PaperlayerView exact revision on arXiv

Source-linked AI summary

Position: Recommender Systems Should Move Beyond Platform-Centric Ranking toward Personal Agent-Mediated Recommendation

Haohan Yuan, Peng He, Dan Zhang, Jianpeng Liang, Junning Zhu

arXiv:2609.11942v1cs.IR

TL;DR

Platform-centric recommenders control candidate access, evidence boundaries, and disclosure, leaving a gap in user-side mediation of recommendation evidence. The paper defines PAMR as a user-facing agent paradigm for discovering, filtering, aggregating, and governing distributed evidence, and reports that selective source choice plus disclosure control gives the strongest observed operating point under a shared LLM ranker. As a position paper, it frames this as a research direction rather than evidence of improved real-world user outcomes.

  • Problem

    Platform-centric recommendation controls evidence access and disclosure, creating a research gap around explicit, traceable, and adaptive user-side mediation.

  • Method

    The paper defines PAMR, specifies its mediation decisions and evaluation framework, and studies source selection and disclosure control in a controlled proof of concept.

  • Results

    Under a shared LLM ranking layer, source selection and controlled disclosure provide the strongest observed utility–traceability–exposure–cost operating point among the multi-source conditions.

  • Takeaways & Limitations

    Recommendation research should evaluate who controls evidence, what is disclosed, how provenance and disagreement are handled, and how mediation changes over time.

  • Takeaways & Limitations

    The paper is a position paper, and its controlled proof of concept uses simulated sources in a single Yelp restaurant domain with fixed candidate sets and hand-designed budgets.

Abstract

from arXiv · show

Recommender systems are usually framed as ranking systems: platforms observe users, construct candidate sets, and select items on their behalf. This framing hides a deeper allocation of control, in which platforms also determine candidate access, evidence boundaries, explanations, and the path from user need to recommended output. We argue that the next bottleneck in recommendation is not only preference modeling, but control over evidence acquisition and disclosure. We argue for \textbf{Personal Agent-Mediated Recommendation} (PAMR), a paradigm in which a user-facing personal agent represents the user in discovering, filtering, aggregating, and governing recommendation evidence across distributed sources. The central shift is not simply from one ranking model to another, but from platform-side item ranking to user-side evidence mediation. As a position paper, we define PAMR as a new recommendation paradigm, establish its boundary criteria, identify its core mediation decisions, and propose a mediation-centered evaluation framework. A proof-of-concept study on hard Yelp restaurant recommendation tasks further shows that, under a shared LLM ranker, source selection and controlled disclosure provide the strongest observed utility--traceability--exposure--cost operating point.

1 Introduction

The paper argues that recommendation must address control over evidence acquisition and disclosure, not only preference modeling. It proposes PAMR, where a user-facing agent mediates distributed evidence through inspectable policies and evaluates both recommendation utility and user control.

  • 1 Introduction: Platforms commonly control candidate access, evidence boundaries, explanations, source selection, and the path from user need to ranked output.The paper identifies this allocation of control as an often-invisible limitation of conventional ranking systems.
  • 1 Introduction: A personal agent can route different constraints to relevant sources, record disclosed context, aggregate evidence, preserve disagreement, and incorporate user trust preferences.The motivating restaurant example separates availability, atmosphere, and dietary-fit evidence across sources.
  • 1 Introduction: PAMR shifts recommendation from platform-side item ranking to user-side mediation of distributed evidence, disclosure, provenance, disagreement, and future policy adaptation.The paradigm changes the evidence boundary, decision locus, traceability, and adaptation target.
  • 1 Introduction: The paper defines PAMR through evidence-boundary, decision-locus, traceability, and mediation-policy-adaptation shifts, distinguishing it from adjacent agentic recommendation paradigms.Its contribution treats mediation control as a first-class research object rather than an implementation detail.
  • 1 Introduction: PAMR’s core mediation decisions are source discovery and routing, privacy-budgeted disclosure, provenance-preserving aggregation, and feedback-driven adaptation.The paper presents these decisions as the operational basis of user-inspectable mediation.
  • 1 Introduction: A proof-of-concept study reports complementary contributions from source selection and disclosure control under a shared LLM ranking layer.The study is introduced as controlled evidence that selective mediation can improve the utility–traceability–exposure–cost operating point.

2 Related Paradigms and Gaps in User-Side Mediation

PAMR is positioned as distinct from ranking, conversational, tool-using, personal-agent, and ecosystem approaches because it makes user-side evidence mediation the shared research object. Its distinguishing criteria concern who controls evidence access and request-time mediation, whether the process is traceable, and what adapts.

  • 2 Related Paradigms and Gaps in User-Side Mediation: PAMR’s boundary is defined by who controls the evidence boundary, who makes request-time mediation decisions, whether the process is traceable, and what behavior adapts.The paper treats allocation of mediation as the key distinction from adjacent paradigms.
  • 2 Related Paradigms and Gaps in User-Side Mediation: Traditional, social, federated, and neural recommenders broaden signals or protect training locality but remain primarily platform-bounded ranking paradigms.These approaches model users and items or expand available signals without making user-side evidence mediation the central object.
  • 2 Related Paradigms and Gaps in User-Side Mediation: LLM reasoning and ranking methods strengthen interpretation and ranking over evidence, but their primary object remains model-side recommendation rather than control of evidence acquisition.The paper contrasts better reasoning over fixed or system-defined evidence with PAMR’s mediation focus.
  • 2 Related Paradigms and Gaps in User-Side Mediation: Conversational and tool-using agents make interaction, external knowledge, delegation, and proactive behavior active decisions, yet usually optimize accuracy, dialogue quality, goal progression, or agent policy.PAMR instead centers source access, disclosure, provenance, disagreement, and mediation-policy adaptation.
  • 2 Related Paradigms and Gaps in User-Side Mediation: iAgent is a close precursor that shields users from platform recommenders through instructions, tools, and feedback memory, while PAMR moves control to evidence acquisition, per-source disclosure, and provenance-bearing aggregation.The distinction is between instruction-aware reranking and explicit evidence mediation.
  • 2 Related Paradigms and Gaps in User-Side Mediation: TRACE, privacy-aware recommendation, decoupled recommender systems, and mechanism-design work reinforce accountability and ecosystem control but do not constitute PAMR’s unified mediation object.These adjacent directions motivate PAMR’s emphasis on accountability, privacy, and strategic influence over evidence and interactions.

3 Defining Personal Agent-Mediated Recommendation

PAMR defines recommendation as a user-facing, user-governable process for acquiring and governing evidence across distributed sources, rather than merely ranking items. Its defining criteria couple evidence scope, mediation control, traceability, and policy adaptation.

  • 3.1 Definition and Core Criteria: PAMR is a recommendation setting where a personal agent acquires, filters, governs, presents, and adapts evidence across distributed sources under inspectable user policies.The system evaluates both recommendation quality and mediation of evidence access, disclosure, provenance, disagreement, and future policy.
  • 3.1 Definition and Core Criteria: PAMR requires mediation to move from platform-controlled ranking to a user-facing process that users can inspect, revise, correct, override, or revoke.Multiple agents alone do not establish PAMR unless user-side evidence mediation is optimized or evaluated.
  • 3.2 Four Defining Shifts: Four shifts distinguish PAMR: broader evidence boundaries, a user-governable decision locus, provenance-rich traceability, and adaptation of mediation policy beyond item preference.Together, these shifts separate PAMR from systems that merely add tools, chat, or LLM reranking to platform-centric recommendation.
  • 3.3 Boundary Cases: Components and Coupled Mediation: PAMR treats evidence boundary, per-source disclosure, provenance-bearing aggregation, disagreement, and policy updates as one coupled recommendation process.Existing systems may supply individual ingredients, but no single capability is sufficient by itself.
  • 3.3 Boundary Cases: Components and Coupled Mediation: Recent systems provide ingredients such as external retrieval, user memory, interaction policies, uncertainty-guided delegation, and collaborative memory, but remain partial PAMR without explicit mediation control.Tool-using and travel-planning agents show infrastructure for distributed evidence access, while shopping assistants show agentic recommendation interfaces.

4 Core Decisions in Personal Agent-Mediated Recommendation

PAMR frames recommendation as constrained information acquisition in which source routing and context disclosure are explicit user-side decisions. Its mediation loop also preserves provenance and uncertainty while adapting source and disclosure policies through feedback.

  • 4 Core Decisions in Personal Agent-Mediated Recommendation: PAMR makes evidence acquisition and disclosure explicit recommendation decisions alongside utility, query cost, privacy cost, latency, and trust or provenance risk.The formulation clarifies what platform-centric ranking typically leaves implicit.
  • 4 Core Decisions in Personal Agent-Mediated Recommendation: Source routing selects peers, services, communities, platforms, marketplaces, or referrals under expertise, trust, cost, latency, privacy, incentive, diversity, and user-control constraints.Source selection is treated as a recommendation variable rather than a backend retrieval detail.
  • 4 Core Decisions in Personal Agent-Mediated Recommendation: Context disclosure determines what each selected source receives, balancing richer relevance-producing context against increased privacy exposure through privacy-budgeted translation.Possible disclosures range from intent alone to task constraints or relationship-sensitive context.
  • 4 Core Decisions in Personal Agent-Mediated Recommendation: Aggregation preserves provenance, disagreement, uncertainty, and freshness instead of collapsing heterogeneous responses into an unsupported ranked list.Stale, duplicated, or self-promotional evidence may be discounted or surfaced as uncertainty.
  • 4 Core Decisions in Personal Agent-Mediated Recommendation: User inspection and feedback extend the mediation loop from item preferences to source preferences, trust estimates, routing weights, referral paths, and disclosure policies.The interface should show consulted sources, disclosed context, and supporting or conflicting evidence.

5 Source Selection and Disclosure Control

A proof-of-concept Yelp study tests source selection and disclosure control under a shared ranker. Selected, controlled mediation achieves the strongest observed balance of utility, traceability, exposure, and evidence cost among multi-source conditions.

  • 5 Source Selection and Disclosure Control: The study varies evidence scope between all five sources and a selected top-three subset, and varies disclosure under a shared LLM ranking pipeline.It is a partial PAMR instantiation intended to illustrate trade-offs rather than provide a benchmark or complete system evaluation.
  • 5 Source Selection and Disclosure Control: The evaluation measures recommendation utility with HR@3, trace verification with Trace Validity, disclosure burden with Exposure, and acquisition cost with Evidence Calls.Mediation metrics are reported only for multi-source conditions.
  • 5 Source Selection and Disclosure Control: .740 HR@3 is reached by selected and controlled mediation, compared with .675 for the platform-only LLM reranker and .230 for platform ranking.The comparison uses 200 hard Yelp restaurant tasks with ten candidates, one target, and constraint-violating hard negatives.
  • 5 Source Selection and Disclosure Control: Source selection improves HR@3 from .705 to .740 under controlled disclosure and from .690 to .715 under full disclosure, while reducing evidence calls by 40%.The study therefore links selective access to both recommendation outcome and evidence cost under the shared ranking pipeline.
  • 5 Source Selection and Disclosure Control: Controlled disclosure reduces exposure by 77–78% and produces near-complete trace validity without an observed utility penalty.Exposure is a within-study weighted disclosure score, while evidence calls count source–candidate evidence requests.

6 Scope Conditions and Failure Modes

PAMR is most applicable when relevant evidence is distributed, contextual, trust-sensitive, or cross-platform, but mediation should remain selective. Additional evidence can be inappropriate when a single authoritative source suffices or when latency, cost, privacy, and interaction burdens dominate.

  • 6 Scope Conditions and Failure Modes: PAMR is most useful when no single platform owns the relevant evidence boundary, including long-tail, context-sensitive, trust-sensitive, and cross-platform decisions.Examples include local services, accessibility-sensitive choices, agent or service selection, and community evidence outside platform inventory.
  • 6 Scope Conditions and Failure Modes: When one source is authoritative, current, and complete, or when a request is low-stakes or time-critical, distributed acquisition may add burden without improving the decision.The agent should estimate whether additional evidence justifies its exposure and interaction costs.
  • 6 Scope Conditions and Failure Modes: Selective use does not reject platforms; it makes reliance on them an explicit and revisable mediation choice.The agent may answer locally, ask before expanding the evidence boundary, or escalate when sources disagree or confidence is low.
  • 6 Scope Conditions and Failure Modes: PAMR introduces risks from missed sources, vague queries, over-disclosure, unsupported or promotional evidence, collapsed disagreement, stale evidence, and coordinated influence.Provenance should track origin, beneficiaries, source independence, and repeated or coordinated evidence.

7 Open Research Challenges

PAMR raises open research challenges around adaptive evidence mediation, user governance, robustness, institutional control, and evaluation of mediation itself.

  • Adaptive source routing and provenance-aware aggregation: PAMR must adapt source routing and aggregate conflicting or uncertain evidence while preserving provenance.Mediators should learn source usefulness and context allocation for each request without erasing evidence origins.
  • Policy learning under user governance: PAMR should let users inspect and revise disclosure, trust, routing, and memory policies without making adaptation opaque.The challenge is learning from feedback at a practical granularity while retaining policy control.
  • Robustness to strategic sources: Strategic sources require source-specific provenance, disclosure logs, disagreement tracking, user-correctable policies, and longitudinal evaluation.Sources may manipulate referrals, confidence, or returned evidence, expanding the manipulation surface beyond final reranking.
  • Institutional user governance: User-facing agents also need institutional safeguards, including policy portability, auditable access, revocation, interoperable provenance, and provider separation.An agent can remain influenced by platforms, marketplaces, operating-system providers, or model vendors despite facing the user.
  • Mediation-centered benchmarks and user studies: PAMR evaluation should vary evidence topology, source reliability, incentives, disclosure budgets, and provider availability, using counterfactual and longitudinal tests.Tests should examine changes after source removal, sensitive-context withholding, or trust-rule revision, as well as users’ understanding and policy calibration.

8 Conclusion

The conclusion frames recommendation as governance of evidence, not only item ranking, and positions PAMR as a first-class research object rather than a finished architecture.

  • 8 Conclusion: PAMR relocates evidence mediation to a user-facing agent that selects sources, controls disclosure, preserves provenance, and adapts future mediation.The shift changes who controls the evidence boundary and what recommendation systems should optimize and evaluate.
  • 8 Conclusion: Selective mediation achieved the strongest observed operating point among multi-source conditions without sacrificing recommendation utility.The proof-of-concept controlled which evidence entered the ranker and what context reached each source.
  • 8 Conclusion: The paper presents PAMR as a conceptual and diagnostic research direction, not a complete system that has demonstrated improved real-user outcomes or practical control shifts.Its partial instantiation supports making user-side evidence mediation a first-class research object.

Limitations

The paper’s evidence is constrained by its position-paper status, diagnostic simulated study, limited privacy measurement, absent longitudinal evaluation, and unresolved deployment governance.

  • Scope: This position paper defines a research space and evaluation protocol but does not establish improved real-user outcomes or practical control shifts.It is not a finished system, deployed agent, or benchmark.
  • Study scope: The proof-of-concept uses simulated sources, one Yelp restaurant domain, 200 hard tasks, fixed candidates, and handdesigned evidence-call budgets.Source selection, disclosure policies, and routing rules are controlled by the setup rather than learned from real users or deployed services.
  • Privacy measurement: The weighted exposure score uses manually specified privacy weights and should be interpreted only as a within-study diagnostic.It is not a dataset-independent privacy measurement.
  • Longitudinal evaluation: The study does not evaluate longitudinal mediation, feedback-driven routing, long-term evidence memory, or users’ ability to inspect and override policies over time.Point estimates also require paired uncertainty tests and cross-dataset validation before small utility differences are treated as reliable.
  • Deployment governance: Current evidence does not show that user-facing agents become user-owned or user-governable in deployments controlled by platforms, enterprises, or marketplaces.The claim remains conceptual and diagnostic rather than evidence that deployed agents solve the control problem.
  • Failure boundaries: PAMR failures span discovery, context, response, aggregation, presentation, adaptation, and referral, so the taxonomy is a diagnostic checklist rather than a closed set.Referral should be depth-limited and trust-aware to avoid runaway cost, noisy paths, manipulation, and feedback loops.

D Proof-of-Concept Study Setup

The diagnostic study evaluates source selection and disclosure control on hard Yelp restaurant tasks using simulated evidence channels, shared ranking conditions, and mediation-oriented metrics.

  • Task and dataset: The study uses 200 hard Yelp tasks with ten candidates, one constraint-satisfying gold target, mediation metadata, and highly rated hard negatives.The gold target is nonobvious under platform rating, averaging rank 5.49 and appearing in the top three in 23% of tasks.
  • Evidence environment: Five simulated sources provide platform metadata, recent reviews, community reviews, peer-like signals, and business-owned information.Selective Mediation chooses roughly three sources using relevance, trust, and cost, then applies persona-specific disclosure rules.
  • Experimental conditions: The comparison includes platform-only ranking, platform LLM reranking, querying all sources, source selection only, and disclosure control conditions.These conditions separate ranking, source-selection, and context-disclosure effects.
  • Evaluation metrics: HR@3 measures whether the gold target reaches the top three, while Mediation Trace Validity verifies candidate validity, evidence support, disclosure compliance, and trace completeness.Trace Validity is averaged across tasks as a binary per-task verifier.
  • Cost and exposure: Exposure sums weighted disclosure over user slots, using slot sensitivity and receiving-source risk; evidence calls quantify source–candidate requests.Query-all conditions issue 50 requests per task, source-selection conditions issue 30, and platform LLM reranking issues 10.
  • Diagnostic scope: The additional diagnostics support implementation analysis but are omitted from the main text because they are not required for the central comparison.The study therefore focuses its main comparison on the core mediation conditions and metrics.

G Full Results and Implementation Mapping

Table 3 reports full diagnostic results for how source scope and disclosure policy shape recommendation utility, exposure, and evidence-request cost, while Table 4 maps condition names to implementation identifiers.

  • Full diagnostic results: Table 3 is a full diagnostic table covering utility, validity, privacy, cost, exposure, and evidence-call measures across study conditions.Exposure is a within-study weighted disclosure score, and Evidence Calls count source–candidate requests rather than LLM inference calls.
  • Full diagnostic results: The study reports point estimates, so small utility differences require paired uncertainty tests before being treated as statistically reliable.
  • Source selection: HR@3 rises from .690 to .715 when Query-All changes to Selection Only under full disclosure, while Exposure falls from 68.2 to 37.7 and Evidence Calls from 50 to 30.
  • Implementation mapping: Table 4 links paper-facing condition names with implementation identifiers used in trace and metric files.
Loading 2609.11942v1…