DARPA's Defense Sciences Office Just Dropped Two AI SBIR Topics — FALCON and Art of Novel Signals — Closing August 19

DARPA does not fund incremental work, and it does not leave its topics open for long. On July 1, 2026, the agency's Defense Sciences Office (DSO) pre-released two FY26 SBIR topics that sit at the center of the current AI moment — one on fusing efficient machine learning with large language models, one on high-confidence forecasting from messy real-world signals. Both opened July 22, 2026 and close August 19, 2026 at 12:00 PM ET. That is a roughly four-week live window, which is exactly the kind of timeline that catches unprepared teams flat-footed.

If your company works in applied AI/ML — particularly at the intersection of structured-data modeling and large language models — these two topics are worth a hard look. Here is what each one actually wants, how DARPA's SBIR economics differ from the civilian agencies', and how to compete when the clock is this short.

FALCON: fusing efficient ML with LLMs (DPA26BZ04-DV016)

FALCON — Fusion of Abstract Learning and Context-Optimized Neural-methods — targets a problem every serious data organization now feels. Large language models are extraordinary at reasoning over unstructured text and at flexible, conversational interfaces, but they are computationally expensive and clumsy with the structured, tabular data that runs enterprises and battlefields. Classical machine learning is the opposite: fast and precise on structured data, but rigid and non-conversational.

FALCON asks small businesses to combine the two: to "combine advanced machine learning (ML) methods that can be computationally efficient in structured data with large language models (LLM)." The goal is "powerful and efficient technology for interactive statistical analysis of large-scale data" that works in both enterprise and battlefield environments.

Read that carefully, because it constrains the solution space in useful ways. DARPA is not asking for a bigger model. It is asking for efficiency — a system that gets the interactive, natural-language flexibility of an LLM without paying full LLM inference costs on every structured-data query. The winning concepts will likely route work intelligently: cheap, exact ML for the structured heavy lifting, and the LLM reserved for reasoning, interpretation, and the human interface. If your team has a genuine architectural insight about that division of labor, FALCON is built for you.

Art of Novel Signals: high-confidence forecasting (DPA26BZ04-DV015)

The second topic, Art of Novel Signals: Predicting and Forecasting with High Confidence, tackles prediction from the kind of data the real world actually produces — fragmentary, multilingual, multimodal, and constantly shifting. The stated aim is to "develop and demonstrate a predictive/forecasting model that leverages Temporal Knowledge Graph Forecasting using In-Context Learning from novel, multilingual, and multimodal data."

Three technical commitments are embedded in that sentence, and each is a filter:

• Temporal Knowledge Graph Forecasting — the system must model how relationships between entities evolve over time, not just correlate features. This is a structured, graph-based approach to prediction, not a black-box time series.
• In-Context Learning — the model must adapt to new situations from examples provided at inference time, rather than requiring retraining. That is a bet on the flexibility of modern foundation models.
• Multilingual and multimodal data — the inputs are not clean English text. They are the messy, cross-language, cross-format signals of real intelligence and operational environments.

The "high confidence" in the title is doing real work. DARPA is not just asking for predictions; it is asking for predictions the system can stand behind — calibrated, defensible forecasts that an operator could actually act on. Teams that can demonstrate not just accuracy but calibrated confidence will stand out.

How DARPA SBIR economics work — and why they're different

If your reference point is a civilian SBIR (NSF, DOE, DOT), recalibrate. DARPA's structure and expectations differ in ways that matter to how you scope and budget.

Phase I awards at DARPA typically run $150,000 to $300,000 for a feasibility study — somewhat richer than the ~$200K civilian norm, reflecting the harder technical bar.

Direct-to-Phase-II (DP2) is the pathway to watch, and DARPA uses it aggressively. DP2 awards run up to roughly $1.8 million (recent DSO and BTO topics have offered $1.5M DP2 slots), and they exist for companies that have already completed the equivalent of Phase I feasibility using non-SBIR funds. If your team has already built and demonstrated the core of a FALCON- or Art-of-Novel-Signals-style system on internal or other-agency money, DP2 lets you skip the $300K feasibility gate that otherwise consumes the first 6–12 months of a program and jump straight to the money that builds a product. The published topic documents will specify whether each topic offers a Phase I, a DP2, or both — and reading that detail is the first thing you should do.

The broader point: DARPA is buying capability for defense, and it expects a clear line from your research to a fielded system. The evaluation weighs technical audacity more heavily than a civilian agency would, but it also weighs your ability to actually deliver. A brilliant idea with no credible execution path loses at DARPA faster than anywhere else.

Competing in a four-week window

Both topics close August 19, 2026 at 12:00 PM ET. That compressed timeline dictates strategy.

1. Decide DP2 eligibility first. Before writing anything, determine whether you qualify for Direct-to-Phase-II — that is, whether you can document that you have already achieved Phase I-equivalent feasibility with non-SBIR funding. If you can, and the topic offers a DP2 slot, that is a dramatically better use of the next four weeks than a Phase I proposal. The prize is roughly $1.5M-$1.8M instead of $300K, and the feasibility evidence you need already exists.

2. Pick one topic and commit. FALCON and Art of Novel Signals are both AI/ML topics, and there is a temptation for a capable team to chase both. In a four-week window, that is a mistake. Each requires a distinct technical narrative — FALCON is about the ML/LLM efficiency fusion; Art of Novel Signals is about temporal knowledge graphs and calibrated forecasting. Concentrate your firepower on the one where your existing work gives you an unfair advantage.

3. Lead with the specific mechanism, not the outcome. DARPA reviewers have seen every "we use AI to solve X" pitch. What separates a fundable DSO proposal is a concrete, novel technical mechanism: how you fuse ML and LLM efficiently, how your temporal knowledge graph handles multilingual input, why your confidence estimates are trustworthy. The abstract should name the mechanism in the first three sentences.

4. Address the dual environment explicitly. FALCON names "enterprise and battlefield environments"; Art of Novel Signals lives in the intelligence/operational world. Show you understand the constraints of the operational setting — bandwidth, latency, adversarial and degraded inputs, the need for on-the-edge inference. A system that only works in a clean data-center demo is not what DSO is buying.

5. Get registered now, not on August 18. DARPA SBIR submissions run through the DoD SBIR/STTR systems, which require current SAM.gov registration, an active SBIR company registration, and topic-specific portal setup. These take days, and the deadline is a hard 12:00 PM ET cutoff. Every cycle, otherwise-competitive companies miss it on registration mechanics. Do the plumbing in week one.

The read

These two topics are a clean snapshot of where DARPA thinks applied AI needs to go: away from ever-larger monolithic models and toward efficient, hybrid systems that fuse the precision of structured ML with the flexibility of foundation models — and toward forecasting that is not just accurate but trustworthy under real-world, multilingual, multimodal conditions.

For a small AI/ML company with genuine technical depth, the August 19 window is short but the prize is real — especially via Direct-to-Phase-II for teams that have already done the hard early work. The companies that win will be the ones that already have a mechanism in hand and spend the next four weeks sharpening the pitch, not inventing the idea.

Working in applied AI/ML and want to know which DARPA, DOE, and NSF topics fit your technology before their deadlines pass? Granted tracks live SBIR/STTR topics across every agency and matches them to your company's capabilities — so a four-week window never closes before you see it.