Beyond Leaderboards: Tokenomics of Agentic Small Language Model Ensembles
Explore the source record for details and available documents.
97 MATHEMATICS AND COMPUTING↗
SEARCH · Search NASA
Results for “Small Language Model”
Search indexed NASA NTRS and DOE OSTI research on propulsion, heat transfer, battery materials and energy systems. Follow report and document links to the original sources.
Quote a phrase for an exact phrase match. Source license links do not imply unrestricted reuse.
Beyond Leaderboards: Tokenomics of Agentic Small Language Model Ensembles
Explore the source record for details and available documents.
97 MATHEMATICS AND COMPUTING↗
Beyond Leaderboards: Tokenomics of Agentic Small Language Model Ensembles
Explore the source record for details and available documents.
97 MATHEMATICS AND COMPUTING↗
Dispersion Loss Counteracts Embedding Condensation and Improves Generalization in Small Language Models
Large language models (LLMs) achieve remarkable performance through ever-increasing parameter counts, but scaling incurs steep computational costs. To better understand LLM scaling, we study representational differences between LLMs and their smaller counterparts, with the goal of replicating the representational qualities of larger models in smaller models. We observe a geometric phenomenon which we term embedding condensation, where token embeddings collapse into a narrow cone-like subspace in some language models. Through systematic analyses across multiple Transformer families, we show that small models such as GPT2 and Qwen3-0.6B exhibit severe condensation, whereas larger models such as GPT2-x1 and Qwen3-32B are more resistant to this phenomenon. Additional observations show that embedding condensation is not reliably mitigated by knowledge distillation from larger models. To fight against it, we formulate a dispersion loss that explicitly encourages embedding dispersion during training. Experiments demonstrate that it mitigates condensation, recovers dispersion patterns seen in larger models, and yields performance gains across 10 benchmarks. We believe this work offers a principled path toward improving smaller Transformers without additional parameters.
Xiao, Xi [ORNL] (ORCID:0009000009316982)↗
MechBERT: Language Models for Extracting Chemical and Property Relationships about Mechanical Stress and Strain
Language models are transforming materials-aware naturallanguage processing by enabling the extraction of dynamic, context-rich information from unstructured text, thus, moving beyond the limitations of traditional information-extraction methods. Moreover, small language models are on the rise because some of them can perform better than large language models (LLMs) when given domain-specific questionanswer tasks, especially about an application area that relies on a highly specialized vernacular, such as materials science. We therefore present a new class of MechBERT language models for understanding mechanical stress and strain in materials. These employ Bidirectional Encoder Representations for transformer (BERT) architectures. We showcase four MechBERT models, all of which were pretrained on a corpus of documents that are textually rich in chemicals and their stress–strain properties and were fine-tuned on question-answering tasks. We evaluated the level of performance of our models on domain-specific as well as general English-language question-answer tasks and also explored the influence of the size and type of BERT architectures on model performance. We find that our MechBERT models outperform BERT-based models of the same size and maintain relevancy better than much larger BERT-based models when tasked with domain-specific question-answering tasks within the stress–strain engineering sector. These small language models also enable much faster processing and require a much smaller fraction of data to pretrain them, affording them greater operational efficiency and energy sustainability than LLMs.
37 INORGANIC, ORGANIC, PHYSICAL, AND ANALYTICAL CH↗
Autogenerating a Domain-Specific Question-Answering Data Set from a Thermoelectric Materials Database to Enable High-Performing BERT Models
We present a method for autogenerating a large domain-specific question-answering (QA) dataset from a thermoelectric materials database. We show that a small language model, BERT, once fine-tuned on this automatically generated dataset of 99,757 QA pairs about thermoelectric materials, affords better performance in the field of thermoelectric materials compared to a BERT model fine-tuned on the generic English-language QA data set, SQuAD-v2. We further show that mixing the two data sets (ours and SQuAD-v2), which have significantly different syntactic and semantic scopes, allows the BERT model to achieve even better performance. The best-performing BERT model fine-tuned on the mixed data set outperforms the models fine-tuned on the other two data sets by scoring an exact match of 67.93% and an F1 score of 72.29% when evaluated on our test data set. This has important implications as it demonstrates the ability to realize high-performing small language models, with modest computational resources, empowered by domain-specific materials data sets which can be generated according to our method.
biological databases↗
Scaling Laws of Graph Neural Networks for Atomistic Materials Modeling
Atomistic materials modeling is a critical task with wide-ranging applications, from drug discovery to materials science, where accurate predictions of the target material property can lead to significant advancements in scientific discovery. Graph Neural Networks (GNNs) represent the state-of-the-art approach for modeling atomistic material data thanks to their capacity to capture complex relational structures. While machine learning performance has historically improved with larger models and datasets, GNNs for atomistic materials modeling remain relatively small compared to large language models (LLMs), which leverage billions of parameters and terabyte-scale datasets to achieve remarkable performance in their respective domains. To address this gap, we explore the scaling limits of GNNs for atomistic materials modeling by developing a foundational model with billions of parameters, trained on extensive datasets in terabytescale. Our approach incorporates techniques from LLM libraries to efficiently manage large-scale data and models, enabling both effective training and deployment of these large-scale GNN models. This work addresses three fundamental questions in scaling GNNs: the potential for scaling GNN model architectures, the effect of dataset size on model accuracy, and the applicability of LLM-inspired techniques to GNN architectures. Specifically, the outcomes of this study include (1) insights into the scaling laws for GNNs, highlighting the relationship between model size, dataset volume, and accuracy, (2) a foundational GNN model optimized for atomistic materials modeling, and (3) a GNN codebase enhanced with advanced LLM-based training techniques. Our findings lay the groundwork for large-scale GNNs with billions of parameters and terabyte-scale datasets, establishing a scalable pathway for future advancements in atomistic materials modeling.
Li, Chaojian [ORNL] (ORCID:0000000340309777)↗
Utilizing Machine Learning to Improve Neutralization Potency of an HIV-1 Antibody Targeting the gp41 N-Heptad Repeat
The N-heptad repeat (NHR) of the HIV-1 gp41 prehairpin intermediate (PHI) is an attractive potential vaccine target with high sequence conservation across diverse strains. However, despite the potency of NHR-targeting peptides and clinical efficacy of the NHR-targeting entry inhibitor enfuvirtide, no potently neutralizing NHR-directed monoclonal antibodies (mAbs) nor antisera have been identified or elicited to date. The lack of potent NHR-binding mAbs both dampens enthusiasm for vaccine development efforts at this target and presents a barrier to performing passive immunization experiments with NHR-targeting antibodies. To address this challenge, we previously developed an improved variant of the NHR-directed mAb D5, called D5_AR, which is capable of neutralizing diverse tier-2 viruses. Building on that work, here we present the 2.7Å-crystal structure of D5_AR bound to NHR mimetic peptide IQN17. We then utilize protein language models and supervised machine learning to generate small (n < 100) libraries of D5_AR variants that are subsequently screened for improved neutralization potency. We identify a variant with 5-fold improved neutralization potency, D5_FI, which is the most potent NHR-directed monoclonal antibody characterized to date and exhibits broad neutralization of tier-2 and −3 pseudoviruses as well as replicating R5 and X4 challenge strains. Additionally, our work highlights the ability of protein language models to efficiently identify improved mAb variants from relatively small libraries.
Biopolymers↗
Leveraging Large Language Models for Understanding Fundamental Principles of Catalysis
Heterogeneous catalysis presents a distinct challenge for artificial intelligence (AI). Data sets are often small and inconsistently reported, catalyst representations are not standardized, and extracting fundamental knowledge requires integrating performance data, spectroscopic characterizations, and mechanistic models across multiple scales. Language offers a unifying representation across these modalities, making catalysis well suited for leveraging large language models (LLMs). By standardizing how catalytic data is represented, LLMs make dispersed experimental results more accessible to downstream statistical modeling. In this perspective, we focus our discussion around three opportunities where LLMs can significantly contribute to catalysis: (1) text to properties; (2) text to structure; and (3) text to mechanistic models. The discussion is followed by a perspective section on LLM-readiness of data, aligning LLM outputs with scientific correctness, and bridging lab-scale discovery to industrial deployment. Across each area, the most productive applications couple dispersed chemical knowledge with physics-grounded validation to produce verifiable hypotheses and actionable representations.
Catalysts↗
Scientific Open-Source Software Is Less Likely to Become Abandoned Than One Might Think! Lessons from Curating a Catalog of Maintained Scientific Software
Scientific software is essential to scientific innovation and in many ways it is distinct from other types of software. Abandoned (or unmaintained), buggy, and hard to use software, a perception often associated with scientific software can hinder scientific progress, yet, in contrast to other types of software, its longevity is poorly understood. Existing data curation efforts are fragmented by science domain and/or are small in scale and lack key attributes. We use large language models to classify public software repositories in World of Code into distinct scientific domains and layers of the software stack, curating a large and diverse collection of over 18,000 scientific software projects. Using this data, we estimate survival models to understand how the domain, infrastructural layer, and other attributes of scientific software affect its longevity. We further obtain a matched sample of non-scientific software repositories and investigate the differences. We find that infrastructural layers, downstream dependencies, mentions of publications, and participants from government are associated with a longer lifespan, while newer projects with participants from academia had shorter lifespan. Against common expectations, scientific projects have a longer lifetime than matched non-scientific open-source software projects. We expect our curated attribute-rich collection to support future research on scientific software and provide insights that may help extend longevity of both scientific and other projects.
Malviya Thakur, Addi [ORNL] (ORCID:000000022681999↗
Improving the Performance of NEML2 with Modern Graph Compilation Backends
NEML2 vectorizes constitutive-model evaluation for large-scale multiphysics simulation, using PyTorch as its tensor backend so that a batch of material-point updates runs on CPU or GPU through a single implementation. In the two prior reports in this series it was a C++-native library, deployed through TorchScript tracing and just-in-time (JIT) compilation; it has since been rewritten from the ground up into a Python-native library deployed through Ahead-of-Time Inductor (AOTInductor), a modern PyTorch graph-compilation backend. The rewrite is driven by a persistent tension, not a language preference: NEML2 composes constitutive models at runtime from a registry of small, independently-authored pieces, and that flexibility is difficult to reconcile with the compile-time knowledge an efficient GPU kernel needs. This report documents the rewrite and the investment that accompanied it: the AOTInductor export pipeline that turns a Python-authored model into a portable, Python-free compiled artifact loadable from pure C++; the eager and compiled runtimes and the new implicit solver layer built on them; a head-to-head benchmark of legacy JIT against AOTInductor; the physics-model catalog and its worked examples; the developer tooling; and the corresponding overhaul of MOOSE’s NEML2 integration that lets MOOSE consume it. A central objective is to examine whether modern PyTorch graph-compilation backends are effective for MOOSE GPU integration. The benchmark answers directly: AOTInductor outperforms legacy JIT on every GPU scenario measured, by 1.0–4.5×. Modern graph-compilation backends are effective for MOOSE GPU integration, and AOTInductor specifically – not compilation in the abstract – is why.
Hu, Gary (Tianchen) [Argonne National Laboratory (↗
Using Artificial Intelligence to Improve Reliability and Operational Efficiency of Small-Scale Hydroelectric Distributed Generation
Reliability and resilience are critical concerns for distributed generation (DG) at the rural electric level. The integration of renewable energy sources, such as small-scale hydroelectric distributed generators (hydro DGs), introduces operational challenges, particularly regarding aging infrastructure and grid stability. Artificial Intelligence (AI)-driven Machine Learning (ML) models and applications of Large Language Models (LLMs) offer promising solutions for optimizing DG operations and enhancing resilience. This paper explores AI-based models for improving efficiency, fault resolution, and outage mitigation in small-scale hydro DGs. Furthermore, it highlights the development of a centralized, AI-powered information portal for rural electric cooperatives and municipalities. The research evaluates hydro DG plant models and discusses the applicability of AI-powered question-answering tools for real-time operations, focusing on statistical data, load flow, voltage regulation, and generation power. The findings demonstrate AI’s potential to transform DG management to ensure greater stability and resilience in rural electric grids.
Bhattacharyya, Arjun [ORNL] (ORCID:000900060976046↗
3_wise_bears
This is a small set of python scripts and HTML that uses OpenAI API to control large language models (LLMs) working agentically to solve a posed question / problem. There are 3 agents and they work to get it "just right" by taking on various "roles" of friendly and adversarial critics. It repeats a number of times specified by the user, and then writes a report.
DeBardeleben, Nathan Andrew [Los Alamos National L↗
Bridging the Gap Between LLMs and LNS with Dynamic Data Format and Architecture Codesign
Deep neural networks (DNNs) have achieved tremendous success in the past few years. However, their training and inference demand exceptional computational and memory resources. Quantization has been shown as an effective approach to mitigate the cost, with the mainstream data types reduced from FP32 to FP16/BF16 and recently FP8 in the latest NVIDIA H100 GPUs. With increasingly aggressive quantization, however, the conventional floating-point formats suffer from limited precision in representing numbers around zero. Recently, NVIDIA demonstrated the potential of using a Logarithmic Number System (LNS) for the next generation of tensor cores. While LNS mitigates the hurdles in representing small numbers, in this work we observed a mismatch between LNS and the emerging Large Language Models (LLM), where LLM exhibits significant outliers when directly adopting the LNS format. In this paper, we present a data-format/architecture codesign to bright this gap. On the format side, we propose a dynamic LNS format to flexibly represent outliers at a higher precision, by exploiting asymmetry in the LNS representation and identifying outliers through a per-vector basis. On the architecture side, for demonstration, we realize the dynamic LNS format in a systolic array, which can handle the irregularity of the outliers at runtime. We implement our approach on an Alveo U280 FPGA as a prototype. Experimental results show that our design can effectively handle the outliers and resolve the mismatch between LNS and LLM, contributing to an accuracy improvement of 15.4% and 16% over the floating-point and the original LNS baselines, using four state-of-the-art LLM models. Our observation and design lay a solid foundation for the large-scale adoption of the LNS format in the next-generation deep learning hardware.
Haghi, Pouya↗
Data imbalance in drug response prediction: multi-objective optimization approach in deep learning setting
Abstract Drug response prediction (DRP) methods tackle the complex task of associating the effectiveness of small molecules with the specific genetic makeup of the patient. Anti-cancer DRP is a particularly challenging task requiring costly experiments as underlying pathogenic mechanisms are broad and associated with multiple genomic pathways. The scientific community has exerted significant efforts to generate public drug screening datasets, giving a path to various machine learning models that attempt to reason over complex data space of small compounds and biological characteristics of tumors. However, the data depth is still lacking compared to application domains like computer vision or natural language processing domains, limiting current learning capabilities. To combat this issue and improves the generalizability of the DRP models, we are exploring strategies that explicitly address the imbalance in the DRP datasets. We reframe the problem as a multi-objective optimization across multiple drugs to maximize deep learning model performance. We implement this approach by constructing Multi-Objective Optimization Regularized by Loss Entropy loss function and plugging it into a Deep Learning model. We demonstrate the utility of proposed drug discovery methods and make suggestions for further potential application of the work to achieve desirable outcomes in the healthcare field.
Biochemistry & Molecular Biology↗
ESAC (EQ-SANS Assisting Chatbot): Application of large language models and retrieval-augmented generation for enhanced user experience at EQ-SANS
Neutron scattering experiments have played vital roles in exploring materials properties in the past decades. While user interfaces have been improved over time, neutron scattering experiments still require specific knowledge or training by an expert due to the complexity of such advanced instrumentation and the limited number of experiments each person may perform each year. This paper introduces an innovative chatbot application that leverages Large Language Models(LLM) and Retrieval-Augmented Generation (RAG) technologies to significantly enhance the user experience at the EQ-SANS, a small-angle neutron scattering instrument at the Spallation Neutron Source of Oak Ridge National Laboratory. Through a user-centric design approach, the EQ-SANS Assisting Chatbot (ESAC) serves as an interactive reference for users, thereby facilitating the use of the instrument by visiting scientists. By bridging the gap between the users of EQ-SANS and the control systems required to perform their experiments, the ESAC sets a new standard for interactive learning and support for the scientific community using large-scale scientific facilities.
97 MATHEMATICS AND COMPUTING↗
Semantic Stealth: Crafting Covert Adversarial Patches for Sentiment Classifiers Using Large Language Models
Deep learning models have been shown to be vulnerable to adversarial attacks, in which perturbations to their inputs cause the model to produce incorrect predictions. As opposed to adversarial attacks in computer vision, where small changes introduced to pixel values can drastically alter a model's output while remaining imperceptible to humans, text-based attacks are difficult to conceal due to the discrete nature of tokens. Consequently, unconstrained gradient-based attacks often produce adversarial examples that lack semantic meaning, rendering them detectable through visual inspection or perplexity filters. In contrast to methods that rely on gradient-based optimization in the embedding space, we propose an approach that leverages a Large Language Model's ability to generate grammatically correct and semantically meaningful text to craft adversarial patches that seamlessly blend in with the original input text. These patches can be used to alter the behavior of a target model, such as a text classifier. Since our approach does not rely on gradient backpropagation, it only requires access to the target model's confidence scores, making it a grey-box attack. We demonstrate the feasibility of our approach using open-source LLMs, including Intel's Neural Chat, Llama2, and Mistral-Instruct, to generate adversarial patches capable of altering the predictions of a distilBERT model fine-tuned on the IMDB reviews dataset for sentiment classification.
Roa Carvajal, Maria↗
Evaluation of the Self Retrieval Augmented Generation Technique on Common Security Advisory Framework Data
This small experimental report evaluates a variation of Retrieval Augmented Generation (RAG), called Self-RAG. This method uses a generative language model that incorporates retrieved facts into its generation and is explicitly trained to be able to determine whether retrieved information is enough to answer the input query, with a user-defined threshold for confidence. We performed an experiment using data from the publicly available CISA Common Security Advisory Framework (CSAF) repository (https://github.com/cisagov/CSAF) as the database of facts to be used in retrieval. Qualitative results from the experiment demonstrate that the Self-RAG method has some ability to provide reasonable answers to queries that are in the dataset and will often ignore irrelevant information when asked outside of domain questions (e.g., general facts). In settings with deliberately confusing questions (the question is within domain, but asks about a fabricated advisory), it was able to refuse 40% of the time without further adjustments to the original framework. While this performance is not sufficient for current practical use, further improvements to data formatting, disambiguating results, and leveraging threshold values could improve performance significantly. However, evaluating this will require more extensive evaluations on larger datasets and potentially better models.
97 MATHEMATICS AND COMPUTING↗