AI Research

AI4AI-Bench: Benchmarking LLM Agents in Algorithmic... | AI Research

Key Takeaways

AI4AI-Bench is a new benchmark designed to test whether AI agents can perform recursive self-improvement (RSI) by redesigning the training algorithms that pr...
Recursive self-improvement (RSI) asks whether an AI system can improve the process that produces AI systems, so that the next system inherits the improvement.
That process is the training algorithm: a better objective or update rule improves the compute\mbox{-}capability exchange rate for every subsequent run, including the one that produces the next agent.
Whether RSI is feasible therefore turns on whether an agent can design training algorithms.
No benchmark isolates that ability: existing suites are won by collecting data or by tuning hyperparameters, and none tells a change to how a run is executed apart from a change to how the model learns.

Paper AbstractExpand

Recursive self-improvement (RSI) asks whether an AI system can improve the process that produces AI systems, so that the next system inherits the improvement. That process is the training algorithm: a better objective or update rule improves the compute\mbox{-}capability exchange rate for every subsequent run, including the one that produces the next agent. Whether RSI is feasible therefore turns on whether an agent can design training algorithms. No benchmark isolates that ability: existing suites are won by collecting data or by tuning hyperparameters, and none tells a change to how a run is executed apart from a change to how the model learns. We present AI4AI\mbox{-}Bench, 10 frozen research repositories spanning 10 training algorithm families. In each task, an agent has 4 hours on one B300 to rewrite the training algorithm; its code is then rerun from scratch for up to 12 hours and scored by a fixed evaluator hidden from the agent, against the repository's original algorithm under the same procedure. Because the 10 metrics are incommensurable, every task is mapped onto one scale on which $0$ is an uninformative model, $0.1$ is the algorithm the repository ships, and $1.0$ is the task optimum. Across 29 configurations of 6 systems on all 10 tasks the mean score is $0.166$, and the best system reaches $0.250$: even the strongest closes under a fifth of the distance between the algorithm that was already there and the optimum. The submissions show where that distance went: most never change how the model learns at all, and the minority that do average $0.226$ against $0.126$ for the rest. More reasoning effort mostly buys the willingness to go there, taking that minority from $8\%$ of submissions to $64\%$ and the mean score from $0.094$ to $0.196$. We release the task suite, the evaluators and every scored submission, so that the measurement can be repeated as these systems change.

AI4AI-Bench is a new benchmark designed to test whether AI agents can perform recursive self-improvement (RSI) by redesigning the training algorithms that produce AI systems. While existing benchmarks focus on data collection or hyperparameter tuning, this project isolates the ability to modify the core objectives, update rules, and regularization methods that dictate how a model learns.

How the Benchmark Works

The benchmark consists of 10 frozen research repositories, each representing a distinct family of training algorithms, such as supervised fine-tuning, preference optimization, and diffusion RL. For each task, an agent is given four hours on a B300 GPU to read the repository and rewrite its training code.
After the four-hour window, the agent’s modified code is executed from scratch for up to 12 hours. A fixed, hidden evaluator then scores the resulting model against the repository’s original algorithm. To ensure a fair comparison, the baseline is the repository’s own code run under the exact same hardware, budget, and evaluation conditions.

Measuring Algorithmic Design

Because the 10 tasks use different, incommensurable metrics (such as perplexity, solve rates, or aesthetic scores), the researchers mapped all results onto a unified scale. On this scale, 0 represents an uninformative model, 0.1 represents the original algorithm shipped with the repository, and 1.0 represents the task optimum.
The researchers found that most agents struggle to improve the core learning process. Out of 263 submissions that made any changes, 141 left the learning procedure untouched, focusing instead on budgets, checkpointing, or hyperparameters. The minority of agents that actually modified the learning algorithm—such as the objective or supervision signal—achieved an average score of 0.226, compared to 0.126 for those that did not.

Performance and Reasoning Effort

Across 29 configurations of six different systems, the mean score was 0.166, with the best-performing system reaching 0.250. Even the strongest systems closed less than one-fifth of the distance between the existing algorithm and the theoretical optimum.
The data suggests that increased reasoning effort primarily correlates with an agent's willingness to modify the learning algorithm itself. When agents were pushed to use more reasoning effort, the share of submissions that actually altered the learning procedure rose from 8% to 64%, which in turn improved the mean score from 0.094 to 0.196.

Franklin Analysis

The results indicate a significant gap between an agent's ability to perform general coding tasks and its capacity for meaningful algorithmic design. The evidence shows that while agents are capable of making technical edits, they rarely diagnose and fix the specific mechanisms—such as policy entropy collapse or reward model saturation—that a human machine learning scientist would address to improve a training run. The benchmark demonstrates that current systems are largely stuck in a regime where they optimize peripheral settings rather than the fundamental training dynamics that drive recursive self-improvement.

Comments (0)

No comments yet

Be the first to share your thoughts!