Source-linked AI summary
AgentWebRec: Compact Evidence Fusion over the Agent Web for Personalized Recommendation
Haoran Qiang, Guannan Liu, Liang Zhang, Junjie Wu
TL;DR
Recommendation over the Agent Web must draw on distributed, opaque memories while selecting evidence relevant to each decision. AgentWebRec progressively acquires and fuses that evidence without centralizing agent memories, and the paper reports consistent gains over baseline recommenders across four InstructRec datasets.
Problem
Agent memories are isolated and contain information from diverse contexts, so task-relevant evidence must be selected without aggregating unrelated records that can dilute preference signals.
Source: §2 · p. 3Method
Results
AgentWebRec achieves consistently strong performance across four InstructRec datasets and outperforms baseline recommenders.
Source: §1 · p. 2Takeaways & Limitations
Target-agent information and external-agent experience provide complementary benefits, while task-oriented acquisition avoids indiscriminately introducing all available agent information.
Source: §1 · p. 2
Abstract
from arXiv · showhide
LLM-based personal agents are emerging as persistent carriers of user semantics and intermediaries between users and recommendation platforms, maintaining richer user knowledge locally. As agents interact with one another, the conventional \textit{User--Platform} relation evolves into a \textit{User--Agent Web--Platform} information pathway, enabling distributed user-side information to complement item-side information. This new pathway, however, defies conventional recommendation: evidence is scattered across mutually opaque agents and reachable only through bounded queries, only a small portion of it is relevant to the current recommendation decision, and the responses returned by different agents are semantically heterogeneous. We therefore recast recommendation over the agent web as a \emph{task-time evidence acquisition and fusion} problem under a finite evidence budget by deciding what to ask and what to keep, rather than learning from aggregated data. We propose AgentWebRec, a user-agent-oriented framework that progressively acquires and fuses distributed evidence for each user-item decision while keeping underlying agent memories local. It grounds each decision in platform-provided item semantics and task-relevant evidence from the target user agent's private memory, and conditionally queries neighboring user agents for complementary preference patterns when local evidence is insufficient. Experiments on four InstructRec datasets show that AgentWebRec consistently outperforms baseline recommenders, and ablations verify that the evidence layers contribute complementary gains.
1 Introduction
Personal agents add persistent, locally maintained user knowledge, but a single agent may lack relevant evidence for a candidate item. AgentWebRec addresses this by acquiring and fusing task-relevant evidence across agents, and the paper reports consistent gains across four InstructRec datasets.
1 Introduction
- A user agent maintains evolving user semantics and local knowledge, yet may lack reusable preference patterns when its prior experience offers little analogous signal for a candidate item.In that case, the agent must look beyond its own memory.Source: §1 · pp. 1–2
- The proposed User–Agent Web–Platform pathway lets user agents interpret preferences from local semantic memory and integrate task-relevant evidence acquired across the Agent Web.The paper frames this as a shift from the conventional User–Platform relation.Source: §1 · p. 2, Figure 1 · p. 2
- Relevant evidence is distributed across opaque agents, only a small fraction may fit the current decision, and responses can differ in semantic perspective.The paper identifies bounded access, selective relevance, and heterogeneous responses as challenges.Source: §1 · p. 2
- AgentWebRec treats recommendation as task-time information acquisition and fusion, using compact responses rather than transferring complete memories or constructing a global user profile.The target agent organizes acquired information around the current user-item context.Source: §1 · p. 2
- AgentWebRec achieves consistently strong performance across four InstructRec datasets, while target-agent information and external-agent experience provide complementary benefits.Task-oriented acquisition also avoids indiscriminately introducing all available agent information.Source: §1 · p. 2
2 Problem Definition
The problem is to predict a user's preference for a candidate item at a particular decision time using distributed, isolated agent memories. The formulation treats recommendation as acquiring task-relevant information and fusing heterogeneous evidence while excluding distracting content.
2 Problem Definition
- An Agent Web is a set of LLM-based agents connected by task-conditioned interactions, and agents query neighbors whose features are similar.A queried agent processes the query using its memory and returns a response.Source: §2 · pp. 2–3
- Recommendation instantiates the Agent Web with a platform agent and user agents; user histories contain item descriptions, feedback or reviews, and interaction times.The item representation includes a semantic description.Source: §2 · p. 3
- Each recommendation task specifies a target user, candidate item, and decision time; isolated memories and unrelated records can make relevant preference evidence difficult to assemble.Indiscriminately incorporating unrelated records or weakly aligned responses can dilute task-relevant signals.Source: §2 · p. 3
- The formulation selects information to acquire as R_t and uses Fuse(·) to identify task-relevant content and align heterogeneous sources into evidence for predicting the user's preference.The stated objective is for the fused evidence to support an approximation of the true preference.Source: §2 · p. 3
3 Methodology
AgentWebRec progressively acquires decision-relevant evidence through platform, private-memory, and neighbor-agent interactions, then fuses it into a recommendation while keeping agent memories local.
3.1 Overview of AgentWebRec
- AgentWebRec interleaves task-time information acquisition with progressive evidence fusion, using decentralized query-response interactions without requiring a global orchestrator.Agents generate queries and responses from their own internal states, which remain mutually inaccessible.Source: §3.1 · p. 3
3.2 Querying the Platform Agent for Item Semantic Enhancement
- To reduce ambiguity from sparse item profiles, the platform agent retrieves semantically similar items and abstracts them with the candidate into enhanced item semantics.The support set has size K_I; its raw-preserving construction supplements the candidate description with neighbor-derived semantics while limiting drift.Source: §3.2 · p. 3
3.3 Retrieving and Projecting Private Memory
- The target agent uses enhanced item semantics to retrieve bounded private-memory records by semantic relevance and recency, then projects them into a preference state and local decision.The weighting balances candidate-specific relevance against temporal recency; only the top K_M records enter task evidence.Source: §3.3 · pp. 4–5
3.4 Confidence-Gated Collaboration in the Agent Web
- When local confidence falls below a threshold, one-hop neighbors return query-relevant preference patterns, which RefineLLM uses only when they provide complementary support.Collaborator contributions are bounded, and supporting interaction records remain local; irrelevant or conflicting patterns leave the local judgment intact.Source: §3.4 · p. 5
4 Experiments
Across four InstructRec datasets, AgentWebRec leads all reported metrics; ablations and sensitivity analyses show the value of complementary, selectively acquired evidence and task-relevant collaboration.
4.1 Experimental Setup
- Experiments use four InstructRec domains, temporally held-out targets, nine sampled negatives per target, and three-run average Hit Rate and NDCG evaluation.The default setting uses five semantic neighbors, up to five retrieved interactions per agent, at most three collaborators, and a confidence threshold of 0.7.Source: §4.1.1 · p. 5, §4.1.2 · p. 6
4.2 Main Results (RQ1)
- AgentWebRec achieves the best results across all reported metrics on four datasets, with gains extending beyond the highest-ranked prediction.The main comparison is against centralized representation-based recommenders.Source: §4.2 · p. 6, Table 2 · p. 7
4.3 Ablation Studies on Agent Actions (RQ2)
- The full model performs best or most reliably on almost all datasets and metrics, indicating that its three sequential agent actions contribute complementary evidence.Source: §4.3 · p. 6, Table 3 · p. 7
- Removing semantic enhancement or private-memory retrieval degrades performance, while removing collaboration hurts all four datasets; the losses vary by domain and ranking metric.The cited ablations identify the semantic anchor, retrieved personalized evidence, and collaborator responses as distinct contributors.Source: §4.3 · p. 6
4.4 Sensitivity Analysis (RQ3)
- A moderate confidence gate preserves reliable local decisions while admitting useful collaborators, but an expanded collaborator set eventually adds redundant or noisy evidence.Source: §4.4 · p. 7
- Recommendation quality follows an inverted-U pattern as semantic-neighbor support grows: moderate neighborhoods help, while excessive expansion dilutes candidate semantics.Source: §4.4 · p. 7
- A moderate target-user memory budget stabilizes task-specific preferences; too little evidence is insufficient, while excessive retrieval dilutes relevant signals and raises reasoning load.Source: §4.4 · p. 7
4.5 Effect of Agent Web Topology (RQ4)
- Structured Agent Web edges outperform random rewiring, with item co-occurrence strongest overall; persona similarity is competitive but less uniform.The results associate task-relevant user proximity with more relevant collaborative responses.Source: §4.5 · p. 8
5 Related Work
Prior recommendation work centers on platform-collected evidence, while LLM and agent approaches add semantic reasoning and agentic capabilities. AgentWebRec instead frames recommendation as task-oriented reasoning over relevant agent-owned evidence.
5.1 Platform-Centric Recommendation
- Platform-centric recommenders learn user and item representations from centrally collected interactions, keeping evidence processing and preference estimation under platform control [12].Sequential methods also model temporal dynamics and contextual dependencies [16] [32].Source: §5.1 · p. 8
5.2 LLM and Agent-Based Recommendation
- LLM recommendation uses language representations, semantic reasoning, and retrieval or graph context, but retrieval usually remains a platform-side operation [9] [18] [29].Collaborative semantics and sparse interaction graphs are also used [37] [45], alongside retrieval-augmented or graph-grounded reasoning [8] [28] [47].Source: §5.2 · pp. 8–9
- Agent recommendation adds planning, tools, memory, reflection, and coordination, while AgentWebRec grounds decisions in relevant agent-owned evidence across the Agent Web [34] [35].Prior work often wraps platform workflows or summarizes histories for downstream LLMs [4] [42] [43] [6] [27] [46].Source: §5.2 · p. 9
6 Conclusion
The paper frames Agent Web recommendation as progressive task-time evidence acquisition and fusion while keeping agent memories local. AgentWebRec combines platform item semantics, target-agent experience, and conditional evidence from neighboring agents; experiments on four InstructRec datasets show strong, consistent performance against baselines.
6 Conclusion
- AgentWebRec formulates Agent Web recommendation as task-time evidence acquisition and progressive fusion over distributed, opaque agent memories.It constructs compact decision evidence while keeping the underlying memories local.Source: §6 · p. 9
- AgentWebRec grounds decisions in platform item semantics and relevant private experience, then supplements insufficient local evidence with preference patterns from related user agents.The framework keeps the underlying agent memories local.Source: §6 · p. 9
- Experiments on four InstructRec datasets demonstrate strong and consistent performance against baseline recommenders.The paper presents task-oriented evidence acquisition and progressive fusion as a foundation for personalized recommendation in the Agent Web.Source: §6 · p. 9
A.1 LLM Prompt Templates
The prompt templates implement a sequence of item-semantic enhancement, target-user preference projection, and candidate matching, with query-driven collaboration and decision refinement when needed. Each template is instantiated as a single user-role message without a separate system prompt.
A.1 LLM Prompt Templates
- Each template uses one user-role message with the model’s chat template and no separate system prompt; identically named placeholders share runtime variables.The prompt sequence includes item enhancement, preference projection, local matching, collaboration-query generation, pattern abstraction, and decision refinement.Source: §A.1 · p. 11
- AbstractLLM uses a candidate item and related-item descriptions to identify recurring semantic and attribute patterns and enhance the candidate description.The prompt asks it to focus on shared patterns and avoid unrelated details.Source: §A.1 · p. 11
- ProjectLLM summarizes the target user’s persona and retrieved recent interaction memories into a task-specific preference state.The prompt asks for a two-to-three-sentence summary of recent preferences.Source: §A.1 · p. 11
- MatchLLM scores each candidate against the preference summary and returns a score, confidence, and rationale for every authoritative candidate ID.The prompt requires one decision per candidate and a non-empty rationale.Source: §A.1 · p. 11
- After collaboration is triggered, QueryLLM uses uncertainty in current decisions to generate a subquery specifying the complementary neighbor evidence needed.The prompt directs the query toward uncertain item features reflected in scores, confidence, and rationales.Source: §A.1 · p. 11
- PatternLLM uses a neighbor’s persona and matched memories to answer the subquery with concise, evidence-grounded text and a preference pattern.The returned patterns are collected for the subsequent refinement step.Source: §A.1 · p. 12
- RefineLLM revises the original candidate decisions using recurring, mutually consistent neighbor signals while preserving the required item IDs and order.If neighbor evidence is irrelevant, the prompt instructs it to keep the original decision.Source: §A.1 · p. 12