RESEARCH / PEER-REVIEWED FOUNDATIONS
AutoSE is not a demo bolted onto a chatbot. Its pipeline, from architecture decision records to model routing and energy-aware execution, is grounded in peer-reviewed work from the SA4S research group at IIIT Hyderabad. These are some of the studies shaping how we build it.
SELECTED PUBLICATIONS
05 PAPERS · 2024-2026
A study of five context strategies for generating Architecture Decision Records, built on a dataset of sequential ADRs from 750 open-source projects. Context-aware prompting improves fidelity more than raw model scale, with the best results from a narrow window of 3–5 prior records.
Read paper ↗An agentic framework that decomposes architecture recovery and documentation into specialized extraction, retrieval, generation, and validation agents. Instantiated to generate ADRs from code repositories and validated in a user study across 29 repositories.
Read paper ↗A controlled evaluation of four agentic frameworks running small language models on SWE-bench Verified Mini, measuring energy, duration, tokens, and memory. Framework architecture, not the model, is the primary driver of energy use, and much of that energy is spent on unproductive reasoning loops.
Read paper ↗An empirical look at whether small language models can drive reliable function and tool calling, the capability that lets AutoSE route work to smaller local models by task complexity while keeping quality high.
Read paper ↗The foundational study on generating Architecture Decision Records with LLMs given a decision context. GPT-4 shows strong capability, and smaller fine-tuned models such as Flan-T5 reach comparable results, motivating AutoSE's local-first, right-model-per-task approach.
Read paper ↗