Source-linked AI summary
Personal-Agent Mediated Recommendation with Cross-Platform User History
Yu Xia, Jiangfan Zhang, Jun Xiao, Julian McAuley, Xiangjun Fan
TL;DR
Personal-agent recommendation asks when cross-platform history justifies changing a strong platform ranking that also reflects population evidence unavailable to the agent. The paper formalizes this mediation task, introduces MediateRec, and proposes PAMO to estimate history-based support and reallocate advantage. Across proxy target platforms and a real cross-platform test, PAMO improves over matched outcome-only RL while achieving a stronger rescue–harm balance.
Problem
Existing agentic recommendation benchmarks mainly assess platform-side agents using within-platform information, leaving open when cross-platform history justifies changing a strong platform ranking.
Source: §1 · pp. 1–2Method
PAMO counterfactually masks cross-platform history to estimate personal mediation support, then reallocates rank-aware advantage under a platform-relative value floor.
Source: §1 · pp. 2–3, §5.3.1 · p. 8Results
Across seen and unseen proxy target platforms and the real cross-platform test, PAMO improves over matched outcome-only RL with a stronger rescue–harm balance.
Source: §7 · p. 14Takeaways & Limitations
Effective mediation requires learning when cross-platform history should override or preserve a platform proposal, rather than relying on history access alone.
Source: §1 · p. 2, §7 · p. 14The optimality result concerns reallocations within a fixed sampled rollout group and does not imply global neural-policy optimality or guaranteed held-out utility improvement.
Source: §5.4 · p. 10
Abstract
from arXiv · showhide
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.
1 Introduction
Personal-Agent Mediated Recommendation gives a user-authorized agent cross-platform history to mediate a platform ranking, balancing rescues against harmful overrides. The paper introduces MediateRec and PAMO to study and train this mediation.
1 Introduction
- A platform ranks candidates using platform-local information, while a personal agent uses cross-platform history to produce the final top-K slate.The platform proposal can encode collaborative patterns across users unavailable to the agent.Source: §1 · p. 1
- Mediation must correct platform misses without overriding correct platform decisions, making selective use of cross-platform history central.An agent may rescue a target the platform misses, but may also drop a target the platform ranks correctly.Source: §1 · p. 2
- MediateRec pairs a platform proposal based on local information with history available only to the agent, using scalable proxy environments and a real cross-platform external test.Strong proprietary LLMs make meaningful corrections but still introduce non-negligible harmful overrides.Source: §1 · p. 2
- PAMO estimates rationale dependence on cross-platform history by masking that history, then uses the resulting support to guide advantage allocation under a platform-relative value floor.Ranking outcomes determine advantage sign and total amount; personal mediation support guides allocation among responses that disagree with the platform.Source: §1 · pp. 2–3
- PAMO preserves positive and negative advantage mass and is locally optimal among first-order reallocations that do not lower average platform-relative value.The stated optimality is among reallocations preserving that mass.Source: §1 · p. 2
- PAMO improves over matched outcome-only RL in recommendation accuracy and rescue–harm balance across seen and unseen target platforms and the real cross-platform test.The paper also characterizes PAMO's local optimality among first-order reallocations that preserve advantage mass without lowering average platform-relative value.Source: §1 · p. 2
2 Related Work
Related work covers platform-side agentic recommenders and user-governed personalization, distinguishing this paper's mediation of a strong platform proposal using cross-platform history.
2.1 Agentic Recommender Systems
- Prior agentic recommender systems add reasoning, tools, memory, or coordination to platform-side recommendation; PAMO instead targets mediation over an existing platform ranking.The related work also describes adaptation through memory and skill evolution, supervised trajectory tuning, and reinforcement learning from recommendation feedback.Source: §2.1 · p. 3
2.2 User-Governed Personalization and Agents
- Unlike systems that rerank sampled candidate lists or build unified cross-source slates, this setting mediates a frozen, strong platform proposal using cross-platform history.The benchmark uses public interaction traces and includes a test-only evaluation on real linked services.Source: §2.2 · p. 3
3 Personal-Agent Mediated Recommendation
Personal-Agent Mediated Recommendation gives a personal agent the platform proposal and user-authorized histories, including cross-platform records unavailable to the platform. Its objective measures platform-relative gains while balancing rescues against harmful changes to the platform slate.
3.1 Task Formulation
- The platform supplies a proposal over candidates, and the agent uses within- and cross-platform history to produce a ranked final slate without access to platform scores or internal state.The platform ranking may encode population-level collaborative evidence the agent cannot observe, making this selective mediation rather than ranking from scratch.Source: §3.1 · p. 4
3.2 Mediation Objective
- Training maximizes expected recommendation utility, while platform-relative mediation value distinguishes beneficial corrections from harmful changes to the fixed platform slate.Because the platform slate is fixed with respect to the agent policy, maximizing expected mediation value is equivalent to maximizing expected utility.Source: §3.2 · p. 4
- Top-K mediation separates preserved hits, rescues, harmful overrides, and mutual misses; NDCG@K also reflects target-item promotions and demotions.Effective mediation therefore seeks high-value corrections without degrading strong platform decisions.Source: §3.2 · p. 4
4 MediateRec Benchmark
MediateRec evaluates personal-agent mediation across scalable Amazon proxy environments and a real cross-platform OpenPlay test. Its construction preserves a platform–agent information boundary while supporting seen-platform, unseen-platform, and external evaluation.
4 MediateRec Benchmark
- MediateRec provides four Amazon proxy environments for scalable training and evaluation plus an OpenPlay real cross-platform external test.The platform uses within-platform information, while the personal agent additionally receives cross-platform history.Source: §4 · p. 4
4.1.1 Amazon Proxy Platforms
- Amazon target categories use local interactions as within-platform history and the same user’s earlier records in other categories as cross-platform history.Movie and Toy each have 5,000 training, 1,000 validation, and 2,000 test episodes; Grocery and Beauty each have 2,000 test episodes for unseen-platform transfer.Source: §4.1.1 · p. 5
4.1.2 OpenPlay: Real Cross-Platform Data
- OpenPlay links Steam, Nintendo, and Xbox records through a shared pseudonymized person identifier, with Steam as the target platform.Its 645-user cohort is used only for external testing and excluded from training, hyperparameter selection, and checkpoint selection.Source: §4.1.2 · p. 5
4.1.3 Temporal Episodes and Candidate Sets
- Each episode holds out a temporally later target, and both agents rank the same 50 candidates: the target and 49 previously unobserved items sampled by popularity.The candidate set is held fixed across agent conditions, so platform-relative changes reflect mediation rather than different retrieval pools.Source: §4.1.3 · p. 5
4.2.1 Target Platform Ranking
- The frozen SASRec-plus-Claude platform ranks candidates using local information, achieving mean HR@10 / NDCG@10 of 49.3%/0.329 across Amazon and 40.8%/0.241 on OpenPlay.Cross-platform history is never exposed during platform-ranking construction, and the resulting ranking is frozen for agent training and evaluation.Source: §4.2.1 · pp. 5–6
4.2.2 Personal-Agent Mediation
- The personal agent receives the platform proposal and both histories, but not platform scores, population interactions, or model states, then returns a ranked top-10 in one call.The prompt treats platform rank as meaningful evidence and asks for revisions only when cross-platform history provides specific support.Source: §4.2.2 · p. 6
4.3 Evaluation Protocol
- Evaluation uses disjoint users and identical users, targets, candidates, and platform rankings across methods, reporting ranking metrics alongside rescue, harmful override, and intervention.Movie and Toy are aggregated as seen platforms, Grocery and Beauty as unseen platforms, and OpenPlay is reported separately as the real cross-platform external test.Source: §4.3 · p. 6
5 Personal Attribution Mediation Optimization
PAMO distinguishes history-supported corrections from generic reranking by estimating cross-platform-history dependence, then reallocates rank-aware advantages under a value floor. It preserves cutoff-level advantage mass and is locally optimal for support gain among admissible first-order reallocations.
5 Personal Attribution Mediation Optimization
- PAMO estimates cross-platform-history dependence, decomposes NDCG into rank-aware group-relative advantages, and reallocates advantage toward supported responses without lowering average platform-relative value.The reallocation preserves the average platform-relative NDCG magnitude of the original uniform allocation.Source: §5 · pp. 6–7
5.1 Counterfactual Personal Mediation Support
- Counterfactual masking measures how strongly each sampled rationale depends on cross-platform history while holding the response and platform proposal fixed.The score is model-relative: it measures history dependence, not whether that history is correct or useful.Source: §5.1 · p. 7
5.2 Rank-Aware Platform-Relative Advantage
- NDCG is decomposed across nested top-k cutoffs, and PAMO reallocates advantages only among rollouts whose cutoff outcome differs from the platform’s.Each marginal cutoff utility represents the NDCG gain from moving the target across that cutoff.Source: §5.2 · pp. 7–8
5.3.1 Value-Constrained Allocation
- The value floor prevents support-based reallocation from lowering the average direction-aligned NDCG magnitude of the original uniform allocation.Direction-aligned magnitude distinguishes stronger rescues from more severe harmful overrides.Source: §5.3.1 · p. 8
5.3.2 PAMO Advantage
- PAMO allocates a fixed signed advantage mass within each platform-disagreement set, with κ controlling how strongly allocation responds to personal mediation support.κ = 0 gives a uniform allocation; larger values increase responsiveness to support differences.Source: §5.3.2 · p. 9
5.4 Theoretical Analysis
- PAMO preserves aggregate positive and negative advantage mass and locally maximizes first-order support gain among reallocations that preserve mass without lowering average direction-aligned NDCG magnitude.This optimality applies within a fixed sampled rollout group with support scores and value magnitudes held fixed; it does not imply global policy optimality or guaranteed held-out utility improvement.Source: §5.4 · pp. 9–10
6 Experiments
The experiments test whether personal-agent mediation and PAMO improve a frozen platform proposal, how much cross-platform history contributes, and when the agent should intervene. Results show that history alone is insufficient, while PAMO improves recommendation and rescue–harm outcomes over matched outcome-only training.
6.1.1 Models and Baselines; 6.1.2 Training; 6.1.3 Evaluation
- The platform proposal is the reference for rescue, harmful override, and intervention; the outcome-only RL baseline shares PAMO’s listed training setup, differing only in advantage reallocation, and evaluation reports HR@{3,5,10} and NDCG@{3,5,10}.The baseline shares PAMO’s SFT initialization, training data, rollout budget, optimizer, and update budget. For each episode, candidate IDs from the response’s FINAL: marker are deduplicated and kept in best-first order.Source: §6.1.1 · p. 11, §6.1.3 · p. 11
6.2.1 Access to Cross-Platform History Is Insufficient
- The base Qwen3-4B-Instruct falls below the platform on all five datasets, while its harmful-override rate exceeds its rescue rate throughout.Sonnet and Opus improve HR@10 over the platform on all five datasets, but still have harmful overrides.Source: §6.2.1 · p. 11
6.2.2 PAMO Improves Outcome-Only Training
- PAMO improves HR@10 and NDCG@10 over GRPO on all five datasets, reduces harmful overrides on four, and achieves strong results on unseen targets and OpenPlay.On MediateRec-OpenPlay, without Open Play training or model selection, PAMO obtains the highest ranking metrics.Source: §6.2.2 · pp. 11–12, Table 2 · p. 12, Table 3 · p. 12
6.3 Cross-Platform History Contribution (RQ2)
- Adding cross-platform history improves HR@10 and NDCG@10 for every evaluated model, with the largest gains for Sonnet, Opus, and PAMO.The comparison masks only cross-platform records while holding the candidate set, platform proposal, and within-platform history fixed.Source: §6.3 · p. 12, Figure 2 · p. 13
6.4.1 Effect of Support Concentration κ
- Increasing κ to 0.5 improves HR@10 and NDCG@10, but performance weakens at κ = 0.75; the main experiments use κ = 0.5.The reported pattern is an inverted U.Source: §6.4.1 · pp. 12–13, Figure 3 · p. 13
6.4.2 Role of the Value Floor
- Without the value floor, support-only reallocation performs below the κ = 0 baseline, whereas full PAMO improves HR@10 and NDCG@10.The floor prevents history-dependent responses from receiving favorable reallocation solely because of that dependence.Source: §6.4.2 · p. 13, Figure 3 · p. 13
6.5.1 Aggregate Mediation Behavior
- Cross-platform user context is the primary driver in 72% of PAMO’s 900 rescues; interests and fandom account for 28%, household and family 19%, and lifestyle, diet, and health 14%.Incidental rescues account for 25%, with only a small fraction attributed primarily to within-platform taste.Source: §6.5.1 · p. 13
6.5.2 Intervening versus Deferring
- PAMO promotes a concert film from platform rank 37 to rank 1 when cross-platform music history specifically supports it, but preserves a top-ranked Zelda collectible when evidence is weak.The two representative rationales illustrate intervention under specific evidence and deference under weak evidence.Source: §6.5.2 · p. 14, Figure 4 · p. 14
7 Conclusion
The paper formalizes personal-agent mediation, introduces MediateRec and PAMO, and reports that PAMO improves over matched outcome-only RL with a stronger rescue–harm balance across proxy and real cross-platform tests.
7 Conclusion
- Across seen and unseen proxy target platforms and the real cross-platform test, PAMO improves over matched outcome-only RL while achieving a stronger rescue–harm balance.The conclusion emphasizes that effective mediation requires learning when cross-platform history should override population evidence encoded by the platform ranking.Source: §7 · p. 14
Appendix
- The appendix provides further benchmark-construction and implementation details, PAMO derivations, and complete proofs of the theoretical results.Source: Appendix · p. 18
A MediateRec Benchmark Details
The appendix details MediateRec’s proxy and external-test construction, platform ranking, and personal-agent interface. Its scope includes candidate-conditioned evaluation and constraints in the Amazon and OpenPlay data.
A.1.1 Source and Episodes; A.2.1 Source and cohort
- The Amazon proxy platforms use rating-only Amazon Reviews 2023 data, while the OpenPlay external test links pseudonymized Steam, Nintendo, and Xbox records.The OpenPlay cohort contains 645 users after target and history eligibility requirements.Source: §A.1.1 · p. 18, §A.2.1 · p. 18
A.1.1 Source and Episodes
- Amazon episodes hold out each eligible user’s final target-platform interaction, require nonempty histories that precede it, and assign users disjointly across targets and splits.A fixed Movie, Toy, Grocery, Beauty assignment order can favor earlier target platforms.Source: §A.1.1 · p. 18
A.1.2 Candidate Sets; A.2.3 Candidate Sets
- Each episode has one target and 49 distinct negatives sampled by target-category popularity, with catalogs and popularity counts drawn from the full population.The OpenPlay episodes likewise use 49 popularity-sampled Steam negatives and exclude previously played titles.Source: §A.1.2 · p. 18, §A.2.3 · p. 19
A.2.2 Temporal Episodes and Histories
- OpenPlay selects each user’s last-adopted Steam title with at least 30 minutes of total playtime; cross-platform histories include Nintendo and Xbox records strictly before adoption, with Xbox records retaining genres.Cross-platform histories are capped at 50 records; mobile activity is omitted because game identities are unavailable.Source: §A.2.2 · pp. 18–19, §A.5 · p. 21
A.3 Target Platform Construction; A.3.1 Platform LLM Reranking Prompt
- SASRec ranks candidates using within-platform information, then Claude Sonnet reranks using within-platform history; cross-platform history is withheld and the resulting platform ranking is frozen.The platform prompt does not contain Nintendo or Xbox history.Source: §A.3 · p. 19, §A.3.1 · p. 19
A.4 Personal-Agent Interface; A.4.1 Default Mediation Prompt
- The personal agent receives the frozen 50-candidate ranking and both histories, then returns ten candidates without observing platform scores, population interactions, or model states.Its prompt treats the platform ranking as meaningful evidence and asks it to intervene only when cross-platform history provides specific support.Source: §A.4 · p. 20, §A.4.1 · p. 20
A.5 Validation and Scope
- MediateRec evaluates candidate-conditioned ranking rather than end-to-end retrieval because the target is inserted into every candidate set.The Amazon construction also uses static catalogs and full-population popularity counts not restricted to items available before each target timestamp.Source: §A.5 · p. 20
B PAMO Derivations, Proofs, and Details
The appendix derives PAMO’s cutoff-level advantage allocation and gives its dual form, including conditions under which the value-floor constraint binds.
B.1 NDCG Decomposition and Group-Relative Consistency
- At each active cutoff, PAMO’s uniform allocation matches centered NDCG supervision and preserves the standard group-relative advantage mass; inactive cutoffs have zero advantages.The derivation identifies successful rollouts when the platform misses and failed rollouts when it succeeds as the relevant groups.Source: §B.1 · p. 21
B.2 Dual Form of the PAMO Allocation
- In the dual allocation, λk is zero when the support-only allocation satisfies the value floor; otherwise the constraint binds, with a unique coefficient found by one-dimensional root finding when values differ.When all group-relative values are equal, the value floor is vacuous and λk is set to zero.Source: §B.2 · p. 22
B.3 Proof of Proposition 1
The proof tracks retained advantages for failed and successful platform rollouts, then invokes the stated feasibility requirement and optimality condition.
B.3 Proof of Proposition 1
- Failed rollouts retain advantage −pk and sum to −mk, while successful rollouts retain advantage 1 −pk and sum to mk.The passage introduces the failed-rollout rescue advantages but does not supply their expression.Source: §B.3 · p. 22
- The first inequality in Eq. (23) is the feasibility requirement in Eq. (17).Source: §B.3 · p. 22
- The second inequality in Eq. (23) follows from the feasibility of uk and the optimality of w⋆, with equality when k = uk.Source: §B.3 · p. 22
B.4 Proof of Theorem 1
The proof derives the constrained allocation’s local behavior and shows that its admissible first-order reallocation maximizes personal mediation support gain.
B.4 Proof of Theorem 1
- At η = 0 and λ = 0, the normalized exponential weights are uniform; when evk = 0, the value floor is met with equality by every allocation and λk(η) may be set to zero.Source: §B.4 · p. 22
- The KKT conditions set λk(η) to zero when the unconstrained support tilt is feasible, and otherwise choose the active boundary λbd.Source: §B.4 · p. 23
- For the boundary case ⟨eck, evk⟩ = 0, λk(η) = o(η).Source: §B.4 · p. 23
- The centered-vector projection qk is the Euclidean projection of eck onto the simplex tangent space.Source: §B.4 · p. 23
- When qk is nonzero, qk/∥qk∥2 maximizes the admissible first-order support gain; when qk is zero, no admissible reallocation yields a positive first-order gain.Source: §B.4 · p. 23
B.5 Training Objective and Cost
PAMO uses response-level advantages in the matched clipped token-level GRPO objective, with one additional scoring pass and no additional rollouts.
B.5 Training Objective and Cost
- PAMO applies the response-level advantage in the same clipped token-level GRPO objective as the matched baseline.Source: §B.5 · p. 24
- The same scalar advantage weights all generated tokens; padding is masked, support scores and allocation weights are detached, and full-history rollout log-probabilities form the denominator.Source: §B.5 · p. 24
- PAMO adds one within-only teacher-forced scoring pass but no rollouts or inference-time component; allocation is solved independently at each active cutoff over at most G responses.Source: §B.5 · p. 24