BizSage: A Self-Evolving Multi-Agent Framework for Business Research with Efficient Knowledge Retrieval
Yuhe Wu, Guangyu Wang, Jiaxin Liu, Guang Zhang
cs.CL · cs.MA
Quick verdict
BizSage claims section-level KG retrieval plus self-evolving agents for business research, with zero-hallucinated-citation evidence but LLM-judge-only scoring.
Summary
BizSage is a two-stage multi-agent framework for economics and business research assistance. Stage 1 extracts section-level knowledge graphs from 134,234 papers across 73 journals (economics, finance, operations management, statistics) with GPT-4o (~$7,300), merges them into a 3.8M-node Lateral Knowledge Graph, and retrieves sections via Personalized PageRank. Stage 2 runs seven specialized agents under a Manager whose Meta-Review loop distills recurring evaluation failures into validated prompt strategies across three iterations, each refined on just 10 development queries. On a 300-query benchmark derived from 4,000 held-out papers and scored by a GPT-4o judge (human ICC>0.70, LLM-human rho>0.75 on 100 outputs), BizSage with GPT-5.5 ranks first on 10 of 12 metrics, leads Grounding most strongly (Survey 4.14 vs 3.57), and wins over 60% of pairwise comparisons. A CrossRef audit found zero hallucinated citations among 894 survey entries, versus 5.1% (InternAgent-1.5) and 2.7% (EvoScientist). Ablations tie Grounding to the section-level graph and iterative refinement, and content quality to the Discussant/Reviewer. Main limitations: every headline score comes from an LLM judge whose rubric structurally rewards outputs that cite; no significance tests accompany the sub-0.1-point margins; the self-evolution evidence rests on tiny development sets; and no code, benchmark, or graph release is stated.
Review complete
Assessment at a glance
6
risk flags identified
● high 0
● medium 4
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Unifying section-level knowledge graphs into a corpus-wide Lateral Knowledge Graph and retrieving with Personalized PageRank yields faster and more precise fine-grained evidence than paper-level or API-based search.
- The Meta-Review mechanism distills recurring evaluation failures into validated strategies that mainly repair each backbone's weakest quality dimension, Grounding.
- BizSage ranks first on the majority of metrics (10 of 12 with GPT-5.5) across three backbones, with pairwise win-rates above 60% against six baselines.
- BizSage produces zero hallucinated citations among 894 CrossRef-verified survey entries, and 'domain drift' — real but topically irrelevant citations — is identified as a failure mode standard hallucination checks miss.
Strengths
Verified citation integrity, not merely claimed
All 894 citations in 20 BizSage survey outputs were checked against CrossRef with human review of unmatched entries: 0% hallucinated, versus 5.1% for InternAgent-1.5 (98 citations, five fabricated) and 2.7% for EvoScientist (75 citations, two fabricated). The paper also names 'domain drift' (~40% of InternAgent-1.5's valid citations from unrelated fields) as a failure mode standard hallucination checks miss (Section 6.3, Table 5).
Leakage-controlled benchmark construction
4,000 papers (1,000 per domain, last five years) are held out before any LKG extraction (Appendix C); queries are GPT-4o-generated then expert-curated on four dimensions with Cohen's weighted kappa = 0.82, and the 100-query-per-task evaluation split stays unseen during self-evolution.
Judge validation and unusual cost transparency
Ablations isolate mechanisms, not just components
Fine-grained retrieval is fast, not just accurate
Weaknesses
Every quality number comes from a single LLM judge
The Grounding advantage is partly definitional and internally inconsistent
No statistical inference anywhere
Self-evolution evidence is thin
Reproducibility and reporting-hygiene gaps
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
BizSage (arXiv 2609.22235, HKUST-Guangzhou) targets research assistance in economics and business: literature surveys, idea formulation, and research planning. It makes three claims: (1) corpus-level section-granularity retrieval beats paper-level retrieval; (2) a Meta-Review self-evolution loop converts evaluation traces into reusable strategies that mostly repair the weakest quality dimension; (3) the full system ranks first on most metrics across three backbones and produces zero hallucinated citations. The evidence base: 134,234 papers from 73 journals (economics 56,213; finance 40,466; operations management 8,094; statistics 29,461; 1920s-2025), a 300-query benchmark built from 4,000 papers held out before LKG construction, and six baselines run under GPT-5.5, Gemini-3-Flash, and DeepSeek-V4-Flash.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 6 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is BizSage?
A multi-agent LLM framework that assists economics and business research with three tasks: literature surveys, idea formulation, and research planning. It pre-builds a knowledge graph over 134,234 journal papers, retrieves section-level evidence with Personalized PageRank, and improves its agents by distilling recurring evaluation failures into reusable prompt strategies.
Does BizSage hallucinate references?
In the authors' audit, all 894 citations from 20 BizSage survey outputs were verified against CrossRef with human review and none were fabricated (0%), versus 5.1% for InternAgent-1.5 and 2.7% for EvoScientist. This is a sampled result on one task type, not a guarantee for every output the system produces.
What data does BizSage rely on?
No model training occurs: the system prompts GPT-5.5, Gemini-3-Flash, or DeepSeek-V4-Flash. Its offline knowledge graph covers 73 peer-reviewed journals across economics, finance, operations management, and statistics (1920s-2025, 134,234 papers). No working papers, conference papers, or preprints are included, and the system is English-only.
Is BizSage's code or benchmark publicly available?
The paper text does not state any code, benchmark, or knowledge-graph release; it lists a project website (bizsage.site) whose contents this review cannot confirm. Full replication would also require roughly $7,300 of GPT-4o extraction over the licensed journal corpus plus commercial LLM API budgets.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Figures 1-4 are images; their content (e.g., per-bar win-rates in Figure 3, the framework diagram in Figure 2) was assessed from captions and surrounding text only.
- The Section 5.1 vs Table 1 discrepancy on Idea-Formulation Relevance/Novelty could be a PDF table-linearization artifact rather than an error in the paper itself; the question to the authors stands either way.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.