A research team splits the work of AI agents that tune recommendation systems
RecHarness lets a simple learning method pick what to change in a recommender while an AI model writes the edit. It was tested live on a video ad platform.
The short version
- A bandit picks the direction of change; the language model writes the specific idea and code edit.
- A jump-basin mechanism triggers a bigger structural change when small edits stop helping.
- The paper reports steadier gains than LLM-directed search across several tasks, datasets and model backbones.
- A 7-day live A/B test raised ADVV by 2.084%, Revenue by 0.534% and Exposure by 0.559%.
- The paper says code is available, and a revised version appeared on 17 Nov 2026.
Why it matters
Recommendation systems decide much of what people see online, and improving them is still slow, manual work. This paper suggests that automating that work goes better when the AI is not given every decision. The evidence comes from one platform and one abstract, so it is a promising direction rather than a settled result.
The problem: tuning is still manual
A recommender system is the software that decides which video, product or ad to show you next. Behind it sits a model that engineers keep adjusting: its architecture, the goal it is trained to hit, and how it is trained. According to the paper, as listed on arxiv.org, this trial-and-error work still depends heavily on people making changes by hand.
AI agents built on large language models (LLMs, the text-generating systems behind chatbots) can in principle take over that loop. They can propose a change, write the code, run the test and read the result. The authors point to a snag, though.
If the same LLM has to choose which kind of change to try and also work out the details, the search can become unstable. That matters most when the budget is small, because every trial costs real computing time and there are few chances to recover from a bad guess.
Two jobs, two decision-makers
RecHarness splits the work in two. The first job is deciding which direction to modify next. The second is turning that direction into a concrete hypothesis and an executable code edit.
The first job goes to a bandit router. A bandit is a method that learns which option to try next. It looks at how earlier attempts scored on validation feedback, meaning results measured on held-out data, and leans toward directions that have paid off. The LLM does the second job, but only inside the direction the router has chosen.
The idea is to give each part the task it handles best. A simple statistical rule keeps track of what has worked, and the language model supplies the creativity and the code. The authors describe this arrangement as a harness, a structure that wraps around the agent and limits where it can wander.
Jumping out of a rut
Any search that makes small improvements can get stuck. The edits keep producing tiny or no gains, because the model has found a local peak and nearby changes cannot lift it higher.
To handle this, RecHarness has what the paper calls a jump-basin mechanism. When local edits stagnate, it activates a structural-jump arm. In bandit language an arm is simply one of the options the router can pick, and this one stands for a larger, more structural change rather than a tweak.
The aim, in the authors' words, is to sustain long-horizon exploration. In plain terms, the system should keep finding improvements over many rounds instead of flattening out early.
What the tests showed
The authors say they tested RecHarness across multiple recommendation tasks, datasets and model backbones, the underlying model designs being tuned. Their summary is that it delivers more consistent improvements than LLM-directed search and makes better use of a limited number of trials. The abstract does not list the individual scores, so the size of those gains is not something we can report.
The most concrete result comes from a live test. Over 7 days, on a large-scale short-video advertising platform, the candidate the system selected was compared in an online A/B test, where one group of users sees the new version and another sees the existing one.
- ADVV rose by 2.084%.
- Revenue rose by 0.534%.
- Exposure rose by 0.559%.
The abstract does not spell out what ADVV stands for, so we use the label as the authors do. The platform is not named either.
How far to trust it
This is a research paper, not a product launch. The revised version, v2, appeared on 17 Nov 2026, after the first submission on 31 Jul 2026. The version history shows the file grew considerably between the two, though the abstract does not say what was added.
Gains of about half a percent on revenue may sound small. On a platform that large, small percentages can still be meaningful, but that judgment depends on figures the abstract does not give. The tests also come from the authors themselves, with one live deployment as the main real-world evidence.
The paper does say code is available. That means other researchers can try to reproduce the claim, and that is the check that will count most.
What to watch
- Will other teams reproduce the steadier gains using the released code, on different datasets or platforms?
- Will the full paper explain how large the offline gains were compared with LLM-directed search?
- Can the router-plus-LLM split work in other kinds of automated tuning beyond recommenders?
Sources and links
- Reportingarxiv.org: the article we sourced
First reported by arxiv.org. Written by the SI Agent News Desk from the reporting above.Written with AI from the source above. Every number is checked against the source. Spot a mistake? Corrections.
Liked it? Say why
Add a line for other readers.
No account needed. We never show your email. Comments with links, ads or abuse are held and not shown. Your comment shows after a quick check.