Community research initiative · 2026

Flex-RSI

Benchmarking Recursive Self‑Improvement of Frontier Agents

Study how agents improve across rounds with a flexible, budget-aware benchmark designed to be practical to run and easy to extend.

Report soon Code Leaderboard

At a glance

Robots, panoramas, puzzles, edits and 3D scenes, each shown by the self-improved solution. Pick one to see it next to where it started.

All demos →
Action · RoboDojoArrange the largest number Reconstruction · Scene from PhotoRebuild a room from one photo Action · Active SearchLook around a 360° scene Action · RoboLabStack three cubes ImageImageCreation · Image EditingBananas, a week later Action · RoboCasaBoil water: kettle on, burner on Reasoning · ARC-AGI-3Learn a game by playing it Reconstruction · 3D TrackingTrack every object across views Perception · Video QualityJudge AI-generated video Action · RoboCasaSet the mug under the coffee machine Reasoning · Ego-Exo MemoryWhere is the wrench usually kept? ImageImageCreation · Image EditingRemove the shadow
ImagePerception · Visual SearchFind small details in large photos Action · RoboDojoInsert the tubes ImageImageCreation · Image EditingSwap in sunflowers Action · RoboLabCube and banana into the bowl ImageReasoning · ARC-AGI-2Infer the rule from examples Action · RoboCasaA bun and a sausage on each plate Action · RoboCasaOpen the cabinet Reconstruction · 3D TrackingKeep each identity across views Perception · Video QualityScore real user videos Action · RoboDojoThrow the bottles into the dustbin Reasoning · ARC-AGI-3Clear levels with no instructions

Overview

How far can a frontier agent improve itself?

Flex-RSI measures how well today's frontier agent systems improve on their own: given a downstream task, the agent iteratively refines its own harness, from its solution to its tools and prompts, to raise its score within a fixed budget. Tasks span visual perception and reasoning, robot control, 3D reconstruction and more.

We are actively expanding Flex-RSI with more tasks, tracks and models.

–frontier models –tasks –model × task runs –total RSI rounds

Shared starting point

Begin from a documented task setup and starter solution, then let the agent improve its own approach.

Declared budgets

Set limits for time, tokens, API spend or GPU hours, and report the budget with each run.

Held-out evaluation

Use validation to guide iteration and reserve test data for the final score.

Traceable progress

Record edits, experiments, model calls and outcomes so each run can be understood and repeated.

Tracks

Where agents improve themselves

The current tasks span six areas, action, reasoning, creation, reconstruction, perception and infra: from robot control and memory over video to image and video generation, 3D scene reconstruction, visual search and a faster video decoder. Flex-RSI is designed to grow as contributors bring new tasks and domains.

Action

Acting in physical and simulated worlds: planning, controlling robots and moving to gather the information a task needs.

Reasoning

Drawing conclusions over several steps: spatial, temporal and logical reasoning about what is observed and what is known.

Creation

Making and editing content: images, scenes and other media produced to precise instructions.

Reconstruction

Recovering the structure of the world: 3D and 4D geometry, correspondence and motion from images and video.

Perception

Seeing and understanding: recognising objects, reading fine detail and making sense of whole scenes in images and video.

Infra

Making the systems that run models faster without changing what they compute: kernels, decoders and serving.

Leaderboard →Bring a task →

Contribute

We welcome more contributions!

Flex-RSI is an evolving benchmark, and it grows with the tasks, baselines and evaluators that the community brings. Contributions of any size are welcome, from a single reproducible run to an entirely new domain.

New tasks

A downstream task with a starter solution, data splits and an evaluator, in a domain the benchmark does not yet cover.

Model baselines

Runs of further models on existing tasks, under the declared budget and protocol, so that the leaderboard keeps up with new releases.

Evaluators and adapters

Stricter or more informative evaluators, and adapters that connect new agent systems to the harness.

Protocol feedback

Suggestions on budgets, splits and acceptance rules that make runs fairer and easier to reproduce.

How to contribute →People →

Affiliations

  • Adobe
  • Carnegie Mellon University
  • Massachusetts Institute of Technology
  • ImageMIT-IBM Watson
    AI Lab
  • National University of Singapore
  • NVIDIA
  • Purdue University
  • Texas A&M University
  • University of Cambridge
  • University of Pennsylvania
  • University of Southern California
  • The University of Texas at Austin
  • University of Toronto

Citation

BibTeX

@misc{flexrsi2026,
  title  = {Flex-RSI: Benchmarking Recursive Self-Improvement
            of Frontier Agents},
  author = {To be announced},
  year   = {2026},
  note   = {Project page: https://flex-rsi.github.io}
}