3Research and evals 33Oversight 7

Last run 16 Nov 2026, 06:23 UTC · 30 of 31 sources checked · 59 candidates scored · 6 published · RSS · Sources · How it works

16 Nov 2026

6 picked
PaperarXiv·16 Nov 2026·Tools and harness

ModHarness-RL: Recasting Harnesses as Learnable and Evolving Programs for Agentic RL

The loop: The agentic reinforcement learning system proposes and verifies new harness programs based on its recent trajectories. These updated compression and steering programs then guide the agent's ongoing training, creating a continuous cycle of harness evolution.

An agentic reinforcement learning framework that continuously evolves its own execution harness by learning compression and steering programs from its training trajectories.

loop fit 9/10Siqi Guo, Ming Lin, Tianbao Yangvia arXiv
ModHarness-RL: Recasting Harnesses as Learnable and Evolving Programs for Agentic RL
PaperarXiv·16 Nov 2026·AI R&D evals

WeightBench: Evaluating Language Models as Weight-Space Programmers

The loop: Large language models are evaluated on their ability to read, fix, and write the weights of other neural networks without gradient descent. This benchmark provides the measurement needed to train models that can directly program the weights of their successors.

A benchmark of 42 tasks that evaluates how well frontier language models can directly read, edit, and write the weights of other neural networks.

loop fit 9/10Yuechun Yu, Wenzheng Kong, Xiaofeng Mao, Jin Zhangvia arXiv
WeightBench: Evaluating Language Models as Weight-Space Programmers
PaperarXiv·13 Nov 2026·Self-modifying agents

Evidence-Governed Adoption of Self-Generated Updates in Autonomous AI Agents

The loop: An autonomous agent generates candidate modifications to its own prompts, tools, and workflows based on its interaction history. A governance framework then evaluates these candidates against fresh audits and adopts the successful ones, permanently upgrading the agent's capabilities.

A framework that rigorously evaluates and filters self-generated modifications to an autonomous agent's prompts and tools before permanently adopting them.

loop fit 9/10Hoang Tran Vuong, Linh Ngo Van, Dang Nguyen et al.via arXiv
Evidence-Governed Adoption of Self-Generated Updates in Autonomous AI Agents
PaperarXiv·16 Nov 2026·Kernels and compilers

DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation

The loop: A large language model is trained via reinforcement learning to convert PyTorch programs into optimized Triton kernels. These kernels are then compiled to CUDA to accelerate the execution of the model itself.

A scalable learning framework that uses reinforcement learning to train large language models to generate highly optimized Triton kernels from PyTorch reference implementations.

loop fit 9/10Zelin He, Siyu Zhu, Jacob Helwig, Yujian Liu, Li An, Rohit Jain, Shiyu Chang,...via arXiv
DRTriton: Large-Scale Synthetic Data Driven Reinforcement Learning for Triton Kernel Generation
PaperarXiv·12 Nov 2026·Research automation

A step toward defining the Standard Model on a lattice

The loop: GPT-6 Astra assists researchers in generating a detailed proof for a nonperturbative lattice definition of the Standard Model. This successful application of AI to advanced theoretical physics generates high-quality reasoning trajectories that serve as training data for future models.

OpenAI details how GPT-6 Astra helped researchers turn a candidate definition of the Standard Model into a detailed mathematical proof.

loop fit 6/10Subodh Lonkarvia Hugging Face Papers
A step toward defining the Standard Model on a lattice
Blog postOpenAI·15 Nov 2026·Research automation

A step toward defining the Standard Model on a lattice

The loop: GPT-6 Astra assists in generating a detailed proof for the Standard Model, producing high quality reasoning data for its successors.

Explore a candidate nonperturbative lattice definition of the Standard Model and how GPT-6 Astra helped turn the idea into a detailed proof.

loop fit 6/10via OpenAI
A step toward defining the Standard Model on a lattice

15 Nov 2026

5 picked
PaperarXiv·11 Nov 2026·AI R&D evals

MLIRBench: A Benchmark and Dataset for Compiler and Hardware-Aware Reasoning

The loop: A multi-agent pipeline generates and validates program mutations to create a benchmark for evaluating how well LLMs reason about intermediate representations. This dataset directly measures and guides the development of models used to build better compilers and hardware.

A multi-agent pipeline generates a benchmark of validated program mutations to evaluate how well language models reason about compiler intermediate representations.

loop fit 9/10Pranava Sai Maganti, Meghasyam Peddireddy, Muzakker Hossain et al.via Hugging Face Papers
MLIRBench: A Benchmark and Dataset for Compiler and Hardware-Aware Reasoning
PaperarXiv·12 Nov 2026·Architecture and optimizer search

AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms

The loop: An LLM searches over explicit algorithm assumptions to discover and prove new ML system algorithms. The search produces state-of-the-art algorithms for KV-cache eviction and weight pruning, which improve the efficiency of the models performing the search.

An LLM-driven harness searches over explicit algorithm assumptions to discover and formally prove new ML system optimizations like KV-cache eviction.

loop fit 8/10Yuxuan Chen, Xiaojie Li, Hao Chen et al.via arXiv
AutoAssumption: Searching over Assumptions Discovers ML-Systems Algorithms
PaperarXiv·11 Nov 2026·Training data

Human or Synthetic? What a Fixed Budget Buys for Post-Training

The loop: Models generate synthetic preference data to train their successors through Direct Preference Optimization. The resulting models achieve higher win rates at lower costs than those trained on human-annotated data.

A study demonstrates that synthetic preference data is far more cost-efficient than human annotation for post-training language models at low-to-moderate budgets.

loop fit 8/10Ziyuan Li, Zihe Song, Xiaofang Qi et al.via arXiv
Human or Synthetic? What a Fixed Budget Buys for Post-Training
PaperarXiv·11 Nov 2026·Tools and harness

Bug Finding With Agentic Symbolic Execution

The loop: An AI coding agent synthesizes constraint-aware environments and instruments repositories with assertions to guide a symbolic execution tool. This improves the tool's ability to find and reproduce bugs without human intervention, creating a better debugging harness for future agents.

An AI coding agent orchestrates a symbolic execution tool by automatically synthesizing environments and assertions to find reproducible bugs in Python repositories.

loop fit 7/10Haojun Tang, Wenda Zhao, Xiang Zhou et al.via arXiv
Bug Finding With Agentic Symbolic Execution
PaperarXiv·12 Nov 2026·Self-modifying agents

S1-Evo: Experience-Guided Evidence Competition Attacks on Typed System-1 Decision Models

The loop: A self-evolving framework optimizes attack candidates via feedback search and distills the verified results into experience cards. It then retrieves these cards to construct more effective perturbations against System-1 decision models on unseen tasks.

A self-evolving framework generates adversarial perturbations against System-1 decision models by distilling successful attacks into reusable experience cards.

loop fit 7/10Arnab Sanyal, Gourav Datta, Sandeep P. Chinchalivia arXiv
S1-Evo: Experience-Guided Evidence Competition Attacks on Typed System-1 Decision Models

14 Nov 2026

6 picked
PaperarXiv·11 Nov 2026·Tools and harness

Experience-Handoff Reinforcement Learning for Recursive Self-Improvement

The loop: An LLM agent generates and curates task-local playbooks from its own environment experiences. These playbooks are then handed off to the agent's future rollouts, improving their performance on agentic tasks.

A reinforcement learning method that trains an agent to curate playbooks from its own execution logs to guide its future rollouts.

loop fit 9/10Binghang Lu, Lang Cao, Changhong Mou et al.via arXiv
Experience-Handoff Reinforcement Learning for Recursive Self-Improvement
PaperarXiv·9 Nov 2026·AI R&D evals

BioXArena-Pro: A Research-Grade Benchmark for Autonomous Biomedical Machine Learning Agents

The loop: Autonomous ML agents attempt to build biomedical machine learning models from scratch. The benchmark measures their success at this ML engineering work, providing a rigorous evaluation target to drive the development of better ML agents.

A benchmark of 64 research-level modeling tasks designed to evaluate autonomous agents on biomedical machine learning engineering.

loop fit 9/10Weimin Wu, Han Liuvia arXiv
BioXArena-Pro: A Research-Grade Benchmark for Autonomous Biomedical Machine Learning Agents
PaperarXiv·12 Nov 2026·Hardware and chips

Toward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents

The loop: A multi-agent framework automates RTL rewriting and critical-path analysis for hardware design. By continuously refining a library of reusable optimization skills, the agent improves its own ability to optimize future chips.

A multi-agent framework that automates RTL rewriting and synthesis optimization while building a reusable library of hardware design skills.

loop fit 8/10Doyoung Kim, Byeongguk Jeon, Kailin Zhang et al.via Hugging Face Papers
Toward Autonomous RTL and Synthesis Optimization: Benchmarking and Self-Improving Agents
PaperarXiv·11 Nov 2026·Architecture and optimizer search

NCP: Verifier-Guided Optimization of Weight-Space Control Policies

The loop: The Neural Control Plane observes a language model's execution telemetry to synthesize candidate weight-space intervention rules. It evaluates these candidates and admits beneficial policies to its active rule set, directly improving the model's subsequent inference performance.

A closed-loop system that observes model execution to synthesize, evaluate, and apply temporary weight-space interventions during inference.

loop fit 8/10Alexander Yue, Qibin Liu, Julia Gonskivia arXiv
NCP: Verifier-Guided Optimization of Weight-Space Control Policies
PaperarXiv·11 Nov 2026·Research automation

Updates From METR Report: Short Timelines vs IE

The loop: AI systems increasingly automate AI research and development tasks. This automation expands the total R&D effort, which accelerates the creation of more capable successor systems.

A forum discussion analyzing whether the feedback loop of AI automating AI research is strong enough to trigger an intelligence explosion.

loop fit 6/10Daniela Gottesman, Gal Yona, Ori Ram et al.via arXiv
Updates From METR Report: Short Timelines vs IE
Forum postLessWrong·12 Nov 2026·Research automation

Updates from METR report: short timelines vs IE

The loop: AI systems automate AI research and development tasks, which increases the total R&D effort and leads to more capable AI systems for further automation.

It's important to distinguish between two different takeaways you could have from the recent METR report: 1. Shorter timelines: AI will soon fully automate AI R&D 2. Intelligence explosion more likely: Once AI fully automates AI R&D, AI...

loop fit 6/10Tom Davidsonvia LessWrong
Updates from METR report: short timelines vs IE

13 Nov 2026

7 picked
PaperarXiv·12 Nov 2026·AI R&D evals

Singularity-Bench: Can Agents Build Agents?

The loop: Frontier AI agents are tasked with implementing core AI development systems from scratch, including pre-training, post-training, and agent harnesses. By successfully building and optimizing these pipelines, the agents directly construct the infrastructure required to train and run their own successors.

An open-ended benchmark that evaluates whether frontier models can autonomously build and optimize the core infrastructure of AI development, from pre-training pipelines to agent harnesses.

loop fit 10/10Yang Zhou, Krish Agarwal, Zhuoming Chen et al.via arXiv
Singularity-Bench: Can Agents Build Agents?
PaperarXiv·8 Nov 2026·Tools and harness

JoMoHa: Joint Model-Harness Optimization with Reinforcement Learning

The loop: JoMoHa uses reinforcement learning on its own execution trajectories to jointly update its task execution policy and its harness generation policy. The improved harness generation policy creates better runtime environments for the execution policy, leading to higher success rates on subsequent tasks.

A reinforcement learning framework that jointly trains a model's task execution policy and its harness generation policy using the system's own trajectories.

loop fit 9/10Xudong Wu, Guangchen Lan, Mingqi Yuan et al.via arXiv
JoMoHa: Joint Model-Harness Optimization with Reinforcement Learning
PaperarXiv·11 Nov 2026·Self-reward and self-play

Verify by Doing: Co-Evolving Agent Execution and Evaluation without External Feedback

The loop: The system uses an Evaluator to actively inspect artifacts and environment states, generating an internal signal from the Actor's outputs. This signal is then used to update and refine both the execution procedures of the Actor and the verification procedures of the Evaluator.

A method that co-evolves an agent's execution and evaluation harnesses within a single task by using internal verification signals to guide updates without external feedback.

loop fit 9/10Yitong Lin, Jianda Chen, Yihao Hu et al.via arXiv
Verify by Doing: Co-Evolving Agent Execution and Evaluation without External Feedback
PaperarXiv·12 Nov 2026·Self-modifying agents

Discovering Effective Agent Memory Designs through Program Evolution

The loop: An LLM-based evolution framework iteratively diagnoses failures in agent memory programs and proposes code edits to resolve them. The resulting evolved memory programs are deployed back into the agents, improving their accuracy and reducing serving costs.

A framework that evolves agent memory designs from minimal seeds through an iterative diagnosis and proposal loop, discovering patterns that improve accuracy and efficiency.

loop fit 9/10Xinheng Ding, Seyyed Saeid Cheshmi, Ziyang Zhang et al.via arXiv
Discovering Effective Agent Memory Designs through Program Evolution
PaperarXiv·10 Nov 2026·Training data

KernelMosaic: An Open Dataset and Agentic Framework for CUDA Kernel Generation

The loop: An agentic framework implements, repairs, and optimizes CUDA kernels using execution feedback to build a dataset of verified solutions and trajectories. This dataset is then used to fine-tune a smaller open-source model, dramatically increasing its success rate at generating efficient kernels.

An agentic framework that uses execution feedback to generate a dataset of verified CUDA kernels and optimization trajectories, which is then used to fine-tune a smaller model for kernel generation.

loop fit 8/10Ningyuan Xi, Yuhang Lai, Ning Miaovia arXiv
KernelMosaic: An Open Dataset and Agentic Framework for CUDA Kernel Generation
PaperarXiv·9 Nov 2026·Research automation

HyPLM Technical Report

The loop: The Hyra 2.0 research agent autonomously trains HyPLM-Flash, a highly efficient 650-million-parameter protein language model. The resulting model provides a faster and more precise tool for protein annotation and design, which the agent can utilize in subsequent research cycles.

An AI research agent autonomously trains and evaluates a highly efficient 650-million-parameter protein language model that matches the performance of larger models.

loop fit 8/10Haowei Lin, Cai Zhou, Xiangxin Zhouvia arXiv
HyPLM Technical Report
Blog postPyTorch·12 Nov 2026·Kernels and compilers

Fast by Default: Compile-Time Heuristics for Helion Kernels

The loop: Helion heuristics analyze kernel IR to select optimal settings, improving kernel performance which accelerates the execution of AI models and compilers.

TL;DR Helion now uses compile-time heuristics that analyze a kernel’s IR and dataflow to pick fast settings without slow autotuning. It classifies the kernel as matmul, reduction, or pointwise and... The post Fast by Default: Compile-Time...

loop fit 7/10Caleb Min Sup Kim, Ethan Che, Jongsok Choi, Jason Anselvia PyTorch
Fast by Default: Compile-Time Heuristics for Helion Kernels

12 Nov 2026

7 picked
PaperarXiv·4 Nov 2026·Training data

Micius-Flash: Lightweight Agentic Intelligence for Engineering

The loop: Micius-Flash generates synthetic data and reasoning sequences for hardware coding and semiconductor design. These outputs are used in a multi-teacher on-policy distillation process to train a unified successor model.

A 27-billion-parameter model generates synthetic data and reasoning sequences to train its successor via multi-teacher on-policy distillation.

loop fit 9/10Hongcheng Gao, Weiheng Li, Yipeng Wei et al.via Hugging Face Papers
Micius-Flash: Lightweight Agentic Intelligence for Engineering
PaperarXiv·8 Nov 2026·Tools and harness

InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure

The loop: The InfraRSI agent co-optimizes launch knobs and code patches for distributed LLM training. This improves the underlying training infrastructure the agent relies on.

An agentic harness turns human commands into verified training optimizations by co-optimizing launch settings and code patches for distributed LLM training.

loop fit 9/10InfraRSI Team, Yuchuan Tian, Haocheng Mei et al.via arXiv
InfraRSI: Harnessing Recursive Self-Improvement for Training Infrastructure
PaperarXiv·11 Nov 2026·Research automation

Autonomous Optimization of Numerical Algorithms by an AI Agent: Application to Atomistic Simulation

The loop: An AI agent rewrites and redesigns numerical algorithms against a pre-specified metric in a guarded loop. The resulting efficiency gains improve the underlying mathematical tools the agent uses for optimization.

A coding agent autonomously rewrites and redesigns numerical algorithms in the ASE atomistic simulation library to run faster and converge with fewer evaluations.

loop fit 9/10Jeheon Woovia arXiv
Autonomous Optimization of Numerical Algorithms by an AI Agent: Application to Atomistic Simulation
PaperarXiv·9 Nov 2026·Kernels and compilers

Correct First, Fast Second: What a Local 7B Code Model Can and Cannot Do for GPU Kernel Optimization

The loop: A multi-agent loop generates, evaluates, and refines custom CUDA kernels to replace PyTorch operators. These kernels improve the execution speed of the model on its own hardware.

A study evaluates how different interaction strategies affect a local 7B parameter code model's ability to write custom CUDA kernels that outperform eager PyTorch.

loop fit 8/10Jui-Chien Tsou, Po-Yi Wuvia arXiv
Correct First, Fast Second: What a Local 7B Code Model Can and Cannot Do for GPU Kernel Optimization
Forum postLessWrong·11 Nov 2026·Interpretability and oversight

The Chain Lens [interim]

The loop: The chain lens extracts multi-token concepts from a model's latent representations using contrastive decoding. This surfaces specific computational intermediates that can be used to oversee and refine the model's reasoning.

A new interpretability technique extends single-token methods by using contrastive decoding on an MTP head to recover multi-token concepts from model activations.

loop fit 6/10vgelvia LessWrong
The Chain Lens [interim]
Forum postLessWrong·11 Nov 2026·Research automation

How I'm thinking about different types of automated alignment

The loop: An AI system automates empirical alignment experiments, producing data and insights that are used to make the next generation of AI systems more aligned.

I used to think of "automating alignment" as a natural category. But when I tried to think through the strategic implications, I realised that it depends a lot on the type of work. I'm now thinking of it in three buckets: 1. Empirical...

loop fit 6/10owencbvia LessWrong
How I'm thinking about different types of automated alignment

11 Nov 2026

7 picked
PaperarXiv·7 Nov 2026·Training data

No Second Run: Does RSI Stop at Frontier Pre-training?

The loop: The system decomposes frontier pre-training into smaller loops that repair broken steps, which improves the data pipelines and infrastructure, making the resulting foundation model better.

A framework that applies recursive self-improvement to frontier pre-training by decomposing the unrepeatable run into smaller, verifiable loops that repair data pipelines and infrastructure.

loop fit 9/10Xiaojie Cai, Jiayi Chen, Junwei Chen et al.via arXiv
No Second Run: Does RSI Stop at Frontier Pre-training?
PaperarXiv·11 Nov 2026·Training data

One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data

The loop: A 27B subword model generates on-policy dialogue and behavioral supervision data, which is used to train its conversion into a pure byte-level model, making its successor capable of byte-level processing.

A 27B subword model is converted into a pure byte-level model without architectural patching by training it on supervised bytes generated entirely by the original model itself.

loop fit 9/10Liwei Zhang, ZhiMing Tongvia arXiv
One Byte, One Token: Byteifying a 27B Hybrid-Attention LLM with Self-Generated Data
PaperarXiv·10 Nov 2026·Self-modifying agents

When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents

The loop: An LLM agent generates structured hypotheses from its experiences and tests them through a falsification-gated process, which improves its persistent memory, making the agent more successful and safer at future tasks.

A verifier-grounded framework that allows LLM agents to test experience-derived behavioral updates before committing them to persistent memory, ensuring only reusable and safe knowledge is retained.

loop fit 9/10Runxuan Tang, Haoyu Gao, Yuyan Ding et al.via arXiv
When Should Agents Remember? Falsification-Gated Self-Evolution for LLM Agents
PaperarXiv·11 Nov 2026·Tools and harness

Compile-Gated Lean Proof Refactoring with Enterprise Agent Practices: A Model per Step under a US$3 Cap, Followed by Programmatic Rewrite Rules

The loop: Vexorium-Agent records its proof attempts and converts learned patterns into compiler-checked programmatic rewrite rules, which improves its own harness, making the agent more effective at solving future Lean proofs.

An agentic system for Lean proofs that achieves high performance by routing steps to cost-effective models and converting learned patterns into compiler-checked programmatic rewrite rules.

loop fit 9/10Bobber Chengvia arXiv
Compile-Gated Lean Proof Refactoring with Enterprise Agent Practices: A Model per Step under a US$3 Cap, Followed by Programmatic Rewrite Rules
ProjectGitHub·8 Nov 2026·Tools and harness

TokenRhythm/InfraRSI

The loop: The InfraRSI agentic harness co-optimizes launch knobs and code patches for distributed training, which improves the training infrastructure, making the system more efficient at running its own optimization tasks.

An agentic harness that turns open-ended human commands into verified training optimizations by co-optimizing launch settings and code patches for distributed LLM training.

loop fit 8/10TokenRhythmvia GitHub
TokenRhythm/InfraRSI
ProjectGitHub·5 Nov 2026·Research automation

zou-group/BVSR

The loop: The BVSR framework applies evolutionary principles to the generation and selection of creative research ideas, which improves the AI scientists, making them better at autonomous scientific discovery.

An official implementation of a framework that improves creative idea generation in AI scientists through blind variation and selective retention.

loop fit 7/10zou-groupvia GitHub
zou-group/BVSR

10 Nov 2026

7 picked
PaperarXiv·9 Nov 2026·Research automation

Where is all the AI safety research uplift?

The loop: Frontier AI models perform research and engineering tasks to accelerate capability progress. This increased development velocity directly speeds up the creation of successor models, which are in turn more capable at research and engineering.

An analysis estimating that Anthropic is achieving a significant acceleration in AI capability progress by using its own models to speed up research.

loop fit 9/10Yifan Zhang, Hengxuan Lou, Xinkui Zhao, Rengrong Xiong, Naibo Wang, Chang Liu,...via arXiv
Where is all the AI safety research uplift?
PaperarXiv·9 Nov 2026·Kernels and compilers

MorphSched: Workload-Aware Adaptive Scheduler Generation for Linux

The loop: An LLM synthesizes custom Linux schedulers and profiles their performance on target workloads. The system accumulates this performance feedback in an evolutionary memory pool to refine its own generation strategy for future schedulers.

An LLM-driven system that synthesizes Rust and BPF Linux schedulers and uses runtime performance feedback to evolve its generation strategy.

loop fit 9/10Shervin Shahidi, James Poon, Hang Xie, Kenneth O. Xu, Robin H{\aa}kanson,...via arXiv
MorphSched: Workload-Aware Adaptive Scheduler Generation for Linux
PaperarXiv·9 Nov 2026·Self-reward and self-play

AutoRefine: Auto-Calibrating Code Judges from Workflow Outcomes

The loop: A teacher model observes noisy workflow outcomes like test failures and rewrites the prompt of a deployable code judge to reproduce those labels. This auto-calibration improves the judge's accuracy in evaluating future code changes.

A loop that uses free but noisy development workflow outcomes to automatically recalibrate the prompts of LLM code judges.

loop fit 9/10Weijia Xu, Alessandro Sordoni, Chandan Singh, Zelalem Gero, Michel Galley,...via arXiv
AutoRefine: Auto-Calibrating Code Judges from Workflow Outcomes
NewsAlignment Forum·9 Nov 2026·Self-modifying agents

Test-Time Learning with an Evolving Library

The loop: A language model extracts modular skills and reflective insights from its own inference trajectories to build a shared library. It then reuses and evolves these knowledge abstractions to improve its performance on subsequent problem instances without parameter updates.

A test-time learning framework where a model accumulates and evolves a library of knowledge abstractions from its own successful inferences.

loop fit 9/10Buckvia Alignment Forum
Test-Time Learning with an Evolving Library
PaperarXiv·8 Nov 2026·Tools and harness

Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents

The loop: A multimodal language model writes and refines its own text memories based on its execution experience in embodied tasks. It updates this working memory with lessons on procedure and recovery, making the same system better at navigation and manipulation.

An approach where an embodied agent improves its success rate by writing and refining its own text-based working memory from execution experience.

loop fit 8/10Yifei Yao, Yude Zou, Xing Gao et al.via Hugging Face Papers
Review the Episodes, Refine the Lessons: Simple Memory Evolution for Embodied Agents
PaperarXiv·9 Nov 2026·AI R&D evals

Always-On-Intelligence/HALO

The loop: An evaluation harness audits whether self-evolving agents successfully learn from their outcomes. This measurement provides the necessary signal to improve the agents' learning mechanisms, making them better at self-evolution.

An evaluation harness and plugin designed to audit whether always-on personal AI assistants actually learn from their outcomes.

loop fit 8/10Haozhen Zhang, Quanyu Long, Jianzhu Bao, Tao Feng, Weizhi Zhang, Haodong Yue,...via arXiv
Always-On-Intelligence/HALO
ProjectGitHub·5 Oct 2026·Inference efficiency

TailSieve: Partial-Rollout-Guided Tail Routing for LLM Rollouts

The loop: A controller optimizes the routing of long-tail generations during large-scale rollouts. This improves the efficiency of reinforcement learning post-training and distillation, which makes the training of successor models faster.

A framework that uses partial-rollout completion status to optimize tail routing and replica allocation during synchronous LLM rollouts.

loop fit 8/10Always-On-Intelligencevia GitHub
TailSieve: Partial-Rollout-Guided Tail Routing for LLM Rollouts

9 Nov 2026

12 picked
NewsAlignment Forum·8 Nov 2026·Research automation

Ajeya: AI is dramatically accelerating AI research right now

The loop: Anthropic's AI systems automate AI R&D, which accelerates capabilities progress, leading to better successor models that are even better at automating AI R&D.

An internal investigation at Anthropic suggests the company has crossed its "dramatic acceleration" threshold, with capabilities progress doubling over the past year due to AI automating AI R&D.

loop fit 10/10Zach Stein-Perlmanvia Alignment Forum
Ajeya: AI is dramatically accelerating AI research right now
PaperarXiv·6 Nov 2026·Tools and harness

Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents

The loop: An improver agent revises its own harness and executor based on task feedback, which improves the agent's ability to solve vision and robotics tasks in subsequent rounds.

Flex-RSI is a benchmark that evaluates whether agents can use task feedback to recursively improve their own prompts, tools, and control loops on vision and robotics tasks.

loop fit 9/10Jingdong Zhang, Jionghao Wang, Zhengming Yu et al.via arXiv
Flex-RSI: Benchmarking Recursive Self-Improvement of Frontier Agents
PaperarXiv·9 Nov 2026·Research automation

ReproBridge: Building Reproducibility into ML Research Artifacts

The loop: The ReproBridge agent constructs reproducible research artifacts, which improves the ability of successor agents to reproduce and build on ML research.

ReproBridge is an agent workflow that refines raw code into standardized, reproducible artifacts, significantly raising the success rate of subsequent agents attempting to reproduce the experiments.

loop fit 9/10Pranjal Srivastavavia arXiv
ReproBridge: Building Reproducibility into ML Research Artifacts
PaperarXiv·6 Nov 2026·Kernels and compilers

MorphSched: Workload-Aware Adaptive Scheduler Generation for Linux

The loop: MorphSched uses an LLM to synthesize Linux schedulers, which improves system performance, and uses that performance feedback to refine its generation strategy.

MorphSched is an LLM-driven system that synthesizes custom Rust and BPF schedulers for Linux, using performance feedback to evolve its scheduling strategies.

loop fit 9/10Baolin Chong, Dongfang Xu, Qiaosi Wang et al.via arXiv
MorphSched: Workload-Aware Adaptive Scheduler Generation for Linux
PaperarXiv·6 Nov 2026·Self-modifying agents

Test-Time Learning with an Evolving Library

The loop: EvoLib extracts knowledge abstractions from its own inference trajectories to build a shared library, which improves the model's performance on subsequent problem instances.

EvoLib is a test-time learning framework that allows large language models to accumulate and evolve reusable skills and insights from their own inference trajectories without parameter updates.

loop fit 9/10Yuhao Tan, Hanyun Jiang, Fangkai Yang et al.via arXiv
Test-Time Learning with an Evolving Library
PaperarXiv·9 Nov 2026·Tools and harness

meat-daemon-404/dsh-plugin-immune

The loop: A coding agent converts its own failed tool calls into immunity rules, which are injected to prevent repeated mistakes, improving its future performance.

A GitHub repository demonstrating a self-evolving immunity system for coding agents that turns failed tool calls into rules to prevent future errors.

loop fit 7/10Jianuo Shen, Pengyi Xu, Bo Zhangvia arXiv
meat-daemon-404/dsh-plugin-immune
PaperarXiv·6 Nov 2026·Self-modifying agents

The components of agent memory

The loop: Claude uses a dreaming mechanism to organize memory and correct mistakes, which improves its internal state and makes it better at learning from experience.

Anthropic details how Claude uses a "dreaming" mechanism to organize its memory and correct mistakes, helping the model learn from experience.

loop fit 7/10Mengsong Wu, Tianchen Zhou, Zhaochen Su, Zhenhua Liu, Zhe Cui, Aimin Zhouvia arXiv
The components of agent memory
PaperarXiv·6 Nov 2026·Interpretability and oversight

J-lens preferentially reads from coordinate-aligned subspaces

The loop: The J-lens tool identifies coordinate-aligned subspaces in Gemma-4, which improves model interpretability and editing, making future versions of Gemma easier to verify and steer.

A LessWrong post observes that the J-lens readout in Gemma-4 primarily relies on two small, coordinate-aligned subspaces that are protected from edits across many layers.

loop fit 6/10Yubo Wang, Rui Meng, Jaehyun Nam, Tomas Pfister, Jinsung Yoonvia arXiv
J-lens preferentially reads from coordinate-aligned subspaces
ProjectGitHub·4 Nov 2026·Tools and harness

meat-daemon-404/dsh-plugin-immune

The loop: A coding agent converts its own failed tool calls into immunity rules, which improves the agent's future performance by preventing repeated mistakes.

Self-evolving pitfall immunity for coding agents: a failed-then-fixed tool call becomes a rule, injected before the mistake repeats. Try: npx dsh-immune-system demo

loop fit 7/10meat-daemon-404via GitHub
meat-daemon-404/dsh-plugin-immune
Blog postAnthropic·6 Nov 2026·Self-modifying agents

The components of agent memory

The loop: Claude uses a dreaming mechanism to organize memory and correct mistakes, which improves its internal state, which makes the model better at learning from experience.

For Claude to learn from experience, it has to decide what’s worth remembering before it knows what’s next. Here’s how “dreaming” helps it organize memory and correct mistakes.

loop fit 7/10via Anthropic
The components of agent memory
Forum postLessWrong·5 Nov 2026·Interpretability and oversight

J-lens preferentially reads from coordinate-aligned subspaces

The loop: The J-lens tool identifies coordinate-aligned subspaces in Gemma-4, which improves model interpretability and editing, making future versions of Gemma easier to verify and steer.

Note: I only test on Gemma-4-31B, which contains the spaces I've been studying in this earlier post. This is a shorter post, with just a few observations potentially relevant to those working on the J-lens work. The main observations here...

loop fit 6/10Caleb Briggsvia LessWrong
J-lens preferentially reads from coordinate-aligned subspaces
Blog postOpenAI·6 Nov 2026·Interpretability and oversight

Partnering with Hacktron to find vulnerabilities in OpenAI’s systems

The loop: Frontier AI systems find vulnerabilities in OpenAI's infrastructure, which improves system security, which makes the deployment of those same AI systems safer.

OpenAI is partnering with Hacktron to find vulnerabilities, test connections with third-party services, and strengthen defenses with frontier AI.

loop fit 6/10via OpenAI
Partnering with Hacktron to find vulnerabilities in OpenAI’s systems

8 Nov 2026

12 picked
PaperarXiv·6 Nov 2026·Inference efficiency

ServerBench: How can your coding agents design specialized LLM serving systems?

The loop: Coding agents design and optimize specialized LLM serving systems, which then run the agents themselves with higher efficiency and lower cost.

A benchmark for agentic design of LLM serving systems where coding agents optimize latency and cost, creating systems that can more efficiently serve the agents themselves.

loop fit 10/10Xiaobo Wang, Zimu Wang, Siyu Ye, Jiaqi Li, Min Tang, Jun Bai, Yuling Shi,...via arXiv
ServerBench: How can your coding agents design specialized LLM serving systems?
PaperarXiv·Research automation

ReproBridge: Building Reproducibility into ML Research Artifacts

The loop: The ReproBridge agent constructs reproducible research artifacts from raw code and papers, which improves the ability of successor agents to reproduce and build on ML research.

An agent workflow and benchmark for converting raw ML research code into standardized, reproducible artifacts that significantly increase downstream reproduction success by other agents.

loop fit 9/10via Awesome-AI4AI
ReproBridge: Building Reproducibility into ML Research Artifacts
PaperarXiv·5 Nov 2026·Tools and harness

Beneath the Surface: What We Overlook When We Talk About Agent Evolution

The loop: Agent harness evolution methods modify the agent's prompts and components, which improves task reliability and coverage, making the agent more effective at its tasks.

An evaluation of four agent harness evolution methods revealing that while evolved agents gain new capabilities, they often lose previously covered tasks and incur higher deployment costs.

loop fit 9/10Shijue Huang, Hangyu Guo, Guanting Dong, Chenxin Li, Junting Lu, Xinyu Geng,...via arXiv
Beneath the Surface: What We Overlook When We Talk About Agent Evolution
PaperarXiv·6 Nov 2026·Tools and harness

TTHE: Test-Time Harness Evolution

The loop: Agentic proposers reason over execution traces to evolve the agent's harness at test time, which then improves the agent's own execution and success rate on subsequent tasks.

A method that treats an executable agent harness as the state of test-time adaptation, iteratively evolving it on unlabeled test batches to handle shifts in task distributions and failure modes.

loop fit 8/10Ellen Xiaoqing Tan, Jack Lanchantin, Zichun Yu, Polina Kirichenko, Shehzaad...via arXiv
TTHE: Test-Time Harness Evolution
PaperarXiv·5 Nov 2026·Self-reward and self-play

Better honeypots, fewer tries: training an optimizer to make PPO cheat

The loop: A learned optimizer searches for gridworld levels that induce reward-hacking, which improves the evaluation of PPO and makes the training process better at identifying and fixing cheating.

A project that trains a small optimizer using evolution strategies to efficiently find reward-hacking honeypots in gridworld levels to test PPO agents.

loop fit 8/10Jun Nie, Yonggang Zhang, Jun Song, Qianshu Cai, Dahai Yu, Yike Guo, Xinmei...via arXiv
Better honeypots, fewer tries: training an optimizer to make PPO cheat
PaperarXiv·6 Nov 2026·AI R&D evals

Auditing Harness Tampering in Self-Improving Agents

The loop: Self-improving agents attempt to modify their own code or environment, while an auditing system monitors for tampering to ensure the safety and reliability of the resulting improved agent.

An evaluation of auditing systems designed to monitor self-improving agents for harness tampering during their attempts to modify their own code.

loop fit 7/10Carel van Niekerk, Renato Vukovic, Benjamin Ruppik, Hsien-chin Lin, Shutong...via arXiv
Auditing Harness Tampering in Self-Improving Agents
PaperarXiv·5 Nov 2026·Self-modifying agents

ThinkFlowLab/agentdescent

The loop: An asynchronous framework treats agents as parameters and diffs as gradients, using an aggregator to optimize agent prompts and harnesses to improve the performance of the swarm.

A parallel framework for self-evolving agents that applies a gradient-descent-like process to optimize agent skills, prompts, and harnesses.

loop fit 7/10Xinrui Zhong, Zhiming Wang, Yongqi Liu, Zhaohui Wang, Shanglin Yang, Zikun Li,...via arXiv
ThinkFlowLab/agentdescent
PaperarXiv·4 Nov 2026·Self-modifying agents

The components of agent memory

The loop: Claude uses a dreaming mechanism to organize memory and correct mistakes, which improves its internal state and makes the model better at learning from experience.

A look into how Claude organizes its memory and uses a dreaming mechanism to decide what to remember and correct mistakes from past experiences.

loop fit 7/10Jizhe Zhang, Ang Li, Chenheng Zhang et al.via arXiv
The components of agent memory
Forum postLessWrong·5 Nov 2026·Self-reward and self-play

Better honeypots, fewer tries: training an optimizer to make PPO cheat

The loop: A learned optimizer searches for gridworld levels that induce reward-hacking, which improves the evaluation of PPO, which makes the training process better at identifying and fixing cheating.

I trained a small optimizer on synthetic functions, and used it to search gridworld levels where PPO cheats. TL;DR * Reward-hacking honeypots are levels that are built by hand where a reward is given to the agent for cheating (e.g. taking...

loop fit 8/10Aniket Ghoshvia LessWrong
Better honeypots, fewer tries: training an optimizer to make PPO cheat
ProjectGitHub·2 Nov 2026·Self-modifying agents

Toyhom/EvoSprig

The loop: EvoSprig uses visual workflows and multi-agent collaboration to enable agents to perform recursive self-improvement on their own tools and harnesses.

Open-source framework for self-evolving agents: visual workflows, multi-agent collaboration, memory, tools and controlled recursive self-improvement (RSI).

loop fit 7/10Toyhomvia GitHub
Toyhom/EvoSprig
ProjectGitHub·24 Jul 2026·Self-modifying agents

ThinkFlowLab/agentdescent

The loop: agentdescent uses a gradient-descent-like process to optimize agent prompts and harnesses, which improves the performance of the agents in the swarm.

Gradient descent, but the parameters are agents - a parallel, asynchronous framework for self-evolving agents (skills, prompts, harnesses). Diffs are the gradients; the aggregator is the optimizer.

loop fit 7/10ThinkFlowLabvia GitHub
ThinkFlowLab/agentdescent
ProjectGitHub·2 Nov 2026·Self-reward and self-play

Tencent-Hunyuan/EmoReward

The loop: EmoReward uses self-evolving reward calibration to improve the reinforcement learning process for emotional video generation.

Official Implementation of EmoReward: Self-Evolving Reward Calibration for Emotional Video Generation RL

loop fit 7/10Tencent-Hunyuanvia GitHub
Tencent-Hunyuan/EmoReward

7 Nov 2026

12 picked
PaperarXiv·2 Nov 2026·Research automation

Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

The loop: Anthropic AI systems perform AI R&D work, which improves Anthropic's models and tools. This makes the systems or their direct successors better at performing further AI R&D work.

METR conducted an independent assessment of Anthropic's internal use of AI to accelerate its own AI research and development.

loop fit 9/10Erkang Zhu, Yin Lin, Xiaorong Shi et al.via arXiv
Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
PaperarXiv·6 Nov 2026·Tools and harness

HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents

The loop: A coding agent uses Monte Carlo Tree Search to iteratively search for and refine its own executable harness. This optimized harness directly governs the agent's interaction with tools and the environment, improving its performance on subsequent tasks.

A framework that automates the design of agent harnesses by using a coding agent and MCTS to search for and refine harness mechanisms.

loop fit 9/10Haoyang Zhang, Qumeng Sun, Xiao Zhang, Xiang Liu, Zhidong Fan, Junjie Wangvia arXiv
HarnessFlow: MCTS-Guided Harness Optimization for LLM Agents
PaperarXiv·6 Nov 2026·AI R&D evals

IdeaTaste: Do LLMs Share the Research Taste of Human Researchers?

The loop: LLM agents write research proposals based on a literature snapshot, which are scored against actual future publications. This benchmark measures AI performance on AI research tasks, providing the necessary evaluation signal to improve future AI research agents.

A benchmark that evaluates whether LLMs can propose research ideas that match the actual future trajectory of human scientific communities.

loop fit 9/10Y. L. Li, Y. H. Wuvia arXiv
IdeaTaste: Do LLMs Share the Research Taste of Human Researchers?
PaperarXiv·6 Nov 2026·Training data

Recursa: Toward Recursively Self-Improving Foundation Model Systems

The loop: The Recursa foundation model system uses parallel learners to execute tasks and synthesize local improvements. These improvements are validated and merged into the system's own knowledge repository, enabling the model to adapt to new domains through recursive self-improvement.

A foundation model system that adapts to new domains by learning from its own experience and storing domain-specific skills in a self-refined knowledge repository.

loop fit 9/10Yiqiao Jin, Yiyang Wang, Lucheng Fu, Yijia Xiao, Yinyi Luo, Haoxin Liu, B....via arXiv
Recursa: Toward Recursively Self-Improving Foundation Model Systems
PaperarXiv·5 Nov 2026·Self-reward and self-play

SEAL: Synergistic Co-Evolution of Agents and Learning Environments

The loop: LLM agents generate execution failures that are used as a shared diagnostic signal to simultaneously reshape the learning environment and adjust how the policy learns. This keeps the training environment aligned with the agent's evolving needs, directly improving the agent's subsequent training.

A framework that uses agent execution failures to co-evolve the agent's policy and its learning environment.

loop fit 9/10Zhi Hong, Qianlong Wang, Hongchen Wei, Aidan Wu, Daohong Ying, Mingze Kong,...via arXiv
SEAL: Synergistic Co-Evolution of Agents and Learning Environments
PaperarXiv·6 Nov 2026·Self-reward and self-play

Better honeypots, fewer tries: training an optimizer to make PPO cheat

The loop: A learned optimizer searches for gridworld levels that induce reward-hacking in PPO. This automated discovery of honeypots improves the evaluation of reinforcement learning algorithms, helping developers build more robust training processes.

A project that trains a small optimizer to automatically discover gridworld levels where PPO agents learn to cheat.

loop fit 8/10Minhua Lin, Juncheng Wu, Zijun Wang, Zhan Shi, Yisi Sang, Bing He, Zewen Liu,...via arXiv
Better honeypots, fewer tries: training an optimizer to make PPO cheat
PaperarXiv·3 Nov 2026·Self-modifying agents

ThinkFlowLab/agentdescent

The loop: A swarm of agents generates diffs to their own skills, prompts, and harnesses. An aggregator acts as an optimizer to apply these changes, improving the performance of the agents in the swarm.

A parallel, asynchronous framework that treats agents as parameters and uses a gradient-descent-like process to optimize their prompts and harnesses.

loop fit 7/10Zijun Huang, Yingjun Du, Cees G. M. Snoekvia arXiv
ThinkFlowLab/agentdescent
PaperarXiv·6 Nov 2026·Self-modifying agents

The components of agent memory

The loop: Claude uses a dreaming mechanism to organize its memory and correct past mistakes. This process improves its internal state, making the model better at learning from experience in subsequent interactions.

An overview of how Claude uses a dreaming mechanism to organize memory and correct mistakes to learn from experience.

loop fit 7/10Jaival Chauhan, Vikram Singh, Pruthwik Mishravia arXiv
The components of agent memory
Blog postMETR·5 Nov 2026·Research automation

Review of the "Risks from automated R&D" section in Anthropic's Risk Report (August 2026)

The loop: AI systems at Anthropic are used to accelerate their own R&D, a process independently verified by METR.

Redaction summary: As part of this assessment, Anthropic had rights to redact any non-public information in this post. Per our approach to transparency , we had the right to disclose if Anthropic exercised redaction rights, and whether or...

loop fit 9/10via METR
Review of the "Risks from automated R&D" section in Anthropic's Risk Report (August 2026)
Blog postMETR·5 Nov 2026·Research automation

Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

The loop: Anthropic AI systems perform AI R&D work, which improves Anthropic's models and tools, which makes the systems better at AI R&D work.

Redaction summary: As part of this assessment, Anthropic had rights to redact any non-public information in this post. Per our approach to transparency , we had the right to disclose if Anthropic exercised redaction rights, and whether or...

loop fit 9/10via METR
Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
Blog postOpenAI·3 Nov 2026·AI R&D evals

Eval-production behavior gaps in agent safety and the role of eval-awareness

The loop: The system measures gaps between agent behavior in evals versus production to improve the reliability of AI safety benchmarks.

We study agent behavior in safety evaluations and in production, and explore how models' interpretations of evaluation objectives relate to eval-production performance gaps.

loop fit 8/10Joy Hsu, Sam Arnesenvia OpenAI Alignment
Eval-production behavior gaps in agent safety and the role of eval-awareness

6 Nov 2026

7 picked
PaperarXiv·6 Nov 2026·Self-reward and self-play

When Can Self-Play Fix a False Belief? Answer Keys, Pass Bands and Replacement Rules in an Execution-Grounded Loop

The loop: Qwen3 4B Base generates output prediction tasks and Python code, then executes the code to derive answer keys. The model trains on these self generated, execution verified keys to correct its own planted false beliefs about task specifications.

An analysis of execution grounded self play reveals that training on self generated code execution only fixes a model's false beliefs if the answer key comes from an unplanted teacher model rather than the model's own flawed implementation.

loop fit 10/10Nirdesh Chauhan, Dushyant Rajput, Siddharth Kosarajuvia arXiv
When Can Self-Play Fix a False Belief? Answer Keys, Pass Bands and Replacement Rules in an Execution-Grounded Loop
NewsAlignment Forum·5 Nov 2026·Research automation

METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026

The loop: Anthropic uses its own AI systems to automate parts of its pretraining, reinforcement learning, and model evaluation workflows. This automated R&D work accelerates the capability improvements of Anthropic's internal models, directly speeding up the development of their successors.

An independent assessment by METR concludes that Anthropic's use of AI for internal R&D is likely doubling the rate of capability improvement for its models on well specified tasks.

loop fit 9/10Zach Stein-Perlmanvia Alignment Forum
METR: Independent Assessment of AI R&D Acceleration at Anthropic in October 2026
PaperarXiv·4 Nov 2026·Inference efficiency

CompressionBench: How can your coding agents compress LLMs?

The loop: Coding agents receive a source checkpoint and a storage limit, then autonomously design and execute compression algorithms guided by perplexity feedback. The resulting compressed models and inference code reduce the operational overhead of the LLMs that drive the agents.

A new benchmark evaluates whether coding agents can autonomously compress large language models by writing and running their own compression algorithms under strict storage and time budgets.

loop fit 9/10Xinrui Zhong, Zhifeng Li, Zhiming Wang et al.via arXiv
CompressionBench: How can your coding agents compress LLMs?
PaperarXiv·6 Nov 2026·Self-modifying agents

WeightLoop: Weight-Space Programming with Experimental Feedback

The loop: GPT-6 Astra writes a parameter editing program that inspects a model's implementation and prediction errors. The program validates candidate edits using experimental feedback and writes the successful changes directly back into the model's weights, improving its accuracy on subsequent tasks.

WeightLoop enables a language model to write reusable programs that persistently edit a neural network's weights based on experimental feedback, without requiring gradient retraining.

loop fit 9/10Mengran Li, Bo Li, Weijing Zhao, Chengyang Zhang, Boyu Zhang, Jinfeng Xu,...via arXiv
WeightLoop: Weight-Space Programming with Experimental Feedback
Blog postTogether AI·4 Nov 2026·Inference efficiency

Polymorphic Inference: self-tuning inference for open models

The loop: AI agents monitor changing traffic patterns and hardware conditions to continuously tune production inference configurations. This optimization directly improves the efficiency and speed of the models running on the platform.

Together AI introduces a compiler that uses AI agents to autonomously adapt and tune production inference for open models based on real time traffic and hardware changes.

loop fit 8/10via Together AI
Polymorphic Inference: self-tuning inference for open models
ProjectGitHub·4 Nov 2026·AI R&D evals

flex-rsi/rsi-bench

The loop: Agents attempt to recursively improve their own external harnesses, prompts, and tool configurations. The benchmark evaluates these updates to measure how effectively the agents can evolve their own operating environments.

An open source benchmark designed specifically to measure and evaluate the recursive self improvement capabilities of agent harnesses.

loop fit 8/10flex-rsivia GitHub
flex-rsi/rsi-bench
Blog postOpenAI·4 Nov 2026·Tools and harness

A thousand fewer paths to root

The loop: An autonomous pipeline searches for, validates, and generates patches for local privilege escalation vulnerabilities in the Linux kernel. Applying these patches secures the underlying operating system infrastructure that hosts the AI's own training and inference workloads.

OpenAI deployed an autonomous pipeline that successfully identified and patched 948 local privilege escalation vulnerabilities in the Linux kernel.

loop fit 7/10via OpenAI
A thousand fewer paths to root

5 Nov 2026

9 picked
PaperarXiv·5 Nov 2026·Training data

EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling

The loop: The model identifies its own weaknesses to guide the generation of structurally diverse training instances. These new instances are then used in a reinforcement learning stage to update the model's own weights.

A framework that co-evolves a language model and its training data by using the model's failure signals to generate new mixed-integer linear programming instances.

loop fit 10/10Qingcan Kang, Mingyang Liu, Xiaojin Fu, Shixiong Kai, Tao Zhong, Mingxuan Yuanvia arXiv
EvoOptiGraph: Weakness-Driven Coevolution via Graph-Based Structural Generation for Optimization Modeling
PaperarXiv·2 Nov 2026·Inference efficiency

SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative Decoding

The loop: Parallel agents explore algorithm and system optimizations for speculative decoding deployments. Successful optimizations are integrated to accelerate the target model's own inference.

An autonomous agentic loop that discovers and integrates deployment-specific algorithm and system optimizations for speculative decoding.

loop fit 9/10Kunjun Li, Wenhao Chai, Yihao Liang et al.via Hugging Face Papers
SpecPilot: Autonomous Algorithm Discovery and Systems Optimization for Speculative Decoding
PaperarXiv·4 Nov 2026·Tools and harness

Reglos: A Self-Evolving Personal Memory Agent with Recursive Self-Improvement

The loop: A coding agent proposes patches to its own memory harness based on failure traces, while a critic screens them for benchmark gaming. Accepted patches permanently update the harness, improving the agent's performance on future conversations.

A personal memory agent that recursively improves its own memory harness by proposing and screening code patches from failure traces.

loop fit 9/10Sen Yang, Jiaduo Lin, Zehong Huang et al.via arXiv
Reglos: A Self-Evolving Personal Memory Agent with Recursive Self-Improvement
PaperarXiv·5 Nov 2026·Architecture and optimizer search

Evolutionary System Prompt Learning for Reinforcement Learning in LLMs

The loop: The language model samples trajectories under multiple system prompts, using them to update its weights via reinforcement learning and evolve its system prompts via self-reflection. Both the updated weights and the evolved prompts are then used by the model in the next iteration.

A method that jointly improves a model's weights through reinforcement learning and its experiential context through system prompt evolution.

loop fit 9/10Lunjun Zhang, Ryan Chen, Bradly C. Stadievia arXiv
Evolutionary System Prompt Learning for Reinforcement Learning in LLMs
PaperarXiv·4 Nov 2026·AI R&D evals

From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?

The loop: Research agents retain insights and persistent state across multiple machine learning tasks. This memory allows them to apply prior discoveries to new problems, improving their own research efficiency and solution quality over time.

An open-ended environment that evaluates whether autonomous research agents can retain and apply prior insights across a stream of machine learning tasks.

loop fit 9/10Jeremy Qin, Ruizhi Deng, Mohamed Osama Ahmed, Frederick Tung, Leo Fengvia arXiv
From AutoResearch to Continual AutoResearch: Do Agents Learn Across Tasks?
Blog postTogether AI·4 Nov 2026·Kernels and compilers

Together Compiler: Continual Compilation for AI inference

The loop: AI agents continuously optimize inference kernels for open models based on changing traffic. The resulting compiled kernels are then deployed to run those same models with higher throughput.

A continual compilation system that uses AI agents to optimize open models for inference in under 24 hours.

loop fit 8/10via Together AI
Together Compiler: Continual Compilation for AI inference
Blog postNVIDIA·4 Nov 2026·Tools and harness

Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes

The loop: Agents analyze their own production failures to generate custom evaluations and propose harness fixes. These fixes are then applied to improve the agent's prompts, memory, and execution code.

A workflow for using NVIDIA NeMo to automatically turn production agent failures into custom evaluations and harness improvements.

loop fit 8/10Patrick Moorheadvia NVIDIA Technical Blog
Use NVIDIA NeMo to Turn Agent Failures into Tests and Harness Fixes

4 Nov 2026

8 picked
PaperarXiv·4 Nov 2026·Inference efficiency

Think Less, Locally: Local Preference Learning for Reasoning Compression

The loop: A reasoning model detects candidate overthinking states in its own traces and regenerates shorter verified continuations. These self-generated preference pairs are then used to train the model via DPO to compress its reasoning.

A local preference learning method that trains a reasoning model to compress its trajectories by learning when to conclude from its own generated traces.

loop fit 10/10Muyao Yuan, Liyan Xu, Weizhan Zhang, Muyan Jiao, Jiangnan Li, Jinchao Zhang,...via arXiv
Think Less, Locally: Local Preference Learning for Reasoning Compression
ProjectGitHub·25 Apr 2026·Self-modifying agents

Protocol-zero-0/evolution-kernel

The loop: A coding agent proposes changes to its own git repository, which are then scored by a kernel. Winning changes are kept, creating an open-source evolution loop that improves the agent's own harness.

An open-source, sandboxed evolution loop that uses a coding agent to iteratively propose and test improvements to its own repository.

loop fit 9/10Protocol-zero-0via GitHub
Protocol-zero-0/evolution-kernel
PaperarXiv·3 Nov 2026·Kernels and compilers

ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents

The loop: A kernel-generation system distills validated architectural decisions into reusable knowledge of hardware mechanisms. It then uses its own optimized kernels as references for subsequent optimization runs, enabling continuous architectural evolution.

A kernel-generation system that accumulates knowledge across runs and uses its own optimized outputs as references to continuously evolve GPU kernel architectures.

loop fit 9/10Yaojian Chen, Dazhi Peng, Yiyun Zheng et al.via arXiv
ForgeKernel: Self-Evolving Kernels Catching and Surpassing Experts and Agents
NewsAlignment Forum·3 Nov 2026·Interpretability and oversight

Automating subtle emergent misalignment research

The loop: A Claude model is instructed to write training data that induces misalignment while bypassing a production safety monitor. The successful attacks it discovers are then used to test and improve the robustness of future monitors.

An automated red-teaming setup where a model generates misaligned training data to bypass a safety monitor, surfacing gaps to improve future defenses.

loop fit 9/10evhubvia Alignment Forum
Automating subtle emergent misalignment research
PaperarXiv·4 Nov 2026·AI R&D evals

Training LLM Agents to Automate LLM Post-Training with RSI-Dojo

The loop: LLM agents use the RSI-Dojo environment to learn post-training skills such as data selection, model training, and failure diagnosis. This internalizes the ability to conduct post-training, improving the agents' capacity to train their own successors.

A collection of executable model training environments designed to teach LLM agents how to autonomously conduct post-training.

loop fit 9/10Shuzheng Gao, Zhaojian Yu, Penghao Yin, Yuling Shi, Kai Cai, Shilin He,...via arXiv
Training LLM Agents to Automate LLM Post-Training with RSI-Dojo
PaperarXiv·31 Oct 2026·Tools and harness

Selecting LLM-Generated Hypotheses Raises Truth Beyond What an LLM Panel Expects: A Pre-registered Study on NAS-Bench-201 and HPOBench

The loop: A harness optimizer searches for agentic steps that improve the truthfulness of LLM-generated hypotheses. It discovers a two-stage drafting and selection step that raises the excess truth of the generated claims.

A testbed for LLM research ideation that uses a harness optimizer to find steps that improve the truthfulness and non-obviousness of generated hypotheses.

loop fit 9/10Shanglin Yang, Zhan Shivia arXiv
Selecting LLM-Generated Hypotheses Raises Truth Beyond What an LLM Panel Expects: A Pre-registered Study on NAS-Bench-201 and HPOBench
ModelHugging Face·3 Nov 2026·AI R&D evals

StateBench: Testing how language models track state changes over long documents

The loop: Language models are evaluated on their ability to track state changes over long documents. The resulting benchmark data identifies reasoning failures, which are used to improve the reasoning capabilities of future models.

A benchmark for evaluating how well language models track state changes over long documents.

loop fit 8/10via Hugging Face
StateBench: Testing how language models track state changes over long documents

3 Nov 2026

11 picked
PaperarXiv·1 Nov 2026·Tools and harness

Exo: An Architecture for Self-Improving Systems at Runtime

The loop: The Exo agent autonomously edits its own harness code and strategies during task execution. These inspectable code changes are preserved and activated to reduce costs and improve performance on subsequent tasks.

An agent runtime architecture that collapses the optimization loop into task execution by allowing agents to revise their own harness and strategies on the fly.

loop fit 9/10Chaoyun Zhang, Bingxuan Li, Pinxue Guo et al.via arXiv
Exo: An Architecture for Self-Improving Systems at Runtime
PaperarXiv·2 Nov 2026·Tools and harness

TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents

The loop: Frontier agents propose and statistically validate edits to a teacher model's harness. The improved harness is then used to collect successful trajectories to distill a compact student agent that operates in that same evolved harness.

A two-stage approach that uses frontier models to evolve a teacher's harness before distilling a compact student agent to run in it.

loop fit 9/10Yanbei Jiang, Chao Lei, Rena Gao et al.via arXiv
TinyAstral: Transferring Self-Evolving Harnesses to Tiny Generalist Manipulation Agents
PaperarXiv·30 Oct 2026·Self-reward and self-play

CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents

The loop: The CARE agent identifies its own production failures and pairs them with independently verified repairs. It then uses on-policy self-distillation to update its own weights, improving its performance on future customer service traffic.

A customer service agent framework that learns continuously in production by diagnosing its own errors and distilling the repaired trajectories.

loop fit 9/10Kehong Gong, Xiaoxu Zheng, Guanli Hou et al.via Hugging Face Papers
CARE: Grounded Recursive Self-Improvement for Production Customer-Service Agents
PaperarXiv·30 Oct 2026·Self-modifying agents

Evolving Memory, Learning to Meta-Control: Test-Time Adaptation for Experience-augmented LLM Agents

The loop: The OMCE meta-controller uses task feedback to update its memory bank and refine its own routing policy via contextual bandit updates. This improves how the agent retrieves and utilizes experience in future tasks.

A framework that allows self-evolving agents to learn a meta-control policy from task feedback to better manage and deploy their accumulated experience.

loop fit 9/10Jun He, Canmiao Fu, Ying Zhang et al.via arXiv
Evolving Memory, Learning to Meta-Control: Test-Time Adaptation for Experience-augmented LLM Agents
PaperarXiv·31 Oct 2026·Research automation

algorithmicsuperintelligence/openevolve: v0.4.1

The loop: The OpenEvolve framework uses mixed evolution modes and competitive checkpoints to iteratively refine programs. It records the mode used on each program in an evolution trace to improve the evolution process.

A new release of an open source evolution framework that adds mixed evolution modes and interoperability with AlphaEvolve.

loop fit 8/10Yandan zheng, Haoran Luo, Xinliang Zhou et al.via arXiv
algorithmicsuperintelligence/openevolve: v0.4.1
PaperarXiv·30 Oct 2026·AI R&D evals

More Rules, Longer Lists? Benchmarking Self-Evolving Agents on a Hidden-Rule Curriculum

The loop: Self-evolving agents rewrite their persistent state from feedback to learn latent rules. CLR-Bench measures this ability, providing a metric to evaluate and improve self-improvement capabilities.

A benchmark that tests whether self-evolving agents can learn hidden rules by rewriting their persistent state from feedback.

loop fit 8/10Alexander Krentsel, Sylvia Ratnasamy, Ion Stoica et al.via arXiv
More Rules, Longer Lists? Benchmarking Self-Evolving Agents on a Hidden-Rule Curriculum
PaperarXiv·1 Nov 2026·Training data

When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications

The loop: A synthesis agent iteratively develops coupled task requirements and executable tests. These synthesized tasks are used to train more capable coding agents for long horizon software development.

A framework where a synthesis agent iteratively develops coupled task requirements and executable tests to train coding agents on long horizon tasks.

loop fit 8/10Zeyuan Zhao, Yuanyuan Shi, Peng Yu et al.via arXiv
When Requirements Interact: Scaling Long-Horizon Coding via Interdependent Specifications
ReleaseGitHub·2 Nov 2026·Interpretability and oversight

Training activation classifiers for safety and alignment

The loop: Activation classifiers monitor a model's internal activity to detect policy violations. This improves the safety and alignment of the model, making it safer to deploy and use for further development.

A method for training efficient monitors that use a model's internal activity to detect and interrupt policy violating behavior in real time.

loop fit 7/10codelionvia GitHub releases
Training activation classifiers for safety and alignment
Blog postOpenAI·2 Nov 2026·Interpretability and oversight

Training activation classifiers for safety and alignment

The loop: Activation classifiers monitor a model's internal activity to detect policy violations, which improves the safety and alignment of the model, which makes the model safer to deploy and use for further development.

We show how we train and deploy efficient monitors that use a model’s internal activity to help detect and interrupt policy-violating behavior in real time.

loop fit 7/10Alex Makelov, Xiaojun Xu, Sundaram Ananthanarayanan, Ethan Mcdonald, Tom Dupre...via OpenAI Alignment
Training activation classifiers for safety and alignment
Blog postOpenAI·2 Nov 2026·Tools and harness

Scaling Git at OpenAI

The loop: Codex generates code at a scale that necessitates a new Git backend, which improves the development infrastructure for OpenAI models, which makes the development of Codex and its successors more efficient.

How OpenAI built Pando, a Git-compatible backend designed for the scale of development that Codex makes possible.

loop fit 6/10via OpenAI
Scaling Git at OpenAI
Blog postNVIDIA·2 Nov 2026·Tools and harness

Keep Distributed Training Running with Automated Workload Recovery

The loop: Automated Workload Recovery detects and recovers from hardware and software failures during training, which improves the efficiency of distributed training runs, which makes the development of large models faster.

As distributed AI training scales, failures across hardware and software dependencies increasingly disrupt training progress, making fast, reliable recovery...

loop fit 6/10Elizabeth Goodmanvia NVIDIA Technical Blog
Keep Distributed Training Running with Automated Workload Recovery

Older editions →