AI evidence extraction

Reasoner-Executor-Synthesizer: Scalable Agentic Architecture with Static O(1) Context Window

Ivan Dobrovolskyi · arXiv (Cornell University) · 2026

AI-generated evidence extraction, verified across multiple analytical personas. Not a substitute for the peer-reviewed original.

9/10

Relevance

1/4

Quality (LMQS)

Evidence

Citations

0.00

FWCI

Methodology & findings

Study design

Formal architectural design with theoretical complexity proof and empirical validation via benchmark testing.

Primary method

Design science methodology with formal architectural specification and empirical validation

Main result

The study demonstrates that the RES architecture achieves O(1) token complexity with respect to dataset size. Specifically, "RES achieves a mean token cost of 1,574 tokens with a standard deviation of only 259 across all 95 successful runs" while "the token cost remains stable whether querying 42,453 articles (cybersecurity) or 16,273,710 articles (medical research), empirically confirming the O(1) behavior established in Theorem 1." At maximum scale, "the naive approach would require approximately 1.9 billion tokens compared to RES's 1,574—a factor of 1.2 million."

Research paradigm

Positivist/engineering - formal theorem proving combined with empirical validation

Author conclusions

The authors conclude: "We presented the Reasoner-Executor-Synthesizer (RES) architecture for building AI agents that are hallucination-free by construction and operate with O(1) LLM token complexity." They further state that "using ScholarSearch, a scholarly research assistant indexing over 130 million articles, we empirically verified Theorem 1 by demonstrating that RES maintains a constant token cost of 1,574±259 tokens across datasets ranging from 42,000 to 16.3 million articles." The authors believe "RES provides a general-purpose pattern for building trustworthy, cost-efficient AI agents across any domain with structured or semi-structured data."

Risk of bias

Limited baseline comparison - only compared against a naive single-prompt baseline with 50 articles, not against state-of-the-art RAG systems; Query selection bias - all 20 queries designed to work with aggregation-based analysis; no evaluation on queries requiring document-level reasoning; Single API source - evaluation limited to Crossref; generalizability to other data sources not empirically tested; Small failure rate - 5 runs failed due to API timeouts (95 of 100 runs successful), but impact not fully analyzed; Single LLM model - only Claude Sonnet tested; no evaluation with other LLM implementations; Selection of query types - all queries amenable to statistical aggregation by design; The evaluation uses a single LLM (Claude Sonnet) for both RES and baseline approaches, which may not generalize to other models. The naive baseline uses only 50 articles, which does not fully explore the scaling characteristics at intermediate dataset sizes. All experiments conducted on a single date (March 14, 2026) without longer-term consistency validation.; Single LLM model tested (Claude Sonnet); Limited baseline comparison (only naive RAG with 50 articles); Single data source (Crossref scholarly articles); Limited query diversity (20 queries across 4 categories); 5 runs per query may be insufficient for statistical robustness; API timeout failures (5 of 100 runs) not fully analyzed

Limitations

"RES is best suited for analytical queries (counts, trends, comparisons, rankings) whose answers can be derived from aggregated statistics
Queries that require deep reading of individual documents (e.g., 'summarize the methodology of paper X') call for a different architecture, such as focused RAG or long-context models." Additionally, "The Reasoner's ability to correctly parse complex queries is bounded by the LLM's instruction-following capability" and "the current implementation relies on Crossref's faceted search capabilities
data sources without built-in aggregation would require the Executor to process records programmatically, though the LLM token cost would remain O(1)."

Open questions raised

Future work includes: (1) extending the Executor layer to support multi-source federation, (2) adding a caching layer for repeated statistical queries, and (3) formally verifying the hallucination-free property via automated fact-checking of the Executor's output.
Multi-source federation in the Executor layer
Caching layer for repeated statistical queries
Formal verification of the hallucination-free property via automated fact-checking
Generalization beyond analytical queries to document summarization and deep reading tasks
Extension to data sources without built-in aggregation capabilities

Data: Crossref REST API (130M+ scholarly articles) - https://api.crossref.org/ (free, no authentication required); Crossref REST API (130M+ scholarly articles) - https://api.crossref.org; Crossref REST APICode: ScholarSearch source code available on GitHub under MIT license (specific URL not provided in paper); ScholarSearch source code available under MIT license on GitHub (URL not explicitly provided in text, but mentioned as open-source); ScholarSearchExtracted from: pdfAgreement 67%

Explore related topics

RAG for Research Literature Discovery Design Science