Freezing-of-Gait (FoG) affects over 50% of mid-to-late stage Parkinson's disease (PD) patients, significantly impairing patients' mobility independence and reducing quality of life. FoG is characterized by sudden episodes where walking cannot start or is interrupted, occurring exclusively during standing or walking, and never while sitting or lying down. Current FoG detection systems require extensive patient-specific training data and lack generalization, limiting clinical deployment. To address these issues, we introduce FM-FoG, a real-time foundation model-based wearable system achieving FoG detection in unseen patients without patient-specific training. Our approach combines self-supervised pretraining on diverse Inertial Measurement Unit (IMU) datasets with sensor context integration. Since FoG occurs only during ambulatory activities, a lightweight CNN-LSTM activity classifier selectively activates the foundation model only during walking or standing, avoiding unnecessary computation. Evaluated on the VCU FoG-IMU dataset with 23 PD patients, FM-FoG achieves a 98.5% F1-score when tested on previously unseen patients, substantially outperforming competitive baseline methods. Deployed on a Google Pixel 8a smartphone, the system extends battery life by up to 72% while maintaining sub-20ms intervention latency. The results indicate that our FM-FoG can enable practical, energy-efficient healthcare applications that generalize across patients without individual training requirements.

The growing prevalence of large language models (LLMs) and vision-language models (VLMs) has heightened the need for reliable techniques to determine whether a model has been fine-tuned from or is even identical to another. Existing similarity-based methods often require access to model parameters or produce heuristic scores without principled thresholds, limiting their applicability. We introduce Random Selection Probing (RSP), a hypothesis-testing framework that formulates model correlation detection as a statistical test. RSP optimizes textual or visual prefixes on a reference model for a random selection task and evaluates their transferability to a target model, producing rigorous p-values that quantify evidence of correlation. To mitigate false positives, RSP incorporates an unrelated baseline model to filter out generic, transferable features. We evaluate RSP across both LLMs and VLMs under diverse access conditions for reference models and test models. Experiments on fine-tuned and open-source models show that RSP consistently yields small p-values for related models while maintaining high p-values for unrelated ones. Extensive ablation studies further demonstrate the robustness of RSP. These results establish RSP as the first principled and general statistical framework for model correlation detection, enabling transparent and interpretable decisions in modern machine learning ecosystems.

Large Language Models (LLMs) can perform new tasks from in-context demonstrations, a phenomenon known as in-context learning (ICL). Recent work suggests that these demonstrations are compressed into task vectors (TVs), compact task representations that LLMs exploit for predictions. However, prior studies typically extract TVs from model outputs or hidden states using cumbersome and opaque methods, and they rarely elucidate the mechanisms by which TVs influence computation. In this work, we address both limitations. First, we propose directly training Learned Task Vectors (LTVs), which surpass extracted TVs in accuracy and exhibit superior flexibility-acting effectively at arbitrary layers, positions, and even with ICL prompts. Second, through systematic analysis, we investigate the mechanistic role of TVs, showing that at the low level they steer predictions primarily through attention-head OV circuits, with a small subset of "key heads" most decisive. At a higher level, we find that despite Transformer nonlinearities, TV propagation is largely linear: early TVs are rotated toward task-relevant subspaces to improve logits of relevant labels, while later TVs are predominantly scaled in magnitude. Taken together, LTVs not only provide a practical approach for obtaining effective TVs but also offer a principled lens into the mechanistic foundations of ICL.

Machine unlearning in large language models (LLMs) is essential for privacy and safety; however, existing approaches remain unstable and unreliable. A widely used strategy, the gradient difference method, applies gradient descent on retained data while performing gradient ascent on forget data, the data whose influence should be removed. However, when combined with cross-entropy loss, this procedure causes unbounded growth of weights and gradients, leading to training instability and degrading both forgetting and retention. We provide a theoretical framework that explains this failure, explicitly showing how ascent on the forget set destabilizes optimization in the feedforward MLP layers of LLMs. Guided by this insight, we propose Bounded Parameter-Efficient Unlearning, a parameter-efficient approach that stabilizes LoRA-based fine-tuning by applying bounded functions to MLP adapters. This simple modification controls the weight dynamics during ascent, enabling the gradient difference method to converge reliably. Across the TOFU, TDEC, and MUSE benchmarks, and across architectures and scales from 125M to 8B parameters, our method achieves substantial improvements in forgetting while preserving retention, establishing a novel theoretically grounded and practically scalable framework for unlearning in LLMs.

We investigate the mechanistic underpinnings of in-context learning (ICL) in large language models by reconciling two dominant perspectives: the component-level analysis of attention heads and the holistic decomposition of ICL into Task Recognition (TR) and Task Learning (TL). We propose a novel framework based on Task Subspace Logit Attribution (TSLA) to identify attention heads specialized in TR and TL, and demonstrate their distinct yet complementary roles. Through correlation analysis, ablation studies, and input perturbations, we show that the identified TR and TL heads independently and effectively capture the TR and TL components of ICL. Using steering experiments with geometric analysis of hidden states, we reveal that TR heads promote task recognition by aligning hidden states with the task subspace, while TL heads rotate hidden states within the subspace toward the correct label to facilitate prediction. We further show how previous findings on ICL mechanisms, including induction heads and task vectors, can be reconciled with our attention-head-level analysis of the TR-TL decomposition. Our framework thus provides a unified and interpretable account of how large language models execute ICL across diverse tasks and settings.

Abstract-- Pretrained large language models (LLMs) can work as high-level robotic planners by reasoning over abstract task descriptions and natural language instructions, etc. However, they have shown a lack of knowledge and effectiveness in planning long-horizon robotic manipulation tasks where the physical properties of the objects are essential. An example is the stacking of containers with hidden objects inside, which involves reasoning over hidden physics properties such as weight and stability. T o this end, this paper proposes to use multimodal LLMs as high-level planners for such long-horizon robotic stacking tasks. The LLM takes multimodal inputs for each object to stack and infers the current best stacking sequence by reasoning over stacking preferences. Furthermore, in order to enable the LLM to reason over multiple preferences at the same time without giving explicit instructions, we propose to create a custom dataset considering stacking preferences including weight, stability, size, and footprint, to fine-tune the LLM. Compared to the pretrained LLM with prompt tuning, we demonstrate the improved stacking completion of the LLM fine-tuned with our custom dataset via large-scale simulation evaluation. Furthermore, we showcase the effectiveness of the proposed framework for the long-horizon stacking task on a real humanoid robot in an online manner .

Large language models (LLMs) are increasingly explored in robot manipulation, but many existing methods struggle to adapt to new environments. Many systems require either environment-specific policy training or depend on fixed prompts and single-shot code generation, leading to limited transferability and manual re-tuning. We introduce Memory Transfer Planning (MTP), a framework that leverages successful control-code examples from different environments as procedural knowledge, using them as in-context guidance for LLM-driven planning. Specifically, MTP (i) generates an initial plan and code using LLMs, (ii) retrieves relevant successful examples from a code memory, and (iii) contextually adapts the retrieved code to the target setting for re-planning without updating model parameters. We evaluate MTP on RLBench, CALVIN, and a physical robot, demonstrating effectiveness beyond simulation. Across these settings, MTP consistently improved success rate and adaptability compared with fixed-prompt code generation, naive retrieval, and memory-free re-planning. Furthermore, in hardware experiments, leveraging a memory constructed in simulation proved effective. MTP provides a practical approach that exploits procedural knowledge to realize robust LLM-based planning across diverse robotic manipulation scenarios, enhancing adaptability to novel environments and bridging simulation and real-world deployment.

Large reasoning models (LRMs) exhibit unprecedented capabilities in solving complex problems through Chain-of-Thought (CoT) reasoning. However, recent studies reveal that their final answers often contradict their own reasoning traces. We hypothesize that this inconsistency stems from two competing mechanisms for generating answers: CoT reasoning and memory retrieval. To test this hypothesis, we conduct controlled experiments that challenge LRMs with misleading cues during reasoning and/or corrupted answers during retrieval. Our results across models and datasets confirm that both mechanisms operate simultaneously, with their relative dominance influenced by multiple factors: problem domains, model scales, and fine-tuning approaches (e.g., reinforcement learning vs. distillation). The findings reveal a critical limitation in current reasoning fine-tuning paradigms: models can exploit the retrieval mechanism as a shortcut, effectively "hacking" the reward signal and undermining genuine reasoning development. To address this challenge, we introduce FARL, a novel fine-tuning framework that integrates memory unlearning with reinforcement learning. By carefully suppressing retrieval shortcuts during the fine-tuning process, FARL promotes reasoning-dominant behavior and enhances generalizable reasoning capabilities.

Current comparisons of large reasoning models (LRMs) focus on macro-level statistics such as task accuracy or reasoning length. Whether different LRMs reason differently remains an open question. To address this gap, we introduce the LLM-proposed Open Taxonomy (LOT), a classification method that uses a generative language model to compare reasoning traces from two LRMs and articulate their distinctive features in words. LOT then models how these features predict the source LRM of a reasoning trace based on their empirical distributions across LRM outputs. Iterating this process over a dataset of reasoning traces yields a human-readable taxonomy that characterizes how models think. We apply LOT to compare the reasoning of 12 open-source LRMs on tasks in math, science, and coding. LOT identifies systematic differences in their thoughts, achieving 80-100% accuracy in distinguishing reasoning traces from LRMs that differ in scale, base model family, or objective domain. Beyond classification, LOT's natural-language taxonomy provides qualitative explanations of how LRMs think differently. Finally, in a case study, we link the reasoning differences to performance: aligning the reasoning style of smaller Qwen3 models with that of the largest Qwen3 during test time improves their accuracy on GPQA by 3.3-5.7%. Following the success of GPT -o1 and DeepSeek-R1, a wave of large reasoning models (LRMs) has recently become available. These models differ in training recipes and report varying benchmark performance, but far less is known about whether they also reason differently from one another. In this paper, we pose a fundamental question: can LRMs be distinguished by their reasoning patterns, and if so, what are the key distinguishing traits? A growing body of work has begun probing the reasoning style of individual LRMs, yielding insights into how artificial thinkers "think". Marjanovi c et al. (2025), for example, finds that DeepSeek-R1's reasoning depth correlates with the human cognitive load when processing complex sentences. Bogdan et al. (2025) annotates functions of DeepSeek-R1-Distill-Qwen's reasoning steps, showing that plan generation and re-evaluation are critical for solving math problems. However, only a few studies attempt broader comparisons across LRMs and on multiple reasoning behaviors. Gandhi et al. (2025) compares the reasoning patterns of base large language models (LLMs) and their RL fine-tuned variants, finding that the habits of base models correlate with fine-tuning gains.

Grounding natural language queries in graphical user interfaces (GUIs) presents a challenging task that requires models to comprehend diverse UI elements across various applications and systems, while also accurately predicting the spatial coordinates for the intended operation. To tackle this problem, we propose GMS: Generalist Scanner Meets Specialist Locator, a synergistic coarse-to-fine framework that effectively improves GUI grounding performance. GMS leverages the complementary strengths of general vision-language models (VLMs) and small, task-specific GUI grounding models by assigning them distinct roles within the framework. Specifically, the general VLM acts as a 'Scanner' to identify potential regions of interest, while the fine-tuned grounding model serves as a 'Locator' that outputs precise coordinates within these regions. This design is inspired by how humans perform GUI grounding, where the eyes scan the interface and the brain focuses on interpretation and localization. Our whole framework consists of five stages and incorporates hierarchical search with cross-modal communication to achieve promising prediction results. Experimental results on the ScreenSpot-Pro dataset show that while the 'Scanner' and 'Locator' models achieve only $2.0\%$ and $3.7\%$ accuracy respectively when used independently, their integration within GMS framework yields an overall accuracy of $35.7\%$, representing a $10 \times$ improvement. Additionally, GMS significantly outperforms other strong baselines under various settings, demonstrating its robustness and potential for general-purpose GUI grounding.