--- title: Can AI Agents Improve the Training Algorithms of Other... description: A recent research paper tested whether AI agents can improve the training algorithms of other AI systems, going beyond hyperparameter tuning into actual... canonical: https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Can AI Agents Improve the Training Algorithms of Other AI Systems? | daily.dev og:description: A recent research paper tested whether AI agents can improve the training algorithms of other AI systems, going beyond hyperparameter tuning into actual... og:url: https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam og:image: https://api.daily.dev/og/posts/JXUiqzoAM.png og:image:alt: Can AI Agents Improve the Training Algorithms of Other AI Systems? og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Can AI Agents Improve the Training Algorithms of Other AI Systems? **[Collections](https://daily.dev/sources/collections)** · 2 min read · 1 upvotes · 0 comments ## Summary A recent research paper tested whether AI agents can improve the training algorithms of other AI systems, going beyond hyperparameter tuning into actual algorithm design. Across 10 research repositories, Opus performed best, but even the top result scored only 0.166 on average. Giving agents more reasoning led to roughly 10x more tokens and 13x more code, but only about 2x better results, showing a steep gap between effort and outcome. The conclusion is that agents can't yet reliably perform algorithm design, though this capability is expected to improve over time. ## Content A new paper from Tsinghua and collaborating Chinese universities takes a hard look at whether AI agents can genuinely improve the algorithms used to train other AI systems — what the field calls recursive self-improvement (RSI). The short answer: not really, at least not yet. The researchers analyzed 1,338 post-training trajectories, covering 3,557 adjacent training experiments. Across all of them, only 74 experiments — about 2.1% — involved a meaningful change to the high-level training strategy. Everything else was local adjustment: tweaking data, tuning hyperparameters, fixing bugs, reformatting inputs. The agents weren't idle. They could train models, evaluate checkpoints, diagnose failures, and iterate. The problem showed up specifically when evidence pointed toward scrapping the current approach entirely. That almost never happened. As the paper puts it, "the agent's training strategy is locked in at the very beginning, and the entire remaining budget is spent on local adjustments within the selected strategy." The researchers tried three escalating fixes to break this pattern: - An experience-driven scaffold — adding an experiment journal, skill library, and evaluator — lifted GSM8K scores by 12.6 points and HumanEval by 40.8 points over the baseline. The strategy still didn't change. - Human guidance could redirect the agent's opening choice, but once training started, the agent drifted back into local tuning. - Giving agents 2–8× more inference compute mostly bought more refinement of the same approach, with almost no reliable gain on harder benchmarks like AIME 2025. Separately, agents were tested across 10 research repositories on the task of improving actual training algorithms. Claude Opus performed best among the models tested, but even the top result was weak — an average score of 0.166. More reasoning helped somewhat: roughly 10× more tokens and 13× more generated code for about 2× better results. That's a poor return. What's missing isn't effort or execution ability. It's a mechanism to explicitly reopen the strategy question while a run is still in progress. Current agents can refine a plan indefinitely but have no reliable trigger to ask whether the plan itself is wrong. The paper is titled "What is Missing from AI Post-Training AI: An Empirical Analysis" and is available at arxiv.org/abs/2608.19072. ## Questions this post answers ### Can AI agents currently improve the training algorithms of other AI models? Not reliably. A benchmark testing agents across 10 research repositories found that even the best-performing model, Opus, only achieved an average score of 0.166 when tasked with improving actual training methods rather than just tuning hyperparameters or optimizing code. The gap between effort and outcome remains too large for agents to reliably perform algorithm design today. _Anyone evaluating AI agent capabilities for research automation can track findings like this on daily.dev._ ### Does giving AI agents more compute or reasoning improve their algorithm design results proportionally? No, the returns are far below proportional. Increasing agent reasoning led to roughly 10x more tokens used and 13x more code generated, but only about 2x better performance on algorithm design tasks. This mismatch between effort scaling and outcome scaling suggests current agents hit diminishing returns quickly on hard research tasks. _Developers weighing compute costs against agent performance gains can follow this kind of analysis on daily.dev._ --- Tags: [#machine-learning](https://daily.dev/tags/machine-learning), [#llm](https://daily.dev/tags/llm), [#ai-agents](https://daily.dev/tags/ai-agents) [View this post on daily.dev](https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Can AI Agents Improve the Training Algorithms of Other AI Systems?","url":"https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam"},"datePublished":"2026-08-21T18:41:10.546Z","dateModified":"2026-08-28T13:23:47.461Z","description":"A recent research paper tested whether AI agents can improve the training algorithms of other AI systems, going beyond hyperparameter tuning into actual...","isAccessibleForFree":true,"articleSection":"Collections","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Collections","logo":"https://media.daily.dev/image/upload/s--fk_6ycEi--/f_auto,q_auto/v1780996001/logos/collections?_a=BAMAMiWQ0","url":"https://daily.dev/sources/collections"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":1},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"machine-learning,llm,ai-agents","timeRequired":"PT2M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Collections","item":"https://daily.dev/sources/collections"},{"@type":"ListItem","position":3,"name":"Can AI Agents Improve the Training Algorithms of Other AI Systems?"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/can-ai-agents-improve-the-training-algorithms-of-other-ai-systems--jxuiqzoam#faq","mainEntity":[{"@type":"Question","name":"Can AI agents currently improve the training algorithms of other AI models?","acceptedAnswer":{"@type":"Answer","text":"Not reliably. A benchmark testing agents across 10 research repositories found that even the best-performing model, Opus, only achieved an average score of 0.166 when tasked with improving actual training methods rather than just tuning hyperparameters or optimizing code. The gap between effort and outcome remains too large for agents to reliably perform algorithm design today. Anyone evaluating AI agent capabilities for research automation can track findings like this on daily.dev."}},{"@type":"Question","name":"Does giving AI agents more compute or reasoning improve their algorithm design results proportionally?","acceptedAnswer":{"@type":"Answer","text":"No, the returns are far below proportional. Increasing agent reasoning led to roughly 10x more tokens used and 13x more code generated, but only about 2x better performance on algorithm design tasks. This mismatch between effort scaling and outcome scaling suggests current agents hit diminishing returns quickly on hard research tasks. Developers weighing compute costs against agent performance gains can follow this kind of analysis on daily.dev."}}]} ```