Accurate diagnosis of Alzheimer's disease (AD) requires handling tabular biomarker data, yet such data are often small and incomplete, where deep learning models frequently fail to outperform classical methods. Pretrained large language models (LLMs) offer few-shot generalization, structured reasoning, and interpretable outputs, providing a powerful paradigm shift for clinical prediction. We propose TAP-GPT Tabular Alzheimer's Prediction GPT, a domain-adapted tabular LLM framework built on TableGPT2 and fine-tuned for few-shot AD classification using tabular prompts rather than plain texts. We evaluate TAP-GPT across four ADNI-derived datasets, including QT-PAD biomarkers and region-level structural MRI, amyloid PET, and tau PET for binary AD classification. Across multimodal and unimodal settings, TAP-GPT improves upon its backbone models and outperforms traditional machine learning baselines in the few-shot setting while remaining competitive with state-of-the-art general-purpose LLMs. We show that feature selection mitigates degradation in high-dimensional inputs and that TAP-GPT maintains stable performance under simulated and real-world missingness without imputation. Additionally, TAP-GPT produces structured, modality-aware reasoning aligned with established AD biology and shows greater stability under self-reflection, supporting its use in iterative multi-agent systems. To our knowledge, this is the first systematic application of a tabular-specialized LLM to multimodal biomarker-based AD prediction, demonstrating that such pretrained models can effectively address structured clinical prediction tasks and laying the foundation for tabular LLM-driven multi-agent clinical decision-support systems. The source code is publicly available on GitHub: https://github.com/sophie-kearney/TAP-GPT.

Opening sterile medical packaging is routine for healthcare workers but remains challenging for robots. Learning from demonstration enables robots to acquire manipulation skills directly from humans, and handheld gripper tools such as the Universal Manipulation Interface (UMI) offer a pathway for efficient data collection. However, the effectiveness of these tools depends heavily on their usability. We evaluated UMI in demonstrating a bandage opening task, a common manipulation task in hospital settings, by testing three conditions: distributed load grippers, concentrated load grippers, and bare hands. Eight participants performed timed trials, with task performance assessed by success rate, completion time, and damage, alongside perceived workload using the NASA-TLX questionnaire. Concentrated load grippers improved performance relative to distributed load grippers but remained substantially slower and less effective than hands. These results underscore the importance of ergonomic and mechanical refinements in handheld grippers to reduce user burden and improve demonstration quality, especially for applications in healthcare robotics.

Large language model (LLM) agents are increasingly used for complex tasks, yet deployed agents often remain static, failing to adapt as user needs evolve. This creates a tension between the need for continuous service and the necessity of updating capabilities to match shifting task distributions. On platforms like OpenClaw, which handle diverse workloads across 20+ channels, existing methods either store raw trajectories without distilling knowledge, maintain static skill libraries, or require disruptive downtime for retraining. We present MetaClaw, a continual meta-learning framework that jointly evolves a base LLM policy and a library of reusable behavioral skills. MetaClaw employs two complementary mechanisms. Skill-driven fast adaptation analyzes failure trajectories via an LLM evolver to synthesize new skills, enabling immediate improvement with zero downtime. Opportunistic policy optimization performs gradient-based updates via cloud LoRA fine-tuning and Reinforcement Learning with a Process Reward Model (RL-PRM). This is triggered during user-inactive windows by the Opportunistic Meta-Learning Scheduler (OMLS), which monitors system inactivity and calendar data. These mechanisms are mutually reinforcing: a refined policy generates better trajectories for skill synthesis, while richer skills provide higher-quality data for policy optimization. To prevent data contamination, a versioning mechanism separates support and query data. Built on a proxy-based architecture, MetaClaw scales to production-size LLMs without local GPUs. Experiments on MetaClaw-Bench and AutoResearchClaw show that skill-driven adaptation improves accuracy by up to 32% relative. The full pipeline advances Kimi-K2.5 accuracy from 21.4% to 40.6% and increases composite robustness by 18.3%. Code is available at https://github.com/aiming-lab/MetaClaw.

Reliable product identification from images is a critical requirement in industrial and commercial applications, particularly in maintenance, procurement, and operational workflows where incorrect matches can lead to costly downstream failures. At the core of such systems lies the visual search component, which must retrieve and rank the exact object instance from large and continuously evolving catalogs under diverse imaging conditions. This report presents a structured benchmark of modern visual embedding models for instance-level image retrieval, with a focus on industrial applications. A curated set of open-source foundation embedding models, proprietary multi-modal embedding systems, and domain-specific vision-only models are evaluated under a unified image-to-image retrieval protocol. The benchmark includes curated datasets, which includes industrial datasets derived from production deployments in Manufacturing, Automotive, DIY, and Retail, as well as established public benchmarks. Evaluation is conducted without post-processing, isolating the retrieval capability of each model. The results provide insight into how well contemporary foundation and unified embedding models transfer to fine-grained instance retrieval tasks, and how they compare to models explicitly trained for industrial applications. By emphasizing realistic constraints, heterogeneous image conditions, and exact instance matching requirements, this benchmark aims to inform both practitioners and researchers about the strengths and limitations of current visual embedding approaches in production-level product identification systems. An interactive companion website presenting the benchmark results, evaluation details, and additional visualizations is available at https://benchmark.nyris.io.

Recovering a dense 3D body mesh from monocular video remains challenging under occlusion from draping and continuously moving camera viewpoints. This configuration arises in surgical augmented reality (AR), where an anesthetized patient lies under surgical draping while a surgeon's head-mounted camera continuously changes viewpoint. Existing human mesh recovery (HMR) methods are typically trained on upright, moving subjects captured from relatively stable cameras, leading to performance degradation under such conditions. To address this, we present Patient4D, a stationarity-constrained reconstruction pipeline that explicitly exploits the stationarity prior. The pipeline combines image-level foundation models for perception with lightweight geometric mechanisms that enforce temporal consistency across frames. Two key components enable robust reconstruction: Pose Locking, which anchors pose parameters using stable keyframes, and Rigid Fallback, which recovers meshes under severe occlusion through silhouette-guided rigid alignment. Together, these mechanisms stabilize predictions while remaining compatible with off-the-shelf HMR models. We evaluate Patient4D on 4,680 synthetic surgical sequences and three public HMR video benchmarks. Under surgical drape occlusion, Patient4D achieves a 0.75 mean IoU, reducing failure frames from 30.5% to 1.3% compared to the best baseline. Our findings demonstrate that exploiting stationarity priors can substantially improve monocular reconstruction in clinical AR scenarios.

Explainable Boosting Machines (EBMs) provide transparent predictions through additive shape functions, enabling direct inspection of feature contributions. However, EBMs can learn non-physical relationships that reduce their reliability in natural hazard applications. This study presents a domain-informed framework to improve the physical consistency of EBMs for lateral spreading prediction. Our approach modifies learned shape functions based on domain knowledge. These modifications correct non-physical behavior while maintaining data-driven patterns. We apply the method to the 2011 Christchurch earthquake dataset and correct non-physical trends observed in the original EBM. The resulting model produces more physically consistent global and local explanations, with an acceptable tradeoff in accuracy (4--5\%).

Iris presentation attack detection (PAD) is critical for secure biometric deployments, yet developing specialized models faces significant practical barriers: collecting data representing future unknown attacks is impossible, and collecting diverse-enough data, yet still limited in terms of its predictive power, is expensive. Additionally, sharing biometric data raises privacy concerns. Due to rapid emergence of new attack vectors demanding adaptable solutions, we thus investigate in this paper whether general-purpose multimodal large language models (MLLMs) can perform iris PAD when augmented with human expert knowledge, operating under strict privacy constraints that prohibit sending biometric data to public cloud MLLM services. Through analysis of vision encoder embeddings applied to our dataset, we demonstrate that pre-trained vision transformers in MLLMs inherently cluster many iris attack types despite never being explicitly trained for this task. However, where clustering shows overlap between attack classes, we find that structured prompts incorporating human salience (verbal descriptions from subjects identifying attack indicators) enable these models to resolve ambiguities. Testing on an IRB-restricted dataset of 224 iris images spanning seven attack types, using only university-approved services (Gemini 2.5 Pro) or locally-hosted models (e.g., Llama 3.2-Vision), we show that Gemini with expert-informed prompts outperforms both a specialized convolutional neural networks (CNN)-based baseline and human examiners, while the locally-deployable Llama achieves near-human performance. Our results establish that MLLMs deployable within institutional privacy constraints offer a viable path for iris PAD.

Large language models (LLMs) are increasingly used as automated judges and synthetic labelers, especially in low-label settings. Yet these systems are stochastic and often overconfident, which makes deployment decisions difficult when external ground truth is limited. We propose a practical calibration protocol based on controlled input interventions: if noise severity increases, task performance should exhibit a statistically significant deterioration trend. We operationalize this with a slope-based hypothesis test over repeated trials, using signal-to-noise-ratio (SNR) perturbations for tabular data and lexical perturbations for text data. Across UCI tabular benchmarks and four text classification datasets, we find clear modality-dependent behavior. Our results reveal a modality gap: while text-based judges degrade predictably, the majority of tabular datasets show a lack of statistically significant performance deterioration even under significant signal-to-noise reduction. Interestingly we find that model performance is lower on datasets that are insensitive to noise interventions. We present a reproducible methodology and reporting protocol for robust LLM-judge calibration under distribution shift.

Evaluating the grammatical competence of second language (L2) learners is essential both for providing targeted feedback and for assessing proficiency. To achieve this, we propose a novel framework leveraging the English Grammar Profile (EGP), a taxonomy of grammatical constructs mapped to the proficiency levels of the Common European Framework of Reference (CEFR), to detect learners' attempts at grammatical constructs and classify them as successful or unsuccessful. This detection can then be used to provide fine-grained feedback. Moreover, the grammatical constructs are used as predictors of proficiency assessment by using automatically detected attempts as predictors of holistic CEFR proficiency. For the selection of grammatical constructs derived from the EGP, rule-based and LLM-based classifiers are compared. We show that LLMs outperform rule-based methods on semantically and pragmatically nuanced constructs, while rule-based approaches remain competitive for constructs that rely purely on morphological or syntactic features and do not require semantic interpretation. For proficiency assessment, we evaluate both rule-based and hybrid pipelines and show that a hybrid approach combining a rule-based pre-filter with an LLM consistently yields the strongest performance. Since our framework operates on pairs of original learner sentences and their corrected counterparts, we also evaluate a fully automated pipeline using automatic grammatical error correction. This pipeline closely approaches the performance of semi-automated systems based on manual corrections, particularly for the detection of successful attempts at grammatical constructs. Overall, our framework emphasises learners' successful attempts in addition to unsuccessful ones, enabling positive, formative feedback and providing actionable insights into grammatical development.

Deducing whodunit proves challenging for LLM agents. In this paper, we implement a text-based multi-agent version of the classic board game Clue as a rule-based testbed for evaluating multi-step deductive reasoning, with six agents drawn from GPT-4o-mini and Gemini-2.5-Flash. We further investigate whether fine-tuning on structured logic puzzles transfers to improved in-game reasoning and gameplay. Across 18 simulated games, agents achieve only four correct wins, indicating difficulty in maintaining consistent deductive reasoning over the course of a full game. Additionally, we find that fine-tuning does not reliably improve performance and, in some cases, appears to increase reasoning volume without improving reasoning precision.

We present SAL (SLAM Adversarial Lab), a modular framework for evaluating visual SLAM systems under adversarial conditions such as fog and rain. SAL represents each adversarial condition as a perturbation that transforms an existing dataset into an adversarial dataset. When transforming a dataset, SAL supports severity levels using easily-interpretable real-world units such as meters for fog visibility. SAL's extensible architecture decouples datasets, perturbations, and SLAM algorithms through common interfaces, so users can add new components without rewriting integration code. Moreover, SAL includes a search procedure that finds the severity level of a perturbation at which a SLAM system fails. To showcase the capabilities of SAL, our evaluation integrates seven SLAM algorithms and evaluates them across three datasets under weather, camera, and video transport perturbations.

We present GazeOnce360, a novel end-to-end model for multi-person gaze estimation from a single tabletop-mounted upward-facing fisheye camera. Unlike conventional approaches that rely on forward-facing cameras in constrained viewpoints, we address the underexplored setting of estimating the 3D gaze direction of multiple people distributed across a 360° scene from an upward fisheye perspective. To support research in this setting, we introduce MPSGaze360, a large-scale synthetic dataset rendered using Unreal Engine, featuring diverse multi-person configurations with accurate 3D gaze and eye landmark annotations. Our model tackles the severe distortion and perspective variation inherent in fisheye imagery by incorporating rotational convolutions and eye landmark supervision. To better capture fine-grained eye features crucial for gaze estimation, we propose a dual-resolution architecture that fuses global low-resolution context with high-resolution local eye regions. Experimental results demonstrate the effectiveness of each component in our model. This work highlights the feasibility and potential of fisheye-based 360° gaze estimation in practical multi-person scenarios. Project page: https://caizhuojiang.github.io/GazeOnce360/.

We present BEV-SLD, a LiDAR global localization method building on the Scene Landmark Detection (SLD) concept. Unlike scene-agnostic pipelines, our self-supervised approach leverages bird's-eye-view (BEV) images to discover scene-specific patterns at a prescribed spatial density and treat them as landmarks. A consistency loss aligns learnable global landmark coordinates with per-frame heatmaps, yielding consistent landmark detections across the scene. Across campus, industrial, and forest environments, BEV-SLD delivers robust localization and achieves strong performance compared to state-of-the-art methods.

Reinforcement Learning (RL) has shown promise in various robotics applications, yet its deployment on real systems is still limited due to safety and operational constraints. The safe RL field has gained considerable attention in recent years, which focuses on imposing safety constraints throughout the learning process. However, real systems often require more complex constraints than just safety, such as periodic recharging or time-bounded visits to specific regions. Imposing such spatio-temporal tasks during learning still remains a challenge. Signal Temporal Logic (STL) is a formal language for specifying temporal properties of real-valued signals and provides a way to express such complex tasks. In this paper, we propose a framework that leverages sequential control barrier functions and model-free RL to ensure that the given STL tasks are satisfied throughout the learning process. Our method extends beyond traditional safety constraints by enforcing rich STL specifications, which can involve visits to dynamic targets with unknown trajectories. We also demonstrate the effectiveness of our framework through various simulations.

Personalized fall detection models can significantly improve accuracy by adapting to individual motion patterns, yet their effectiveness is often limited by the scarcity of real-world fall data and the dominance of non-fall feedback samples. This imbalance biases the model toward routine activities and weakens its sensitivity to true fall events. To address this challenge, we propose a personalization framework that combines semi-supervised clustering with contrastive learning to identify and balance the most informative user feedback samples. The framework is evaluated under three retraining strategies, including Training from Scratch (TFS), Transfer Learning (TL), and Few-Shot Learning (FSL), to assess adaptability across learning paradigms. Real-time experiments with ten participants show that the TFS approach achieves the highest performance, with up to a 25% improvement over the baseline, while FSL achieves the second-highest performance with a 7% improvement, demonstrating the effectiveness of selective personalization for real-world deployment.

Decision-making problems often feature uncertainty stemming from heterogeneous and context-dependent human preferences. To address this, we propose a sequential learning-and-optimization pipeline to learn preference distributions and leverage them to solve downstream problems, for example risk-averse formulations. We focus on human choice settings that can be formulated as (integer) linear programs. In such settings, existing inverse optimization and choice modelling methods infer preferences from observed choices but typically produce point estimates or fail to capture contextual shifts, making them unsuitable for risk-averse decision-making. Using a bounded-variance score function gradient estimator, we train a predictive model mapping contextual features to a rich class of parameterizable distributions. This approach yields a maximum likelihood estimate. The model generates scenarios for unseen contexts in the subsequent optimization phase. In a synthetic ridesharing environment, our approach reduces average post-decision surprise by up to 114$\times$ compared to a risk-neutral approach with perfect predictions and up to 25$\times$ compared to leading risk-averse baselines.

Creating high-fidelity, animatable 3D dog avatars remains a formidable challenge in computer vision. Unlike human digital doubles, animal reconstruction faces a critical shortage of large-scale, annotated datasets for specialized applications. Furthermore, the immense morphological diversity across species, breeds, and crosses, which varies significantly in size, proportions, and features, complicates the generalization of existing models. Current reconstruction methods often struggle to capture realistic fur textures. Additionally, ensuring these avatars are fully editable and capable of performing complex, naturalistic movements typically necessitates labor-intensive manual mesh manipulation and expert rigging. This paper introduces SMAL-pets, a comprehensive framework that generates high-quality, editable animal avatars from a single input image. Our approach bridges the gap between reconstruction and generative modeling by leveraging a hybrid architecture. Our method integrates 3D Gaussian Splatting with the SMAL parametric model to provide a representation that is both visually high-fidelity and anatomically grounded. We introduce a multimodal editing suite that enables users to refine the avatar's appearance and execute complex animations through direct textual prompts. By allowing users to control both the aesthetic and behavioral aspects of the model via natural language, SMAL-pets provides a flexible, robust tool for animation and virtual reality.

Many wireless vision applications, such as autonomous driving, require preservation of global structural information rather than only per-pixel fidelity. However, existing Deep joint source-channel coding (DeepJSCC) schemes mainly optimize pixel-wise losses and provide no explicit protection of connectivity or topology. This letter proposes TopoJSCC, a topology-aware DeepJSCC framework that integrates persistent-homology regularizers to end-to-end training. Specifically, we enforce topological consistency by penalizing Wasserstein distances between cubical persistence diagrams of original and reconstructed images, and between Vietoris--Rips persistence of latent features before and after the channel to promote a robust latent manifold. TopoJSCC is based on end-to-end learning and requires no side information. Experiments show improved topology preservation and peak signal-to-noise ratio (PSNR) in low signal-to-noise ratio (SNR) and bandwidth-ratio regimes.

Video diffusion models are moving beyond short, plausible clips toward world simulators that must remain consistent under camera motion, revisits, and intervention. Yet spatial memory remains a key bottleneck: explicit 3D structures can improve reprojection-based consistency but struggle to depict moving objects, while implicit memory often produces inaccurate camera motion even with correct poses. We propose Mosaic Memory (MosaicMem), a hybrid spatial memory that lifts patches into 3D for reliable localization and targeted retrieval, while exploiting the model's native conditioning to preserve prompt-following generation. MosaicMem composes spatially aligned patches in the queried view via a patch-and-compose interface, preserving what should persist while allowing the model to inpaint what should evolve. With PRoPE camera conditioning and two new memory alignment methods, experiments show improved pose adherence compared to implicit memory and stronger dynamic modeling than explicit baselines. MosaicMem further enables minute-level navigation, memory-based scene editing, and autoregressive rollout.

Ensembling Vision-Language Models (VLMs) from different providers maximizes benchmark accuracy, yet models from the same architectural family share correlated errors that standard voting ignores. We study this structure across 17 VLMs from 8 families on VQAv2, TextVQA, and GQA. Family-correlated errors reduce effective ensemble dimensionality to 2.5-3.6 independent voters and create a Misleading tier (1.5-6.5% of questions) where correlated majority errors destroy accuracy to 0% despite the best model being correct. We propose three family-aware methods. Hierarchical Family Voting (HFV) aggregates within families before voting across them, recovering +18-26 pp on the Misleading tier. QualRCCV, a training-free method weighting models by calibration, family quality, and inverse family size, is the first to beat calibrated voting on all three benchmarks (p<0.05). Learned Candidate Scoring (LCS) trains a cross-validated classifier to re-rank candidate answers using support breadth, family diversity, and model quality, achieving the largest gains: +0.68% VQAv2, +0.61% TextVQA, +2.45% GQA -- all significant -- and is the only learned method that never degrades any benchmark. On VQAv2 test-standard (EvalAI), LCS reaches 87.83% with 12 models, confirming generalization.

Image segmentation relies on large annotated datasets, which are expensive and slow to produce. Silver-standard (AI-generated) labels are easier to obtain, but they risk introducing bias. Self-supervised learning, needing only images, has become key for pre-training. Recent work combining contrastive learning with counterfactual generation improves representation learning for classification but does not readily extend to pixel-level tasks. We propose a pipeline combining counterfactual generation with dense contrastive learning via Dual-View (DVD-CL) and Multi-View (MVD-CL) methods, along with supervised variants that utilize available silver-standard annotations. A new visualisation algorithm, the Color-coded High Resolution Overlay map (CHRO-map) is also introduced. Experiments show annotation-free DVD-CL outperforms other dense contrastive learning methods, while supervised variants using silver-standard labels outperform training on the silver-standard labeled data directly, achieving $\sim$94% DSC on challenging data. These results highlight that pixel-level contrastive learning, enhanced by counterfactuals and silver-standard annotations, improves robustness to acquisition and pathological variations.

Decoding brain activity into natural language is a major challenge in AI with important applications in assistive communication, neurotechnology, and human-computer interaction. Most existing Brain-Computer Interface (BCI) approaches rely on memory-intensive fine-tuning of Large Language Models (LLMs) or encoder-decoder models on raw EEG signals, resulting in expensive training pipelines, limited accessibility, and potential exposure of sensitive neural data. We introduce SENSE (SEmantic Neural Sparse Extraction), a lightweight and privacy-preserving framework that translates non-invasive electroencephalography (EEG) into text without LLM fine-tuning. SENSE decouples decoding into two stages: on-device semantic retrieval and prompt-based language generation. EEG signals are locally mapped to a discrete textual space to extract a non-sensitive Bag-of-Words (BoW), which conditions an off-the-shelf LLM to synthesize fluent text in a zero-shot manner. The EEG-to-keyword module contains only ~6M parameters and runs fully on-device, ensuring raw neural signals remain local while only abstract semantic cues interact with language models. Evaluated on a 128-channel EEG dataset across six subjects, SENSE matches or surpasses the generative quality of fully fine-tuned baselines such as Thought2Text while substantially reducing computational overhead. By localizing neural decoding and sharing only derived textual cues, SENSE provides a scalable and privacy-aware retrieval-augmented architecture for next-generation BCIs.

Modern LLMs continue to exhibit significant variance in behavior across languages, such as being able to recall factual information in some languages but not others. While typically studied as a problem to be mitigated, in this work, we propose leveraging this cross-lingual inconsistency as a tool for interpretability in mixture-of-experts (MoE) LLMs. Our knowledge localization framework contrasts routing for sets of languages where the model correctly recalls information from languages where it fails. This allows us to isolate model components that play a functional role in answering about a piece of knowledge. Our method proceeds in two stages: (1) querying the model with difficult factual questions across a diverse set of languages to generate "success" and "failure" activation buckets and then (2) applying a statistical contrastive analysis to the MoE router logits to identify experts important for knowledge. To validate the necessity of this small number of experts for answering a knowledge question, we deactivate them and re-ask the question. We find that despite only deactivating about 20 out of 6000 experts, the model no longer answers correctly in over 40% of cases. Generally, this method provides a realistic and scalable knowledge localization approach to address increasingly complex LLMs.

The convolutional neural networks (CNNs) are not inherently shift invariant or equivariant. The downsampling operation, used in CNNs, is one of the key reasons which breaks the shift invariant property of a CNN. Conversely, downsampling operation is important to improve computational efficiency and increase the area of the receptive field for more contextual information. In this work, we propose Gaussian-Hermite Sampling (GHS), a novel downsampling strategy designed to achieve accurate shift invariance. GHS leverages Gaussian-Hermite polynomials to perform shift-consistent sampling, enabling CNN layers to maintain invariance to arbitrary spatial shifts prior to training. When integrated into standard CNN architectures, the proposed method embeds shift invariance directly at the layer level without requiring architectural modifications or additional training procedures. We evaluate the proposed approach on CIFAR-10, CIFAR-100, and MNIST-rot datasets. Experimental results demonstrate that GHS significantly improves shift consistency, achieving 100% classification consistency under spatial shifts, while also improving classification accuracy compared to baseline CNN models.

Simulating human conversations using large language models (LLMs) has emerged as a scalable methodology for modeling human social interaction. However, simulating human conversations is challenging because they inherently involve inconsistent and uncollaborative behaviors, such as misunderstandings and interruptions. Analysis comparing inconsistent and uncollaborative behaviors in human- and LLM-generated conversations remains limited, although reproducing these behaviors is integral to simulating human-like and complex social interaction. In this work, we introduce CoCoEval, an evaluation framework that analyzes LLM-simulated conversations by detecting 10 types of inconsistent and uncollaborative behaviors at the turn level using an LLM-as-a-Judge. Using CoCoEval, we evaluate GPT-4.1, GPT-5.1, and Claude Opus 4 by comparing the frequencies of detected behaviors in conversations simulated by each model and in human conversations across academic, business, and governmental meetings, as well as debates. Our analysis shows that (1) under vanilla prompting, LLM-simulated conversations exhibit far fewer inconsistent and uncollaborative behaviors than human conversations; (2) prompt engineering does not provide reliable control over these behaviors, as our results show that different prompts lead to their under- or overproduction; and (3) supervised fine-tuning on human conversations can lead LLMs to overproduce a narrow set of behaviors, such as repetition. Our findings highlight the difficulty of simulating human conversations, raising concerns about the use of LLMs as a proxy for human social interaction.

Sim-to-real transfer of locomotion policies often leads to performance degradation due to the inevitable sim-to-real gap. Naively fine-tuning these policies directly on hardware is problematic, as it poses risks of mechanical failure and suffers from high sample inefficiency. In this paper, we address the challenge of safely and efficiently fine-tuning reinforcement learning (RL) policies for dynamic locomotion tasks. Specifically, we focus on fine-tuning policies learned in simulation directly on hardware, while explicitly enforcing safety constraints. In doing so, we introduce SLowRL, a framework that combines Low-Rank Adaptation (LoRA) with training-time safety enforcement via a recovery policy. We evaluate our method both in simulation and on a real Unitree Go2 quadruped robot for jump and trot tasks. Experimental results show that our method achieves a $46.5\%$ reduction in fine-tuning time and near-zero safety violations compared to standard proximal policy optimization (PPO) baselines. Notably, we find that a rank-1 adaptation alone is sufficient to recover pre-trained performance in the real world, while maintaining stable and safe real-world fine-tuning. These results demonstrate the practicality of safe, efficient fine-tuning for dynamic real-world robotic applications.

We present an ensemble-driven self-training framework for unsupervised neural machine translation (UNMT). Starting from a primary language pair, we train multiple UNMT models that share the same translation task but differ in an auxiliary language, inducing structured diversity across models. We then generate pseudo-translations for the primary pair using token-level ensemble decoding, averaging model predictions in both directions. These ensemble outputs are used as synthetic parallel data to further train each model, allowing the models to improve via shared supervision. At deployment time, we select a single model by validation performance, preserving single-model inference cost. Experiments show statistically significant improvements over single-model UNMT baselines, with mean gains of 1.7 chrF when translating from English and 0.67 chrF when translating into English.

The success of CLIP-like vision-language models (VLMs) on natural images has inspired medical counterparts, yet existing approaches largely fall into two extremes: specialist models trained on single-domain data, which capture domain-specific details but generalize poorly, and generalist medical VLMs trained on multi-domain data, which retain broad semantics but dilute fine-grained diagnostic cues. Bridging this specialization-generalization trade-off remains challenging. To address this problem, we propose ACE-LoRA, a parameter-efficient adaptation framework for generalist medical VLMs that maintains robust zero-shot generalization. ACE-LoRA integrates Low-Rank Adaptation (LoRA) modules into frozen image-text encoders and introduces an Attention-based Context Enhancement Hypergraph Neural Network (ACE-HGNN) module that captures higher-order contextual interactions beyond pairwise similarity to enrich global representations with localized diagnostic cues, addressing a key limitation of prior Parameter-Efficient Fine-Tuning (PEFT) methods that overlook fine-grained details. To further enhance cross-modal alignment, we formulate a label-guided InfoNCE loss to effectively suppress false negatives between semantically related image-text pairs. Despite adding only 0.95M trainable parameters, ACE-LoRA consistently outperforms state-of-the-art medical VLMs and PEFT baselines across zero-shot classification, segmentation, and detection benchmarks spanning multiple domains. Our code is available at https://github.com/icon-lab/ACE-LoRA.

Motivated by auto-proof generation and Valiant's VP vs. VNP conjecture, we study the problem of discovering efficient arithmetic circuits to compute polynomials, using addition and multiplication gates. We formulate this problem as a single-player game, where an RL agent attempts to build the circuit within a fixed number of operations. We implement an AlphaZero-style training loop and compare two approaches: Proximal Policy Optimization with Monte Carlo Tree Search (PPO+MCTS) and Soft Actor-Critic (SAC). SAC achieves the highest success rates on two-variable targets, while PPO+MCTS scales to three variables and demonstrates steady improvement on harder instances. These results suggest that polynomial circuit synthesis is a compact, verifiable setting for studying self-improving search policies.

In this work, we analyze shortcomings in cross-lingual knowledge transfer in large, modern reasoning LLMs. We demonstrate that the perceived gap in knowledge transfer is primarily a script barrier. First, we conduct an observational data analysis on the performance of thinking models on two datasets with local knowledge from around the world, ECLeKTic and MultiLoKo. Our regression analysis shows that script match - not language or family - is the primary predictor of knowledge transfer failure once model capability and question difficulty are accounted for. We further this finding by providing the LLMs with the key entities of the questions in their source language and find that this disproportionately improves cross-script questions. We then posit that these LLMs could be reasoning better at test-time. To evaluate this, we develop a synthetic generation pipeline to design SFT samples to encourage the model to better reason about transliteration ambiguities when trying to fetch parametric knowledge at inference-time. We show that teaching two models to reason better reduces the cross-script transfer gap. As a result, we conclude that there is potential to improve cross-lingual parametric knowledge transfer during post-training.