Figure 1: Semantically complex language instructions, such as those involving the relative positions of objects, pose a difficult challenge for vision-language-action models (VLAs). To address this problem, we propose IA-VLA, a framework for augmenting the input to VLAs, that offloads the semantic understanding to a larger vision language model (VLM) with greater semantic understanding. We use semantic segmentation to label image regions which a VLM then uses to identify the masks of the task-relevant objects. The task-relevant objects are highlighted in the VLA input, together with the language instruction which can optionally be simplified. Abstract-- Vision-language-action models (VLAs) have become an increasingly popular approach for addressing robot manipulation problems in recent years. However, such models need to output actions at a rate suitable for robot control, which limits the size of the language model they can be based on, and consequently, their language understanding capabilities. Manipulation tasks may require complex language instructions, such as identifying target objects by their relative positions, to specify human intention. Therefore, we introduce IA-VLA, a framework that utilizes the extensive language understanding of a large vision language model as a pre-processing stage to generate improved context to augment the input of a VLA. We evaluate the framework on a set of semantically complex tasks which have been underexplored in VLA literature, namely tasks involving visual duplicates, i.e., visually indistinguishable objects.

Zero-shot object navigation in unknown environments presents significant challenges, mainly due to two key limitations: insufficient semantic guidance leads to inefficient exploration, while limited spatial memory resulting from environmental structure causes entrapment in local regions. To address these issues, we propose SSR-ZSON, a spatial-semantic relative zero-shot object navigation method based on the TARE hierarchical exploration framework, integrating a viewpoint generation strategy balancing spatial coverage and semantic density with an LLM-based global guidance mechanism. The performance improvement of the proposed method is due to two key innovations. First, the viewpoint generation strategy prioritizes areas of high semantic density within traversable sub-regions to maximize spatial coverage and minimize invalid exploration. Second, coupled with an LLM-based global guidance mechanism, it assesses semantic associations to direct navigation toward high-value spaces, preventing local entrapment and ensuring efficient exploration. Deployed on hybrid Habitat-Gazebo simulations and physical platforms, SSR-ZSON achieves real-time operation and superior performance. On Matterport3D and Habitat-Matterport3D datasets, it improves the Success Rate(SR) by 18.5\% and 11.2\%, and the Success weighted by Path Length(SPL) by 0.181 and 0.140, respectively, over state-of-the-art methods.

The quadratic complexity of attention mechanisms limits the efficiency of Large Language Models (LLMs) on long-text tasks. Recently, methods that dynamically estimate block importance have enabled efficient block sparse attention, leading to significant acceleration in long-text pre-filling of LLMs. However, their coarse-grained estimation inevitably leads to performance degradation at high sparsity rates. In this work, we propose ProxyAttn, a training-free sparse attention algorithm that achieves more precise block estimation by compressing the dimension of attention heads. Based on our observation of the similarity among multiple attention heads, we use the scores of pooled representative heads to approximate the scores for all heads. To account for the varying sparsity among heads, we also propose a block-aware dynamic budget estimation method. By combining the scores from representative proxy heads with multi-head dynamic budgets, we achieve a more fine-grained block importance evaluation at low computational cost. Experiments on a variety of mainstream models and extensive benchmarks confirm the underlying similarity among attention heads. Leveraging a fine-grained estimation, the proposed method achieves substantial gains in performance and efficiency compared to existing methods. More precisely, ProxyAttn can achieve up to 10.3x attention acceleration and 2.4x prefilling acceleration without significant performance loss. Our code is available at https://github.com/wyxstriker/ProxyAttn.

Recent breakthroughs in large language models (LLMs) on reasoning tasks rely heavily on massive, high-quality datasets-typically human-annotated and thus difficult to scale. While data synthesis or distillation offers a promising alternative, existing methods struggle with inconsistent data quality and an inability to dynamically adapt to the evolving capabilities of the model, leading to suboptimal training signals. To address these limitations, we introduce Socratic-Zero, a fully autonomous framework that generates high-quality training data from minimal seed examples through the co-evolution of three agents: the Teacher, the Solver, and the Generator. The Solver continuously refines its reasoning by learning from preference feedback on both successful and failed trajectories; the Teacher adaptively crafts increasingly challenging questions based on the Solver's weaknesses; and the Generator distills the Teacher's question-design strategy to enable scalable, high-fidelity curriculum generation. This closed-loop system produces a self-improving curriculum-requiring no pre-existing tasks or labels. Remarkably, starting from only 100 seed questions, our Socratic-Solver-8B achieves an average gain of +20.2 percentage points over prior data synthesis methods across seven mathematical reasoning benchmarks (AMC23, AIME24-25, Olympiad, MATH-500, Minerva, and GSM8K), with consistent gains on both Qwen3 and GLM4 series models. Even more surprisingly, synthetic data from Socratic-Generator-32B enables student LLMs to achieve superior performance compared to other state-of-the-art (SOTA) commercial LLMs on these benchmarks, including Qwen3-235B-A22B, DeepSeek-V3.1-671B, GPT-5, Gemini-2.5-Pro, Grok-4, and Claude-4.1-Opus.

Abstract--Effective human-robot collaboration depends on task-oriented handovers, where robots present objects in ways that support the partner's intended use. T o address this gap, we propose LLM-Handover, a novel framework that integrates large language model (LLM)-based reasoning with part segmentation to enable context-aware grasp selection and execution. Given an RGB-D image and a task description, our system infers relevant object parts and selects grasps that optimize post-handover usability. T o support evaluation, we introduce a new dataset of 60 household objects spanning 12 categories, each annotated with detailed part labels. We first demonstrate that our approach improves the performance of the used state-of-the-art part segmentation method, in the context of robot-human handovers. Next, we show that LLM-Handover achieves higher grasp success rates and adapts better to post-handover task constraints. During hardware experiments, we achieve a success rate of 83% in a zero-shot setting over conventional and unconventional post-handover tasks. Finally, our user study underlines that our method enables more intuitive, context-aware handovers, with participants preferring it in 86% of cases. S robots become more common in everyday settings, their ability to collaborate with humans on joint tasks becomes increasingly important. Recent research in human-robot interaction explores these challenges, with object handovers being a key component for successful collaboration [1]. These handovers form the basis for many joint activities that require both physical coordination and contextual understanding [2]. For instance, observations of human-to-human handovers reveal that people often anticipate each other's intended use of an object by interpreting the surrounding context [2], [3]. This ability, known as task-orientation, becomes especially critical in environments like factories, surgeries, or construction sites, where one partner may have limited mobility or freedom to adjust their actions.

The performance of large language models (LLMs) is highly sensitive to the input prompt, making prompt optimization a critical task. However, real-world application is hindered by three major challenges: (1) the black-box nature of powerful proprietary LLMs, (2) the need for high sample efficiency due to query costs, and (3) the desire for privacy-preserving collaboration among multiple users. To address these challenges simultaneously, we introduce a novel framework for sample-efficient federated prompt optimization based on multi-armed bandits (MABs). The MAB framework is uniquely suited for this problem as it is (1) inherently a black-box optimization method, (2) practically sample-efficient, and (3) enables collaborative learning with theoretically guaranteed benefit from more participating agents. We first propose the Federated Prompt Optimization via Bandits (FedPOB) algorithm, a federated variant of the Linear UCB algorithm, where agents collaborate by sharing model parameters instead of raw data. We then extend our approach to the practical setting of comparative user feedback by introducing FedPOB with Preference Feedback (FedPOB-Pref), an efficient algorithm based on federated dueling bandits. Extensive experiments demonstrate that both FedPOB and FedPOB-Pref significantly outperform existing baselines and that their performance consistently improves as more agents participate in the collaboration, validating the effectiveness of our federated approach. Large language models (LLMs) have achieved impressive performance in a variety of real-world applications (Guo et al., 2025). However, the performance of LLMs has been shown to be highly sensitive to the input prompt (Zhou et al., 2023; Lin et al., 2024b). Consequently, prompt optimization, in which we aim to find the best prompt for a task, has emerged as a critical research area. Despite its growing popularity, the widespread real-world adoption of prompt optimization is still hindered by three important challenges. The first challenge is black-box access. Some of the most powerful LLMs, such as ChatGPT and Gemini (OpenAI, 2023b; Team et al., 2023), are proprietary, black-box models that are only accessible via API queries.

Personalizing large language models (LLMs) to individual user preferences is a critical step beyond generating generically helpful responses. However, current personalization methods are ill-suited for new users, as they typically require either slow, resource-intensive fine-tuning or a substantial amount of pre-existing user data, creating a significant cold-start problem. To address this challenge, we introduce a new paradigm for real-time personalization by learning from online pairwise preference feedback collected during text generation. We propose T-POP (T est-Time P ersonalization with O nline P reference Feedback), a novel algorithm that synergistically combines test-time alignment with dueling bandits. Without updating the LLM parameters, T-POP steers the decoding process of a frozen LLM by learning a reward function online that captures user preferences. By leveraging dueling bandits, T-POP intelligently queries the user to efficiently balance between exploring their preferences and exploiting the learned knowledge to generate personalized text. Extensive experiments demonstrate that T-POP achieves rapid and data-efficient personalization, significantly outperforming existing baselines and showing consistent improvement with more user interactions. While large language models (LLMs) have achieved remarkable success in generating human-like text, a critical frontier remains: moving from generic, one-size-fits-all responses to deeply personalized interactions.

How reliable are single-response LLM-as-a-judge ratings without references, and can we obtain fine-grained, deterministic scores in this setting? We study the common practice of asking a judge model to assign Likert-scale scores to free-text responses and show two systematic issues: scores are unstable under sampling and poorly calibrated, leading to compression near the top of the scale and frequent ties. We then propose and evaluate Latent Judges, which derive scalar ratings from internal model signals: (i) probability-weighted scores over integer ratings, (ii) verifier-style probabilities of "yes", and (iii) linear probes trained on model activations at the rating position. Across a broad suite of pairwise and single-rating benchmarks, latent methods match or surpass standard prompting, with consistent gains on pairwise accuracy and listwise ranking relevant to Best-of-N selection. Probability-weighted scores achieve the strongest single-rating correlations, while probes recover useful signals when output logits are miscalibrated. These results indicate that latent information provides deterministic and more discriminative signals for reference-free evaluation, and can improve selection and training approaches like Best-of-$N$, multi-teacher distillation, and routing.

Unlearning seeks to remove specific knowledge from large language models (LLMs), but its effectiveness remains contested. On one side, "forgotten" knowledge can often be recovered through interventions such as light fine-tuning; on the other side, unlearning may induce catastrophic forgetting that degrades general capabilities. Despite active exploration of unlearning methods, interpretability analyses of the mechanism are scarce due to the difficulty of tracing knowledge in LLMs' complex architectures. We address this gap by proposing unPact, an interpretable framework for unlearning via prompt attribution and contribution tracking. Typically, it quantifies each prompt token's influence on outputs, enabling pre- and post-unlearning comparisons to reveal what changes. Across six mainstream unlearning methods, three LLMs, and three benchmarks, we find that: (1) Unlearning appears to be effective by disrupting focus on keywords in prompt; (2) Much of the knowledge is not truly erased and can be recovered by simply emphasizing these keywords in prompts, without modifying the model's weights; (3) Catastrophic forgetting arises from indiscriminate penalization of all tokens. Taken together, our results suggest an unlearning dilemma: existing methods tend either to be insufficient - knowledge remains recoverable by keyword emphasis, or overly destructive - general performance collapses due to catastrophic forgetting, still leaving a gap to reliable unlearning.

Long-sequence processing is a critical capability for modern large language models. However, the self-attention mechanism in the standard Transformer architecture faces severe computational and memory bottlenecks when processing long sequences. While trainable sparse attention methods offer a promising solution, existing approaches such as NSA introduce excessive extra parameters and disrupt the conventional \textit{pretrain-on-short, finetune-on-long} workflow, resulting in slow convergence and difficulty in acceleration. To overcome these limitations, we introduce dense-sparse switchable attention framework, termed as InfLLM-V2. InfLLM-V2 is a trainable sparse attention that seamlessly adapts models from short to long sequences. Specifically, InfLLM-V2 reuses dense attention parameters through parameter-free architecture modification, maintaining consistency between short and long sequence processing. Additionally, InfLLM-V2 ensures computational efficiency across all sequence lengths, by using dense attention for short inputs and smoothly transitioning to sparse attention for long sequences. To achieve practical acceleration, we further introduce an efficient implementation of InfLLM-V2 that significantly reduces the computational overhead. Our experiments on long-context understanding and chain-of-thought reasoning demonstrate that InfLLM-V2 is 4$\times$ faster than dense attention while retaining 98.1% and 99.7% of the performance, respectively. Based on the InfLLM-V2 framework, we have trained and open-sourced MiniCPM4.1 (https://huggingface.co/openbmb/MiniCPM4.1-8B), a hybrid reasoning model, providing a reproducible implementation for the research community.