Hallucination is a long-standing problem that has been actively investigated in Vision-Language Models (VLMs). Existing research commonly attributes hallucinations to technical limitations or sycophancy bias, where the latter means the models tend to generate incorrect answers to align with user expectations. However, these explanations primarily focus on technical or externally driven factors, and may have neglected the possibility that hallucination behaviours might mirror cognitive biases observed in human psychology. In this work, we introduce a psychological taxonomy, categorizing VLMs' cognitive biases that lead to hallucinations, including sycophancy, logical inconsistency, and a newly identified VLMs behaviour: appeal to authority. To systematically analyze these behaviours, we design AIpsych, a scalable benchmark that reveals psychological tendencies in model response patterns. Leveraging this benchmark, we investigate how variations in model architecture and parameter size influence model behaviour when responding to strategically manipulated questions. Our experiments reveal that as model size increases, VLMs exhibit stronger sycophantic tendencies but reduced authority bias, suggesting increasing competence but a potential erosion of response integrity. A human subject study further validates our hypotheses and highlights key behavioural differences between VLMs and human respondents. This work suggests a new perspective for understanding hallucination in VLMs and highlights the importance of integrating psychological principles into model evaluation. The benchmark and codes are tested and available in the anonymous link https://anonymous.4open.science/r/AIpsych-666.Figure 1: Left: a VLM exhibits sycophancy by favouring the questioner's options despite recognising it is a pink cup. Right: a human demonstrates authority bias by accepting the question's framing, also yielding the wrong answer. However, to distinguish between them, we will need to ask more questions. VLMs have made remarkable progress, achieving increasingly higher accuracy in visual reasoning tasks and enhancing real-world applications such as image captioning, visual question answering, and multimodal retrieval (Chen et al., 2023).

In addition to matching the performance of Transformers (Vaswani et al. 2017) across many tasks, the recurrent mechanism brings two benefits: the ability to efficiently process long sequences thanks to its linear complexity, and the capacity to easily process tokens beyond their training context by simply rolling out the state. Nevertheless, in practice these benefits are often unrealized, given that their performance can drop considerably when the sequence length exceeds their training context (Ben-Kish et al. 2024; Waleffe et al. 2024; Ye et al. 2025; Yuan et al. 2024). This naturally leads to two questions: (1) why do these models fail to length generalize? and (2) how can we efficiently enable length generalization across several recurrent models? Recently, some works have studied the length generalization of Mamba (Dao and Gu 2024) and have proposed solutions such as forcing the model to forget previous context (Yingfa Chen et al. 2024) or skipping tokens in the state update to reduce the effective context of the processed sequence (Ben-Kish et al. 2024; Ye et al. 2025). However, these methods require changing the internal mechanism of Mamba and might not be easily transferable to other architectures. Other works have linked length generalization to state capacity and overfitting (Yingfa Chen et al. 2024; S. Wang 2024), proposing training on longer sequences and with Truncated Backpropagation Through Time (TBTT) (Sutskever 2013; Williams and J. Peng 1990) as a way to enable length generalization. In this work, we reason about the distribution of states that the model is trained on to introduce a precise hypothesis that explains why recurrent models fail to length generalize. Moreover, we perform comprehensive interventions that elucidate on what distributions recurrent models need to be trained to enable length generalization. 1

Gradients of neural networks encode valuable information for optimization, editing, and analysis of models. Therefore, practitioners often treat gradients as inputs to task-specific algorithms, e.g. for pruning or optimization. Recent works explore learning algorithms that operate directly on gradients but use architectures that are not specifically designed for gradient processing, limiting their applicability. In this paper, we present a principled approach for designing architectures that process gradients. Our approach is guided by three principles: (1) equivariant design that preserves neuron permutation symmetries, (2) processing sets of gradients across multiple data points to capture curvature information, and (3) efficient gradient representation through rank-1 decomposition. Based on these principles, we introduce GradMetaNet, a novel architecture for learning on gradients, constructed from simple equivariant blocks. We prove universality results for GradMetaNet, and show that previous approaches cannot approximate natural gradient-based functions that GradMetaNet can. We then demonstrate GradMetaNet's effectiveness on a diverse set of gradient-based tasks on MLPs and transformers, such as learned optimization, INR editing, and estimating loss landscape curvature.

Figure 1: TriVLA is a unified Vision-Language-Action framework that adopts a triple-system architecture inspired by the episodic world model. Image and language inputs are processed by a Vision-Language Model for multimodal perception. A Video Diffusion Model provides dynamic world modeling and future prediction. The policy module integrates sequential outputs, robot state, and action history and generates real-time actions for complex manipulation tasks. Recent advances in vision-language models (VLMs) have enabled robots to follow open-ended instructions and demonstrate impressive commonsense reasoning. However, current vision-language-action (VLA) frameworks primarily rely on static representations and limited temporal context, restricting agents to short-horizon, reactive behaviors and hindering robust generalization in dynamic embodied environments. Inspired by cognitive neuroscience theories of episodic memory, we are, to our knowledge, among the first to introduce a formalized episodic world model in VLA, enabling embodied robots to accumulate, recall, and predict sequential experiences. As an instantiation of this concept, our unified TriVLA realizes the episodic world model through a triple-system architecture: integrating multimodal grounding from a pretrained VLM (System 2) and temporally rich dynamics perception from a video diffusion model (System 3). This enables the agent to accumulate and recall sequential experiences, interpret current contexts, and predict future environmental evolution. Guided by episodic representations that span both the past and anticipated future, the downstream policy (System 1) generates coherent, context-aware action sequences through flow-matching and cross-modal attention mechanisms. It demonstrates strong long-horizon planning and open-ended intent understanding, showcasing the advantages of episodic world model-inspired reasoning for robust, generalizable robot intelligence. "Episodic memory is the only memory system that allows mental time travel--backward into the past and forward into the future. " -- Endel Tulving Building on this cognitive foundation, we advocate that robotic agents, require an internal episodic world model: a representational system that not only recalls past interactions but also anticipates future dynamics, thereby enabling robust generalization in embodied environments. Decades of cognitive neuroscience provide compelling evidence for this perspective.

Self-improvement is among the most prominent techniques within the realm of large language models (LLM), aiming to enhance the LLM performance without relying on external data. Despite its significance, generally how LLM performances evolve during the self-improvement process remains underexplored. In this paper, we theoretically model the training dynamics of self-improvement via the concept of solver-verifier gap. This is inspired by the conjecture that the performance enhancement of self-improvement stems from the gap between LLM's solver capability and verifier capability. Based on the theoretical framework, we further show how to model the entire training trajectory. This framework allows quantifying the capability limit of self-improvement by fitting the theoretical model to the experiment results. We empirically validate the effectiveness of the theoretical framework on various LLMs and datasets. Beyond self-improvement, we extend our analysis to investigate how external data influences these dynamics within the framework. Notably, we find that under limited external data regimes, such external data can be utilized at any stage without significantly affecting final performances, which accords with the empirical observations.

Extracting molecular structure-activity relationships (SARs) from scientific literature and patents is essential for drug discovery and materials research. However, this task remains challenging due to heterogeneous document formats and limitations of existing methods. Specifically, rule-based approaches relying on rigid templates fail to generalize across diverse document layouts, while general-purpose multimodal large language models (MLLMs) lack sufficient accuracy and reliability for specialized tasks, such as layout detection and optical chemical structure recognition (OCSR). To address these challenges, we introduce DocSAR-200, a rigorously annotated benchmark of 200 scientific documents designed specifically for evaluating SAR extraction methods. Additionally, we propose Doc2SAR, a novel synergistic framework that integrates domain-specific tools with MLLMs enhanced via supervised fine-tuning (SFT). Extensive experiments demonstrate that Doc2SAR achieves state-of-the-art performance across various document types, significantly outperforming leading end-to-end baselines. Specifically, Doc2SAR attains an overall Table Recall of 80.78% on DocSAR-200, exceeding end2end GPT -4o by 51.48%. Furthermore, Doc2SAR demonstrates practical usability through efficient inference and is accompanied by a web app. The code and data are provided in the supplementary materials.

Ensuring the moral reasoning capabilities of Large Language Models (LLMs) is a growing concern as these systems are used in socially sensitive tasks. Nevertheless, current evaluation benchmarks present two major shortcomings: a lack of annotations that justify moral classifications, which limits transparency and interpretability; and a predominant focus on English, which constrains the assessment of moral reasoning across diverse cultural settings. In this paper, we introduce MFTCXplain, a multilingual benchmark dataset for evaluating the moral reasoning of LLMs via multi-hop hate speech explanation using the Moral Foundations Theory. MFTCXplain comprises 3,000 tweets across Portuguese, Italian, Persian, and English, annotated with binary hate speech labels, moral categories, and text span-level rationales. Our results show a misalignment between LLM outputs and human annotations in moral reasoning tasks. While LLMs perform well in hate speech detection (F1 up to 0.836), their ability to predict moral sentiments is notably weak (F1 < 0.35). Furthermore, rationale alignment remains limited mainly in underrepresented languages. Our findings show the limited capacity of current LLMs to internalize and reflect human moral reasoning

Scaling large language models typically involves three dimensions: depth, width, and parameter count. In this work, we explore a fourth dimension, \textbf{virtual logical depth} (VLD), which increases effective algorithmic depth without changing parameter count by reusing weights. While parameter reuse is not new, its role in scaling has been underexplored. Unlike recent test-time methods that scale token-wise, VLD alters the internal computation graph during training and inference. Through controlled experiments, we obtain three key insights. (1) \textit{Knowledge capacity vs. parameters}: at fixed parameter count, VLD leaves knowledge capacity nearly unchanged, while across models capacity still scales with parameters. (2) \textit{Reasoning vs. reuse}: properly implemented VLD substantially improves reasoning ability \emph{without} more parameters, decoupling reasoning from size. This suggests a new scaling path beyond token-wise test-time methods. (3) \textit{Robustness and generality}: reasoning gains persist across architectures and reuse schedules, showing VLD captures a general scaling behavior. These results provide insight into future scaling strategies and raise a deeper question: does superintelligence require ever-larger models, or can it be achieved by reusing parameters and increasing logical depth? We argue many unknown dynamics in scaling remain to be explored. Code is available at https://anonymous.4open.science/r/virtual_logical_depth-8024/.

The Cyc project, started in 1984, created the first large-scale database of commonsense knowledge. The initiative continues to this day with its aim to provide a comprehensive ontology and knowledge base of commonsense knowledge to enable human-like reasoning for AI systems. In the concluding paragraph of his Communications of the Association of Computing Machinery (CACM) 1995 article A Large-Scale Investment in Knowledge Infrastructure [52], Cyc's founder Douglas B. Lenat wrote: Is Cyc necessary? How far would a user get with something simpler than Cyc but that lacks everyday commonsense knowledge? Nobody knows; the question will be settled empirically. Our guess is most of these applications will eventually tap the synergy in a suite of sources (including neural nets and decision theory), one of which will be Cyc. Although 30 years have passed since the above article was written, AI research community has not conclusively settled [10] the question "How far would a user get with something simpler than Cyc but that lacks everyday commonsense knowledge?" However, it is clear that significant strides have been made in addressing many of the tasks that were original Cyc use cases, including information retrieval, semi-automatically linking multiple heterogeneous external information sources, spelling and grammar correction, machine translation, natural language understanding and speech understanding.

Sparse Mixture of Experts (MoE) models offer a scalable and efficient architecture for training large neural networks by activating only a subset of parameters ("experts") for each input. A learned router computes a distribution over these experts, and assigns input tokens to a small subset. However, without auxiliary balancing mechanisms, routers often converge to using only a few experts, severely limiting model capacity and degrading performance. Most current load balancing mechanisms encourage a distribution over experts that resembles a roughly uniform distribution of experts per token. During training, this can result in inconsistent routing behavior, resulting in the model spending its capacity to learn redundant knowledge. We address this by introducing a novel load balancing loss that preserves token-wise relational structure, encouraging consistent expert choices for similar inputs during training. Our experimental results show that applying our loss to the router results in 36% faster convergence and lower redundancy compared to a popular load balancing loss.