Visual-Language Navigation (VLN) is a fundamental challenge in robotic systems, with broad applications for the deployment of embodied agents in real-world environments. Despite recent advances, existing approaches are limited in long-range spatial reasoning, often exhibiting low success rates and high inference latency, particularly in long-range navigation tasks. To address these limitations, we propose FSR-VLN, a vision-language navigation system that combines a Hierarchical Multi-modal Scene Graph (HMSG) with Fast-to-Slow Navigation Reasoning (FSR). The HMSG provides a multi-modal map representation supporting progressive retrieval, from coarse room-level localization to fine-grained goal view and object identification. Building on HMSG, FSR first performs fast matching to efficiently select candidate rooms, views, and objects, then applies VLM-driven refinement for final goal selection. We evaluated FSR-VLN across four comprehensive indoor datasets collected by humanoid robots, utilizing 87 instructions that encompass a diverse range of object categories. FSR-VLN achieves state-of-the-art (SOTA) performance in all datasets, measured by the retrieval success rate (RSR), while reducing the response time by 82% compared to VLM-based methods on tour videos by activating slow reasoning only when fast intuition fails. Furthermore, we integrate FSR-VLN with speech interaction, planning, and control modules on a Unitree-G1 humanoid robot, enabling natural language interaction and real-time navigation.

Drug shortages pose critical risks to patient care and healthcare systems worldwide, yet the effectiveness of regulatory interventions remains poorly understood due to information asymmetries in pharmaceutical supply chains. We propose \textbf{ShortageSim}, addresses this challenge by providing the first simulation framework that evaluates the impact of regulatory interventions on competition dynamics under information asymmetry. Using Large Language Model (LLM)-based agents, the framework models the strategic decisions of drug manufacturers and institutional buyers, in response to shortage alerts given by the regulatory agency. Unlike traditional game theory models that assume perfect rationality and complete information, ShortageSim simulates heterogeneous interpretations on regulatory announcements and the resulting decisions. Experiments on self-processed dataset of historical shortage events show that ShortageSim reduces the resolution lag for production disruption cases by up to 84\%, achieving closer alignment to real-world trajectories than the zero-shot baseline. Our framework confirms the effect of regulatory alert in addressing shortages and introduces a new method for understanding competition in multi-stage environments under uncertainty. We open-source ShortageSim and a dataset of 2,925 FDA shortage events, providing a novel framework for future research on policy design and testing in supply chains under information asymmetry.

Large Language Models (LLMs) are increasingly deployed in high-stakes domains such as science, law, and healthcare, where accurate expressions of uncertainty are essential for reliability and trust. However, current LLMs are often observed to generate incorrect answers with high confidence, a phenomenon known as "overconfidence". Recent efforts have focused on calibrating LLMs' verbalized confidence: i.e., their expressions of confidence in text form, such as "I am 80% confident that...". Existing approaches either rely on prompt engineering or fine-tuning with heuristically generated uncertainty estimates, both of which have limited effectiveness and generalizability. Motivated by the notion of proper scoring rules for calibration in classical machine learning models, we introduce ConfTuner, a simple and efficient fine-tuning method that introduces minimal overhead and does not require ground-truth confidence scores or proxy confidence estimates. ConfTuner relies on a new loss function, tokenized Brier score, which we theoretically prove to be a proper scoring rule, intuitively meaning that it "correctly incentivizes the model to report its true probability of being correct". ConfTuner improves calibration across diverse reasoning tasks and generalizes to black-box models such as GPT-4o. Our results further show that better-calibrated confidence enables downstream gains in self-correction and model cascade, advancing the development of trustworthy LLM systems. The code is available at https://github.com/liushiliushi/ConfTuner.

Bold claims about AI's role in science-from "AGI will cure all diseases" to promises of radically accelerated discovery-raise a central epistemic question: do large language models (LLMs) truly generate new knowledge, or do they merely remix memorized fragments? We propose unlearning-as-ablation as a falsifiable probe of constructive scientific discovery. The idea is to systematically remove a target result together with its forget-closure (supporting lemmas, paraphrases, and multi-hop entailments) and then evaluate whether the model can re-derive the result from only permitted axioms and tools. Success would indicate generative capability beyond recall; failure would expose current limits. Unlike prevailing motivations for unlearning-privacy, copyright, or safety-our framing repositions it as an epistemic probe for AI-for-Science. We outline a minimal pilot in mathematics and algorithms to illustrate feasibility, and sketch how the same approach could later be extended to domains such as physics or chemistry. This is a position paper: our contribution is conceptual and methodological, not empirical. We aim to stimulate discussion on how principled ablation tests could help distinguish models that reconstruct knowledge from those that merely retrieve it, and how such probes might guide the next generation of AI-for-Science benchmarks.

Large Language Models (LLMs) are increasingly used as evaluators in natural language processing tasks, a paradigm known as LLM-as-a-Judge (LaaJ). The analysis of a LaaJ software, commonly refereed to as meta-evaluation, pose significant challenges in domain-specific contexts. In such domains, in contrast to general domains, annotated data is scarce and expert evaluation is costly. As a result, meta-evaluation is often performed using metrics that have not been validated for the specific domain in which they are applied. Therefore, it becomes difficult to determine which metrics effectively identify LaaJ quality, and further, what threshold indicates sufficient evaluator performance. In this work, we introduce LaaJMeter, a simulation-based framework for controlled meta-evaluation of LaaJs. LaaJMeter enables engineers to generate synthetic data representing virtual models and judges, allowing systematic analysis of evaluation metrics under realistic conditions. This helps practitioners validate LaaJs for specific tasks: they can test whether their metrics correctly distinguish between high and low quality (virtual) LaaJs, and estimate appropriate thresholds for evaluator adequacy. We demonstrate the utility of LaaJMeter in a code translation task involving a legacy programming language, showing how different metrics vary in sensitivity to evaluator quality. Our results highlight the limitations of common metrics and the importance of principled metric selection. LaaJMeter provides a scalable and extensible solution for assessing LaaJs in low-resource settings, contributing to the broader effort to ensure trustworthy and reproducible evaluation in NLP.

With the increasing presence of intelligent robots in everyday life, the demand for reliable and straightforward Human-Robot Interaction (HRI) interfaces is rapidly rising. Traditional robot control paradigms require users to learn particular commands [1] or interact with the robots through rigid user interfaces, especially in unstructured environments [2]. However, recent works target more flexible and adaptive communication strategies, unlocking the full potential of autonomous agents in human-centered environments. Accordingly, advances in generative AI and Large Language Models (LLMs) reveal new opportunities for enabling seamless communication between humans and robots, where natural language is the primary means of communication [3]. Such models are powerful enough to comprehend given instructions and even "reason" about the demanded tasks, intentions, and environmental context [4]. When paired with robotic perception and control systems, LLMs enable users to intuitively instruct the robot to perform complex tasks such as following multiple objects [5], navigating through dynamic scenes [6], or interacting with specific items [7], all using natural dialogue. Furthermore, integrating multimodal capabilities, including vision and speech, enhances HRI by enabling more natural, context-aware communication and improving adaptability across tasks and environments [8].

Large language models (LLMs) have been widely deployed across numerous fields. Reinforcement Learning from Human Feedback (RLHF) leverages reward models (RMs) as proxies for human preferences to align LLM behaviors with human values, making the accuracy, reliability, and interpretability of RMs critical for effective alignment. However, traditional RMs lack interpretability, offer limited insight into the reasoning behind reward assignments, and are inflexible toward user preference shifts. While recent multidimensional RMs aim for improved interpretability, they often fail to provide feature-level attribution and require costly annotations. To overcome these limitations, we introduce the Sparse Autoencoder-enhanced Reward Model (SARM), a novel architecture that integrates a pretrained Sparse Autoencoder (SAE) into a reward model. SARM maps the hidden activations of LLM-based RM into an interpretable, sparse, and monosemantic feature space, from which a scalar head aggregates feature activations to produce transparent and conceptually meaningful reward scores. Empirical evaluations demonstrate that SARM facilitates direct feature-level attribution of reward assignments, allows dynamic adjustment to preference shifts, and achieves superior alignment performance compared to conventional reward models. Our code is available at https://github.com/schrieffer-z/sarm.

Modern large language models (LLMs) demonstrate exceptional performance on knowledge-intensive tasks, yet the theoretical mechanisms underlying knowledge acquisition (storage and memorization) during pre-training and extraction (retrieval and recall) during inference after fine-tuning remain poorly understood. Although prior theoretical studies have explored these processes through analyses of training dynamics, they overlook critical components essential for a comprehensive theory: (1) the multi-layer perceptron (MLP), empirically identified as the primary module for knowledge storage; (2) out-of-distribution (OOD) adaptivity, which enables LLMs to generalize to unseen scenarios post-pre-training; and (3) next-token prediction, the standard autoregressive objective that encodes knowledge as conditional probabilities. In this work, we introduce, to the best of our knowledge, the first theoretical framework that addresses these limitations by examining the training dynamics of one-layer transformers. Under regularity assumptions, we establish that: (i) transformers attain near-optimal training loss during pre-training, demonstrating effective knowledge acquisition; (ii) given a sufficiently large fine-tuning dataset and appropriate data multiplicity conditions, transformers achieve low generalization error on factual knowledge acquired during pre-training but not revisited in fine-tuning, indicating robust knowledge extraction; and (iii) violation of these conditions leads to elevated generalization error, manifesting as hallucinations. Our analysis encompasses both full fine-tuning and low-rank fine-tuning, yielding insights into the efficacy of practical low-rank adaptation methods. We validate our theoretical findings through experiments on synthetic datasets and the real-world PopQA benchmark, employing GPT-2 and Llama-3.2-1B models.

The rapid scaling of large language models (LLMs) has significantly increased GPU memory pressure, which is further aggravated by training optimization techniques such as virtual pipeline and recomputation that disrupt tensor lifespans and introduce considerable memory fragmentation. Such fragmentation stems from the use of online GPU memory allocators in popular deep learning frameworks like PyTorch, which disregard tensor lifespans. As a result, this inefficiency can waste as much as 43% of memory and trigger out-of-memory errors, undermining the effectiveness of optimization methods. To address this, we introduce STAlloc, a GPU memory allocator for deep learning frameworks that reduces fragmentation by exploiting the spatial and temporal regularity in memory allocation behaviors of training workloads. STAlloc introduces a novel paradigm that combines offline planning with online allocation. The offline planning leverages spatio-temporal regularities to generate a near-optimal allocation plan, while the online allocation handles complex and dynamic models such as Mixture-of-Experts (MoE). Built as a pluggable PyTorch memory allocator, STAlloc reduces fragmentation ratio on average by 85.1% (up to 100%) across both dense and MoE models, with negligible overhead. This enables more efficient, high-throughput training configurations and improves throughput performance by up to 32.5%.

Large language models (LLMs) have emerged as powerful tools for simulating complex social phenomena using human-like agents with specific traits. In human societies, value similarity is important for building trust and close relationships; however, it remains unexplored whether this principle holds true in artificial societies comprising LLM agents. Therefore, this study investigates the influence of value similarity on relationship-building among LLM agents through two experiments. First, in a preliminary experiment, we evaluated the controllability of values in LLMs to identify the most effective model and prompt design for controlling the values. Subsequently, in the main experiment, we generated pairs of LLM agents imbued with specific values and analyzed their mutual evaluations of trust and interpersonal closeness following a dialogue. The experiments were conducted in English and Japanese to investigate language dependence. The results confirmed that pairs of agents with higher value similarity exhibited greater mutual trust and interpersonal closeness. Our findings demonstrate that the LLM agent simulation serves as a valid testbed for social science theories, contributes to elucidating the mechanisms by which values influence relationship building, and provides a foundation for inspiring new theories and insights into the social sciences.