Large Language Models (LLMs) excel when a solution can be planned from the prompt, but many sequential decision-making tasks require actions to depend on observations acquired during interaction. We study two central challenges in this regime: learning reusable interaction policies for history-dependent search, and acting under partial observability when task-relevant state is revealed only through interaction. We instantiate these challenges with two diagnostic benchmarks: Calc-Bench for reusable, history-dependent tool interaction, and partially observed sorting for action selection under hidden state. We introduce the Expanded Action space (ExpA), which decouples language reasoning from environment interaction by adding routing and environment-specific actions beyond vocabulary tokens. ExpA is trained with EARL, an RL algorithm using counterfactual policy optimization to explore rarely invoked but useful actions. EARL improves Countdown by 26.3% with 3B models and achieves perfect Sort-4 accuracy while discovering an efficient sorting strategy, providing a proof-of-concept for learned non-language interaction in controlled sequential decision-making diagnostics.

Vision-Language Models (VLMs) show immense promise in medicine, yet surgical scene understanding remains formidable due to complex visual obstructions and rigorous multi-step clinical logic. Standard Supervised Fine-Tuning (SFT) imparts domain knowledge, but we identify a critical flaw—the "Reasoning Collapse" phenomenon: during answer-only SFT, models progressively lose explicit step-by-step reasoning capabilities, regressing into "black boxes" that generate uninterpretable shortcut answers—a severe risk for clinical traceability. Constructing high-quality Chain-of-Thought (CoT) datasets to mitigate this is prohibitively expensive in the surgical domain. We propose SurgMind-R1, a surgical VLM with a three-stage training pipeline. We construct Cholec-Think, a reasoning-enhanced dataset combining expert and VLM-assisted CoT annotations; apply domain-adaptive SFT and reasoning-enhanced SFT on Cholec-Think to internalize surgical logic; and introduce Reinforcement Learning via Group Relative Policy Optimization (GRPO) with rule-based rewards to enforce structured thinking paths and eliminate reasoning collapse. On SurgLite-Bench, our comprehensive benchmark, SurgMind-R1 significantly outperforms state-of-the-art VLMs on complex inference tasks. It internalizes structured thinking, maintaining high accuracy in a direct no-think mode that resolves the high-latency bottleneck of reasoning models for real-time clinical applications, and exhibits strong zero-shot out-of-distribution (OOD) generalization across unseen surgical scenarios and novel tasks, paving the way for transparent, trustworthy computer-assisted interventions. The code, benchmark, and models will be available at https://github.com/YifanLuo0103/SurgMind-R1.

Vision-Language-Action (VLA) models have shown promising progress in language-conditioned robotic manipulation. However, existing datasets and benchmarks mainly evaluate task completion under predefined settings, offering limited insight into model reasoning under increasing spatial and procedural complexity. We introduce RoboSPA (Robot Spatial-Procedural Assessment), a large-scale robotic manipulation dataset and benchmark for diagnosing embodied reasoning in VLA models. RoboSPA focuses on two core dimensions, Fine-Grained Spatial Reasoning and Long-Horizon Procedural Planning, covering 10 task categories and 56 base tasks. Each task is instantiated across five difficulty levels, yielding 280 variants with increasing spatial ambiguity and procedural complexity. We collect 527K trajectories across multiple embodiments and diverse scenes. Beyond binary success rate, RoboSPA introduces diagnostic metrics for more detailed evaluation. Experiments on representative VLA models show that current systems still struggle with complex spatial relations, precise low-level execution, and memory-intensive planning. These results establish RoboSPA as a challenging diagnostic benchmark for developing more capable, reliable, and generalizable embodied agents. Our data and code are available at https://github.com/fanzhenxuan/RoboSPA.

Recent studies integrate Low-Rank Adaptation (LoRA) and Mixture-of-Experts (MoE) to further enhance the performance of parameter-efficient fine-tuning (PEFT) methods in Large Language Model (LLM) applications. Existing methods employ homogeneous MoE-LoRA architectures composed of LoRA experts with either similar or identical structures and capacities. However, these approaches often suffer from representation collapse and expert load imbalance, which negatively impact the potential of LLMs. To address these challenges, we propose a heterogeneous Mixture-of-Adapters (MoA) approach. This method dynamically integrates PEFT adapter experts with diverse structures, leveraging their complementary representational capabilities to foster expert specialization, thereby enhancing the effective transfer of pre-trained knowledge to downstream tasks. MoA supports two variants: (i) Soft MoA achieves fine-grained integration by performing a weighted fusion of all expert outputs; (ii) Sparse MoA activates adapter experts sparsely based on their contribution, achieving this with negligible performance degradation. Experimental results demonstrate that heterogeneous MoA outperforms homogeneous MoE-LoRA methods in both performance and parameter efficiency. Our project is available at https://github.com/DCDmllm/MoA.

Recent advances in Multi-modal Large Language Models (MLLMs), such as LLaVA-series models, are driven by massive machine-generated instruction-following data tuning. Such automatic instruction collection pipelines, however, inadvertently introduce significant variability in data quality. This paper introduces a novel instruction curation algorithm, derived from two unique perspectives, human and LLM preference alignment, to compress this vast corpus of machine-generated multimodal instructions to a compact and high-quality form: (i) For human preference alignment, we have collected a machine-generated multimodal instruction dataset and established a comprehensive set of both subjective and objective criteria to guide the data quality assessment critically from human experts. By doing so, a reward model was trained on the annotated dataset to internalize the nuanced human understanding of instruction alignment. (ii) For LLM preference alignment, given the instruction selected by the reward model, we propose leveraging the inner LLM used in MLLM to align the writing style of visual instructions with that of the inner LLM itself, resulting in LLM-aligned instruction improvement. Extensive experiments demonstrate that we can maintain or even improve model performance by compressing synthetic multimodal instructions by up to 90%. Impressively, by aggressively reducing the training instructions from 158k to 14k (9× smaller), our model consistently outperforms its full-size dataset counterpart across various MLLM benchmarks. Our project is available at https://github.com/DCDmllm/Align2LLaVA.

Inspired by human cognitive behavior, we introduce visual modality to enhance the performance of pure text-based question-answering tasks with the development of multimodal models. However, obtaining corresponding images through manual annotation often entails high costs. Faced with this challenge, an intuitive strategy is to use search engines or use web scraping techniques to automatically obtain relevant image information. However, the images obtained by this strategy may be of low quality and may not match the context of the original task, which could fail to improve or even decrease performance on downstream tasks. In this paper, we propose a novel framework named “ITERATE”, aimed at retrieving and optimizing the quality of images to improve the alignment between text and images. Inspired by evolutionary algorithms in reinforcement learning and driven by the synergy of large language models (LLMs) and multimodal models, ITERATE employs a series of strategic actions such as filtering, optimizing, and retrieving to acquire higher quality images, and repeats this process over multiple generations to enhance the quality of the entire image cluster. Our experimental results on the ScienceQA, ARC-Easy, and OpenDataEval datasets also verify the effectiveness of our method, showing improvements of 3.5%, 5%, and 7%, respectively.

While Parameter-Efficient Fine-Tuning (PEFT) methods like Low-Rank Adaptation (LoRA) effectively address resource constraints during fine-tuning, their performance often falls short, especially in multidimensional task scenarios. To address this issue, one straightforward solution is to introduce task-specific LoRA as domain experts, leveraging the modeling of multiple capabilities of experts and thus enhancing the general capability of multi-task learning.Although promising, these additional components often add complexity to the training and inference process, contravening the efficiency that PEFT is designed to deliver. Considering this, we introduce an innovative PEFT method, TeamLoRA, consisting of a collaboration and competition module for LoRA experts, thus achieving the right balance of effectiveness and efficiency:(i) For collaboration, we introduce a novel knowledge sharing and organization mechanism designed to optimize hierarchical learning while enhancing the efficiency of model training and inference.(ii) For competition, we propose leveraging a game-theoretic interaction mechanism for experts, encouraging experts to transfer their domain-specific knowledge while facing diverse downstream tasks, thus enhancing the performance.By doing so, TeamLoRA elegantly connects the experts as a “Team” with internal collaboration and competition, enabling a faster and more accurate PEFT paradigm. Meanwhile, we curate a Comprehensive Multi-Task Evaluation (CME) benchmark to thoroughly assess the capability of multi-task learning. Experiments conducted on our CME and other benchmarks indicate the effectiveness and efficiency of TeamLoRA. Our project is available at https://github.com/DCDmllm/TeamLoRA.

Prompt tuning is a parameter-efficient method, which learns soft prompts and conditions frozen language models to perform specific downstream tasks. Though effective, prompt tuning under few-shot settings on the one hand heavily relies on a good initialization of soft prompts. On the other hand, it can easily overfit to few-shot training samples, thereby undermining generalizability. Existing works leverage pre-training or supervised meta-learning to initialize soft prompts but they fail to data-efficiently generalize to unseen downstream tasks. To address the above problems, this paper proposes a novel Self-sUpervised meta-Prompt learning framework with MEta-gradient Regularization for few-shot generalization (SUPMER). SUPMER leverages self-supervised meta-learning with a diverse set of well-designed meta-tasks to learn a universal prompt initialization for efficient adaptation using only unlabeled data. Additionally, it jointly meta-learns a gradient regularization function to transform raw gradients into a domain-generalizable direction, thus alleviating the problem of overfitting. Extensive experiments show that SUPMER achieves better performance for different few-shot downstream tasks, and also exhibits a stronger domain generalization ability. The code for SUPMER will be available at https://github.com/beepkh/SUPMER.

Relation extraction is often challenged by insufficient labeled data. Previous methods exploit knowledge from unlabeled data by generating pseudo labels in a self-training pipeline, which suffers a gradual drift problem. Logic rules, a transferable and explainable form of expert knowledge, have achieved promising success by improving the model with weak labels. But manually writing comprehensive rules set is challenging and tedious. To alleviate the human labor of writing high-quality rules, in this work, we propose ARIA, an Automatic task-specific Rules distilling framework. Specifically, we guide the pre-trained language model to reason rules as experts and compose them into robust compound rules for data labeling. Besides, ARIA could continuously enrich the rules set to power the labeling ability by discovering reliable model-labeled data for distinguishable rules generation. Experiments on two public datasets demonstrate the effectiveness of ARIA in a low-resource scenario.

Visual Relation Extraction (VRE) is a powerful means of discovering relationships between entities within visually-rich documents. Existing methods often focus on manipulating entity features to find pairwise relations, yet neglect the more fundamental structural information that links disparate entity pairs together. The absence of global structure information may make the model struggle to learn long-range relations and easily predict conflicted results. To alleviate such limitations, we propose a GlObal Structure knowledge-guided relation Extraction (GOSE) framework. GOSE initiates by generating preliminary relation predictions on entity pairs extracted from a scanned image of the document. Subsequently, global structural knowledge is captured from the preceding iterative predictions, which are then incorporated into the representations of the entities. This “generate-capture-incorporate” cycle is repeated multiple times, allowing entity representations and global structure knowledge to be mutually reinforced. Extensive experiments validate that GOSE not only outperforms existing methods in the standard fine-tuning setting but also reveals superior cross-lingual learning capabilities; indeed, even yields stronger data-efficient performance in the low-resource setting.