Open-sourcing AstaBrief, the fast report-generation model in Asta

AstaBrief 8B is an open-weights model released to speed up generation of long-form, cited scientific reports within Asta. Built from Qwen3-8B and optimized with supervised fine-tuning (SFT) and direct preference optimization (DPO) rather than reinforcement learning, it generates full reports in one pass from a research question plus retrieved snippets, cutting end-to-end report time from 178.5s to 51.1s on average (about 3.5× faster) while remaining usable for sensitive on-premises workflows. Weights, training data, and an example local PDF-to-report workflow are published so institutions can run and adapt the system themselves.

Training prioritized real researcher behavior and citation grounding: 90K filtered user queries produced 47K SFT examples via the ScholarQA retrieval-and-synthesis pipeline (backed by a mix of proprietary models), and a cleaned DPO set of ~6K pairwise comparisons where GPT-4.1 and DeepSeek-R1 judges agreed. Development focused on relevance, coverage, citation precision/recall, and preserving the scope and strength of source claims, recognizing that correct citations can still enable subtle overgeneralization. Evaluations used SQABench-CS2, DeepScholarBench, LLM-judged comparisons, and a small human study; results show that careful data selection and system design can yield faster, citation-grounded report generation without complex RL training, though comparisons reflect the frontier as of 2025.