PaperlayerView exact revision on arXiv

Source-linked AI summary

SocialBuddy: Tailoring Search Agent for Social Scenarios

Mingxuan Li, Yirong Mao, FaZhan Zhang, Haibiao Yao, Runze Hu, Wenhui Que

arXiv:2609.01641v2cs.SIcs.LG

TL;DR

Existing agentic search systems are not designed for social feeds with heterogeneous queries and multi-dimensional constraints. SocialBuddy addresses this gap with a social-search framework, SocialEnv, SocialPO, and SocialSearch Benchmark; SocialBuddy-35B surpasses significantly larger frontier LLMs.

  • Problem

    Existing agentic search frameworks overlook social-media search, where users need to retrieve and synthesize friends’ posts under relational, temporal, and other multi-dimensional constraints.

  • Method

    SocialBuddy combines a large-scale simulated SocialEnv, curated query-trajectory data, and SocialPO, which jointly uses outcome rewards and localized token-level supervision.

  • Results

    SocialBuddy-35B surpasses significantly larger frontier LLMs on the paper’s social-search evaluation.

  • Takeaways & Limitations

    The framework extends agentic search toward daily social intelligence by pairing social-adapted tools with dedicated training and evaluation resources.

Abstract

from arXiv · show

In the era of digital social interaction, searching friends' posts from massive social streams has become a fundamental user need. However, while modern agentic search frameworks have achieved remarkable success in conventional retrieval tasks, they break down when confronted with heterogeneous user queries and multi-dimensional social feeds, resulting in severe performance degradation in complex social search. To bridge this gap, we introduce SocialBuddy, the first agentic search framework tailored for social scenarios. Specifically, we construct SocialEnv, the first large-scale simulated environment for social search. Powered by an automated data and trajectory synthesis pipeline, SocialEnv includes 200K user profiles, 10 million social posts, and 50K reasoning trajectories, establishing a solid foundation for the development of social search agents. To tackle the credit assignment dilemma caused by sparse rewards in social search, we design SocialPO, a hybrid-granularity optimization framework. It macroscopically reinforces successful reasoning paths via multi-dimensional rewards, while microscopically rectifying deviated trajectories through fine-grained prefix truncation and token-level supervision. This hybrid-granularity design delivers multi-scale guidance in complex long-sequence scenarios. Finally, we construct SocialSearch Benchmark to provide a quantitative evaluation scheme for assessing the social search capabilities of SocialBuddy. Extensive experiments demonstrate that SocialBuddy-35B surpasses significantly larger frontier LLMs. Code and dataset will be released upon article acceptance.

Introduction

SocialBuddy addresses the mismatch between conventional agentic search and social-media queries that combine relational, temporal, and other dimensions. It introduces a social-specific framework, simulated environment, optimization method, and benchmark, with experiments showing strong performance against larger LLMs.

  • Motivation: Existing agentic search frameworks overlook social media, where users need to search, track, and synthesize friends’ multi-dimensional posts.Social feeds contain heterogeneous posts with relational and temporal constraints that conventional open-web search systems do not target.
  • Motivation: Embedding-based retrieval can miss specific users and time windows in social queries, causing semantic drift and irrelevant context.The paper illustrates this with a query combining person, time, topic, and popularity constraints.
  • SocialBuddy: SocialBuddy couples semantic understanding with multi-dimensional context parsing through tools that search, filter, and aggregate social posts.The framework is tailored specifically for social-media environments.
  • Environment and optimization: SocialEnv models 200K users, 10 million posts, and 50K interaction trajectories for social-agent training and multi-turn reasoning optimization.The environment is designed as a large-scale simulation of social networks.
  • Environment and optimization: SocialPO combines outcome-level rewards with prefix truncation and token-level supervision to reinforce successful paths and correct localized errors.Its hybrid-granularity design targets credit assignment in sparse-feedback, long-horizon search.
  • Evaluation: SocialBuddy-35B consistently outperforms significantly larger top-tier LLM baselines, while SocialSearch Benchmark evaluates social search and summarization capabilities.The benchmark provides a systematic evaluation protocol for the framework.

Related Work

Prior search-agent research focuses primarily on iterative reasoning and retrieval over open-web pages, while social platforms present heterogeneous posts with temporal and relational constraints. SocialBuddy’s framework organizes a dedicated social-search pipeline around an interactive environment, trajectory curation, and SocialPO optimization.

  • Retrieval-augmented generation: Retrieval-augmented generation evolved from static retrieve-then-generate systems toward adaptive retrieval mechanisms.The cited related work includes dense and late-interaction retrievers as well as adaptive retrieval.
  • Search agents: Search agents use iterative reasoning and tool-mediated interaction, with later systems extending retrieval to multi-turn and long-horizon web search.Examples include WebGPT, ReAct, Search-o1, Search-R1, and DeepResearcher.
  • Social search gap: Existing search agents largely ignore social platforms containing billions of users and heterogeneous, multi-attribute posts.These platforms also impose strict temporal and relational constraints.

Method

SocialBuddy models social search as iterative reasoning and tool use over a personalized, time-varying environment. Its method combines privacy-aware SocialEnv construction, structured multi-dimensional retrieval, anchor-guided trajectory synthesis, and SocialPO training for sparse-reward, long-horizon search.

  • System architecture: SocialBuddy interacts with an environment conditioned on the target user and query timestamp, producing observations that vary across users and time.The formulation captures segregated social circles and time-sensitive feeds.
  • System architecture: Five specialized tools sequentially filter and rank an active post pool by author, time, location, content, and presentation requirements.Each filtering step updates the candidate pool using a tool-specific constraint indicator.
  • SocialEnv construction: SocialEnv enforces privacy-aware visibility by restricting each central user to first-degree connections and their posts, with no access to second-degree entities.The environment synthesizes personas, posts, and localized friend circles to support social-agent training.
  • Trajectory synthesis: Anchor-guided inverse trajectory synthesis improves sampling efficiency by deriving query intents and target trajectories from observable posts within historical, ego-centric circles.The strategy uses an intent skeleton over dimensions such as person, time, location, and content, then verifies retrieved targets and response quality.
  • SocialPO optimization: SocialPO addresses sparse rewards and credit assignment by combining trajectory-level rewards with localized token-level supervision on the first erroneous step and its suffix.The framework reinforces efficient macro decisions while correcting micro-level reasoning errors.

Experiment

The experiments evaluate SocialBuddy across benchmark composition, baseline comparisons, case studies, and SocialPO ablations. SocialBuddy-35B leads competing models, while ablations show gains from both macro rewards and targeted error supervision.

  • Implementation: SocialBuddy is trained in 9B and 35B parameter sizes using SFT followed by SocialPO optimization.The models are built on Qwen3.5-9B and Qwen3.5-35B-A3B base models, respectively.
  • Benchmark and Environment: 4,893 validated queries comprise 2,001 easy and 2,892 hard tasks spanning person, time, location, and content dimensions.SocialEnv includes nearly 200,000 personas, 10 million posts, and 1,000 ego-centric circles.
  • Overall Performance: SocialBuddy-35B outperforms open-weights and proprietary baselines across Precision, Recall, and Exact Match on both benchmark subsets.It leads competitors despite using significantly fewer parameters.
  • Overall Performance: 7.8% Precision, 8.2% Recall, and 9.7% Exact Match are SocialBuddy-35B’s average absolute improvements over GLM-5.2.Against GPT-5, the corresponding average improvements are 6.9%, 6.5%, and 8.0%.
  • Case Study: In a five-dimensional query, SocialBuddy-35B retrieves and ranks target posts, whereas Qwen3.5-397B-A17B over-constrains the temporal window and returns no results.The case study compares semantic and engagement constraints.
  • Ablation Study: SocialBuddy-9B reaches 0.935 EM after targeted error supervision, exceeding macro-reward and standard OPD variants in the ablation.Macro composite reward reaches 0.894 EM, while standard OPD and its combination with macro rewards reach 0.909 and 0.921 EM.

Conclusion

SocialBuddy is presented as an agentic search framework for social scenarios, combining a simulated environment, hybrid-granularity optimization, and a dedicated benchmark. Experiments report that SocialBuddy-35B surpasses significantly larger frontier LLMs.

  • Framework: SocialBuddy is introduced as an agentic search framework tailored for social scenarios.The paper positions it as an extension of agentic search to social settings.
  • Framework: SocialEnv provides a large-scale simulated environment, while SocialPO reinforces valid decisions and rectifies localized errors.These components address social-search development and policy optimization.
  • Evaluation: The SocialSearch benchmark covers diverse intent complexity and search trajectories for systematic capability assessment.It is used to evaluate social search capabilities.
  • Results: SocialBuddy-35B surpasses significantly larger frontier LLMs in extensive experiments.The reported result extends conventional agentic search toward daily social intelligence.

Appendix I: User Experience Evaluation

The evaluation tests four models on users’ private social feeds through 1,000 double-blind searches. It measures hallucination and user preference, while addressing the semi-synthetic benchmark’s real-world generalization boundary.

  • Scope: The SocialBenchmark environment remains semi-synthetic because it uses publicly sourced posts, motivating evaluation on genuine private social feeds.
  • Evaluation Setup: 50 volunteers conducted 20 searches each on their personal social histories, producing 1,000 comparative trials.Four target models were evaluated under explicit consent and privacy-preservation protocols.
  • Evaluation Setup: The double-blind protocol anonymized and randomized model responses to reduce evaluation bias.
  • Metrics: The study evaluates both hallucination and Top-k Preference Rate across the real-world trials.Top-k Preference Rate is the proportion of trials in which a response ranked within the top-k models.
  • Metrics: Hallucination Rate counts fabricated posts or descriptions and citations contradicting retrieved content.

Appendix II: Tool Design and Interaction Protocol

This appendix specifies the implementation contract for SocialBuddy’s five-tool action space and describes its structural design, parameters, and execution flow.

  • Implementation Contract: The appendix defines the implementation contract underlying the five-tool action space introduced in Section 3.
  • Evaluation Reference: Table 3 reports user-experience results across 1,000 comparative trials in real-world personal social environments.HR, Top-1, and Top-2 denote Hallucination Rate, Top-1 Preference Rate, and Top-2 Preference Rate.
  • Implementation Contract: It details the interaction protocol’s structural design, parameter definitions, and execution flow.

Design Principles

SocialBuddy decomposes multi-dimensional social queries into selectors over a shared candidate pool, then orders execution to reduce the cost of expensive content relevance assessment.

  • Query Decomposition: Social queries can jointly constrain author, time, location, content, and result order.
  • Query Decomposition: Each retrieval dimension uses a selector to retain posts satisfying its criterion in a shared candidate pool.The satisfaction set contains posts for which the dimension-specific predicate evaluates to 1.
  • Pool Update: Evaluated constraints compose conjunctively, so successful selection contracts the pool monotonically; only query-supported dimensions become predicates.
  • Execution Order: SocialBuddy applies requested structured selectors before content selection because open-vocabulary relevance requires batched model inference whose cost grows with the surviving pool.Within content selection, exact keyword matches precede dense semantic scoring.

Tool Roles and Semantic Boundaries

The tool design assigns distinct retrieval responsibilities while preserving all earlier constraints through updates to one shared candidate pool.

  • Tool Responsibilities: The first three tools operate on structured author or post metadata, while select_by_content handles modality and open-vocabulary relevance.
  • Tool Responsibilities: The finalize tool orders, truncates, and projects surviving evidence without adding another relevance condition.
  • Shared State: Because every selector updates the same pool, constraints applied earlier remain active in later steps.

Selection Grounded in Explicit Attributes

The priority-1 selectors apply explicit person, time, location, and content constraints using structured matching rules. These rules preserve query semantics across demographic, temporal, spatial, modality, and relevance dimensions.

  • Person selection: select_by_person conjunctively matches relationship, closeness, gender, and age-cohort attributes, with normalized substring name matching.Public age cohorts map to finer SocialEnv categories without inferring numeric age.
  • Time selection: select_by_time evaluates inclusive date bounds, weekdays, festival annotations, and fixed half-open time-of-day buckets.Relative temporal expressions are resolved to absolute dates before invocation.
  • Location selection: select_by_location normalizes case and apostrophe variants, removes standard administrative suffixes, and matches poi_city or poi_name by substring.The contract supports locations at multiple administrative granularity levels.
  • Content selection: select_by_content applies exact modality matching before semantic relevance assessment and serializes surviving posts with content, metadata, and engagement fields.Identifiers, author attributes, raw timestamps, weekdays, and generated topic labels are excluded from the serialized summaries.

Hybrid Content Selection

Content selection combines deterministic filtering with batched model-based semantic matching. Structured predicates reduce the candidate pool before the sole model-based stage, lowering inference cost while retaining relevance filtering.

  • Semantic selection: Semantic selection is the sole model-based filtering stage and evaluates candidate posts in batches of up to 100 with Qwen-3.5-397B-A17B.A pool of size m requires ⌈m/100⌉ forward passes.
  • Terminal action: finalize is the terminal action and does not alter relevance membership after content selection.It only orders the surviving records for output.

Terminal Ranking and Public Output

After relevance filtering, the system ranks surviving posts and exposes a bounded public projection. The output order supplies citation indices while limiting records and withholding sensitive author attributes.

  • Ranking: Importance ranking uses candidate-local min–max recency, normalized likes, and normalized comments, with deterministic tie-breaking.Equal primary scores are resolved by ascending publication time and then post identifier.
  • Public output: The output is capped at five records and exposes no agent-controlled count argument.Returned records retain content, publication time, POI, engagement counts, and minimal contact descriptors while excluding author age, gender, residence, and internal annotations.

Interaction, Feedback, and Failure Semantics

SocialBuddy decomposes queries into explicit dimensions, executes corresponding tools with constrained feedback, and synthesizes responses from terminal evidence. SocialEnv supplies privacy-preserving, persona-conditioned histories and multimodal records for this interaction setting.

  • Feedback: Intermediate feedback exposes tool identity, applied conditions, and candidate-count transitions, while only finalize returns ordered public evidence for response composition and citation.Intermediate observations omit record identifiers.
  • Policy interaction: The policy decomposes queries into WHO, WHEN, WHERE, WHAT, and SORT facets before applying one supported retrieval dimension per action.It uses low-cost structured predicates before semantic content matching and ends with a single ranking call.
  • Policy specification: The deployed prompt separates task specification, decision policy, and machine-checkable serialization, with tool schemas factored into dedicated tables.This structure makes tool calls and citations explicitly checkable.
  • Environment construction: SocialEnv generates multimodal records from persona context, timestamps, topics, situational cues, and style controls within ego-centric first-degree circles.Supported modalities include text, image, video, music share, and article share.
  • Persona representation: The persona representation contains 14 top-level blocks and 37 leaf fields spanning demographics, routines, values, interests, expression controls, narrative grounding, and relationships.These dimensions jointly condition associated social histories.
  • Narrative grounding: Compact narrative anchors connect static persona attributes to recurrent behaviors without allowing narrative length to dominate profile comparison.Examples include resolving outages, mediating disputes, and rescheduling delayed shipments.
  • Persona–post consistency: The examples show that demographic tags alone do not specify social behavior, because profiles combine life context, routines, responsibilities, and expression controls.GreenThumbChef, Grandma Li, and Steady Hearth illustrate distinct behaviorally grounded histories.
  • Privacy boundary: SocialEnv partitions persona information into public attributes and private generation context, exposing only authorized attributes and observable records to search.The complete profile conditions post synthesis, while private fields remain unavailable as selector predicates.
Loading 2609.01641v2…