A 4-billion-parameter personal agent that used a gamer’s Nintendo and Xbox history put the right Steam title in its top 10 picks 47.9% of the time, against 40.8% for the ranking it was given, according to a paper by researchers at Meta AI and the University of California San Diego. That is 7.1 points, or 17.4% more hits, by our arithmetic.
The paper tests what it calls personal-agent mediated recommendation: an app ranks items from its own records, and an agent acting for the user, with “user-authorized cross-platform history”, edits that ranking before the user sees it. The team’s training method, PAMO, improved on the matched reinforcement-learning baseline on all five test sets, the paper says.
Short explainer · Steam top-10 hits 47.9%
A vertical news Short summarising this story. Figures come from the Meta AI and UC San Diego paper on arXiv; our arithmetic is labelled.
Business Pill · The Model and the Harness
A short Business Engineer explainer on the difference between an AI model and the harness of tools and workflow around it. It teaches the general idea only and does not discuss this paper.
The key insight: In the Meta AI and UC San Diego paper, a 4B personal agent using a gamer’s Nintendo and Xbox history raised Steam top-10 hits from 40.8% to 47.9%, 17.4% more hits by our arithmetic, with fewer harmful overrides than Claude Opus 4.6.
The real-world test uses the Open Play dataset, which “links Steam, Nintendo, and Xbox activity through a shared pseudonymized person identifier”. Steam is the target platform, and the cohort has 645 users, the paper says. Four Amazon product categories serve as larger proxy platforms for training and further tests.
The ranking the agent receives is built by the researchers to be strong: a SASRec model ranks 50 candidates, and Claude Sonnet 4.6 reranks them using only the user’s history on that platform. Every personal agent sees that frozen ranking plus the user’s history elsewhere, but not the platform’s scores or other users’ data.
Top-10 hit rate and harmful-override rate on the paper’s real cross-platform Steam test (645 users), for the platform ranking, Claude Opus 4.6 and the 4B PAMO agent. Figures from the Meta AI and UC San Diego paper on arXiv; chart by FourWeekMBA.
What Was Posted on X
New Meta paper shows that a personal AI agent can improve an app’s recommendations with your history from other apps, if it learns when to defer.
Each app ranks items using only what you did there, plus patterns from other users. A personal agent sees activity across apps, but it can also break a good ranking by dropping the right item.
They built PAMO, a training method that hides the cross-app history to see which of the agent’s answers depend on it. Answers that lean on that history get more credit, but only if ranking accuracy does not drop.
On real Steam users with Nintendo and Xbox history, a 4B PAMO agent reached 47.9% top-10 hit rate, versus 40.8% for the platform. It also beat standard RL training on all 5 test sets.
Personal agents that edit other apps’ results should defer by default and override only when the user’s history gives specific evidence.
โ arxiv. org/abs/2610.07588
Title: “Personal-Agent Mediated Recommendation with Cross-Platform User History”
Rohan Paul posted a summary of the paper on X on 31 October, which is how the paper reached us; the figures in this piece come from the paper itself.
What the Agent Learned
The paper says access to outside history is not enough by itself. An agent that overrides a good ranking can drop the right item, so the authors count “harmful overrides” as well as rescues. PAMO, built on Qwen3-4B-Instruct, masks the cross-platform history during training to see which answers depend on it, and gives those answers more credit only when ranking quality holds.
On the Steam test, PAMO’s harmful-override rate was 6.0%, against 9.0% for Claude Opus 4.6, while its hit rate was 47.9% against Opus 4.6’s 45.4%. PAMO had 2.2 rescues per harmful override, against 1.5 for Opus 4.6, by our arithmetic. The model was not trained on any Open Play data, the paper says.
The Business Lever: Who Edits the Feed
Recommendation feeds are where platforms decide what users see. In this set-up, a user’s own agent sits between the platform’s ranking and the user, carrying data the platform does not hold. The paper concludes that effective mediation requires “learning when that history should override population evidence encoded by a strong platform ranking”.
As we read the paper, the result turns on restraint: Claude Opus 4.6 also had the outside history, but dropped more correct items than the smaller trained agent.
As we read it, the set-up moves part of the ranking decision from the app to an agent the user controls, carrying data the app does not hold.
As we read it, the open question is deployment: the test covers 645 users on one target platform, and the paper does not describe a Metaproduct using the method.
Xia et al., arXiv 2610.07588, 6 Oct 2026
“These results show that effective mediation requires not only access to cross-platform user history, but also learning when that history should override population evidence encoded by a strong platform ranking.”
Three Implications
FOR PLATFORMS The paper models an agent that edits a platform’s ranking using history from other services, and counts how often it removes a correct item.
FOR AGENT BUILDERS The paper reports that its 4B PAMO agent had a 6.0% harmful-override rate on the Steam test, against 9.0% for Claude Opus 4.6.
FOR USERS’ DATA The paper describes the history as user-authorized and keeps platform scores and other users’ data out of the agent’s view.
The framework says “The surface, not the model, governs autonomy.”
As we read it, the paper tests a new surface between an app and its user: the personal agent that decides what the user sees. The framework is the owner’s argument and does not discuss this paper or its figures.
What Is Not Established
The Steam test covers 645 users and Steam as the only target, and Xbox history is available only at the genre level, the paper says. The platform ranking is the researchers’ own construction, not Steam’s recommender, and the paper does not describe any Metaproduct using the method. It is a preprint posted on arXiv on 6 October.
This Week In The AI Supercycle
Anthropic’s Reality Check, The Capital Squeeze, & The New World Order
This week’s issue: for two years the AI build-out was priced as a growthstory. This week it started being priced as a credit story.
Researchers at Meta AI and UC San Diego report that a 4B personal agent using a gamer’s outside history raised top-10 hits on a Steam test from 40.8% to 47.9%, with fewer harmful overrides than Claude Opus 4.6. The paper’s finding is that the agent has to learn when to leave a good ranking alone.
A note on sourcing. All facts and figures come from the paper by Xia and colleagues at Meta AI and UC San Diego on arXiv (6 October 2026) and its abstract, read on 31 October; the results are the authors’. We did not contact the authors or Meta.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
