Personal-Agent Mediated Recommendation with Cross-Platform User History

Yu Xia; Jiangfan Zhang; Jun Xiao; Julian McAuley; Xiangjun Fan

cs.AIOct 6, 2026

Personal-Agent Mediated Recommendation with Cross-Platform User History

Authors: Yu Xia, Jiangfan Zhang, Jun Xiao, Julian McAuley, Xiangjun Fan

Organizations: University of California San Diego · Meta · Meta AI

Abstract

Modern recommendation is shifting from platform-centric personalization toward user-governed personalization, where a personal LLM agent can act on the user's behalf across services. We formalize this emerging paradigm as Personal-Agent Mediated Recommendation: a platform recommender ranks a candidate set using platform-local information, and a personal agent uses user-authorized cross-platform history to mediate the resulting ranking and produce the final top-K slate. Such mediation is nontrivial: the platform ranking can encode strong population evidence that the personal agent cannot observe, so effective mediation must therefore balance beneficial rescues against harmful overrides. To study this trade-off, we introduce MediateRec, a benchmark that includes scalable proxy cross-platform environments and a real cross-platform test under a controlled platform-agent information boundary. To train the agent to use cross-platform history effectively, we further propose Personal Attribution Mediation Optimization (PAMO), which counterfactually masks that history to estimate personal mediation support and reallocates rank-aware advantage mass under a platform-relative value floor. We theoretically prove that PAMO preserves cutoff-level advantage mass and is locally optimal among first-order reallocations that preserve this mass without lowering average platform-relative value. Experiments on MediateRec show that personal-agent mediation enables meaningful platform corrections, yet even strong proprietary LLMs introduce non-negligible harmful overrides. PAMO consistently improves over matched outcome-only RL across seen and unseen target platforms and on the real cross-platform test, while achieving a better rescue-harm balance.

Figures & tables

Figure 1: Personal-Agent Mediated Recommendation. The platform proposal summarizes platform-local population evidence, while the personal agent additionally uses cross-platform history to selectively revise or preserve the ranking.

MediateRec	Train	Val.	Test	Within	Cross	Platform H@10	Common Cross-Platform Sources
Movie	5,000	1,000	2,000	4.9	26.7	56.0	Books (59.7%); Home & Kitchen (53.4%); Electronics (50.3%)
Toy	5,000	1,000	2,000	2.8	24.0	47.7	Home & Kitchen (71.3%); Clothing, Shoes & Jewelry (68.5%); Electronics (52.4%)
Grocery	–	–	2,000	2.9	31.6	45.0	Home & Kitchen (77.7%); Clothing, Shoes & Jewelry (71.4%); Health & Household (60.3%)
Beauty	–	–	2,000	3.2	26.6	48.6	Clothing, Shoes & Jewelry (74.1%); Home & Kitchen (71.4%); Electronics (52.9%)
OpenPlay	–	–	645	11.9	11.0	40.8	Nintendo (85.1%); Xbox (19.8%)

Table 1: MediateRec statistics. Train/Val./Test report episode counts; Within/Cross are mean pre-target history lengths; Platform H@10 is the frozen platform’s HR@10; Common cross-platform sources are ranked by user coverage.

	MediateRec-Movie									MediateRec-Toy
Method	H@3	H@5	H@10	N@3	N@5	N@10	Resc.	Harm	Int.	H@3	H@5	H@10	N@3	N@5	N@10	Resc.	Harm	Int.
Platform	37.0	44.5	56.0	0.319	0.349	0.386	–	–	–	31.2	36.8	47.7	0.268	0.291	0.326	–	–	–
Claude Haiku 4.5	35.3	42.4	56.8	0.304	0.333	0.380	10.9	10.1	3.77	30.4	37.3	48.2	0.261	0.290	0.325	9.6	9.1	3.96
Claude Sonnet 4.6	40.4	49.0	63.1	0.340	0.376	0.421	12.2	5.0	3.27	32.6	39.5	52.3	0.281	0.308	0.350	12.1	7.5	3.71
Claude Opus 4.6	42.1	50.2	63.2	0.353	0.387	0.428	11.6	4.2	2.92	33.5	40.4	52.4	0.285	0.313	0.352	11.7	7.0	3.62
Qwen3-4B-Instruct	34.1	41.0	50.1	0.298	0.326	0.355	4.1	9.9	2.90	29.4	34.8	45.5	0.254	0.276	0.311	5.5	7.7	3.14

Table 2: Recommendation and mediation results on the four proxy cross-platform test sets. H@X and N@X denote HR@X and NDCG@X. Best results are in bold ; second-best results are underlined .

	MediateRec-OpenPlay
Method	H@3	H@5	H@10	N@3	N@5	N@10	Resc.	Harm	Int.
Platform	20.5	27.3	40.8	0.169	0.198	0.241	–	–	–
Claude Haiku 4.5	20.5	28.2	40.8	0.160	0.191	0.231	12.2	12.2	4.52
Claude Sonnet 4.6	23.7	30.5	44.0	0.180	0.208	0.251	13.5	10.2	4.40
Claude Opus 4.6	21.1	29.9	45.4	0.174	0.211	0.260	13.6	9.0	3.75
Qwen3-4B-Instruct	20.2	28.2	38.9	0.165	0.198	0.232	8.5	10.4	3.58

Table 3: Recommendation and mediation results on the real cross-platform test set.

Figure 2: Effect of cross-platform history averaged over all five test sets in MediateRec. Open markers use within-platform history only and filled markers use the full-context input. The dashed vertical line marks the frozen platform baseline.

Figure 3: Effect of support concentration

\kappa

and the value floor, averaged over the four proxy cross-platform test sets.

Panel of Figure 4 — Figure 4: Representative PAMO rationales illustrating intervention under specific cross-platform evidence and deference under weak cross-platform evidence.

Figure 5: Primary evidence behind PAMO’s 900 rescues.

Explore similar work

Jul 16, 2026cs.CY

PersonaMem-v3: Toward Omni-Platform Personal Intelligence for Holistic User Understanding, Recommendation, and Agentic Tasks

Personal intelligence is becoming a central frontier for user-facing AI agents. To be helpful in everyday life, agents must understand users across the digital contexts where their preferences, intents, habits, social relationships, and needs unfold over time. Today's systems can personalize within individual apps or tasks, but personal intelligence as a whole remains under-measured: how agents build cross-context user understanding, support steerable recommendation systems, act proactively across platforms, and avoid over-personalization. We introduce PersonaMem-v3, a real-world-grounded benchmark and evaluation harness for omni-platform personal intelligence. PersonaMem-v3 is seeded from more than one million anonymized real-world engagement histories, most of which are implicit signals, and uses them to construct time-indexed user digital worlds across social media, chatbot, calendar, and AI-companion with preference evolvement over time. The benchmark brings personalization, LLM-powered recommendation, proactiveness, agentic tool use, and geo-temporal reasoning into one framework, anchored in psychology, social-linguistics, and user-behavior theories. It evaluates whether AI agents can infer holistic user understanding from cross-platform evidence, personalize responses, rerank recommendations on social media, follow user steering through natural language, and hold back when personalization would be inappropriate, repetitive, outdated, or unnecessary. PersonaMem-v3 points toward LLM-powered personal intelligent agents that work with existing scalable recommendation infrastructure while making personalization more interactive, agentic, and aligned with how real users experience their digital lives.

Bowen Jiang, Yuan Yuan, Zhuoqun Hao +12

University of Pennsylvania · Meta Recommendation Systems · MIT

Personalized Large Language Model Agents

Ask PDF

Jul 28, 2026cs.AI

The User Asks, Platforms Compete: How Agentic Recommendation Markets Take Shape

Online recommendation has traditionally taken place after a user enters a platform, which determines the candidate pool and the ranking shown to the user. LLM-based user agents enable a different recommendation process: a user specifies a need before choosing a platform, leaving platforms to compete for the user's attention, which we refer to as an agentic recommendation market. In our controlled LLM-based experiments across three product domains, we find this new setting of recommendation creates a tension between access and attention. Compared with traditional platform-centric recommendation, user-centric recommendation greatly expands the opportunity for relevant items to enter comparison; yet broader participation does not translate directly into effective exposure. Competition directly triggers platforms' strategic play: selectively positive explanations occupy 73--78% of first-ranked positions. When the user agent relates platforms' actions to subsequent user feedback, this share falls to 36--41%, while the chance of a user purchasing the relevant item increases. A user agent is therefore more than a ranker over a larger pool of candidates: its querying, ranking, and feedback mechanism governing who can compete, how scarce attention is allocated, and how earlier outcomes shape the evaluation of platforms directly affect user utility. Designing agentic recommendation therefore requires treating access, attention, and accountability as a joint mechanism design problem.

Deyao Hong, Kehan Zheng, Qian Li +3

Tsinghua University · Tencent Inc.

Agentic Commerce Agentic Feedback Loop

Ask PDF

Jul 22, 2026cs.IR

Personalized Recommendation Tool Learning via Autonomous Language Agents

Although large language models (LLMs) have recently gained traction in recommender systems due to their strong reasoning capabilities and extensive world knowledge, previous LLM-based agents suffer from hallucination and context-length limitations, and thus are not suitable for full-ranking recommendation tasks. To circumvent these limitations through architectural design rather than modifying the LLM itself, we propose an agent-based recommendation framework, memory-based

\textbf{P}

ersonalized

\textbf{R}

ecommendation

\textbf{T}

ool learning via autonomous language

\textbf{A}

gents (PRTA), in which an LLM acts as a central planner interacting with multiple recommendation models as tools. The LLM-based agent is responsible for high-level reasoning and personalized tool selection, while traditional recommendation models perform full-ranking scoring, leveraging their scalability in modeling behavioral patterns. To support personalized tool selection, we design reflection mechanisms that enable the agent to evaluate and compare tools for each user based on user profiles and candidate ranked lists. Extensive experiments across three public datasets demonstrate the superiority of \modelname over traditional recommendation and LLM-based baselines in improving full-ranking recommendation performance.

Mingdai Yang, Zhiwei Liu, Weizhi Zhang +3

Univ. of Illinois Chicago Chicago, IL, USA · Microsoft Mountain View, USA · Beihang University, & Hangzhou Innovation Institute of BUAA Beijing & Hangzhou, China

Multi-Llm Agents Large Language Model Personalization Recommendation Multilingual Agents Language Pairs

Ask PDF