HomepageDistyl Research: Systems to Build Systems
Distyl’s research arm is a systems shop built on the old intuition that intelligence emerges not only from the cognitive unit but also from the surrounding structure. Our work spans topics like system self-improvement, structured world-models, system self-construction, use case discovery, boundary-case synthetic data generation, and plenty in between.
Our Work
A Systems View of the Space
A brief retrospective on the emergence of LLM systems, and a few of the frontier problems Distyl finds most compelling.
Lattice: Generative Guardrails for Conversational Agents
Lattice is a self-constructing, continuously improving guardrail framework that builds guardrails through iterative simulation and then autonomously adapts them via risk assessment and adversarial testing, achieving strong gains over existing methods and further improving through closed-loop optimization.
How Many Instructions Can LLMs Follow at Once?
IFScale shows that as instruction density climbs to 500 simultaneous directives, even top frontier models fall to about 68% accuracy, revealing distinct degradation patterns, positional biases, and error types—insights that clarify performance–latency tradeoffs and guide the design of instruction-dense prompts.
GenEdit: Compounding Operators and Continuous Improvement to Tackle Text-to-SQL in the Enterprise
A feedback-driven Text-to-SQL system that builds a company-specific knowledge base, decomposes SQL generation into structured operators guided by retrieved context and step-by-step plans, and updates its knowledge set through an interactive copilot to steadily improve complex SQL generation in enterprise settings.
Distyl ranks 1st on BIRD-SQL benchmark
Distyl’s fine-tuned GPT-4o ranked first on the BIRD-SQL benchmark with 71.83% execution accuracy, showcasing strong performance in SQL generation and related tasks like reformulation, intent classification, chain-of-thought, and self-correction.
The Death of Schema Linking? Text-to-SQL in the Age of Well-Reasoned Language Models
Given the recall fragility of schema linking and the contextual robustness of frontier models, forgoing filtering in favor augmentation, selection, and correction becomes both safer and more effective.
End-to-end Text-to-SQL Generation within an Analytics Insight Engine
A proposed architecture for enterprise Text-to-SQL engines, capable of high-complexity, low-latency, domain-aware SQL generation by extracting and retrieving external knowledge, decomposing queries via hierarchical CTEs, and continually adapting through feedback.