Fusion of Abstract Learning and Context-Optimized Neural-methods (FALCON)

Fusion of Abstract Learning and Context-Optimized Neural-methods (FALCON) is sponsored by Defense Advanced Research Projects Agency (DARPA). This DARPA SBIR topic aims to combine advanced machine learning methods that are computationally efficient in structured data with large language models that are general and can extract context from data.

The effort must survey and research emerging ML methods, develop an architecture to combine select ML methods with LLM models, determine metrics for accuracy, insights, efficiency, and generalization, and evaluate the architecture in datasets from multiple applications. It also requires developing and demonstrating methods to mitigate hallucination and demonstrate verifiable and reproducible analytic traces.

Department of War organization.

Fusion of Abstract Learning and Context-Optimized Neural-methods Partner with Us to Accelerate DARPA Innovation OUSD (R&E) critical technology area(s) : Advanced Computing and Software, Trusted AI and Autonomy Objective: The goal of this effort is to combine advanced machine learning (ML) methods that can be computationally efficient in structured data with large language models (LLM) that are general and can extract context from data.

The aim is to derive powerful and efficient technology for interactive statistical analysis of large-scale data seen such as in enterprise or battlefield. Description: By integrating the contextualization power of LLMs with the statistical power of ML, the program aims to leverage the benefits of both to provide domain-specific statistical contextualization of structured data.

Survey and research the emerging ML methods suitable for large scale data containing both structured and unstructured data. Develop an architecture to combine select ML methods with LLM models. Determine a set of metrics that encompass accuracy, new insights, computational efficiency, and ability to generalize across datasets.

Evaluate the combined architecture and methods in datasets drawn from multiple applications. This may focus on tabular data in enterprise or engineering applications. Develop and demonstrate methods to mitigate possible hallucination in the workflow and demonstrate verifiable and reproducible analytic traces.

Demonstrate interactive analysis by incorporating new insights as they develop during the course of analysis. This topic is soliciting Direct to Phase II (DP2) proposals only. Phase I feasibility should be demonstrated by documenting in the technical proposal the research team’s prior comprehensive research experience on emerging ML methods for structured data.

Proposals must demonstrate the research team understands their functionality, ability to scale, and advantages relative to SOTA ML methods. Prior research should have been conducted in the last three years. Reports which provide data, clearly present the analysis done, and provide evidence of scholarly impact will be strongly preferred.

Identify one or more promising ML methods and develop the software for combining it with one or more available LLMs, preferably open source. The implementation plan should incorporate an initial demonstration of the analysis functionality (in the first six months), and a plan to scale up to enterprise level data and interactive analysis (at the end of the first year).

The final demonstrations are to be done with at least two datasets from different areas. The evaluation of the combined method should demonstrate accuracy relative to ground truth as well document the improvement over SOTA methods (ML only, and LLM only). This Phase should have a parallel effort on the commercialization strategy implementation and a go-to-market plan.

Phase II fixed payable milestones for this program should include: Month 1: Kick off and technical report on the ML methods to be implemented Month 3: Quarterly meeting presentation material, including demonstration of progress to date and plans Month 6: Quarterly meeting presentation material, including demonstration of progress to date, plans overall, and an initial demonstration of the analysis functionality Month 11: Phase deliverable to demonstrate ability to scale up to enterprise level data and interactive analysis with at least two datasets from different areas Month 12: Quarterly/Phase meeting presentation material on progress over base year.

Phase II Final report and software delivery, with suitable documentation. Month 15: Quarterly meeting presentation material, including demonstration of technical and commercialization progress to date and plans Month 18: Quarterly/Phase meeting presentation material on progress over Option. Option Final report and software delivery, with suitable documentation.

Phase III dual use applications The effort should deliver technology that is quantitatively superior to SOTA methods in both civilian and military sectors. Modern commercial and scientific applications from business enterprises to biology labs which have structured data along with unstructured text are potential applications.

Military strategic and tactical applications also must deal with structured tabular data in design and manufacturing as well as in analyzing economic, social, and geographic data. See, e.g. the papers accepted/published to Neural and Information Systems Processing 2024 and 2025: https://neurips.

cc/ Radhakrishan, Beaglehole, Pandit, Belkin, “Mechanism for feature learning in neural networks and backpropogation-free machine learning models,” Science, 2024. DARPA Artificial Intelligence Quantified program: https://www. darpa.

mil/research/programs/aiq-artificial-intelligence-quantified . Neuro-symbolic AI, large scale structured data analysis, decision theory For additional information and to submit your full proposal package, visit the DSIP Portal . Publication: July 1, 2026 Closes: Aug.

19, 2026 12:00 PM ET DoW SBIR 2026 BAA | Release 4