Federated Recommender Agents

• Centralized processing of profiles, interaction histories, and preferences raises privacy concerns, while independent local agents cannot leverage other users’ behavioral patterns and decision experiences.Source: §1 · p. 1
• The framework addresses challenges of accumulating personalized knowledge, representing reusable knowledge without exposing private details, and integrating reasoning when similar contexts lead to different choices.Source: §1 · p. 2

Structured Reasoning Federation

• Local agents maintain semantic memory for user preferences and decision memory for choice tendencies across behavioral contexts, guiding retrieval and selection of shared reasoning.Source: §1 · p. 2

Federated Recommender Agents

• FedRecAgent enables local agents to collaborate by sharing natural-language knowledge rather than model parameters, while keeping private user information local.Both client and server LLM parameters remain fixed.Source: §1 · p. 2

Structured Reasoning Federation

• The framework structures shareable experiences as Condition–Transition–Reasoning units and integrates reasoning hierarchically by condition and transition.The units abstract away specific item identities and user-specific preference details.Source: §1 · p. 2

Experiments

• Experiments compare FedRecAgent with federated and LLM-based recommendation methods across three domains and dense and sparse settings.Ablation studies and additional analyses examine the proposed components.Source: §1 · p. 2

• Federated recommendation has largely relied on learned representations or model updates rather than explicit, reusable reasoning knowledge.Prior approaches include parameter or item-representation sharing, personalized adaptation, aggregation, continual learning, and cold-start methods.Source: §2 · p. 2
• LLM recommender-agent research explores planning, deliberative reasoning, multi-agent collaboration, and persistent or collaborative memory, but does not primarily address sharing reusable reasoning while keeping histories and personal memories local.Source: §2 · p. 3
• Federated LLM-agent work exchanges parameters, policies, or natural-language knowledge, including tool-use knowledge, textual feedback, reasoning insights, and reusable skills.Source: §2 · p. 3
• For recommendation, similar behavioral contexts can lead to different decision transitions, so contextual similarity alone cannot determine which reasoning to integrate.FedRecAgent organizes reasoning into condition–transition branches and uses local private memories for recommendation.Source: §2 · p. 3

• An observed episode pairs a recent interaction history with the observed next item; training episodes slide the history window forward one interaction at a time.Source: §3 · p. 3
• At inference, the agent reranks a candidate set given the current history, aiming to rank the ground-truth next item as highly as possible.Source: §3 · p. 3
• Each client represents one user and keeps raw interaction histories and private knowledge on the client device.Agents use private knowledge together with server-shared knowledge for recommendation.Source: §3 · p. 3
• The server integrates client uploads into reusable shared knowledge and returns it, while client and server LLM parameters remain fixed.Clients update private knowledge chronologically within each round; the server integrates uploads at round end.Source: §3 · p. 3

4.1 Knowledge Representation

• Each shareable C–T–R unit describes a behavioral condition, an observed decision transition, and reasoning for that choice without specific item identities or user preferences.Conditions and transitions are free-form descriptions; the hierarchical bank stores different transition branches and their reasoning under each condition.Source: §4.1.1 · p. 4, §4.1.2 · p. 4

4.2 Client-side Knowledge Update

• Clients derive conditions from preceding history and transitions from observed next items, while keeping episode-specific evidence local and updating separate semantic and decision memories.Semantic memory captures preferences; decision memory captures choice tendencies under different conditions. For a matched branch, the client preserves supported claims and revises unsupported ones before uploading the unit.Source: §4.2.1 · pp. 4–5, §4.2.2 · p. 5

4.3 Server-side Hierarchical Knowledge Integration

• Condition–transition alignment lets the server consolidate reasoning within branches while preserving distinct decision directions under similar conditions.It first aligns conditions and groups contributions by transition; it also compares reasoning across sibling branches to clarify when each explanation applies.Source: §4.3 · p. 5, §4.3.3 · p. 5, §4.3.2 · p. 5

4.4 Recommendation by Local Agents

• At inference, clients retrieve shared reasoning using a condition and memory-based query, select a branch, and rank candidate items using both private memories.The three query variants use semantic memory, decision memory, or both; candidate items are not used during branch retrieval or selection.Source: §4.4 · p. 6, §4.4.1 · p. 6, §4.4.2 · p. 6

5.1 Experimental Setup

• The evaluation uses CDs, Music, and Office in dense and sparse subsets, chronological leave-one-out testing, and 11 federated and adapted baselines.Metrics are NDCG@K for K∈{1, 5, 10} and MRR; results are generally averaged over users and five seeds.Source: §5.1.1 · p. 6, §5.1.2 · p. 6, §5.1.3 · p. 6

5.2 Performance Comparison

• FedRecAgent’s best variants improve N@10 by 1.54%–9.19% and MRR by 3.01%–16.29% over the strongest baselines across six datasets.At least one variant achieves the highest mean across all datasets and metrics; all three outperform the strongest baselines in N@10 and MRR on every dataset.Source: §5.2 · p. 6, Table 1 · p. 7
• Under FedSASRec-based hard negatives, FedRecAgent achieves the highest N@10 and MRR on five of six datasets; FedI2Agent leads on Music-Dense.The authors report that these results support FedRecAgent’s effectiveness under hard-negative sampling.Source: §5.2 · pp. 6–7, Table 2 · p. 7

5.3 Ablation Study

• Removing either private memory reduces performance, removing both causes the largest degradation, and omitting episode evidence extraction also degrades performance.The unified-memory comparison favors separately updating preferences and conditional choice tendencies.Source: §5.3.1 · p. 7, Table 3 · p. 7
• Global-knowledge ablations reduce performance, supporting shared knowledge, cross-round consolidation, transition discrimination, and condition–transition alignment.The tested alternatives include no shared knowledge, no global-guided update, no distillation, no discrimination, no merge, and condition-only grouping.Source: §5.3 · p. 7, §5.3.2 · p. 8

5.4 Study of FedRecAgent

• On CDs-Dense and Music-Dense, higher thresholds create finer condition and transition partitions, and multiple transition branches can coexist under one condition.Most uploads belong to branches containing experiences from multiple users.Source: §5.4.1 · p. 8, Figure 3 · p. 8
• Transition discrimination consistently improves matching accuracy across the evaluated dataset–round pairs.The evaluation compares matching before and after refinement using the same candidates and ordering.Source: §5.4.2 · p. 8, Table 4 · p. 8
• Dual memory has higher preservation rates than unified memory for both preferences and conditional choice tendencies on both datasets.The assessment uses LLM-evaluated claims extracted from training histories and checks whether memories preserve them.Source: §5.4.3 · p. 8, Table 5 · p. 8

• Federated Recommender Agents let personalized LLM agents collaborate through natural-language knowledge while keeping private user information local.Source: §6 · p. 9
• FedRecAgent maintains private memories of preferences and choice tendencies and shares C–T–R knowledge units integrated by condition and transition, without updating LLM parameters.Source: §6 · p. 9
• Experiments show improved recommendation performance over baselines, and further analyses support private memories and structured reasoning integration.Source: §6 · p. 9

• Table 7 reports statistics for the datasets.Source: Table 7 · p. 11
• The table’s paired values correspond to Dense and Sparse subsets.Source: Table 7 · p. 11
• Density is defined as interaction count divided by the number of possible user–item pairs.Source: Table 7 · p. 11

• The data protocol orders interactions chronologically, reserving the last two for validation and testing and using preceding interactions for training episodes.Source: §A.1 · p. 11
• Training episodes are processed chronologically once, with natural-language memories and shared knowledge updated while LLM parameters remain fixed.Clients process timestamp-grouped episodes sequentially using the previous round’s shared knowledge, and the server aggregates contributions at each interval’s end.Source: §A.1 · p. 11
• Baseline configurations preserve the core methods within the client–server setting; FedAgentCF exchanges item memories, while FedI2Agent uses interest memory and a feedback-based profile.Source: §A.1 · p. 11
• Figure 4 presents a hyperparameter study of FedRecAgent.Source: Figure 4 · p. 11
• Table 8 measures input-item reconstruction from client uploads using Episode Recall@10, with 100 users and 10 uploads per user.Higher values indicate greater reconstruction success.Source: Table 8 · p. 11
• Table 9 reports post-filtered detection of explicit item-title disclosure in uploaded C–T–R records.Detected entries are reported as upload counts and percentages of all evaluated uploads.Source: Table 9 · p. 11

A.2.1 Hyperparameter Study

• Increasing the branch budget does not uniformly improve performance, and preferred condition and transition thresholds vary by dataset.The branch budget controls how many retrieved branches reach the selector; the thresholds control knowledge-unit grouping.Source: §A.2.1 · p. 11

A.2.2 Input-Item Reconstruction

• Complete-record uploads yield more Recall@10 than random retrieval, but all field-based attacks remain below popularity ranking.The reconstruction analysis compares catalog retrieval from upload fields with random and popularity controls.Source: §A.2.2 · p. 12

A.2.3 Explicit Item-Title Detection

• No confirmed title disclosures were detected by the 8B judge; the 14B judge found rare cases only on CDs-Dense.Rates use all evaluated uploads as the denominator.Source: §A.2.3 · p. 12

A.2.4 Server LLM Size

• With the client backbone fixed as Qwen3-4B, larger server models generally improve N@10 and MRR, with modest, non-monotonic gains on CDs-Dense.The default server performs close to larger configurations.Source: §A.2.4 · p. 12, Figure 5 · p. 12

A.2.5 Cost Analysis

• FedRecAgent has higher recommendation performance at intermediate inference cost: slower than FedLLMRank and FedAgentCF, but faster than FedAFL and FedI2Agent.Figure 6 compares per-user latency with main-table recommendation performance.Source: §A.2.5 · p. 12, Figure 6 · p. 12
• LLM weights dominate storage across methods; FedRecAgent’s additional storage is primarily its sentence encoder, while memories and shared bank are relatively small.Table 10 compares client-side storage under a common client LLM backbone.Source: §A.2.5 · p. 12, Table 10 · p. 12

A.2.6 Continual Recommendation

• FedRecAgent achieves the highest reported N@10 on both continual-evaluation datasets; removing either personalized memory reduces performance.Private memories persist across evaluation blocks, and the shared bank remains fixed within each block.Source: §A.2.6 · p. 12, Figure 7 · p. 12

A.3.1 Evolution of Personalized Memory

• After two Christmas-themed selections, User 69’s semantic memory emphasizes a reinforced seasonal preference; later observations make it the leading emphasis.Decision memory shifts from maintaining calming soundscapes to balancing familiar holiday selections with newer choices.Source: §A.3.1 · p. 12, Figure 8 · p. 13

A.3.2 Cross-Branch Discrimination

• Across branches sharing a condition, discrimination separates reasoning about an unfamiliar-era and genre shift from reasoning about continuity and artist-specific depth.The refinement clarifies the transition each explanation supports without introducing a new observed transition.Source: §A.3.2 · p. 13, Figure 9 · p. 13, §A.3.1 · p. 13

A.3.3 Reasoning Distillation

• Within one condition–transition branch, distillation consolidates three users’ explanations while retaining shared decision logic and cautious variation.The shared explanation connects long-term tendencies with recent choice signals.Source: §A.3.3 · p. 13, Figure 10 · p. 13