ResearchBoost

An open research initiative · AI4AI

ResearchBoost Agent (RBA)

AI accelerates the research that advances AI.

Better research starts with better tools. We build AI agents that profile, optimize, and verify the systems behind the next generation of AI.

Explore the research Read the blog
Fig. 1baseline
Every idea in AI waits on infrastructure.

Real research workloads across multiple domains.

More cases being collected
workloads
12
domains
5
mean speedup
2.51×
max speedup
9.59×

Speedup on a log scale, each against its own baseline (1×). Source-reported measurements—not a head-to-head ranking. See each repository for its evaluation setup.

§ 01The premise

Our path to RSI starts with infrastructure.

AI research depends on the infrastructure beneath it. Improving that infrastructure gives researchers more room to experiment.

The work spans architecture, execution and kernels; every step is checked for correctness, speed and training quality.

  • long-horizon
  • technically demanding
  • clear feedback
Explore the ResearchBoost agent
Fig. 2the research stack
Faster infrastructure → more experiments → better research. The agent probes every layer, down to the kernel.
§ 02The agent at work

Profile. Optimize. Verify.

One agent, one frozen task. Each round tests one idea; a judge the agent cannot edit decides.

Inside the workflow
Fig. 3 · The Research Loop · one agent, isolated sandbox Select a step to explore
Round 01 1000 ms

Step 01 / 06Profile

Find the real bottleneck.

Profile the best current version, narrowing from stage to module to operation to kernel. Name the top hotspot before changing code.

OutputA localized bottleneck and a testable hypothesis.
Target reached: Screen → Check → Log → Next.
Timing branch

Interactive illustration · not a live agent run. Failed and no-gain rounds are recorded, too.

A frozen judge decides. Structure → benefit → bitwise → statistical checks. The agent cannot edit the judge.

A supervisor reviews. Rules, facts and stalled runs. The agent chooses what to optimize.

Fastrapid evaluation feedbackFaithfulpreserved task semanticsLocalpinpointed diagnosticsDenseactionable multi-stage feedback
CaseA real agent session · RAE task · ResearchBoost-Bench

Faster is a hypothesis.
Evidence is the standard.

Fig. 4 · ms per training step, lower is faster
    989.38 → 357.19 ms2.77× · final frozen-judge check PASS over 6 seeds · faster-but-wrong changes rejected
    § 03Wishing List

    What would you like to run faster?

    Bring a research workload. Tell us what slows you down, what hardware you can provide, and what a useful improvement would look like.

    We select cases based on available resources and a clear, reproducible evaluation. A submission starts a conversation; it does not reserve compute.

    ydove1031@gmail.com

    Opens your email app with a draft. Nothing is sent automatically.

    § 04Mini Bench Leaderboard

    A common ground for better agents.

    Shared tasks. Fixed judges. Comparable evidence.

    Leaderboard coming soon
    § 05The long-term direction

    AI that improves the tools that improve AI.

    ResearchBoost also offers a concrete starting point for studying recursive self-improvement (RSI): long-horizon, technically demanding tasks with relatively clear feedback, spanning system-level architecture and low-level GPU kernels.

    For us, this is a practical starting point: learning how AI can improve the tools and workflows that advance AI research.

    § 06An open invitation

    The next breakthrough could use your help.

    Build an agent. Bring a workload. Question a result. Share what you learn. There is more than one way to move research forward.

    01 · Build with usContribute to the agent 02 · Open a conversationIdeas, issues & workloads 03 · Follow the researchRead the project story