Interpreting and controlling the internal mechanisms of large language models (LLMs) is crucial for improving their trustworthiness and utility. Recent efforts have primarily focused on identifying and manipulating neurons by establishing discrete mappings between neurons and semantic concepts. However, such mappings struggle to handle the inherent polysemanticity in LLMs, where individual neurons encode multiple, distinct concepts. This makes precise control challenging and complicates downstream interventions. Through an in-depth analysis of both encoder and decoder-based LLMs across multiple text classification datasets, we uncover that while individual neurons encode multiple concepts, their activation magnitudes vary across concepts in distinct, Gaussian-like patterns. Building on this insight, we introduce NeuronLens, a novel range-based interpretation and manipulation framework that provides a finer view of neuron activation distributions to localize concept attribution within a neuron. Extensive empirical evaluations demonstrate that NeuronLens significantly reduces unintended interference, while maintaining precise control for manipulation of targeted concepts, outperforming existing methods.

The prevalence of drawing causal conclusions from observational studies has raised concerns about potential exaggeration in science communication. While some believe causal language should only apply to randomized controlled trials, others argue that rigorous methods can justify causal claims in observational studies. Ideally, causal language should align with the strength of the evidence. However, through the analysis of over 80,000 observational study abstracts using computational linguistic and regression methods, we found that causal language is more frequently used by less experienced authors, smaller research teams, male last authors, and authors from countries with higher uncertainty avoidance indices. These findings suggest that the use of causal language may be influenced by external factors such as the sociocultural backgrounds of authors and the dynamics of research collaboration. This newly identified link deepens our understanding of how such factors help shape scientific conclusions in causal inference and science communication.

Evaluations of large language model (LLM) risks and capabilities are increasingly being incorporated into AI risk management and governance frameworks. Currently, most risk evaluations are conducted by designing inputs that elicit harmful behaviors from the system. However, a fundamental limitation of this approach is that the harmfulness of the behaviors identified during any particular evaluation can only lower bound the model's worst-possible-case behavior. As a complementary method for eliciting harmful behaviors, we propose evaluating LLMs with model tampering attacks which allow for modifications to latent activations or weights. We pit state-of-the-art techniques for removing harmful LLM capabilities against a suite of 5 input-space and 6 model tampering attacks. In addition to benchmarking these methods against each other, we show that (1) model resilience to capability elicitation attacks lies on a low-dimensional robustness subspace; (2) the attack success rate of model tampering attacks can empirically predict and offer conservative estimates for the success of held-out input-space attacks; and (3) state-of-the-art unlearning methods can easily be undone within 16 steps of fine-tuning. Together these results highlight the difficulty of removing harmful LLM capabilities and show that model tampering attacks enable substantially more rigorous evaluations than input-space attacks alone. We release models at https://huggingface.co/LLM-GAT

Inferring physical laws from data is a central challenge in science and engineering, including but not limited to healthcare, physical sciences, biosciences, social sciences, sustainability, climate, and robotics. Deep networks offer high-accuracy results but lack interpretability, prompting interest in models built from simple components. The Sparse Identification of Nonlinear Dynamics (SINDy) method has become the go-to approach for building such modular and interpretable models. SINDy leverages sparse regression with L1 regularization to identify key terms from a library of candidate functions. However, SINDy's choice of candidate library and optimization method requires significant technical expertise, limiting its widespread applicability. This work introduces Al-Khwarizmi, a novel agentic framework for physical law discovery from data, which integrates foundational models with SINDy. Leveraging LLMs, VLMs, and Retrieval-Augmented Generation (RAG), our approach automates physical law discovery, incorporating prior knowledge and iteratively refining candidate solutions via reflection. Al-Khwarizmi operates in two steps: it summarizes system observations-comprising textual descriptions, raw data, and plots-followed by a secondary step that generates candidate feature libraries and optimizer configurations to identify hidden physics laws correctly. Evaluating our algorithm on over 198 models, we demonstrate state-of-the-art performance compared to alternatives, reaching a 20 percent increase against the best-performing alternative.

Multimodal fusion leverages information across modalities to learn better feature representations with the goal of improving performance in fusion-based tasks. However, multimodal datasets, especially in medical settings, are typically smaller than their unimodal counterparts, which can impede the performance of multimodal models. Additionally, the increase in the number of modalities is often associated with an overall increase in the size of the multimodal network, which may be undesirable in medical use cases. Utilizing smaller unimodal encoders may lead to sub-optimal performance, particularly when dealing with high-dimensional clinical data. In this paper, we propose the Modality-INformed knowledge Distillation (MIND) framework, a multimodal model compression approach based on knowledge distillation that transfers knowledge from ensembles of pre-trained deep neural networks of varying sizes into a smaller multimodal student. The teacher models consist of unimodal networks, allowing the student to learn from diverse representations. MIND employs multi-head joint fusion models, as opposed to single-head models, enabling the use of unimodal encoders in the case of unimodal samples without requiring imputation or masking of absent modalities. As a result, MIND generates an optimized multimodal model, enhancing both multimodal and unimodal representations. It can also be leveraged to balance multimodal learning during training. We evaluate MIND on binary and multilabel clinical prediction tasks using time series data and chest X-ray images. Additionally, we assess the generalizability of the MIND framework on three non-medical multimodal multiclass datasets. Experimental results demonstrate that MIND enhances the performance of the smaller multimodal network across all five tasks, as well as various fusion methods and multimodal architectures, compared to state-of-the-art baselines.

The rise of Agent AI and Large Language Model-powered Multi-Agent Systems (LLM-MAS) has underscored the need for responsible and dependable system operation. Tools like LangChain and Retrieval-Augmented Generation have expanded LLM capabilities, enabling deeper integration into MAS through enhanced knowledge retrieval and reasoning. However, these advancements introduce critical challenges: LLM agents exhibit inherent unpredictability, and uncertainties in their outputs can compound across interactions, threatening system stability. To address these risks, a human-centered design approach with active dynamic moderation is essential. Such an approach enhances traditional passive oversight by facilitating coherent inter-agent communication and effective system governance, allowing MAS to achieve desired outcomes more efficiently.

Recent studies show that in supervised fine-tuning (SFT) of large language models (LLMs), data quality matters more than quantity. While most data cleaning methods concentrate on filtering entire samples, the quality of individual tokens within a sample can vary significantly. After pre-training, even in high-quality samples, patterns or phrases that are not task-related can be redundant or uninformative. Continuing to fine-tune on these patterns may offer limited benefit and even degrade downstream task performance. In this paper, we investigate token quality from a noisy-label perspective and propose a generic token cleaning pipeline for SFT tasks. Our method filters out uninformative tokens while preserving those carrying key task-specific information. Specifically, we first evaluate token quality by examining the influence of model updates on each token, then apply a threshold-based separation. The token influence can be measured in a single pass with a fixed reference model or iteratively with self-evolving reference models. The benefits and limitations of both methods are analyzed theoretically by error upper bounds. Extensive experiments show that our framework consistently improves performance across multiple downstream tasks.

Large language models (LLMs) are prone to hallucinations in question-answering (QA) tasks when faced with ambiguous questions. Users often assume that LLMs share their cognitive alignment, a mutual understanding of context, intent, and implicit details, leading them to omit critical information in the queries. However, LLMs generate responses based on assumptions that can misalign with user intent, which may be perceived as hallucinations if they misalign with the user's intent. Therefore, identifying those implicit assumptions is crucial to resolve ambiguities in QA. Prior work, such as AmbigQA, reduces ambiguity in queries via human-annotated clarifications, which is not feasible in real application. Meanwhile, ASQA compiles AmbigQA's short answers into long-form responses but inherits human biases and fails capture explicit logical distinctions that differentiates the answers. We introduce Conditional Ambiguous Question-Answering (CondAmbigQA), a benchmark with 200 ambiguous queries and condition-aware evaluation metrics. Our study pioneers the concept of ``conditions'' in ambiguous QA tasks, where conditions stand for contextual constraints or assumptions that resolve ambiguities. The retrieval-based annotation strategy uses retrieved Wikipedia fragments to identify possible interpretations for a given query as its conditions and annotate the answers through those conditions. Such a strategy minimizes human bias introduced by different knowledge levels among annotators. By fixing retrieval results, CondAmbigQA evaluates how RAG systems leverage conditions to resolve ambiguities. Experiments show that models considering conditions before answering improve performance by $20\%$, with an additional $5\%$ gain when conditions are explicitly provided. These results underscore the value of conditional reasoning in QA, offering researchers tools to rigorously evaluate ambiguity resolution.

Pre-trained transformer models with extended context windows are notoriously expensive to run at scale, often limiting real-world deployment due to their high computational and memory requirements. In this paper, we introduce Hamming Attention Distillation (HAD), a novel framework that binarizes keys and queries in the attention mechanism to achieve significant efficiency gains. By converting keys and queries into {-1, +1} vectors and replacing dot-product operations with efficient Hamming distance computations, our method drastically reduces computational overhead. Additionally, we incorporate attention matrix sparsification to prune low-impact activations, which further reduces the cost of processing long-context sequences. \par Despite these aggressive compression strategies, our distilled approach preserves a high degree of representational power, leading to substantially improved accuracy compared to prior transformer binarization methods. We evaluate HAD on a range of tasks and models, including the GLUE benchmark, ImageNet, and QuALITY, demonstrating state-of-the-art performance among binarized Transformers while drastically reducing the computational costs of long-context inference. \par We implement HAD in custom hardware simulations, demonstrating superior performance characteristics compared to a custom hardware implementation of standard attention. HAD achieves just $\mathbf{1.78}\%$ performance losses on GLUE compared to $9.08\%$ in state-of-the-art binarization work, and $\mathbf{2.5}\%$ performance losses on ImageNet compared to $12.14\%$, all while targeting custom hardware with a $\mathbf{79}\%$ area reduction and $\mathbf{87}\%$ power reduction compared to its standard attention counterpart.

The understanding of physical processes has been a long-standing effort for scientists and engineers. A key step in this endeavor is to translate physical insights (laws) into precise mathematical relationships that capture the underlying phenomena. These relationships are then tested through experiments, which either validate the proposed hypothesis or suggest refinements. Among such mathematical formulations, differential equations (DEs) are especially ubiquitous across disciplines, as they describe how physical quantities evolve over time and space. Finding analytical (also referred to as explicit or closed-form) solutions to these equations, that is, a mathematical expression that satisfies the differential equation along with the given initial and boundary conditions, provides a structured way to compare theoretical predictions with experimental measurements. Moreover, analytical solutions often reveal intrinsic properties of physical systems, such as stability, periodicity, underlying symmetries and asymptotic behavior. Thus, analytical solutions provide deep insight into how these systems behave in time and space. Despite intense efforts over centuries, there are very few methods to construct analytical solutions of differential equations. All of them can be viewed as fundamentally compositional: They break complex equations into simpler, more manageable pieces and then systematically recombine those pieces into a final solution.