Social interaction depends on both language and visible social signals, such as facial expressions, posture, gaze, and emotional shifts. Yet existing social-agent benchmarks are largely text-based and rarely test whether multimodal agents can use visual cues to guide interaction. We introduce AgentViSS, a benchmark evaluating visual social intelligence in multimodal social simulation. It contains 240 scenarios, 585 role instances, and 2,340 role-task instances, combining aligned textual-visual evidence, structured role profiles, and four role-level tasks: expression task, characteristic task, interaction regulation task, and interaction outcome task. Evaluating seven recent MLLMs under verbalized-vision and direct-vision reveals a clear gap between local role enactment and interaction management: role-specific expression and conflict handling are near saturation, whereas interaction regulation and visually grounded outcome achievement remain substantially more difficult.

Large vision-language models (LVLMs) remain prone to hallucination. While recent visual Direct Preference Optimization (DPO) methods try to fix this by contrasting positive images with heuristic or offline negatives for the same correct response. We show that such negatives are mostly "trivial": the model already assigns near-zero probability to the correct response, so the resulting gradient is uninformative or even spuriously misdirected. We propose OViP, an on-policy framework that synthesizes negative images on the fly from the model’s own hallucinated responses via a diffusion prior, producing semantically misaligned yet still high-likelihood negatives. Across five hallucination and four general benchmarks, OViP achieves better hallucination mitigation on LLaVA-1.5 series and Qwen2.5-VL-3B, while being 2 times more compute-efficient than SIMA and incurring no degradation on general capabilities.

Typical large vision-language models (LVLMs) apply autoregressive supervision primarily to textual responses, without fully exploiting causal learning over rich visual inputs. As a result, these models often emphasize vision-to-language alignment while potentially overlooking fine-grained visual information. While prior work has explored autoregressive image generation, effectively leveraging autoregressive visual supervision to enhance image understanding remains an open challenge. In this paper, we introduce Autoregressive Semantic Visual Reconstruction (ASVR), which enables joint learning of visual and textual modalities within a unified autoregressive framework. ASVR trains models to autoregressively reconstruct the semantic content of input images, which consistently enhances multimodal comprehension. Notably, we show that even when provided with continuous image features as input, models can effectively reconstruct discrete semantic tokens, resulting in stable and consistent improvements across various multimodal understanding benchmarks. ASVR delivers significant performance gains and scalability across varying data scales, visual input, visual supervision and model architectures. In particular, ASVR generally improves baselines by 2-3% across 14 multimodal benchmarks.

Communication protocol design is a central challenge in large language model-based multi-agent systems. Existing single-channel approaches face an inherent communication trilemma: text-based methods are interpretable but verbose, while latent-space methods are efficient but opaque and limited to single-round workflows. Inspired by multi-channel communication theory, we propose HyLaT, a hybrid latent-text communication protocol that transmits elaborate cognitive signals through a latent channel for efficiency, while expressing concise critical signals in natural language to preserve interpretability and precision. We introduce a two-stage training framework combining single-agent hybrid generation learning and multi-agent interactive co-training, enabling agents to generate and interpret hybrid messages across multiple rounds of interaction. Experiments demonstrate that HyLaT reduces communication overhead significantly while maintaining competitive task performance, with strong generalization and robustness across diverse settings.

Evaluating multimodal large language models (MLLMs) is becoming increasingly expensive as benchmarks grow in scale and cross-modality complexity. Inspired by structuralism in cognitive psychology, we tackle this difficulty with an adaptive evaluation framework for efficient benchmarking, namely AutoJudger. Instead of passively scoring on a fixed test set, AutoJudger treats evaluation as an interview-like process by keeping a hypothesized ability structure of the evaluated model and actively selecting the informative questions so as to refine these ability boundaries. Specifically, AutoJudger has three core components: ability decomposition to organize evaluation along meaningful capability dimensions, ability estimation to maintain an up-to-date quantitative profile of the model competence, and adaptive question selection to choose the most informative questions. To operationalize this paradigm, we introduce A²-Judger, a novel MLLM-based Agentic instantiation of AutoJudger equipped with semantic-aware retrieval and dynamic memory. Experiments on four representative multimodal benchmarks show that A²-Judger significantly improves sample efficiency while maintaining reliable evaluation results.

The gap between existing benchmarks and the dynamic nature of real-world legal practice poses a key barrier to advancing legal intelligence. To this end, we introduce J1-ENVS, the first interactive and dynamic legal environment tailored for LLM-based agents. Guided by legal experts, it comprises six representative scenarios from Chinese legal practices at three levels of environmental complexity. We further introduce J1-EVAL, a dual-metric evaluation framework, designed to assess both task performance and procedural compliance across varying levels of legal proficiency. Extensive experiments on 17 LLM agents reveal that while many models demonstrate solid legal knowledge, they struggle with procedural execution in dynamic settings. Even the SOTA model is below 60% overall performance . These findings highlight persistent challenges in achieving dynamic legal intelligence and offer valuable insights to guide future research.

Multimodal Large Language Models (MLLMs) have achieved strong performance on vision-language tasks, yet often fail to preserve and effectively leverage visual evidence throughout generation. We identify three fundamental types of visual grounding failures: Long-Context Grounding Error, where visual information gradually decays over long sequences; Fine-Grained Grounding Error, where low-resolution or degraded inputs hinder the recovery of detailed visual information; and Regional Grounding Error, where spatially diffuse attention weakens region-level vision-language alignment. To address these issues, we propose a tool-augmented reasoning framework with three targeted compensation strategies: reuse, which re-injects the original image to mitigate visual forgetting; focus_area, which constrains attention to task-relevant regions; and zoom_in, which enhances visual resolution for fine-grained perception. We further construct the TWI-Tools-146K dataset and develop Simple-VGC, a tool-augmented MLLM that interleaves visual and textual tokens. Extensive experiments show that each tool yields targeted improvements for its corresponding grounding error, while their combination produces synergistic gains in visual reasoning. Beyond performance, our analysis provides mechanistic insights into how tool-based interventions improve visual grounding, pointing toward more reliable multimodal reasoning.

Interleaved reasoning paradigms enhance Multimodal Large Language Models (MLLMs) with visual feedback but are hindered by the prohibitive computational cost of re-encoding pixel-dense images. A promising alternative, latent visual reasoning, circumvents this bottleneck yet faces limitations: methods either fail to capture intermediate state evolution due to single-step, non-interleaved structures, or sacrifice precise perceptual modeling by over-compressing features. We introduce Interleaved Latent Visual Reasoning (ILVR), a framework that unifies dynamic state evolution with precise perceptual modeling. ILVR interleaves textual generation with latent visual representations that act as specific, evolving cues for subsequent reasoning. Specifically, we employ a self-supervision strategy where a momentum teacher model selectively distills relevant features from ground-truth intermediate images into sparse supervision targets. This adaptive selection mechanism guides the model to autonomously generate context-aware visual signals. Extensive experiments on multimodal reasoning benchmarks demonstrate that ILVR outperforms existing approaches, effectively bridging the gap between fine-grained perception and sequential multimodal reasoning.

Mobile GUI agents powered by large foundation models enable autonomous task execution in applications, but frequent updates that alter UI appearance and reorganize workflows cause agents trained on historical data to fail. Despite these surface changes, we observe that functional semantics and task intents remain fundamentally stable. Building on this insight, we introduce MAGNET, a memory-driven adaptive agent framework with dual-level memory: stationary memory that links diverse visual features to stable functional semantics for robust action grounding and procedural memory that captures stable task intents across varying workflows. Furthermore, we propose a dynamic memory evolution mechanism that continuously refines both memories by prioritizing frequently accessed knowledge. Evaluations on the online benchmark AndroidWorld demonstrate substantial improvements over memory-augmented baselines, while offline benchmarks confirm consistent gains under distribution shifts. These results validate that leveraging stable structures across interface changes improves agent performance and generalization in evolving software environments.

Large Language Models (LLMs) have significantly advanced legal intelligence, but the scarcity of scenario data impedes the progress toward interactive legal scenarios. This paper introduces a Multi-agent Legal Simulation Driver (MASER) to scalably generate synthetic data by simulating interactive legal scenarios. Leveraging real-legal case sources, MASER ensures the consistency of legal attributes between participants and introduces a supervisory mechanism to align participants’ characters and behaviors as well as addressing distractions. A Multi-stage Interactive Legal Evaluation (MILE) benchmark is further constructed to evaluate LLMs’ performance in dynamic legal scenarios. Extensive experiments confirm the effectiveness of our framework.

Jailbreak attacks, where harmful prompts bypass generative models’ built-in safety, raise serious concerns about model vulnerability. While many defense methods have been proposed, the trade-offs between safety and helpfulness, and their application to Large Vision-Language Models (LVLMs), are not well understood. This paper systematically examines jailbreak defenses by reframing the standard generation task as a binary classification problem to assess model refusal tendencies for both harmful and benign queries. We identify two key defense mechanisms: safety shift, which increases refusal rates across all queries, and harmfulness discrimination, which improves the model’s ability to differentiate between harmful and benign inputs. Using these mechanisms, we develop two ensemble defense strategies—inter-mechanism and intra-mechanism ensembles—to balance safety and helpfulness. Experiments on the MM-SafetyBench and MOSSBench datasets with LLaVA-1.5 models show that these strategies effectively improve model safety or optimize the trade-off between safety and helpfulness.

Large Language Models (LLMs) perform well on reasoning benchmarks but often fail when inputs alter slightly, raising concerns about the extent to which their success relies on memorization. This issue is especially acute in Chain-of-Thought (CoT) reasoning, where spurious memorized patterns can trigger intermediate errors that cascade into incorrect final answers. We introduce STIM, a novel framework for Source-aware Token-level Identification of Memorization, which attributes each token in a reasoning chain to one of multiple memorization sources – local, mid-range, or long-range – based on their statistical co-occurrence with the token in the pretraining corpus. Our token-level analysis across tasks and distributional settings reveals that models rely more on memorization in complex or long-tail cases, and that local memorization is often the dominant driver of errors, leading to up to 67% of wrong tokens. We also show that memorization scores from STIM can be effective in predicting the wrong tokens in the wrong reasoning step. STIM offers a powerful tool for diagnosing and improving model reasoning and can generalize to other structured step-wise generation tasks.

Advances in Large Language Models (LLMs) have improved multi-step reasoning by generating free-text rationales, but these models tend to lose focus over the middle of long contexts. This raises concerns that as reasoning progresses, LLMs may overlook information in earlier steps when decoding subsequent steps, leading to unreliable and redundant rationales. To address this, we propose guiding LLMs to generate more accurate and concise rationales by (1) proactively referencing information from underutilized prior steps, and (2) minimizing redundant information between new and existing steps. We introduce stepwise informativeness search, an inference-time tree search framework incorporating two selection heuristics: grounding-guided selection which prioritizes steps paying higher attention over underutilized steps; and novelty-guided selection which encourages steps with novel conclusions. We further utilize a self-grounding strategy that prompts LLMs to explicitly reference relevant prior steps as premises before deduction at each step, mitigating distraction from irrelevant content. Experiments on five reasoning datasets across five LLMs show the effectiveness and efficiency of our approach to improve reasoning with reduced errors and redundancy.

The reward model has become increasingly important in alignment, assessment, and data construction for large language models (LLMs). Most existing researchers focus on enhancing reward models through data improvements, following the conventional training framework for reward models that directly optimizes the predicted rewards.In this paper, we propose a hybrid alignment framework HAF-RM for reward model training by introducing an additional constraint on token-level policy probabilities in addition to the reward score. It can simultaneously supervise the internal preference model at the token level and optimize the mapping layer of the reward model at the sequence level.Experiment results on five datasets sufficiently show the validity and effectiveness of our proposed hybrid framework for training a high-quality reward model.By decoupling the reward modeling procedure and incorporating hybrid supervision, our HAF-RM framework offers a principled and effective approach to enhancing the performance and alignment of reward models, a critical component in the responsible development of powerful language models. We release our code at https://haf-rm.github.io.

Large Vision-Language Models (LVLMs) typically learn visual capacity through visual instruction tuning, involving updates to both a projector and their LLM backbones. Inspired by the concept of a visual region in the human brain, we investigate the existence of an analogous visual region within LLMs that functions as a cognitive core, and explore the potential of efficient training of LVLMs via selective layers tuning. Using Bunny-Llama-3-8B-V for detailed analysis and other three LVLMs for validation across diverse visual and textual tasks, we find that selectively updating 25% of LLMs layers, when sparsely and uniformly distributed, can preserve nearly 99% of visual performance and maintain or improve textual task results, while effectively reducing training time. Based on this targeted training approach, we further propose a novel visual region-based pruning paradigm, removing non-critical layers outside the visual region, which can achieve minimal performance loss. This study offers an effective and efficient strategy for LVLM training and inference by activating a layer-wise visual region within LLMs, which proves consistently effective across different models.

We introduce ALaRM, the first framework modeling hierarchical rewards in reinforcement learning from human feedback (RLHF), which is designed to enhance the alignment of large language models (LLMs) with human preferences. The framework addresses the limitations of current alignment approaches, which often struggle with the inconsistency and sparsity of human supervision signals, by integrating holistic rewards with aspect-specific rewards. This integration enables more precise and consistent guidance of language models towards desired outcomes, particularly in complex and open text generation tasks. By employing a methodology that filters and combines multiple rewards based on their consistency, the framework provides a reliable mechanism for improving model alignment. We validate our approach through applications in long-form question answering and machine translation tasks, employing gpt-3.5-turbo for pairwise comparisons, and demonstrate improvements over existing baselines. Our work underscores the effectiveness of hierarchical rewards modeling in refining LLM training processes for better human preference alignment. We release our code at https://ALaRM-fdu.github.io.

Large Language Models (LLMs) have shown remarkable reasoning performance but struggle with multi-step deductive reasoning involving a series of rule application steps, especially when rules are presented non-sequentially. Our preliminary analysis shows that while LLMs excel in single-step rule application, their performance drops significantly in multi-step scenarios due to the challenge in rule grounding. It requires anchoring the applicable rule and supporting facts at each step, amidst multiple input rules, facts, and inferred facts. To address this, we propose augmenting LLMs with external working memory and introduce a neurosymbolic framework for rule application. The memory stores facts and rules in both natural language and symbolic forms, enabling precise tracking. Utilizing this memory, our framework iteratively performs symbolic rule grounding and LLM-based rule implementation. The former matches predicates and variables of symbolic rules and facts to ground applicable rules at each step. Experiments indicate our framework’s effectiveness in rule application and its robustness across various steps and settings.

The rapid development of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has exposed vulnerabilities to various adversarial attacks. This paper provides a comprehensive overview of jailbreaking research targeting both LLMs and MLLMs, highlighting recent advancements in evaluation benchmarks, attack techniques and defense strategies. Compared to the more advanced state of unimodal jailbreaking, multimodal domain remains underexplored. We summarize the limitations and potential research directions of multimodal jailbreaking, aiming to inspire future research and further enhance the robustness and security of MLLMs.

To effectively use large language models (LLMs) for real-world queries, it is imperative that they generalize to the long-tail distribution, i.e. rare examples where models exhibit low confidence. In this work, we take the first step towards evaluating LLMs in the long-tail distribution of inferential knowledge. We exemplify long-tail evaluation on the Natural Language Inference task. First, we introduce Logic-Induced-Knowledge-Search (LINK), a systematic long-tail data generation framework, to obtain factually-correct yet long-tail inferential statements. LINK uses variable-wise prompting grounded on symbolic rules to seek low-confidence statements while ensuring factual correctness. We then use LINK to curate Logic-Induced-Long-Tail (LINT), a large-scale long-tail inferential knowledge dataset that contains 108K statements spanning four domains. We evaluate popular LLMs on LINT; we find that state-of-the-art LLMs show significant performance drop (21% relative drop for GPT4) on long-tail data as compared to on head distribution data, and smaller models show even more generalization weakness. These results further underscore the necessity of long-tail evaluation in developing generalizable LLMs.

Logical reasoning over incomplete knowledge graphs to answer complex logical queries is a challenging task. With the emergence of new entities and relations in constantly evolving KGs, inductive logical reasoning over KGs has become a crucial problem. However, previous PLMs-based methods struggle to model the logical structures of complex queries, which limits their ability to generalize within the same structure. In this paper, we propose a structure-modeled textual encoding framework for inductive logical reasoning over KGs. It encodes linearized query structures and entities using pre-trained language models to find answers. For structure modeling of complex queries, we design stepwise instructions that implicitly prompt PLMs on the execution order of geometric operations in each query. We further separately model different geometric operations (i.e., projection, intersection, and union) on the representation space using a pre-trained encoder with additional attention and maxout layers to enhance structured modeling. We conduct experiments on two inductive logical reasoning datasets and three transductive datasets. The results demonstrate the effectiveness of our method on logical reasoning over KGs in both inductive and transductive settings.

Matching model is essential for Image-Text Retrieval framework. Existing research usually train the model with a triplet loss and explore various strategy to retrieve hard negative sentences in the dataset. We argue that current retrieval-based negative sample construction approach is limited in the scale of the dataset thus fail to identify negative sample of high difficulty for every image. We propose our TAiloring neGative Sentences with Discrimination and Correction (TAGS-DC) to generate synthetic sentences automatically as negative samples. TAGS-DC is composed of masking and refilling to generate synthetic negative sentences with higher difficulty. To keep the difficulty during training, we mutually improve the retrieval and generation through parameter sharing. To further utilize fine-grained semantic of mismatch in the negative sentence, we propose two auxiliary tasks, namely word discrimination and word correction to improve the training. In experiments, we verify the effectiveness of our model on MS-COCO and Flickr30K compared with current state-of-the-art models and demonstrates its robustness and faithfulness in the further analysis.

Analytical reasoning is an essential and challenging task that requires a system to analyze a scenario involving a set of particular circumstances and perform reasoning over it to make conclusions. However, current neural models with implicit reasoning ability struggle to solve this task. In this paper, we study the challenge of analytical reasoning of text and collect a new dataset consisting of questions from the Law School Admission Test from 1991 to 2016. We analyze what knowledge understanding and reasoning abilities are required to do well on this task, and present an approach dubbed ARM. It extracts knowledge such as participants and facts from the context. Such knowledge are applied to an inference engine to deduce legitimate solutions for drawing conclusions. In our experiments, we find that ubiquitous pre-trained models struggle to deal with this task as their performance is close to random guess. Results show that ARM outperforms pre-trained models significantly. Moreover, we demonstrate that ARM has better explicit interpretable reasoning ability.

Logical reasoning of text requires identifying critical logical structures in the text and performing inference over them. Existing methods for logical reasoning mainly focus on contextual semantics of text while struggling to explicitly model the logical inference process. In this paper, we not only put forward a logic-driven context extension framework but also propose a logic-driven data augmentation algorithm. The former follows a three-step reasoning paradigm, and each step is respectively to extract logical expressions as elementary reasoning units, symbolically infer the implicit expressions following equivalence laws and extend the context to validate the options. The latter augments literally similar but logically different instances and incorporates contrastive learning to better capture logical information, especially logical negative and conditional relationships. We conduct experiments on two benchmark datasets, ReClor and LogiQA. The results show that our method achieves state-of-the-art performance on both datasets, and even surpasses human performance on the ReClor dataset.