Cambridge Consultants

Sequential decision making with imperfect knowledge of the environment and adversary lies at the heart of defense and security. As such, contemporary reinforcement learning methods which are well adapted to these sorts of problems have the potential to provide the key tools to support decision makers in a military context. This paper presents a literature review of reinforcement learning for decision support in a defense context with a focus on recent work produced by NATO member states. To aid the reader, we provide the prerequisite background in military decision making; decision support systems in general; reinforcement learning basics; simulation; and explainability. Through a systematic review of the literature and using tools such as UMAP, we construct the research landscape and study the emergent trends and research gaps that are present. This allows us to examine the challenges for exploiting this work in reality and finally, we look to the future and how these challenges can be overcome.

Recent work has demonstrated state-of-the-art results in large language model (LLM) hallucination detection and mitigation through consistency-based approaches which involve aggregating multiple responses sampled from a single LLM for a given prompt. These approaches help offset limitations stemming from the imperfect data on which LLMs are trained, which includes biases and under-representation of information required at deployment time among other limitations which can lead to hallucinations. We show that extending these single-model consistency methods to combine responses from multiple LLMs with different training data, training schemes and model architectures can result in substantial further improvements in hallucination detection and mitigation capabilities beyond their single-model consistency counterparts. We evaluate this \emph{consortium consistency} approach across many model teams from a pool of 15 LLMs and explore under what conditions it is beneficial to team together different LLMs in this manner. Further, we show that these performance improvements often come with reduced inference costs, offsetting a significant drawback with single-model consistency methods.

Enzyme-constrained genome-scale metabolic models (ecGEMs) have improved Flux Balance Analysis (FBA) by incorporating enzyme turnover numbers ( k cat s). Since in-vivo k cat data is costly to obtain and therefore scarce, we present a novel multi-modal transformer-based approach with cross-attention to predict k cat values for Escherichia coli using enzyme amino acid sequences and SMILES annotations of reaction substrates. For heteromeric enzymes, we evaluate multiple subunit k cat aggregation strategies. We benchmark ecGEMs constructed with these strategies against current state-of-the-art models using experimental growth rates, 13C fluxes, and enzyme abundances, and prior to any calibration outperform or match existing methods. We also devise a new calibration method using flux control coefficients (derivatives of log flux with respect to log k cat ), which we show to be identical to enzyme cost at the FBA optimum. Using these coefficients, we identify 8 key k cat values to recalibrate using experimental data, subsequently achieving superior performance to the current state-of-the-art with 81% fewer calibrations. ### Competing Interest Statement The authors have declared no competing interest.

3055 Background: Risk of recurrence for patients (pts) with HR+/HER2- breast cancer persists for decades. Most distant recurrences occur in the ‘late’ adjuvant setting, > 5 years (yrs) from diagnosis. In CHiRP (ASCO 2022), we showed that minimal residual disease (MRD) was detectable in the late adjuvant setting: ctDNA was detected in 8/83 (9.6%) pts in the cohort and 6/8 (75%) pts with positive ctDNA (+ctDNA) had developed distant recurrence when initially reported (median follow-up 2 years from first plasma sample collected on study). Here, we report updated clinical outcomes and investigate the meaning of a ctDNA test result during surveillance with longer follow-up. Methods: In CHiRP, pts with stage II-III HR+/HER2- breast cancer at high risk of recurrence diagnosed > 5 yrs prior with no evidence of recurrence were prospectively identified. All pts provided informed consent for prospective plasma collection every 6-12 months at routine follow-up visits for batched, retrospective ctDNA testing using RaDaR, a tumor-informed whole exome sequencing-based assay. Pts were followed at the discretion of the clinical provider without any routine surveillance imaging, as per guideline-concordant care. See CHiRP ASCO 2022 presentation for additional methods. Results: Of 83 pts in the analytic cohort, 57 (68.7%) pts had stage III disease, and most (n = 75, 90.4%) underwent (neo)adjuvant chemotherapy. All pts received endocrine therapy. In this update, median follow-up from first sample collection was 4.4 yrs (interquartile range 4.0, 4.9). 214 plasma samples were collected prior to any known recurrences and included in this analysis. 8/83 (9.6%) pts had +ctDNA at any timepoint including 4/83 (4.8%) with +ctDNA on first study plasma sample. In pts initially ctDNA-negative (-ctDNA; n = 4), median time from first sample collection to MRD detection was 1.29 yrs (range, 0.72 – 3.05). During follow-up, 8 (9.6%) pts developed distant recurrence and 1 (1.2%) pt had a local recurrence. With additional follow-up included in this update, all 8/8(100%) pts with +ctDNA developed distant recurrence with a median lead time of 1.39 years (range 0.01 – 4.24). Among -ctDNA plasma samples with > 2 yrs of follow-up (n = 185), the negative predictive value (NPV) of a -ctDNA test for lack of clinical recurrence for > 2 yrs post-test was 98.4% (3/185). The NPV indicating freedom from recurrence > 1 and > 3 yrs was 100% (0/196) and 96.6% (5/147), respectively. Conclusions: In pts with high-risk HR+/HER2- breast cancer in the late adjuvant setting, all pts with +ctDNA developed distant metastasis. A -ctDNA test was strongly associated with lack of recurrence over a 3 yr follow-up period. Future studies are needed to determine if ctDNA-guided intervention can impact clinical outcomes for early-stage breast cancer and to determine the optimal role of MRD surveillance during follow-up.