Personal Agents Learn When to Override Recommendation Rankings

Introduction

Recommendation systems are increasingly moving beyond personalization controlled by a single platform. A personal language-model agent could, in principle, act on a user’s behalf across services, bringing information from one platform to another and helping the user avoid fragmented profiles. Yet this idea raises a basic question: should the agent always trust its own cross-platform view over the platform’s ranking?

A new paper featured in Hugging Face Daily Papers formalizes this setting as Personal-Agent Mediated Recommendation. The proposed workflow is deliberately divided into two stages. First, a platform recommender ranks a candidate set using information available only within that platform. Then, a personal agent uses user-authorized history from other services to mediate the ranking and produce the final Top-K slate.

Key ideas

• Separate responsibilities. The platform contributes local interactions, content signals, and population-level evidence. The personal agent contributes a broader view of the individual’s interests.
• Intervention must be selective. Cross-platform history can rescue items that a platform missed, but it can also lead the agent to replace a good recommendation without sufficient evidence.
• Evaluate under an information boundary. The proposed MediateRec benchmark includes scalable proxy cross-platform environments and a real cross-platform test conducted under a controlled boundary between platform and agent information.
• Attribute the value of personal evidence. Personal Attribution Mediation Optimization, or PAMO, counterfactually masks cross-platform history to estimate how much support that history provides for mediation. It then reallocates rank-aware advantage while enforcing a platform-relative value floor.

Why the training objective matters

A straightforward outcome-only reinforcement-learning objective can reward a final result without explaining whether the personal history actually justified the intervention. PAMO instead asks a more targeted question: what changed because the agent had access to the user’s cross-platform evidence? This counterfactual view helps distinguish a useful rescue from an unnecessary override.

The paper theoretically shows that PAMO preserves cutoff-level advantage mass and is locally optimal among first-order reallocations that preserve this mass without reducing average platform-relative value. In practical terms, the agent is encouraged to improve the ranking while retaining the valuable information embedded in the platform’s original ordering.

Experiments on MediateRec suggest that personal agents can make meaningful corrections to platform recommendations. However, even strong proprietary language models produce a non-negligible number of harmful overrides. PAMO consistently improves over matched outcome-only reinforcement-learning baselines across seen and unseen target platforms, as well as on the real cross-platform test described by the authors.

The broader implication is that a personal agent should not be treated simply as a more personalized recommender. It is better understood as a governance layer between the user and multiple platforms: it needs authorization boundaries, evidence-sensitive intervention, and a fallback to platform results when its own information is weak. Future systems will also need clear policies for revoking cross-platform access, resolving conflicts between platform and user interests, and auditing why a ranking was changed.

Source: Hugging Face Daily Papers