We identify semantically coherent, context-consistent network components in large language models (LLMs) using coactivation of sparse autoencoder (SAE) features collected from just a handful of prompts. Focusing on concept-relation prediction tasks, we show that ablating these components for concepts (e.g., countries and words) and relations (e.g., capital city and translation language) changes model outputs in predictable ways, while amplifying these components induces counterfactual responses. Notably, composing relation and concept components yields compound counterfactual outputs. Further analysis reveals that while most concept components emerge from the very first layer, more abstract relation components are concentrated in later layers. Lastly, we show that extracted components more comprehensively capture concepts and relations than individual features while maintaining specificity. Overall, our findings suggest a modular organization of knowledge accessed through compositional operations, and advance methods for efficient, targeted LLM manipulation.

Large Language Models have been shown to contain extensive world knowledge in their parameters, enabling impressive performance on many knowledge intensive tasks. However, when deployed in novel settings, LLMs often encounter situations where they must integrate parametric knowledge with new or unfamiliar information. In this work, we explore whether LLMs can combine knowledge in-context with their parametric knowledge through the lens of counterfactual reasoning. Through synthetic and real experiments in multi-hop reasoning problems, we show that LLMs generally struggle with counterfactual reasoning, often resorting to exclusively using their parametric knowledge. Moreover, we show that simple post-hoc finetuning can struggle to instill counterfactual reasoning ability - often leading to degradation in stored parametric knowledge. Ultimately, our work reveals important limitations of current LLM's abilities to re-purpose parametric knowledge in novel settings. Benchmarks like NaturalQuestions and HotpotQA have driven progress on recall-based and multi-hop reasoning, but they primarily evaluate a model's ability to regurgitate stored facts or compose chains of parametric knowledge without new external inputs (Y ang et al., 2018; Kwiatkowski et al., 2019). In contrast, many real-world scenarios require LLMs to integrate their pretrained knowledge with novel or hypothetical information provided at inference time. For example, consider a counterfactual query: "If Paris were located in Italy, in which country would the Eiffel T ower stand?"

The rapid advancement of large language models (LLMs) has exacerbated the memory bottleneck due to the widening gap between model parameter scaling and hardware capabilities. While post-training quantization techniques effectively reduce memory overhead, existing methods predominantly rely on static quantization strategies, which struggle to adapt to dynamic workloads. To address this, we propose FlexQuant, a dynamic precision-switching framework that optimizes the trade-off between inference speed and accuracy. Leveraging model perplexity entropy and Kullback-Leibler divergence, FlexQuant enables fine-grained, layer-wise mixed-precision quantization and dynamically adjusts bit-widths during each token generation. FlexQuant provides a comprehensive analysis of quantization strategies, introduces a precision requirement model for optimal switching, and implements efficient fine-grained precision management. Evaluations demonstrate that FlexQuant achieves a 1.3x end-to-end speedup across diverse language tasks with negligible accuracy loss introduced. This framework offers a flexible and adaptive solution for efficient LLM deployment. Code is released at https://github.com/ZongwuWang/FlexQuant.git.

While in-context learning (ICL) has achieved remarkable success in natural language and vision domains, its theoretical understanding-particularly in the context of structured geometric data-remains unexplored. This paper initiates a theoretical study of ICL for regression of Hölder functions on manifolds. We establish a novel connection between the attention mechanism and classical kernel methods, demonstrating that transformers effectively perform kernel-based prediction at a new query through its interaction with the prompt. This connection is validated by numerical experiments, revealing that the learned query-prompt scores for Hölder functions are highly correlated with the Gaussian kernel. Building on this insight, we derive generalization error bounds in terms of the prompt length and the number of training tasks. When a sufficient number of training tasks are observed, transformers give rise to the minimax regression rate of Hölder functions on manifolds, which scales exponentially with the intrinsic dimension of the manifold, rather than the ambient space dimension. Our result also characterizes how the generalization error scales with the number of training tasks, shedding light on the complexity of transformers as in-context kernel algorithm learners. Our findings provide foundational insights into the role of geometry in ICL and novels tools to study ICL of nonlinear models.

The real world is messy and unstructured. Uncovering critical information often requires active, goal-driven exploration. It remains to be seen whether Vision-Language Models (VLMs), which recently emerged as a popular zero-shot tool in many difficult tasks, can operate effectively in such conditions. In this paper, we answer this question by introducing FlySearch, a 3D, outdoor, photorealistic environment for searching and navigating to objects in complex scenes. We define three sets of scenarios with varying difficulty and observe that state-of-the-art VLMs cannot reliably solve even the simplest exploration tasks, with the gap to human performance increasing as the tasks get harder. We identify a set of central causes, ranging from vision hallucination, through context misunderstanding, to task planning failures, and we show that some of them can be addressed by finetuning. We publicly release the benchmark, scenarios, and the underlying codebase.

We introduce EvaLearn, a pioneering benchmark designed to evaluate large language models (LLMs) on their learning capability and efficiency in challenging tasks, a critical, yet underexplored aspect of model potential. EvaLearn contains 648 challenging problems across six task types, grouped into 182 sequences, each sequence dedicated to one task type. Diverging from most existing benchmarks that evaluate models in parallel, EvaLearn requires models to solve problems sequentially, allowing them to leverage the experience gained from previous solutions. EvaLearn provides five comprehensive automated metrics to evaluate models and quantify their learning capability and efficiency. We extensively benchmark nine frontier models and observe varied performance profiles: some models, such as Claude-3.7-sonnet, start with moderate initial performance but exhibit strong learning ability, while some models struggle to benefit from experience and may even show negative transfer. Moreover, we investigate model performance under two learning settings and find that instance-level rubrics and teacher-model feedback further facilitate model learning. Importantly, we observe that current LLMs with stronger static abilities do not show a clear advantage in learning capability across all tasks, highlighting that EvaLearn evaluates a new dimension of model performance. We hope EvaLearn provides a novel evaluation perspective for assessing LLM potential and understanding the gap between models and human capabilities, promoting the development of deeper and more dynamic evaluation approaches. All datasets, the automatic evaluation framework, and the results studied in this paper are available at the GitHub repository.

Current parameter-efficient fine-tuning methods for adapting pre-trained language models to downstream tasks are susceptible to interference from noisy data. Conventional noise-handling approaches either rely on laborious data pre-processing or employ model architecture modifications prone to error accumulation. In contrast to existing noise-process paradigms, we propose a noise-robust adaptation method via asymmetric LoRA poisoning experts (LoPE), a novel framework that enhances model robustness to noise only with generated noisy data. Drawing inspiration from the mixture-of-experts architecture, LoPE strategically integrates a dedicated poisoning expert in an asymmetric LoRA configuration. Through a two-stage paradigm, LoPE performs noise injection on the poisoning expert during fine-tuning to enhance its noise discrimination and processing ability. During inference, we selectively mask the dedicated poisoning expert to leverage purified knowledge acquired by normal experts for noise-robust output. Extensive experiments demonstrate that LoPE achieves strong performance and robustness purely through the low-cost noise injection, which completely eliminates the requirement of data cleaning.

Although inspired by neuronal systems in the brain, artificial neural networks generally employ point-neurons, which offer far less computational complexity than their biological counterparts. Neurons have dendritic arbors that connect to different sets of synapses and offer local non-linear accumulation - playing a pivotal role in processing and learning. Inspired by this, we propose a novel neuron design based on a multi-gate ferroelectric field-effect transistor that mimics dendrites. It leverages ferroelectric nonlinearity for local computations within dendritic branches, while utilizing the transistor action to generate the final neuronal output. The branched architecture paves the way for utilizing smaller crossbar arrays in hardware integration, leading to greater efficiency. Using an experimentally calibrated device-circuit-algorithm co-simulation framework, we demonstrate that networks incorporating our dendritic neurons achieve superior performance in comparison to much larger networks without dendrites ($\sim$17$\times$ fewer trainable weight parameters). These findings suggest that dendritic hardware can significantly improve computational efficiency, and learning capacity of neuromorphic systems optimized for edge applications.

The demand for high-fidelity test data is paramount in industrial settings where access to production data is largely restricted. Traditional data generation methods often fall short, struggling with low-fidelity and the ability to model complex data structures and semantic relationships that are critical for testing complex SQL code generation services like Natural Language to SQL (NL2SQL). In this paper, we address the critical need for generating syntactically correct and semantically relevant high-fidelity mock data for complex data structures that includes columns with nested structures that we frequently encounter in Google workloads. We highlight the limitations of existing approaches used in production, particularly their inability to handle large and complex data structures, as well as the lack of semantically coherent test data that lead to limited test coverage. We demonstrate that by leveraging Large Language Models (LLMs) and incorporating strategic pre- and post-processing steps, we can generate syntactically correct and semantically relevant high-fidelity test data that adheres to complex structural constraints and maintains semantic integrity to the SQL test targets (queries/functions). This approach supports comprehensive testing of complex SQL queries involving joins, aggregations, and even deeply nested subqueries, ensuring robust evaluation of SQL code generation services, like NL2SQL and SQL Code Assistant. Our results demonstrate the practical utility of an LLM (\textit{gemini}) based test data generation for industrial SQL code generation services where generating high-fidelity test data is essential due to the frequent unavailability and inaccessibility of production datasets for testing.

Respiratory diseases account for a significant portion of global mortality. Affordable and early detection is an effective way of addressing these ailments. To this end, a low-cost commercial off-the-shelf (COTS), IEEE 802.15.4z standard compliant impulse-radio ultra-wideband (IR-UWB) radar system is used to estimate human respiration rates. We propose a convolutional neural network (CNN) specifically adapted to predict breathing rates from ultra-wideband (UWB) channel impulse response (CIR) data, and compare its performance with both other rule-based algorithms and model-based solutions. The study uses a diverse dataset, incorporating various real-life environments to evaluate system robustness. To facilitate future research, this dataset will be released as open source. Results show that the CNN achieves a mean absolute error (MAE) of 1.73 breaths per minute (BPM) in unseen situations, significantly outperforming rule-based methods (3.40 BPM). By incorporating calibration data from other individuals in the unseen situations, the error is further reduced to 0.84 BPM. In addition, this work evaluates the feasibility of running the pipeline on a low-cost embedded device. Applying 8-bit quantization to both the weights and input/ouput tensors, reduces memory requirements by 67% and inference time by 64% with only a 3% increase in MAE. As a result, we show it is feasible to deploy the algorithm on an nRF52840 system-on-chip (SoC) requiring only 46 KB of memory and operating with an inference time of only 192 ms. Once deployed, an analytical energy model estimates that the system, while continuously monitoring the room, can operate for up to 268 days without recharging when powered by a 20 000 mAh battery pack. For breathing monitoring in bed, the sampling rate can be lowered, extending battery life to 313 days, making the solution highly efficient for real-world, low-cost deployments.