Ai2 Open-Sources AstaBrief 8B for Faster Scientific Report Generation
Allen Institute for AI releases AstaBrief 8B, a specialized open-weights model that produces cited scientific reports from queries and literature excerpts. It delivers reports…
- AstaBrief 8B was built from Qwen3-8B using supervised fine-tuning and DPO on 47K real-researcher queries and 6K preference pairs filtered for citation density and relevance.
- The model enables one-pass report generation that skips multi-stage summarization, cutting average time from 178.5 s to 51.1 s per report in Asta’s Fast mode.
- Institutions can now run the model locally on their own infrastructure for sensitive or unpublished work, with an example workflow provided for custom PDF-based reports.
Training Approach and Data Curation
Ai2 started with the Qwen3-8B base and focused primarily on post-training. They collected tens of thousands of genuine research queries submitted to the Asta platform and ScholarQA framework. After filtering for quality, relevance, non-English content, and privacy, the team retained 90K queries.
For supervised fine-tuning they generated target reports using a multi-step retrieval and synthesis pipeline backed by a mixture of frontier models available in 2025, ultimately producing 47K usable examples. Preference data for DPO was created from a held-out query set by generating contrasting reports with different models and having two judge models select winners, keeping only pairs where both judges agreed. This yielded about 6K clean preference pairs.
The strongest performance gains came from a straightforward citation-density filter that removed synthetic reports leaving too many claims uncited. More complex statistic combinations and aggressive filtering produced diminishing returns.
Evaluation and Performance Gains
Development used the SQABench-CS2 benchmark of 200 computer-science questions plus secondary checks on DeepScholarBench. Metrics tracked rubric coverage, answer precision, citation precision, and citation recall. Human and LLM pairwise comparisons showed AstaBrief competitive with the prior Claude-powered pipeline on overall preference and especially strong on citation accuracy.
The redesigned one-pass generation pipeline eliminated separate snippet summarization and section-by-section writing. This produced nearly an order-of-magnitude speed improvement in the full Asta system. Early usage data from 374 users indicates 29.1 percent returned for multiple days and 23 percent stayed exclusively with the Fast mode.
Results reflect the 2025 frontier models used for training and comparison. Ai2 notes the data-construction, filtering, and system-design lessons are expected to generalize even as newer models appear.
Availability and Future Directions
AstaBrief 8B is now the Fast mode inside Asta’s Generate a report feature, running alongside the more compute-intensive Claude-powered Thinking mode. The model weights, training data, and an example workflow for turning personal PDFs into reports are released on Hugging Face, allowing institutions to deploy behind their own firewalls when handling unpublished or sensitive material.
Ai2 plans further work on finer-grained preference learning, RAG-plus-RL methods, multi-turn capabilities, and richer evaluations that test whether models preserve the exact scope and strength of scientific claims rather than broadening them into unsupported generalizations.
Official records readers should consult include the Allen Institute for AI blog post dated October 2 2026, the model card and weights on Hugging Face, and the referenced papers on ScholarQA and DR Tulu for full methodological detail.
