Research studies
Documented findings from AnovaAI Labs. Each study follows structured methodology with reproducible results — covering model performance, agent behavior, efficiency, and safety.
Task-Specific Distillation: Retaining 87% Accuracy at 40x Lower Cost
We distilled a 70B-parameter teacher model into a 1.3B student focused on binary classification tasks. After targeted fine-tuning on domain data, the student retained 87% accuracy while reducing inference cost by a factor of 40. This study documents the distillation pipeline, dataset preparation, and evaluation methodology.
All studies
Context Position Sensitivity in Long-Window Models
Instructions placed at the beginning and end of 128k-token context windows achieved 23% higher compliance than mid-context placement. We tested across five model families and three task types, finding consistent positional bias that impacts production prompt design.
Self-Correction in Multi-Step Agent Tasks
Adding a structured reflection step after each agent action improved multi-step task completion by 31%. We compared agents with and without self-correction across 500 tasks spanning code generation, data analysis, and research workflows.
Memory Retrieval Latency at Scale in Vector Databases
We benchmarked DomeAI's retrieval performance from 10k to 10M stored entries across three vector database backends. Sub-100ms retrieval was achievable up to 5M entries with proper indexing and cache warming strategies.
Prompt Injection Resistance in Production Agent Systems
We tested 12 prompt injection patterns against production-configured agents to measure resistance rates. Agents using our defense pipeline blocked 96% of injection attempts, compared to 71% for unprotected baselines.
Synthetic Data Quality: When Generated Training Samples Help (and Hurt)
Not all synthetic training data improves model performance. We developed a quality scoring rubric that predicts whether a synthetic dataset will improve or degrade fine-tuning outcomes, achieving 89% prediction accuracy across 40 training runs.
Research methodology
Every study we publish follows consistent standards for transparency, reproducibility, and practical applicability.
Clear Hypotheses
Each study starts with testable questions tied to real production challenges, not abstract curiosity.
Documented Datasets
Data sources, preparation steps, and sampling methods are fully documented for every experiment.
Reproducible Methods
We describe our process in enough detail that other teams could replicate our results independently.
Honest Reporting
Negative results, limitations, and failure modes are reported alongside successes. No cherry-picking.
Questions about our research
Interested in our research?
Our published studies are open for discussion. If you'd like to collaborate or have questions about our findings, reach out.