01ImportantGoogle + Google DeepMind + UMD + UVA
To know what “Google RSI” actually is
Dream-RSI
Weights stay frozen. The system rewrites exploration-policy code using past discovery trees.
THE SEPTEMBER READING LIST
September research papers on Google, DeepMind, and some other AI worth reads.
Updated 24 September 2026
Loading 3D
09 / 2026
Find your next read
LINKS CHECKED 24 SEP12 picks · 13 reads
The selected reads
02ImportantGoogle Research + Virginia Tech
The skill/wiki paper everyone cites
WikiSkill
Agent experience becomes a persistent wiki that informs reusable, evolving skills.
03Stanford Medicine (Miao, Zou)
A new way to use papers
Paper2Agent
Stanford’s Nature paper turns research and code into interactive agents. Works with Claude Code.
Sources & code (3)
04Google + Georgia Tech + Peking (Sercan Ö. Arık et al.)
Why long-horizon agents lose the plot
Procedural Graphs
An approach to planning with execution structures that agents can inspect and evolve.
05KAIST + Google DeepMind
Long context without another RAG lecture
Declarative Attention
Language models explicitly declare where to attend, making attention more selective.
06Yan / Shanghai AI Lab et al.
Agents that run for days
Harness-of-Harness
A framework for multi-day autonomous software development with continual improvement.
07Google Research (Lin, Woodruff, Deng, Mao, Zuo, Mirrokni)
What a research-agent harness looks like
Stellar Colosseum
Google Research’s many-agent approach to long-horizon mathematics and theoretical computer science.
08Microsoft
How to RL an agent that already has a harness
Agent Lightning v1.0
Microsoft’s reinforcement-learning framework trains agents while retaining their existing harnesses.
09ImportantAnthropic
What real-world misuse of Claude looks like
Claude misuse
Anthropic’s September report documents disrupted operations across cyberattacks, surveillance, fraud, and other harm areas.
10NVIDIA
How NVIDIA improves harnesses and mathematical reasoning
The NVIDIA pair
SoL-Pi evolves efficient agent harnesses. An Open Recipe for IMO Gold describes Nemotron’s training and proof-search pipeline.
11Research paper
“The harness is the model” made concrete
JIT-Agent
A model generates task-specific harnesses just in time, then repairs and improves them.
12survey
Vocabulary so RSI threads don’t eat you
Toward Genuine Recursive Self-Improvement
A survey of recursive self-improvement: definitions, mechanisms, and the gap between current systems and the larger claim.
All research reads
The Past Frames the Future: Memory for Autoregressive Video Generation
SpeakerMem-R1
RewardVerse
Introducing MentalHealthBench
HyperAgents / DGM-H listing
Capable yet Parsimonious: Hidden Chain-of-Thought in Frontier Models
SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness
Accelerating open-source biomolecular models with Claude
Measurements for understanding the pace of AI development inside frontier labs
Scaling Discovery through Test-Time Communication (Team-of-N vs Best-of-N)
Reimagining research papers as interactive and reliable AI agents (Paper2Agent)
Read now
Sources & code (3)
Cognitive Extensions for Dual-Process Language Agents: Memory and Self-Reflection
CERA-MoA: Co-Evolving Routing Mechanisms with Continually Learning LLM Agents
Ask the Tool, Don't Guess: Agent Tool Calls Hold Their Progress
ASLEval: Measuring Privacy Exposure Displacement in LLM Agent Sessions
Confidence Comes from Experience: Experiential Confidence Estimation from Reasoning to Agents
Amplio: A Lightweight Agent Harness for Robust and Long-Horizon Runs
DeepMind Institute launch essays
HPOQuest: Rare-Disease Diagnostic Agent
Compiled Agency: coding agents build game players
Social Laws for Multi-agent Coordination
Where Should Agents Live? Energy-Memory on the Edge-Cloud Continuum
Routing Multiple Agents Below the Sum of Distances
ReFigBench
Taming the Agentic RAN
Bypassing inference bottlenecks: Retrieve-for-Train / Efficient Property-Aligned Fan-Out Retrieval via RL-Compiled Diffusion
PrimeScientist: Strategic Allocation of Research Effort in Autonomous Research
RepoAtlas: Guiding Coding Agents via Evolving Multimodal Repository Views
OPEN-1B: A Fully Auditable Training Run
EvolveTrade: Experience-Driven Policy Refinement for Self-Evolving LLM Trading Agents
Gemini 3.8 Live and 3.8 Live Extended Thinking
Mo Models Mo Problems: selecting model pools
Fuse: Verifiable social reasoning
Skill-Based Agentic Evaluation for Real-Time Data Science
Dream-RSI: Recursive Self-Improvement through Evolving Worlds
Stellar Colosseum: A Many-Agent Harness for Long-Horizon Research in Mathematics and TCS
AlgoEvo: Self-Evolving Agentic Search for Automated Algorithm Discovery
HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
Discovery Foundation Models: Toward Open-Ended Discovery Intelligence
The Router Within: Eliciting Native Skill Routing from a Frozen LLM
BLINDSPOT: A Benchmark for Safety and Refusal Calibration in Long-Horizon Tool-Using Agents
LongAgent: History-Guided Agentic Search for Longitudinal Outcome Prediction
OpenAI4S: Code as Action, Science as Sessions
Salesforce Koa
Capability Laundering
Recursive Reasoning or Statistical Extrapolation?
OpenResearcher v2 (trajectory synthesis)
ZGCM-1 open 7B math + agentic search
ToolGrad: Efficient tool-use dataset generation with textual gradients
COBRA-Skills: Contextual Bandit-Guided Evolution for Agent Skill Optimization
T1: Terminal Agent Reinforcement Learning for Long-Horizon Tasks
Detecting and countering misuse of AI: September 2026
Byte Model Scaling
GAUGE: When not to trust LLM-as-a-judge
Is Bash All You Need?
When Synthetic Data Hurts: On Catastrophic Forgetting in Skill Retrieval for LLM Agents
The Last AI Built by Humans: Toward Genuine Recursive Self-Improvement
An Open Recipe for IMO Gold: Training Nemotron for Olympiad Mathematics
AlphaGenome Atlas: A predictive map of every possible DNA letter change in the human genome
Procedural Graphs: Self-Evolving Execution Structures for LLM Agents
Google at ECCV 2026 (41 accepted papers presented 8–12 Sep)
An OpenAI model proposes a solution to the Navier–Stokes problem
FrogNano: Training a 4B coding agent via online task synthesis
Co-Evolving Harnesses and Models
Research acceleration: The view inside OpenAI
PARSER: Read in Parallel, Reason in Depth
Substrate-Aware AI Agents: Execution Context as a First-Class Input
AI-Native Design Docs
WeatherNext 3: Increasing resolution and performance of global weather models with raw observations
Sexual dimorphism in the complete connectome of the Drosophila male central nervous system
Read now
Sources & code (2)
Transfer learning for genomic prediction in underrepresented populations
A Case Study on Emergent Cheating and Whistleblowing in Autonomous Research Swarms
SimSkill: A Self-Evolving LLM Agent for Skill and Knowledge Accumulation in Traffic Simulation
GPT-6 Astra System Card
STAIR: Structure-Aware Information Retriever
Language Models Can Control Their Own Attention (Declarative Attention)
MASkills: Continual Skills Optimization for Multi-Agent LLM Systems
Act More, Decide Less: Skill-Guided Adaptive Action Chunking (SPACE)
Gemini 3.8 Flash and 3.8 Flash Cyber
Proactive cyber defense for governments and enterprises
Trace as State
CORAL: LLM-native harness for recommenders
Codebook Agent
Automated Researchers Can Mitigate Well-characterized Alignment Failures
Designing Proactive Thought Partners for Writing
Read now
Sources & code (2)
Mapping global methane emissions from space with deep learning
Skill Following: Evaluating Actual Skill Use in Retrieval-Enabled LLM Agents
Defense-as-Skill: Evolving Runtime Guard Skill for Skill-Augmented Agents
Harness-of-Harness: Multi-Day Autonomous Software Development with Continual Improvement
Introducing agentic video understanding with Gemini
Runtime-Independent Persistent Agents
E-Commerce Bench
AI Research Preference Models
Selective Forgetting: Graph Memory for Long-Term LLM Agents
Scientific Agent Skills: A Library of Procedural Knowledge for Research Agents
LongWoF-Bench / EvoMap Genes
LongRCA Bench: Diagnosing long-horizon agent failures
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
Accelerating Scientific Research with Gemini in the Real-World (Co-Scientist labs)
SKILL.state: Explicit Execution State Replaces Append-Only History
Prime Agent
What Compaction Destroys
Recuris
Meta^n
EvoMal
JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution
Context Management as Code
Visual General Intelligence: A White Paper
Enabling independent research on how people use Claude
The Hugging Face incident and the road ahead
Thinkingbox: Agents in Stateful Business Workflows
ClawProBench
AgentOPSD: Recursive Self-Distillation for Agentic RL
Skill Lift
Filesystem-Based Memory for LLM Agents
Judges as a Lifecycle
Can AI agents conduct open-ended AI research?
OneDayAgent
Toward Skill-Native LLMs / Skill Entropy
Agent Lightning v1.0: Towards Harnessed Agentic RL
The Skill Trigger Bottleneck
MemHarness: Memory Is Reconstructed, Not Replayed
Discovering cryptographic weaknesses with Claude
RL Towards Broadly and Persistently Beneficial Models
Superintelligent Retrieval Agent (SIRA)
MagenticLite / MagenticBrain / Fara1.5
Autoregressive Ranking
No matching reads
Try a shorter search or remove a filter.
Research indexes
KEEP READING- arXiv cs.AI new listingsFirehose
- arXiv cs.CL new listingsFirehose
- arXiv cs.MA recent submissionsFirehose
- papers.cool cs.AICurated daily
- DAIR AI Papers of the Week 2026Weekly picks
- Awesome AI Agent Papers 2026Curated year list
- Google DeepMind publicationsOfficial
- Google Research publicationsOfficial
- Google Research blogOfficial
- masamasa59 AI agent papersResearch index
- Hugging Face Daily PapersResearch index
- Anthropic researchResearch index
- Meta AI publicationsResearch index
- Microsoft Research AI blogResearch index