Long-term memory is a critical capability for multimodal large language model (MLLM) agents, particularly in conversational settings where information accumulates and evolves over time. However, existing benchmarks either evaluate multi-session memory in text-only conversations or assess multimodal understanding within localized contexts, failing to evaluate how multimodal memory is preserved, organized, and evolved across long-term conversational trajectories. Thus, we introduce Mem-Gallery, a new benchmark for evaluating multimodal long-term conversational memory in MLLM agents. Mem-Gallery features high-quality multi-session conversations grounded in both visual and textual information, with long interaction horizons and rich multimodal dependencies. Building on this dataset, we propose a systematic evaluation framework that assesses key memory capabilities along three functional dimensions: memory extraction and test-time adaptation, memory reasoning, and memory knowledge management. Extensive benchmarking across twelve memory systems reveals several key findings, highlighting the necessity of explicit multimodal information retention and memory organization, the persistent limitations in memory reasoning and knowledge management, as well as the efficiency bottleneck of current models. Our benchmark and dataset are available at https://github.com/YuanchenBei/Mem-Gallery.

Retrieval-Augmented Generation (RAG) enhances Large Language Models (LLMs) by integrating them with an external knowledge base to improve the answer relevance and accuracy. In real-world scenarios, beyond pure text, a substantial amount of knowledge is stored in tables, and user questions often require retrieving answers that are distributed across multiple tables. Retrieving knowledge from a table corpora (i.e., various individual tables) for a question remains nascent, for (i) how to understand intra- and inter-table knowledge effectively, (ii) how to filter unnecessary tables and retrieve the most relevant tables efficiently, (iii) how to organize complex retrieved contexts for LLMs’ reasoning, and (iv) how to evaluate the corresponding performance in a realistic setting. Facing the above challenges, in this paper, we first propose a table-corpora-aware RAG framework, named T-RAG, which consists of the hierarchical memory index, multi-stage retrieval, and graph-aware context organization for effective and efficient table knowledge retrieval and inference. Then, we develop a multi-table question answering benchmark named MultiTableQA, which spans 3 different task types, 57,193 tables, and 23,758 questions in total, and the sources are all from real-world scenarios. Based on MultiTableQA, we perform a comprehensive comparison of table retrieval methods, RAG-based approaches, and table-to-graph representation learning methods. T-RAG consistently achieves state-of-the-art accuracy, recall, and runtime performance, with improvements of up to 9.4%. Moreover, T-RAG yields an average inference gain of 11.8% across different downstream backbone LLMs. Our code and data are available at https://github.com/jiaruzouu/T-RAG.

Large language models (LLMs) exhibit complementary strengths arising from differences in pretraining data, model architectures, and decoding behaviors. Inference-time ensembling provides a practical way to combine these capabilities without retraining. However, existing ensemble approaches suffer from fundamental limitations. Most rely on fixed fusion granularity, which lacks the flexibility required for mid-generation adaptation and fails to adapt to different generation characteristics across tasks. To address these challenges, we propose AdaFuse, an adaptive ensemble decoding framework that dynamically selects semantically appropriate fusion units during generation. Rather than committing to a fixed granularity, AdaFuse adjusts fusion behavior on the fly based on the decoding context, with words serving as basic building blocks for alignment. To be specific, we introduce an uncertainty-based criterion to decide whether to apply ensembling at each decoding step. Under confident decoding states, the model continues generation directly. In less certain states, AdaFuse invokes a diversity-aware scaling strategy to explore alternative candidate continuations and inform ensemble decisions. This design establishes a synergistic interaction between adaptive ensembling and test-time scaling, where ensemble decisions guide targeted exploration, and the resulting diversity in turn strengthens ensemble quality. Experiments on open-domain QA, arithmetic reasoning, and machine translation demonstrate that AdaFuse consistently outperforms strong ensemble baselines, achieving an average relative improvement of 6.88%.

Different large language models (LLMs) exhibit diverse strengths and weaknesses, and LLM ensemble serves as a promising approach to integrate their complementary capabilities. Despite substantial progress in improving ensemble quality, limited attention has been paid to the robustness of ensembles against potential erroneous signals, which often arise from heterogeneous tokenization schemes and varying model expertise. Our analysis shows that ensemble failures typically arise from both the token level and the model level: the former reflects severe disagreement in token predictions, while the latter involves low confidence and pronounced disparities among models. In light of this, we propose CoRE, a plug-and-play technique that harnesses model consistency for robust LLM ensemble, which can be seamlessly integrated with diverse ensemble methods. *Token-level consistency* captures fine-grained disagreements by applying a low-pass filter to downweight uncertain tokens with high inconsistency, often due to token misalignment, thereby improving robustness at a granular level. *Model-level consistency* models global agreement by promoting model outputs with high self-confidence and minimal divergence from others, enhancing robustness at a coarser level. Extensive experiments across diverse benchmarks, model combinations, and ensemble strategies demonstrate that CoRE consistently improves ensemble performance and robustness. Our code is available at https://github.com/zhichenz98/CoRE-EACL26.

Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs). However, methods such as GRPO and DAPO suffer from substantial computational cost, since they rely on sampling many rollouts for each prompt. Moreover, in RLVR the relative advantage is often sparse: many samples become nearly all-correct or all-incorrect, yielding low within-group reward variance and thus weak learning signals. In this paper, we introduce ARRoL (**A**ccelerating **R**LV**R** via **o**nline Ro**L**lout Pruning), an online rollout pruning method that prunes rollouts during generation while explicitly steering the surviving ones more correctness-balanced to enhance learning signals. Specifically, ARRoL trains a lightweight quality head on-the-fly to predict the success probability of partial rollouts and uses it to make early pruning decisions. The learned quality head can further weigh candidates to improve inference accuracy during test-time voting. To improve efficiency, we present a system design that prunes rollouts inside the inference engine and re-batches the remaining ones for log-probability computation and policy updates. Across GRPO and DAPO on Qwen-3 and LLaMA-3.2 models (1B-8B), ARRoL improves average accuracy by +2.30 to +2.99 while achieving up to 1.7× training speedup, and yielding up to +8.33 additional gains in average accuracy in test-time voting.

Understanding scientific papers requires more than answering isolated questions or summarizing content. It involves an integrated reasoning process that grounds textual and visual information, interprets experimental evidence, synthesizes information across sources, and critically evaluates scientific claims. However, existing benchmarks typically assess these abilities in isolation, making it difficult to evaluate scientific paper understanding as a unified set of interacting cognitive abilities. In this work, we introduce PaperMind , a benchmark designed to evaluate integrated and agent-oriented scientific reasoning over research papers. PaperMind is constructed from real scientific papers across seven domains, including agriculture, biology, chemistry, computer science, medicine, physics, and economics. It comprises four complementary task families that collectively operationalize distinct cognitive facets of scientific paper reasoning, including multimodal grounding, experimental interpretation, cross-source evidence reasoning, and critical assessment. By analyzing model behavior across multiple tasks, PaperMind enables a diagnostic evaluation of integrated scientific reasoning behaviors that are difficult to assess through isolated task evaluations. Extensive experiments on both open-source and closed-source multimodal LLMs reveal consistent performance gaps across tasks, highlighting persistent challenges in integrated scientific reasoning and critique. Our benchmark and dataset are available at https://github.com/Yanjun-Zhao/PaperMind.

The performance of Large Language Models (LLMs) critically depends on designing effective instructions, which is particularly challenging for black-box LLMs with inaccessible internal states. To this end, we introduce Learning to Instruct, a novel paradigm that formulates instruction optimization as an LLM fine-tuning objective for a white-box “instruction engineer” LLM, leveraging its rich learning capacity and vast pre-trained knowledge to enable efficient and effective instruction optimization. Within this paradigm, we propose Automatic Instruction Optimizer (AIO), a novel framework that fine-tunes a white-box LLM into a capable instruction engineer. AIO learns to optimize task-aware, human-comprehensible instructions by incorporating task nuances and feedback from the task-solving black-box LLM. To overcome the challenges of inaccessible black-box gradients and high API costs, AIO introduces a novel zeroth-order (ZO) gradient approximation mechanism guided by Thompson Sampling (TS), which reuses informative black-box LLM feedback for improved query efficiency. Extensive experiments show that AIO generally outperforms strong baselines in both effectiveness and efficiency, establishing Learning to Instruct as a promising new direction for black-box LLM instruction optimization.

The advancement of Large Language Models (LLMs) has made ensuring their trustworthiness increasingly critical, especially in terms of fairness across diverse human groups. While modern LLMs are aligned with user preferences through Reinforcement Learning from Human Feedback (RLHF), the reward models used for alignment are trained on preference data that may both reflect societal biases and suffer from demographic skewness, as labeler populations are often uneven due to systemic accessibility or participation gaps. In this work, we reveal that reward models can exhibit significant discrepancies across different demographic groups, posing a fundamental challenge to fair and robust alignment. Using real-world datasets, we conduct the most comprehensive study to date, auditing various state-of-the-art reward models across nine sensitive attributes, including age, gender, ethnicity, etc. Our evaluation spans both (1) the agreement level between reward models and specific user groups, and (2) the reward model’s preference toward responses associated with different groups. Based on these findings, we propose the first method to mitigate group disparities in reward modeling. Code is available at https://github.com/Violet24K/FaRM.

While great success has been achieved in building vision models with Contrastive Language-Image Pre-training (CLIP) over Internet-scale image-text pairs, building transferable Graph Neural Networks (GNNs) with CLIP pipeline is challenging because of the scarcity of labeled data and text supervision, different levels of downstream tasks, and the conceptual gaps between domains. In this work, to address these issues, we propose a multi-modal prompt learning paradigm to effectively adapt pre-trained GNN to downstream tasks and data, given only a few semantically labeled samples, each with extremely weak text supervision. Our new paradigm embeds the graphs directly in the same space as the Large Language Models (LLMs) by learning both graph prompts and text prompts simultaneously. We demonstrate the superior performance of our paradigm in few-shot, multi-task-level, and cross-domain settings. Moreover, we build the first CLIP-style zero-shot classification prototype that can generalize GNNs to unseen classes with extremely weak text supervision.

Missing data imputation is a critical challenge in various domains, such as healthcare and finance, where data completeness is vital for accurate analysis. Large language models (LLMs), trained on vast corpora, have shown strong potential in data generation, making them a promising tool for data imputation. However, challenges persist in designing effective prompts for a finetuning-free process and in mitigating biases and uncertainty in LLM outputs. To address these issues, we propose a novel framework, LLM-Forest, which introduces a “forest” of few-shot learning LLM “trees” with their outputs aggregated via confidence-based weighted voting based on LLM self-assessment, inspired by the ensemble learning (Random Forest). This framework is established on a new concept of bipartite information graphs to identify high-quality relevant neighboring entries with both feature and value granularity. Extensive experiments on 9 real-world datasets demonstrate the effectiveness and efficiency of LLM-Forest. The implementation is available at https://github.com/Xinrui17/LLM-Forest