The NLP research community has made publicly available numerous instruments for measuring representational harms caused by large language model (LLM)-based systems. These instruments have taken the form of datasets, metrics, tools, and more. In this paper, we examine the extent to which such instruments meet the needs of practitioners tasked with evaluating LLM-based systems. Via semi-structured interviews with 12 such practitioners, we find that practitioners are often unable to use publicly available instruments for measuring representational harms. We identify two types of challenges. In some cases, instruments are not useful because they do not meaningfully measure what practitioners seek to measure or are otherwise misaligned with practitioner needs. In other cases, instruments - even useful instruments - are not used by practitioners due to practical and institutional barriers impeding their uptake. Drawing on measurement theory and pragmatic measurement, we provide recommendations for addressing these challenges to better meet practitioner needs.
Automated counter-narratives (CN) offer a promising strategy for mitigating online hate speech, yet concerns about their affective tone, accessibility, and ethical risks remain. We propose a framework for evaluating Large Language Model (LLM)-generated CNs across four dimensions: persona framing, verbosity and readability, affective tone, and ethical robustness. Using GPT-4o-Mini, Cohere's CommandR-7B, and Meta's LLaMA 3.1-70B, we assess three prompting strategies on the MT-Conan and HatEval datasets. Our findings reveal that LLM-generated CNs are often verbose and adapted for people with college-level literacy, limiting their accessibility. While emotionally guided prompts yield more empathetic and readable responses, there remain concerns surrounding safety and effectiveness.
The misuse of Large Language Models (LLMs) to infer emotions from text for malicious purposes, known as emotion inference attacks, poses a significant threat to user privacy. In this paper, we investigate the potential of Apple Intelligence's writing tools, integrated across iPhone, iPad, and MacBook, to mitigate these risks through text modifications such as rewriting and tone adjustment. By developing early novel datasets specifically for this purpose, we empirically assess how different text modifications influence LLM-based detection. This capability suggests strong potential for Apple Intelligence's writing tools as privacy-preserving mechanisms. Our findings lay the groundwork for future adaptive rewriting systems capable of dynamically neutralizing sensitive emotional content to enhance user privacy. To the best of our knowledge, this research provides the first empirical analysis of Apple Intelligence's text-modification tools within a privacy-preservation context with the broader goal of developing on-device, user-centric privacy-preserving mechanisms to protect against LLMs-based advanced inference attacks on deployed systems.
Modern transportation systems rely on cyber-physical systems (CPS), where cyber systems interact seamlessly with physical systems like transportation-related sensors and actuators to enhance safety, mobility, and energy efficiency. However, growing automation and connectivity increase exposure to cyber vulnerabilities. Existing threat modeling frameworks for transportation CPS are often limited in scope, resource-intensive, and dependent on significant cybersecurity expertise. To address these gaps, we present TraCR-TMF (Transportation Cybersecurity and Resiliency Threat Modeling Framework), a large language model (LLM)-based framework that minimizes expert intervention. TraCR-TMF identifies threats, potential attack techniques, and corresponding countermeasures by leveraging the MITRE ATT&CK matrix through three LLM-based approaches: (i) a retrieval-augmented generation (RAG) method requiring no expert input, (ii) an in-context learning approach requiring low expert input, and (iii) a supervised fine-tuning method requiring moderate expert input. TraCR-TMF also maps attack paths to critical assets by analyzing vulnerabilities using a customized LLM. The framework was evaluated in two scenarios. First, it identified relevant attack techniques across transportation CPS applications, with 90% precision as validated by experts. Second, using a fine-tuned LLM, it successfully predicted multiple exploitations including lateral movement, data exfiltration, and ransomware-related encryption that occurred during a major real-world cyberattack incident. These results demonstrate TraCR-TMF's effectiveness in CPS threat modeling, its reduced reliance on cybersecurity expertise, and its adaptability across CPS domains.
This study uses Jordanian law as a case study to explore the fine-tuning of the Llama-3.1 large language model for Arabic question-answering. Two versions of the model- Llama-3.1-8B-bnb-4bit and Llama-3.1-8B-Instruct-bnb-4bit-were fine-tuned using parameter-efficient fine-tuning (PEFT) with LoRA adapters and 4-bit quantized models, leveraging the Unsloth framework for accelerated and resource-efficient training. A custom dataset of 6000 legal question-answer pairs was curated from Jordanian laws and formatted into structured prompts. Performance was evaluated using the BLEU and the ROUGE metrics to compare the fine-tuned models to their respective base versions. Results demonstrated improved legal reasoning and accuracy while achieving resource efficiency through quantization and optimized fine-tuning strategies. This work underscores the potential of adapting large language models for Arabic legal domains and highlights effective techniques for fine-tuning domain-specific tasks.
Is there really much more to say about sparse autoencoders (SAEs)? Autoencoders in general, and SAEs in particular, represent deep architectures that are capable of modeling low-dimensional latent structure in data. Such structure could reflect, among other things, correlation patterns in large language model activations, or complex natural image manifolds. And yet despite the wide-ranging applicability, there have been relatively few changes to SAEs beyond the original recipe from decades ago, namely, standard deep encoder/decoder layers trained with a classical/deterministic sparse regularizer applied within the latent space. One possible exception is the variational autoencoder (VAE), which adopts a stochastic encoder module capable of producing sparse representations when applied to manifold data. In this work we formalize underappreciated weaknesses with both canonical SAEs, as well as analogous VAEs applied to similar tasks, and propose a hybrid alternative model that circumvents these prior limitations. In terms of theoretical support, we prove that global minima of our proposed model recover certain forms of structured data spread across a union of manifolds. Meanwhile, empirical evaluations on synthetic and real-world datasets substantiate the efficacy of our approach in accurately estimating underlying manifold dimensions and producing sparser latent representations without compromising reconstruction error. In general, we are able to exceed the performance of equivalent-capacity SAEs and VAEs, as well as recent diffusion models where applicable, within domains such as images and language model activation patterns.
Large Language Models (LLMs) have shown potential in generating hypothetical documents for query expansion, thereby enhancing information retrieval performance. However, the efficacy of this method is highly dependent on the quality of the generated documents, which often requires complex prompt strategies and the integration of advanced dense retrieval techniques. This can be both costly and computationally intensive. To mitigate these limitations, we explore the use of zero-shot LLM-based query expansion to improve sparse retrieval, particularly for learned sparse retrievers. We introduce a novel fusion ranking framework, Exp4Fuse, which enhances the performance of sparse retrievers through an indirect application of zero-shot LLM-based query expansion. Exp4Fuse operates by simultaneously considering two retrieval routes-one based on the original query and the other on the LLM-augmented query. It then generates two ranked lists using a sparse retriever and fuses them using a modified reciprocal rank fusion method. We conduct extensive evaluations of Exp4Fuse against leading LLM-based query expansion methods and advanced retrieval techniques on three MS MARCO-related datasets and seven low-resource datasets. Experimental results reveal that Exp4Fuse not only surpasses existing LLM-based query expansion methods in enhancing sparse retrievers but also, when combined with advanced sparse retrievers, achieves SOTA results on several benchmarks. This highlights the superior performance and effectiveness of Exp4Fuse in improving query expansion for sparse retrieval.
Pretrained VLMs achieve strong performance on downstream tasks when adapted with just a few labeled examples. As the adapted models inevitably encounter out-of-distribution (OOD) test data that deviates from the in-distribution (ID) task-specific training data, enhancing OOD generalization in few-shot adaptation is critically important. We study robust few-shot VLM adaptation, aiming to increase both ID and OOD accuracy. By comparing different adaptation methods (e.g., prompt tuning, linear probing, contrastive finetuning, and full finetuning), we uncover three key findings: (1) finetuning with proper hyperparameters significantly outperforms the popular VLM adaptation methods prompt tuning and linear probing; (2) visual encoder-only finetuning achieves better efficiency and accuracy than contrastively finetuning both visual and textual encoders; (3) finetuning the top layers of the visual encoder provides the best balance between ID and OOD accuracy. Building on these findings, we propose partial finetuning of the visual encoder empowered with two simple augmentation techniques: (1) retrieval augmentation which retrieves task-relevant data from the VLM's pretraining dataset to enhance adaptation, and (2) adversarial perturbation which promotes robustness during finetuning. Results show that the former/latter boosts OOD/ID accuracy while slightly sacrificing the ID/OOD accuracy. Yet, perhaps understandably, naively combining the two does not maintain their best OOD/ID accuracy. We address this dilemma with the developed SRAPF, Stage-wise Retrieval Augmentation-based Adversarial Partial Finetuning. SRAPF consists of two stages: (1) partial finetuning the visual encoder using both ID and retrieved data, and (2) adversarial partial finetuning with few-shot ID data. Extensive experiments demonstrate that SRAPF achieves the state-of-the-art ID and OOD accuracy on the ImageNet OOD benchmarks.
The rapid growth of large-language models (LLMs) is driving a new wave of specialized hardware for inference. This paper presents the first workload-centric, cross-architectural performance study of commercial AI accelerators, spanning GPU-based chips, hybrid packages, and wafer-scale engines. We compare memory hierarchies, compute fabrics, and on-chip interconnects, and observe up to 3.7x performance variation across architectures as batch size and sequence length change. Four scaling techniques for trillion-parameter models are examined; expert parallelism offers an 8.4x parameter-to-compute advantage but incurs 2.1x higher latency variance than tensor parallelism. These findings provide quantitative guidance for matching workloads to accelerators and reveal architectural gaps that next-generation designs must address.
The choice of tokenizer can profoundly impact language model performance, yet accessible and reliable evaluations of tokenizer quality remain an open challenge. Inspired by scaling consistency, we show that smaller models can accurately predict significant differences in tokenizer impact on larger models at a fraction of the compute cost. By systematically evaluating both English-centric and multilingual tokenizers, we find that tokenizer choice has negligible effects on tasks in English but results in consistent performance differences in multilingual settings. We propose new intrinsic tokenizer metrics inspired by Zipf's law that correlate more strongly with downstream performance than text compression when modeling unseen languages. By combining several metrics to capture multiple aspects of tokenizer behavior, we develop a reliable framework for intrinsic tokenizer evaluations. Our work offers a more efficient path to informed tokenizer selection in future language model development.
Foundation models have achieved transformative success across biomedical domains by enabling holistic understanding of multimodal data. However, their application in surgery remains underexplored. Surgical intelligence presents unique challenges - requiring surgical visual perception, temporal analysis, and reasoning. Existing general-purpose vision-language models fail to address these needs due to insufficient domain-specific supervision and the lack of a large-scale high-quality surgical database. To bridge this gap, we propose SurgVLM, one of the first large vision-language foundation models for surgical intelligence, where this single universal model can tackle versatile surgical tasks. To enable this, we construct a large-scale multimodal surgical database, SurgVLM-DB, comprising over 1.81 million frames with 7.79 million conversations, spanning more than 16 surgical types and 18 anatomical structures. We unify and reorganize 23 public datasets across 10 surgical tasks, followed by standardizing labels and doing hierarchical vision-language alignment to facilitate comprehensive coverage of gradually finer-grained surgical tasks, from visual perception, temporal analysis, to high-level reasoning. Building upon this comprehensive dataset, we propose SurgVLM, which is built upon Qwen2.5-VL, and undergoes instruction tuning to 10+ surgical tasks. We further construct a surgical multimodal benchmark, SurgVLM-Bench, for method evaluation. SurgVLM-Bench consists of 6 popular and widely-used datasets in surgical domain, covering several crucial downstream tasks. Based on SurgVLM-Bench, we evaluate the performance of our SurgVLM (3 SurgVLM variants: SurgVLM-7B, SurgVLM-32B, and SurgVLM-72B), and conduct comprehensive comparisons with 14 mainstream commercial VLMs (e.g., GPT-4o, Gemini 2.0 Flash, Qwen2.5-Max).
Computational Fluid Dynamics (CFD) is essential for scientific and engineering advancements but is limited by operational complexity and the need for extensive expertise. This paper presents ChatCFD, a large language model-driven pipeline that automates CFD workflows within the OpenFOAM framework. It enables users to configure and execute complex simulations from natural language prompts or published literature with minimal expertise. The innovation is its structured approach to database construction, configuration validation, and error reflection, integrating CFD and OpenFOAM knowledge with general language models to improve accuracy and adaptability. Validation shows ChatCFD can autonomously reproduce published CFD results, handling complex, unseen configurations beyond basic examples, a task challenging for general language models.
Table reasoning, encompassing tasks such as table question answering, fact verification, and text-to-SQL, requires precise understanding of structured tabular data, coupled with numerical computation and code manipulation for effective inference. Supervised fine-tuning (SFT) approaches have achieved notable success but often struggle with generalization and robustness due to biases inherent in imitative learning. We introduce Reasoning-Table, the first application of reinforcement learning (RL) to table reasoning, achieving state-of-the-art performance. Through rigorous data preprocessing, reward design, and tailored training strategies, our method leverages simple rule-based outcome rewards to outperform SFT across multiple benchmarks. Unified training across diverse tasks enables Reasoning-Table to emerge as a robust table reasoning large language model, surpassing larger proprietary models like Claude-3.7-Sonnet by 4.0% on table reasoning benchmarks. The approach also achieves excellent performance on text-to-SQL tasks, reaching 68.3% performance on the BIRD dev dataset with a 7B model. Further experiments demonstrate that Reasoning-Table enhances the model's generalization capabilities and robustness.
Neural audio codecs, used as speech tokenizers, have demonstrated remarkable potential in the field of speech generation. However, to ensure high-fidelity audio reconstruction, neural audio codecs typically encode audio into long sequences of speech tokens, posing a significant challenge for downstream language models in long-context modeling. We observe that speech token sequences exhibit short-range dependency: due to the monotonic alignment between text and speech in text-to-speech (TTS) tasks, the prediction of the current token primarily relies on its local context, while long-range tokens contribute less to the current token prediction and often contain redundant information. Inspired by this observation, we propose a \textbf{compressed-to-fine language modeling} approach to address the challenge of long sequence speech tokens within neural codec language models: (1) \textbf{Fine-grained Initial and Short-range Information}: Our approach retains the prompt and local tokens during prediction to ensure text alignment and the integrity of paralinguistic information; (2) \textbf{Compressed Long-range Context}: Our approach compresses long-range token spans into compact representations to reduce redundant information while preserving essential semantics. Extensive experiments on various neural audio codecs and downstream language models validate the effectiveness and generalizability of the proposed approach, highlighting the importance of token compression in improving speech generation within neural codec language models. The demo of audio samples will be available at https://anonymous.4open.science/r/SpeechTokenPredictionViaCompressedToFinedLM.
Periodic or quasi-periodic phenomena reveal intrinsic characteristics in various natural processes, such as weather patterns, movement behaviors, traffic flows, and biological signals. Given that these phenomena span multiple modalities, the capabilities of Multimodal Large Language Models (MLLMs) offer promising potential to effectively capture and understand their complex nature. However, current MLLMs struggle with periodic tasks due to limitations in: 1) lack of temporal modelling and 2) conflict between short and long periods. This paper introduces Period-LLM, a multimodal large language model designed to enhance the performance of periodic tasks across various modalities, and constructs a benchmark of various difficulty for evaluating the cross-modal periodic capabilities of large models. Specially, We adopt an "Easy to Hard Generalization" paradigm, starting with relatively simple text-based tasks and progressing to more complex visual and multimodal tasks, ensuring that the model gradually builds robust periodic reasoning capabilities. Additionally, we propose a "Resisting Logical Oblivion" optimization strategy to maintain periodic reasoning abilities during semantic alignment. Extensive experiments demonstrate the superiority of the proposed Period-LLM over existing MLLMs in periodic tasks. The code is available at https://github.com/keke-nice/Period-LLM.
Visual hallucinations in Large Language Models (LLMs), where the model generates responses that are inconsistent with the visual input, pose a significant challenge to their reliability, particularly in contexts where precise and trustworthy outputs are critical. Current research largely emphasizes post-hoc correction or model-specific fine-tuning strategies, with limited exploration of preprocessing techniques to address hallucination issues at the input stage. This study presents a novel ensemble-based preprocessing framework that adaptively selects the most appropriate filtering approach -- noise reduced (NR), edge enhanced (EE), or unaltered input (org) based on the type of question posed, resulting into reduced hallucination without requiring any modifications to the underlying model architecture or training pipeline. Evaluated on the `HaloQuest' dataset -- a benchmark designed to test multimodal reasoning on visually complex inputs, our method achieves a 44.3% reduction in hallucination rates, as measured by Natural Language Inference (NLI) scores using SelfCheckGPT. This demonstrates that intelligent input conditioning alone can significantly enhance factual grounding in LLM responses. The findings highlight the importance of adaptive preprocessing techniques in mitigating hallucinations, paving the way for more reliable multimodal systems capable of addressing real-world challenges.
Although chain-of-thought reasoning and reinforcement learning (RL) have driven breakthroughs in NLP, their integration into generative vision models remains underexplored. We introduce ReasonGen-R1, a two-stage framework that first imbues an autoregressive image generator with explicit text-based "thinking" skills via supervised fine-tuning on a newly generated reasoning dataset of written rationales, and then refines its outputs using Group Relative Policy Optimization. To enable the model to reason through text before generating images, We automatically generate and release a corpus of model crafted rationales paired with visual prompts, enabling controlled planning of object layouts, styles, and scene compositions. Our GRPO algorithm uses reward signals from a pretrained vision language model to assess overall visual quality, optimizing the policy in each update. Evaluations on GenEval, DPG, and the T2I benchmark demonstrate that ReasonGen-R1 consistently outperforms strong baselines and prior state-of-the-art models. More: aka.ms/reasongen.
We propose a new method for estimating how much a model ``knows'' about a datapoint and use it to measure the capacity of modern language models. Prior studies of language model memorization have struggled to disentangle memorization from generalization. We formally separate memorization into two components: \textit{unintended memorization}, the information a model contains about a specific dataset, and \textit{generalization}, the information a model contains about the true data-generation process. When we completely eliminate generalization, we can compute the total memorization, which provides an estimate of model capacity: our measurements estimate that GPT-style models have a capacity of approximately 3.6 bits per parameter. We train language models on datasets of increasing size and observe that models memorize until their capacity fills, at which point ``grokking'' begins, and unintended memorization decreases as models begin to generalize. We train hundreds of transformer language models ranging from $500K$ to $1.5B$ parameters and produce a series of scaling laws relating model capacity and data size to membership inference.
Dataset diversity plays a pivotal role for the successful training of many machine learning models, particularly in the supervised fine-tuning (SFT) stage of large language model (LLM) development. Despite increasing recognition of its importance, systematic analyses of dataset diversity still remain underexplored. To address this gap, this work presents a systematic taxonomy of existing diversity-control strategies, which primarily focus on the instruction component, operating at either macroscopic (entire instruction semantics) or mesoscopic levels (instruction units), and furthermore introduces a novel analysis of microscopic diversity within the response component, specifically analyzing the statistical distribution of tokens in SFT training samples. In the experimental evaluation, we construct fixed-size datasets (e.g., 10,000 samples each) from a corpus of 117,000 open-source SFT samples, incorporating six distinct diversity-control strategies spanning macro-, meso-, and microscopic levels applied to both instructions and responses. We then fine-tune LLMs on these datasets to assess the six diversity-control strategies. Results reveal that while macroscopic and mesoscopic strategies lead to higher performance with increasing diversity, the microscopic strategy in responses exhibits both a stronger correlation between model performance and the degree of diversity and superior performance with maximum diversity across all strategies. These findings offer actionable insights for constructing high-performance SFT datasets.
Extensively evaluating the capabilities of (large) language models is difficult. Rapid development of state-of-the-art models induce benchmark saturation, while creating more challenging datasets is labor-intensive. Inspired by the recent developments in mechanistic interpretability, we introduce circuit stability as a new way to assess model performance. Circuit stability refers to a model's ability to apply a consistent reasoning process-its circuit-across various inputs. We mathematically formalize circuit stability and circuit equivalence. Then, through three case studies, we empirically show that circuit stability and the lack thereof can characterize and predict different aspects of generalization. Our proposed methods offer a step towards rigorously relating the generality of models to their interpretability.