The substantial memory bandwidth and computational demands of large language models (LLMs) present critical challenges for efficient inference. To tackle this, the literature has explored heterogeneous systems that combine neural processing units (NPUs) with DRAM-based processing-in-memory (PIM) for LLM acceleration. However, existing high-precision (e.g., FP16) PIM compute units incur significant area and power overhead in DRAM technology, limiting the effective computation throughput. In this paper, we introduce P3-LLM, a novel NPU-PIM integrated accelerator for LLM inference using hybrid numerical formats. Our approach is threefold: First, we propose a flexible mixed-precision quantization scheme, which leverages hybrid numerical formats to quantize different LLM operands with high compression efficiency and minimal accuracy loss. Second, we architect an efficient PIM accelerator for P3-LLM, featuring enhanced compute units to support hybrid numerical formats. Our careful choice of numerical formats allows to co-design low-precision PIM compute units that significantly boost the computation throughput under iso-area constraints. Third, we optimize the low-precision dataflow of different LLM modules by applying operator fusion to minimize the overhead of runtime dequantization. Evaluation on a diverse set of representative LLMs and tasks demonstrates that P3-LLM achieves state-of-the-art accuracy in terms of both KV-cache quantization and weight-activation quantization. Combining the proposed quantization scheme with PIM architecture co-design, P3-LLM yields an average of $4.9\times$, $2.0\times$, and $3.4\times$ speedups over the state-of-the-art LLM accelerators HBM-PIM, Ecco, and Pimba, respectively. Our quantization code is available at https://github.com/yc2367/P3-LLM.git

Building trustworthy clinical AI systems requires not only accurate predictions but also transparent, biologically grounded explanations. We present \texttt{DiagnoLLM}, a hybrid framework that integrates Bayesian deconvolution, eQTL-guided deep learning, and LLM-based narrative generation for interpretable disease diagnosis. DiagnoLLM begins with GP-unmix, a Gaussian Process-based hierarchical model that infers cell-type-specific gene expression profiles from bulk and single-cell RNA-seq data while modeling biological uncertainty. These features, combined with regulatory priors from eQTL analysis, power a neural classifier that achieves high predictive performance in Alzheimer's Disease (AD) detection (88.0\% accuracy). To support human understanding and trust, we introduce an LLM-based reasoning module that translates model outputs into audience-specific diagnostic reports, grounded in clinical features, attribution signals, and domain knowledge. Human evaluations confirm that these reports are accurate, actionable, and appropriately tailored for both physicians and patients. Our findings show that LLMs, when deployed as post-hoc reasoners rather than end-to-end predictors, can serve as effective communicators within hybrid diagnostic pipelines.

Reasoning over procedural sequences, where the order of steps directly impacts outcomes, is a critical capability for large language models (LLMs). In this work, we study the task of reconstructing globally ordered sequences from shuffled procedural steps, using a curated dataset of food recipes, a domain where correct sequencing is essential for task success. We evaluate several LLMs under zero-shot and few-shot settings and present a comprehensive evaluation framework that adapts established metrics from ranking and sequence alignment. These include Kendall's Tau, Normalized Longest Common Subsequence (NLCS), and Normalized Edit Distance (NED), which capture complementary aspects of ordering quality. Our analysis shows that model performance declines with increasing sequence length, reflecting the added complexity of longer procedures. We also find that greater step displacement in the input, corresponding to more severe shuffling, leads to further degradation. These findings highlight the limitations of current LLMs in procedural reasoning, especially with longer and more disordered inputs.

Pre-trained robot policies serve as the foundation of many validated robotic systems, which encapsulate extensive embodied knowledge. However, they often lack the semantic awareness characteristic of foundation models, and replacing them entirely is impractical in many situations due to high costs and the loss of accumulated knowledge. To address this gap, we introduce GUIDES, a lightweight framework that augments pre-trained policies with semantic guidance from foundation models without requiring architectural redesign. GUIDES employs a fine-tuned vision-language model (Instructor) to generate contextual instructions, which are encoded by an auxiliary module into guidance embeddings. These embeddings are injected into the policy's latent space, allowing the legacy model to adapt to this new semantic input through brief, targeted fine-tuning. For inference-time robustness, a large language model-based Reflector monitors the Instructor's confidence and, when confidence is low, initiates a reasoning loop that analyzes execution history, retrieves relevant examples, and augments the VLM's context to refine subsequent actions. Extensive validation in the RoboCasa simulation environment across diverse policy architectures shows consistent and substantial improvements in task success rates. Real-world deployment on a UR5 robot further demonstrates that GUIDES enhances motion precision for critical sub-tasks such as grasping. Overall, GUIDES offers a practical and resource-efficient pathway to upgrade, rather than replace, validated robot policies.

Large Language Models have gained remarkable interest in industry and academia. The increasing interest in LLMs in academia is also reflected in the number of publications on this topic over the last years. For instance, alone 78 of the around 425 publications at ICSE 2024 performed experiments with LLMs. Conducting empirical studies with LLMs remains challenging and raises questions on how to achieve reproducible results, for both researchers and practitioners. One important step towards excelling in empirical research on LLM and their application is to first understand to what extent current research results are eventually reproducible and what factors may impede reproducibility. This investigation is within the scope of our work. We contribute an analysis of the reproducibility of LLM-centric studies, provide insights into the factors impeding reproducibility, and discuss suggestions on how to improve the current state. In particular, we studied the 85 articles describing LLM-centric studies, published at ICSE 2024 and ASE 2024. Of the 85 articles, 18 provided research artefacts and used OpenAI models. We attempted to replicate those 18 studies. Of the 18 studies, only five were sufficiently complete and executable. For none of the five studies, we were able to fully reproduce the results. Two studies seemed to be partially reproducible, and three studies did not seem to be reproducible. Our results highlight not only the need for stricter research artefact evaluations but also for more robust study designs to ensure the reproducible value of future publications.

Large Language Models (LLMs) demonstrate remarkable capabilities, but their training on massive corpora poses significant risks from memorized sensitive information. To mitigate these issues and align with legal standards, unlearning has emerged as a critical technique to selectively erase specific knowledge from LLMs without compromising their overall performance. This survey provides a systematic review of over 180 papers on LLM unlearning published since 2021. First, it introduces a novel taxonomy that categorizes unlearning methods based on the phase in the LLM pipeline of the intervention. This framework further distinguishes between parameter modification and parameter selection strategies, thus enabling deeper insights and more informed comparative analysis. Second, it offers a multidimensional analysis of evaluation paradigms. For datasets, we compare 18 existing benchmarks from the perspectives of task format, content, and experimental paradigms to offer actionable guidance. For metrics, we move beyond mere enumeration by dividing knowledge memorization metrics into 10 categories to analyze their advantages and applicability, while also reviewing metrics for model utility, robustness, and efficiency. By discussing current challenges and future directions, this survey aims to advance the field of LLM unlearning and the development of secure AI systems.

We introduce VeriStruct, a novel framework that extends AI-assisted automated verification from single functions to more complex data structure modules in Verus. VeriStruct employs a planner module to orchestrate the systematic generation of abstractions, type invariants, specifications, and proof code. To address the challenge that LLMs often misunderstand Verus' annotation syntax and verification-specific semantics, VeriStruct embeds syntax guidance within prompts and includes a repair stage to automatically correct annotation errors. In an evaluation on eleven Rust data structure modules, VeriStruct succeeds on ten of the eleven, successfully verifying 128 out of 129 functions (99.2%) in total. These results represent an important step toward the goal of automatic AI-assisted formal verification.

The rapid evolution of large language models (LLMs) creates complex bidirectional expectations between users and AI systems that are poorly understood. We introduce the concept of "mutual wanting" to analyze these expectations during major model transitions. Through analysis of user comments from major AI forums and controlled experiments across multiple OpenAI models, we provide the first large-scale empirical validation of bidirectional desire dynamics in human-AI interaction. Our findings reveal that nearly half of users employ anthropomorphic language, trust significantly exceeds betrayal language, and users cluster into distinct "mutual wanting" types. We identify measurable expectation violation patterns and quantify the expectation-reality gap following major model releases. Using advanced NLP techniques including dual-algorithm topic modeling and multi-dimensional feature extraction, we develop the Mutual Wanting Alignment Framework (M-WAF) with practical applications for proactive user experience management and AI system design. These findings establish mutual wanting as a measurable phenomenon with clear implications for building more trustworthy and relationally-aware AI systems.

Function calling (FC) empowers large language models (LLMs) and autonomous agents to interface with external tools, a critical capability for solving complex, real-world problems. As this ability becomes increasingly central to advanced AI systems, the need for high-quality, multi-turn training data to develop and refine it cannot be overstated. Existing data synthesis methods, such as random environment sampling or multi-agent role-playing, are not powerful enough to generate high-quality data in real-world environments. Practical challenges come in three folds: targeted data synthesis, hard query construction, and multi-turn logical dependency. To address these structural deficiencies, we present FunReason-MT, a novel data synthesis framework for real-world multi-turn tool use. FunReason-MT resolves the complexity barrier in multi-turn FC data by employing 1) Environment-API Graph Interactions to gather varied high-quality trajectories with targeted tool, 2) Advanced Tool-Query Synthesis to simplify hard query construction, and 3) Guided Iterative Chain for sophisticated CoT generation. Evaluations on Berkeley Function-Calling Leaderboard (BFCLv3) demonstrate the power of our framework: a 4B model built upon FunReason-MT generated data achieves state-of-the-art performance among comparable-sized models. Further performance improvements on BFCLv4 confirm that FunReason-MT provides a reliable and robust source for agentic learning.

Knowledge distillation (KD) is a standard route to compress Large Language Models (LLMs) into compact students, yet most pipelines uniformly apply token-wise loss regardless of teacher confidence. This indiscriminate supervision amplifies noisy, high-entropy signals and is especially harmful under large teacher-student capacity gaps. We introduce SelecTKD, a plug-and-play Selective Token-Weighted distillation framework that shifts the focus from "how to measure divergence" to "where to apply learning". At each step, the student proposes tokens that are verified by the teacher through a robust propose-and-verify procedure with two variants: greedy Top-k and non-greedy Spec-k. Accepted tokens receive full loss, while rejected tokens are masked or down-weighted. This objective-agnostic design works with on- and off-policy data, induces an implicit curriculum quantified by Token Acceptance Rate (TAR), and stabilizes optimization. Across instruction following, mathematical reasoning, code generation, and a VLM setting, SelecTKD consistently improves strong baselines and achieves state-of-the-art results for small models without architectural changes or extra reference models.