Although multimodal large language models (MLLMs) exhibit remarkable reasoning capabilities on complex multimodal understanding tasks, they still suffer from the notorious hallucination issue: generating outputs misaligned with obvious visual or factual evidence. Currently, training-based solutions, like direct preference optimization (DPO), leverage paired preference data to suppress hallucinations. However, they risk sacrificing general reasoning capabilities due to the likelihood displacement. Meanwhile, training-free solutions, like contrastive decoding, achieve this goal by subtracting the estimated hallucination pattern from a distorted input. Yet, these handcrafted perturbations (e.g., add noise to images) may poorly capture authentic hallucination patterns. To avoid these weaknesses of existing methods, and realize robust hallucination mitigation (i.e., maintaining general reasoning performance), we propose a novel framework: Decoupling Contrastive Decoding (DCD). Specifically, DCD decouples the learning of positive and negative samples in preference datasets, and trains separate positive and negative image projections within the MLLM. The negative projection implicitly models real hallucination patterns, which enables vision-aware negative images in the contrastive decoding inference stage. Our DCD alleviates likelihood displacement by avoiding pairwise optimization and generalizes robustly without handcrafted degradation. Extensive ablations across hallucination benchmarks and general reasoning tasks demonstrate the effectiveness of DCD, i.e., it matches DPO's hallucination suppression while preserving general capabilities and outperforms the handcrafted contrastive decoding methods.

Trustworthy AI encompasses many aspirational aspects for aligning AI systems with human values, including fairness, privacy, robustness, explainability, and uncertainty quantification. Ultimately the goal of Trustworthy AI research is to achieve all aspects simultaneously. However, efforts to enhance one aspect often introduce unintended trade-offs that negatively impact others. In this position paper, we review notable approaches to these five aspects and systematically consider every pair, detailing the negative interactions that can arise. For example, applying differential privacy to model training can amplify biases, undermining fairness. Drawing on these findings, we take the position that current research practices of improving one or two aspects in isolation are insufficient. Instead, research on Trustworthy AI must account for interactions between aspects and adopt a holistic view across all relevant axes at once. To illustrate our perspective, we provide guidance on how practitioners can work towards integrated trust, examples of how interactions affect the financial industry, and alternative views.

Chain-of-Thought (CoT) prompting has been widely recognized for its ability to enhance reasoning capabilities in large language models (LLMs). However, our study reveals a surprising contradiction to this prevailing perspective within the fundamental domain of pattern-based in-context learning (ICL). Through extensive experiments involving 16 state-of-the-art LLMs and nine diverse pattern-based ICL datasets, we demonstrate that CoT and its reasoning variants consistently underperform direct answering across varying model scales and benchmark complexities. To systematically investigate this unexpected phenomenon, we designed extensive experiments to validate several hypothetical explanations. Our analysis uncovers a fundamental hybrid mechanism of explicit-implicit reasoning driving CoT's performance in pattern-based ICL: while explicit reasoning falters due to LLMs' struggles to infer underlying patterns from demonstrations, implicit reasoning-disrupted by the increased contextual distance of CoT rationales-often compensates, delivering correct answers despite flawed rationales. This hybrid mechanism explains CoT's relative underperformance, as noise from weak explicit inference undermines the process, even as implicit mechanisms partially salvage outcomes. Notably, even long-CoT reasoning models, which excel in abstract and symbolic reasoning, fail to fully overcome these limitations despite higher computational costs. Our findings challenge existing assumptions regarding the universal efficacy of CoT, yielding novel insights into its limitations and guiding future research toward more nuanced and effective reasoning methodologies for LLMs.

Large language models (LLMs) have shown potential in supporting decision-making applications, particularly as personal assistants in the financial, healthcare, and legal domains. While prompt engineering strategies have enhanced the capabilities of LLMs in decision-making, cognitive biases inherent to LLMs present significant challenges. Cognitive biases are systematic patterns of deviation from norms or rationality in decision-making that can lead to the production of inaccurate outputs. Existing cognitive bias mitigation strategies assume that input prompts only contain one type of cognitive bias, limiting their effectiveness in more challenging scenarios involving multiple cognitive biases. To fill this gap, we propose a cognitive debiasing approach, self-adaptive cognitive debiasing (SACD), that enhances the reliability of LLMs by iteratively refining prompts. Our method follows three sequential steps - bias determination, bias analysis, and cognitive debiasing - to iteratively mitigate potential cognitive biases in prompts. We evaluate SACD on finance, healthcare, and legal decision-making tasks using both open-weight and closed-weight LLMs. Compared to advanced prompt engineering methods and existing cognitive debiasing techniques, SACD achieves the lowest average bias scores in both single-bias and multi-bias settings.

CrowdVLM-R1: Expanding R1 Ability to Vision Language Model for Crowd Counting using Fuzzy Group Relative Policy Reward 1 st Zhiqiang Wang Florida Atlantic University Boca Raton, USA zwang2022@fau.edu 2 nd Pengbin Feng University of Southern California Los Angeles, USA fengpengbin.apply@gmail.com Abstract--We propose CrowdVLM-R1, which expands the R1 base model for accurate crowd counting, using a novel framework that integrates the fuzzy group relative policy optimization reward function (FGRPR) to enhance learning efficiency. Unlike the conventional binary (0/1) accuracy reward, our fuzzy reward model, FGRPR, which contains both format and precision rewards, provides nuanced incentives to encourage the R1 model to learn to adjust policies towards precise outputs. Supervised fine-tuning (SFT) is also integrated for the CrowdVLM-R1 model to learn from a handful of inputs to enable both in-domain and out-of-domain counting. Experimental results demonstrate that GRPO with a standard binary accuracy reward underperforms compared to SFT . In contrast, FGRPR, applied to Qwen2.5-VL-(3B/7B), surpasses all baseline models, including GPT -4o, LLaMA2-70B and SFT, in five domain datasets. For out-of-domain datasets, FGRPR achieves performance comparable to SFT but excels when target values are larger, as its fuzzy reward function assigns higher rewards to closer approximations. This approach is broadly applicable to tasks where the precision of the answer is critical. I. INTRODUCTION Recently, DeepSeek R1 [1] has drawn much attention among advances in large language models (LLMs), as it demonstrates how reinforcement learning (RL) can be the primary driver of reasoning.

GraphRAG enhances large language models (LLMs) to generate quality answers for user questions by retrieving related facts from external knowledge graphs. However, current GraphRAG methods are primarily evaluated on and overly tailored for knowledge graph question answering (KGQA) benchmarks, which are biased towards a few specific question patterns and do not reflect the diversity of real-world questions. To better evaluate GraphRAG methods, we propose a complete four-class taxonomy to categorize the basic patterns of knowledge graph questions and use it to create PolyBench, a new GraphRAG benchmark encompassing a comprehensive set of graph questions. With the new benchmark, we find that existing GraphRAG methods fall short in effectiveness (i.e., quality of the generated answers) and/or efficiency (i.e., response time or token usage) because they adopt either a fixed graph traversal strategy or free-form exploration by LLMs for fact retrieval. However, different question patterns require distinct graph traversal strategies and context formation. To facilitate better retrieval, we propose PolyG, an adaptive GraphRAG approach by decomposing and categorizing the questions according to our proposed question taxonomy. Built on top of a unified interface and execution engine, PolyG dynamically prompts an LLM to generate a graph database query to retrieve the context for each decomposed basic question. Compared with SOTA GraphRAG methods, PolyG achieves a higher win rate in generation quality and has a low response latency and token cost. Our code and benchmark are open-source at https://github.com/Liu-rj/PolyG.

The rapid escalation from elementary school-level to frontier problems of the difficulty for LLM benchmarks in recent years have weaved a miracle for researchers that we are only inches away from surpassing human intelligence. However, is the LLMs' remarkable reasoning ability indeed comes from true intelligence by human standards, or are they simply reciting solutions witnessed during training at an Internet level? To study this problem, we propose RoR-Bench, a novel, multi-modal benchmark for detecting LLM's recitation behavior when asked simple reasoning problems but with conditions subtly shifted, and conduct empirical analysis on our benchmark. Surprisingly, we found existing cutting-edge LLMs unanimously exhibits extremely severe recitation behavior; by changing one phrase in the condition, top models such as OpenAI-o1 and DeepSeek-R1 can suffer 60 percent performance loss on elementary school-level arithmetic and reasoning problems. Such findings are a wake-up call to the LLM community that compels us to re-evaluate the true intelligence level of cutting-edge LLMs.

Accurate segmentation of nodules in both 2D breast ultrasound (BUS) and 3D automated breast ultrasound (ABUS) is crucial for clinical diagnosis and treatment planning. Therefore, developing an automated system for nodule segmentation can enhance user independence and expedite clinical analysis. Unlike fully-supervised learning, weakly-supervised segmentation (WSS) can streamline the laborious and intricate annotation process. However, current WSS methods face challenges in achieving precise nodule segmentation, as many of them depend on inaccurate activation maps or inefficient pseudo-mask generation algorithms. In this study, we introduce a novel multi-agent reinforcement learning-based WSS framework called Flip Learning, which relies solely on 2D/3D boxes for accurate segmentation. Specifically, multiple agents are employed to erase the target from the box to facilitate classification tag flipping, with the erased region serving as the predicted segmentation mask. The key contributions of this research are as follows: (1) Adoption of a superpixel/supervoxel-based approach to encode the standardized environment, capturing boundary priors and expediting the learning process. (2) Introduction of three meticulously designed rewards, comprising a classification score reward and two intensity distribution rewards, to steer the agents' erasing process precisely, thereby avoiding both under- and over-segmentation. (3) Implementation of a progressive curriculum learning strategy to enable agents to interact with the environment in a progressively challenging manner, thereby enhancing learning efficiency. Extensively validated on the large in-house BUS and ABUS datasets, our Flip Learning method outperforms state-of-the-art WSS methods and foundation models, and achieves comparable performance as fully-supervised learning algorithms.

Earthquakes, as natural phenomena, have continuously caused damage and loss of human life historically. Earthquake prediction is an essential aspect of any society's plans and can increase public preparedness and reduce damage to a great extent. Nevertheless, due to the stochastic character of earthquakes and the challenge of achieving an efficient and dependable model for earthquake prediction, efforts have been insufficient thus far, and new methods are required to solve this problem. Aware of these issues, this paper proposes a novel prediction method based on attention mechanism (AM), convolution neural network (CNN), and bi-directional long short-term memory (BiLSTM) models, which can predict the number and maximum magnitude of earthquakes in each area of mainland China-based on the earthquake catalog of the region. This model takes advantage of LSTM and CNN with an attention mechanism to better focus on effective earthquake characteristics and produce more accurate predictions. Firstly, the zero-order hold technique is applied as pre-processing on earthquake data, making the model's input data more proper. Secondly, to effectively use spatial information and reduce dimensions of input data, the CNN is used to capture the spatial dependencies between earthquake data. Thirdly, the Bi-LSTM layer is employed to capture the temporal dependencies. Fourthly, the AM layer is introduced to highlight its important features to achieve better prediction performance. The results show that the proposed method has better performance and generalize ability than other prediction methods.

Large Language Models (LLMs) are increasingly used to simulate human users in interactive settings such as therapy, education, and social role-play. While these simulations enable scalable training and evaluation of AI agents, off-the-shelf LLMs often drift from their assigned personas, contradict earlier statements, or abandon role-appropriate behavior. We introduce a unified framework for evaluating and improving persona consistency in LLM-generated dialogue. We define three automatic metrics: prompt-to-line consistency, line-to-line consistency, and Q&A consistency, that capture different types of persona drift and validate each against human annotations. Using these metrics as reward signals, we apply multi-turn reinforcement learning to fine-tune LLMs for three user roles: a patient, a student, and a social chat partner. Our method reduces inconsistency by over 55%, resulting in more coherent and faithful simulated users.