Reward Design for Physical Reasoning in Vision-Language Models

Physical reasoning over visual inputs demands tight integration of visual perception, domain knowledge, and multi-step symbolic inference. Yet even state-of-the-art Vision Language Models (VLMs) fall far short of human performance on physics benchmarks. While post-training algorithms such as Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO) have demonstrated strong reasoning gains in language models, how reward design shapes VLM physical reasoning behavior remains poorly understood. We present a systematic reward ablation study for GRPO-based VLM training on physical reasoning. We compare four reward signals of increasing semantic richness: format compliance, answer accuracy, a composite rubric reward (answer correctness, physics principle identification, and unit consistency), and a novel internal reward derived from model attention weights over input image regions. We evaluate on PhyX, a 3,000-problem benchmark spanning six physics domains and six reasoning types across multiple-choice and open-ended formats, using IBM Granite Vision 3.3 (2B). Across both formats, GRPO with accuracy-based rewards outperforms SFT on most domains, though gains vary substantially by reward type and domain. Reward design does not uniformly improve performance. Instead, it induces domain-specific reasoning behaviors. Accuracy-based rewards provide the strongest overall gains. Rubric rewards improve structured reasoning quality without consistent accuracy improvements. Attention-based rewards enhance spatial reasoning while degrading performance in symbolic domains. Our internal attention-weight reward requires no spatial annotations and improves spatial relation accuracy from 0.27 to 0.50, suggesting that supervising where the model attends during generation is a promising direction for visually grounded physical reasoning.

Back to the Barn with LLAMAs: Evolving Pretrained LLM Backbones in Finetuning Vision Language Models

Vision-Language Models (VLMs) have rapidly advanced by leveraging powerful pre-trained Large Language Models (LLMs) as core reasoning backbones. As new and more capable LLMs emerge with improved reasoning, instruction-following, and generalization, there is a pressing need to efficiently update existing VLMs to incorporate these advancements. However, the integration of new LLMs into VLMs, particularly how the evolving LLMs contribute to multimodal reasoning, alignment, and task-specific performance remains underexplored. Addressing this gap is important for VLM development, given the rapid evolution of pretrained LLM backbones. This study presents a controlled and systematic investigation of how changes in the pretrained LLM backbone affect downstream VLM task performance. By having the vision encoder, training data, and post-training algorithm remain same across LLAMA-1, LLAMA-2, and LLAMA-3 based VLMs, we find that newer LLM backbones do not always lead to better VLMs, but the performance depends on the downstream VLM task. For example, in visual question and answering tasks, newer LLM backbones tend to solve different questions rather than just more questions, and our analysis shows this is driven by differences in how the models process information, including better calibrated confidence and more stable internal representations. We also find that some VLM capabilities appear only in the newest LLM generation, while tasks that depend mainly on visual understanding see little benefit from a newer LLM backbone.

Surprisingly Fragile: Assessing and Addressing Prompt Instability in Multimodal Foundation Models

Multimodal foundation models (MFMs) such as OFASys show the potential to unlock analysis of complex data such as images, videos, and audio data via text prompts alone. However, their performance may suffer in the face of text input that differs even slightly from their training distribution, which is surprising considering the use of modality-specific data to "ground" the text input. This study demonstrates that prompt instability is a major concern for MFMs, leading to a consistent drop in performance across all modalities, but that instability can be mitigated with additional training with augmented data. We evaluate several methods for grounded prompt perturbation, where we generate perturbations and filter based on similarity to text and/or modality data. After re-training the models on the augmented data, we find improved accuracy and more stable performance on the perturbed test data regardless of perturbation condition, suggesting that the data augmentation strategy helps the models handle domain shifts more effectively. In error analysis, we find consistent patterns of performance improvement across domains, suggesting that retraining on prompt perturbations tends to help general reasoning capabilities in MFMs.

PermitQA: A Benchmark for Retrieval Augmented Generation in Wind Siting and Permitting domain

In the rapidly evolving landscape of Natural Language Processing (NLP) and text generation, the emergence of Retrieval Augmented Generation (RAG) presents a promising avenue for improving the quality and reliability of generated text by leveraging information retrieved from user specified database. Benchmarking is essential to evaluate and compare the performance of the different RAG configurations in terms of retriever and generator, providing insights into their effectiveness, scalability, and suitability for the specific domain and applications. In this paper, we present a comprehensive framework to generate a domain relevant RAG benchmark. Our framework is based on automatic question-answer generation with Human (domain experts)-AI Large Language Model (LLM) teaming. As a case study, we demonstrate the framework by introducing PermitQA, a first-of-its-kind benchmark on the wind siting and permitting domain which comprises of multiple scientific documents/reports related to environmental impact of wind energy projects. Our framework systematically evaluates RAG performance using diverse metrics and multiple question types with varying complexity level. We also demonstrate the performance of different models on our benchmark.

RAG vs. Long Context: Examining Frontier Large Language Models for Environmental Review Document Comprehension

Large Language Models (LLMs) have been applied to many research problems across various domains. One of the applications of LLMs is providing question-answering systems that cater to users from different fields. The effectiveness of LLM-based question-answering systems has already been established at an acceptable level for users posing questions in popular and public domains such as trivia and literature. However, it has not often been established in niche domains that traditionally require specialized expertise. To this end, we construct the NEPAQuAD1.0 benchmark to evaluate the performance of three frontier LLMs -- Claude Sonnet, Gemini, and GPT-4 -- when answering questions originating from Environmental Impact Statements prepared by U.S. federal government agencies in accordance with the National Environmental Environmental Act (NEPA). We specifically measure the ability of LLMs to understand the nuances of legal, technical, and compliance-related information present in NEPA documents in different contextual scenarios. For example, we test the LLMs' internal prior NEPA knowledge by providing questions without any context, as well as assess how LLMs synthesize the contextual information present in long NEPA documents to facilitate the question/answering task. We compare the performance of the long context LLMs and RAG powered models in handling different types of questions (e.g., problem-solving, divergent). Our results suggest that RAG powered models significantly outperform the long context models in the answer accuracy regardless of the choice of the frontier LLM. Our further analysis reveals that many models perform better answering closed questions than divergent and problem-solving questions.

Generalist Multimodal AI: A Review of Architectures, Challenges and Opportunities

Multimodal models are expected to be a critical component to future advances in artificial intelligence. This field is starting to grow rapidly with a surge of new design elements motivated by the success of foundation models in natural language processing (NLP) and vision. It is widely hoped that further extending the foundation models to multiple modalities (e.g., text, image, video, sensor, time series, graph, etc.) will ultimately lead to generalist multimodal models, i.e. one model across different data modalities and tasks. However, there is little research that systematically analyzes recent multimodal models (particularly the ones that work beyond text and vision) with respect to the underling architecture proposed. Therefore, this work provides a fresh perspective on generalist multimodal models (GMMs) via a novel architecture and training configuration specific taxonomy. This includes factors such as Unifiability, Modularity, and Adaptability that are pertinent and essential to the wide adoption and application of GMMs. The review further highlights key challenges and prospects for the field and guide the researchers into the new advancements.

ATLANTIC: Structure-Aware Retrieval-Augmented Language Model for Interdisciplinary Science

Large language models record impressive performance on many natural language processing tasks. However, their knowledge capacity is limited to the pretraining corpus. Retrieval augmentation offers an effective solution by retrieving context from external knowledge sources to complement the language model. However, existing retrieval augmentation techniques ignore the structural relationships between these documents. Furthermore, retrieval models are not explored much in scientific tasks, especially in regard to the faithfulness of retrieved documents. In this paper, we propose a novel structure-aware retrieval augmented language model that accommodates document structure during retrieval augmentation. We create a heterogeneous document graph capturing multiple types of relationships (e.g., citation, co-authorship, etc.) that connect documents from more than 15 scientific disciplines (e.g., Physics, Medicine, Chemistry, etc.). We train a graph neural network on the curated document graph to act as a structural encoder for the corresponding passages retrieved during the model pretraining. Particularly, along with text embeddings of the retrieved passages, we obtain structural embeddings of the documents (passages) and fuse them together before feeding them to the language model. We evaluate our model extensively on various scientific benchmarks that include science question-answering and scientific document classification tasks. Experimental results demonstrate that structure-aware retrieval improves retrieving more coherent, faithful and contextually relevant passages, while showing a comparable performance in the overall accuracy.

Empirical evaluation of Uncertainty Quantification in Retrieval-Augmented Language Models for Science

Large language models (LLMs) have shown remarkable achievements in natural language processing tasks, producing high-quality outputs. However, LLMs still exhibit limitations, including the generation of factually incorrect information. In safety-critical applications, it is important to assess the confidence of LLM-generated content to make informed decisions. Retrieval Augmented Language Models (RALMs) is relatively a new area of research in NLP. RALMs offer potential benefits for scientific NLP tasks, as retrieved documents, can serve as evidence to support model-generated content. This inclusion of evidence enhances trustworthiness, as users can verify and explore the retrieved documents to validate model outputs. Quantifying uncertainty in RALM generations further improves trustworthiness, with retrieved text and confidence scores contributing to a comprehensive and reliable model for scientific applications. However, there is limited to no research on UQ for RALMs, particularly in scientific contexts. This study aims to address this gap by conducting a comprehensive evaluation of UQ in RALMs, focusing on scientific tasks. This research investigates how uncertainty scores vary when scientific knowledge is incorporated as pretraining and retrieval data and explores the relationship between uncertainty scores and the accuracy of model-generated outputs. We observe that an existing RALM finetuned with scientific knowledge as the retrieval data tends to be more confident in generating predictions compared to the model pretrained only with scientific knowledge. We also found that RALMs are overconfident in their predictions, making inaccurate predictions more confidently than accurate ones. Scientific knowledge provided either as pretraining or retrieval corpus does not help alleviate this issue. We released our code, data and dashboards at https://github.com/pnnl/EXPERT2.

Evaluating the Effectiveness of Retrieval-Augmented Large Language Models in Scientific Document Reasoning

Despite the dramatic progress in Large Language Model (LLM) development, LLMs often provide seemingly plausible but not factual information, often referred to as hallucinations. Retrieval-augmented LLMs provide a non-parametric approach to solve these issues by retrieving relevant information from external data sources and augment the training process. These models help to trace evidence from an externally provided knowledge base allowing the model predictions to be better interpreted and verified. In this work, we critically evaluate these models in their ability to perform in scientific document reasoning tasks. To this end, we tuned multiple such model variants with science-focused instructions and evaluated them on a scientific document reasoning benchmark for the usefulness of the retrieved document passages. Our findings suggest that models justify predictions in science tasks with fabricated evidence and leveraging scientific corpus as pretraining data does not alleviate the risk of evidence fabrication.

NuclearQA: A Human-Made Benchmark for Language Models for the Nuclear Domain

As LLMs have become increasingly popular, they have been used in almost every field. But as the application for LLMs expands from generic fields to narrow, focused science domains, there exists an ever-increasing gap in ways to evaluate their efficacy in those fields. For the benchmarks that do exist, a lot of them focus on questions that don't require proper understanding of the subject in question. In this paper, we present NuclearQA, a human-made benchmark of 100 questions to evaluate language models in the nuclear domain, consisting of a varying collection of questions that have been specifically designed by experts to test the abilities of language models. We detail our approach and show how the mix of several types of questions makes our benchmark uniquely capable of evaluating models in the nuclear domain. We also present our own evaluation metric for assessing LLM's performances due to the limitations of existing ones. Our experiments on state-of-the-art models suggest that even the best LLMs perform less than satisfactorily on our benchmark, demonstrating the scientific knowledge gap of existing LLMs.