We demonstrate that large language models (LLMs) exhibit consistent value orientations despite adopting diverse personas, revealing a persistent inertia in their responses that remains stable across the variety of roles they are prompted to assume. To systematically explore this phenomenon, we introduce the role-play-at-scale methodology, which involves prompting LLMs with randomized, diverse personas and analyzing the macroscopic trend of their responses. Unlike previous works that simply feed these questions to LLMs as if testing human subjects, our role-play-at-scale methodology diagnoses inherent tendencies in a systematic and scalable manner by: (1) prompting the model to act in different random personas and (2) asking the same question multiple times for each random persona. This approach reveals consistent patterns in LLM responses across diverse role-play scenarios, indicating deeply encoded inherent tendencies. Our findings contribute to the discourse on value alignment in foundation models and demonstrate the efficacy of role-play-at-scale as a diagnostic tool for uncovering encoded biases in LLMs.

The application of the Multi-modal Large Language Models (MLLMs) in medical clinical scenarios remains underexplored. Previous benchmarks only focus on the capacity of the MLLMs in medical visual question-answering (VQA) or report generation and fail to assess the performance of the MLLMs on complex clinical multi-modal tasks. In this paper, we propose a novel Medical Personalized Multi-modal Consultation (Med-PMC) paradigm to evaluate the clinical capacity of the MLLMs. Med-PMC builds a simulated clinical environment where the MLLMs are required to interact with a patient simulator to complete the multi-modal information-gathering and decision-making task. Specifically, the patient simulator is decorated with personalized actors to simulate diverse patients in real scenarios. We conduct extensive experiments to access 12 types of MLLMs, providing a comprehensive view of the MLLMs' clinical performance. We found that current MLLMs fail to gather multimodal information and show potential bias in the decision-making task when consulted with the personalized patient simulators. Further analysis demonstrates the effectiveness of Med-PMC, showing the potential to guide the development of robust and reliable clinical MLLMs. Code and data are available at https://github.com/LiuHC0428/Med-PMC.

Reinforcement learning (RL) has seen significant success across various domains, but its adoption is often limited by the black-box nature of neural network policies, making them difficult to interpret. In contrast, symbolic policies allow representing decision-making strategies in a compact and interpretable way. However, learning symbolic policies directly within on-policy methods remains challenging. In this paper, we introduce SYMPOL, a novel method for SYMbolic treebased on-POLicy RL. SYMPOL employs a tree-based model integrated with a policy gradient method, enabling the agent to learn and adapt its actions while maintaining a high level of interpretability. We evaluate SYMPOL on a set of benchmark RL tasks, demonstrating its superiority over alternative tree-based RL approaches in terms of performance and interpretability. To the best of our knowledge, this is the first method, that allows a gradient-based end-to-end learning of interpretable, axis-aligned decision trees within existing on-policy RL algorithms. Therefore, SYMPOL can become the foundation for a new class of interpretable RL based on decision trees. Existing methods for symbolic, tree-based RL (see Figure 1b and 1c) suffer from severe information loss when converting the differentiable policy (high train reward) into the symbolic policy (low test reward). Using SYMPOL (Figure 1a), we can directly optimize the symbolic policy with PPO and therefore have no information loss during the application (high train and test reward). Reinforcement learning (RL) has achieved remarkable success in solving complex sequential decision-making problems, ranging from robotics and autonomous systems to game playing and recommendation systems. However, the policies learned by traditional RL algorithms, represented by neural networks, often lack interpretability and transparency, making them difficult to understand, trust, and deploy in safety-critical or high-stakes scenarios (Landajuela et al., 2021). These symbolic representations do not only facilitate human understanding and analysis but also ensure predictable and explainable behavior, which is crucial for building trust and enabling effective human-AI collaboration. Moreover, the deployment of symbolic policies in safety-critical systems, such as autonomous vehicles or industrial robots, could significantly improve their reliability and trustworthiness.

Diffusion models, which leverage stochastic processes to capture complex data distributions effectively, have shown their performance as generative models, achieving notable success in image-related tasks through iterative denoising processes. Recently, diffusion models have been further applied and show their strong abilities in planning tasks, leading to a significant growth in related publications since 2023. To help researchers better understand the field and promote the development of the field, we conduct a systematic literature review of recent advancements in the application of diffusion models for planning. Specifically, this paper categorizes and discusses the current literature from the following perspectives: (i) relevant datasets and benchmarks used for evaluating diffusion modelbased planning; (ii) fundamental studies that address aspects such as sampling efficiency; (iii) skill-centric and condition-guided planning for enhancing adaptability; (iv) safety and uncertainty managing mechanism for enhancing safety and robustness; and (v) domain-specific application such as autonomous driving. Finally, given the above literature review, we further discuss the challenges and future directions in this field.

Understanding and characterising quantum many-body dynamics remains a significant challenge due to both the exponential complexity required to represent quantum many-body Hamiltonians, and the need to accurately track states in time under the action of such Hamiltonians. This inherent complexity limits our ability to characterise quantum many-body systems, highlighting the need for innovative approaches to unlock their full potential. To address this challenge, we propose a novel method to solve the Hamiltonian Learning (HL) problem-inferring quantum dynamics from many-body state trajectories-using Neural Differential Equations combined with an Ansatz Hamiltonian. Our method is reliably convergent, experimentally friendly, and interpretable, making it a stable solution for HL on a set of Hamiltonians previously unlearnable in the literature. In addition to this, we propose a new quantitative benchmark based on power laws, which can objectively compare the reliability and generalisation capabilities of any two HL algorithms. Finally, we benchmark our method against state-of-the-art HL algorithms with a 1D spin-1/2 chain proof of concept.

The impressive performance of Large Language Models (LLMs) has consistently surpassed numerous human-designed benchmarks, presenting new challenges in assessing the shortcomings of LLMs. Designing tasks and finding LLMs' limitations are becoming increasingly important. In this paper, we investigate the question of whether an LLM can discover its own limitations from the errors it makes. To this end, we propose a Self-Challenge evaluation framework with human-in-the-loop. Starting from seed instances that GPT-4 fails to answer, we prompt GPT-4 to summarize error patterns that can be used to generate new instances and incorporate human feedback on them to refine these patterns for generating more challenging data, iteratively. We end up with 8 diverse patterns, such as text manipulation and questions with assumptions. We then build a benchmark, SC-G4, consisting of 1,835 instances generated by GPT-4 using these patterns, with human-annotated gold responses. The SC-G4 serves as a challenging benchmark that allows for a detailed assessment of LLMs' abilities. Our results show that only 44.96\% of instances in SC-G4 can be answered correctly by GPT-4. Interestingly, our pilot study indicates that these error patterns also challenge other LLMs, such as Claude-3 and Llama-3, and cannot be fully resolved through fine-tuning. Our work takes the first step to demonstrate that LLMs can autonomously identify their inherent flaws and provide insights for future dynamic and automatic evaluation.

Large Language Models (LLMs) have achieved remarkable success and have been applied across various scientific fields, including chemistry. However, many chemical tasks require the processing of visual information, which cannot be successfully handled by existing chemical LLMs. This brings a growing need for models capable of integrating multimodal information in the chemical domain. In this paper, we introduce \textbf{ChemVLM}, an open-source chemical multimodal large language model specifically designed for chemical applications. ChemVLM is trained on a carefully curated bilingual multimodal dataset that enhances its ability to understand both textual and visual chemical information, including molecular structures, reactions, and chemistry examination questions. We develop three datasets for comprehensive evaluation, tailored to Chemical Optical Character Recognition (OCR), Multimodal Chemical Reasoning (MMCR), and Multimodal Molecule Understanding tasks. We benchmark ChemVLM against a range of open-source and proprietary multimodal large language models on various tasks. Experimental results demonstrate that ChemVLM achieves competitive performance across all evaluated tasks. Our model can be found at https://huggingface.co/AI4Chem/ChemVLM-26B.

Multilingual large language models (MLLMs) struggle to answer questions posed in non-dominant languages, even though they have acquired the relevant knowledge from their dominant language corpus. In contrast, human multilinguals can overcome such non-native language context limitations through Positive Native Language Transfer (PNLT). Inspired by the process of PNLT, we analogize the dominant language of MLLMs to the native language of human multilinguals, and propose Native Language Prompting (NatLan) to simulate the PNLT observed in human multilinguals. It explicitly creates native language contexts for MLLMs to facilitate the elicitation of the rich native language knowledge during question-answering, unlocking the limitations imposed by non-native language contexts. By employing multi-MLLM collaboration, NatLan reduces the workload on each MLLM in simulating PNLT and refines semantic transfer. On the C-Eval benchmark, NatLan provides up to a 10.1% average accuracy improvement and up to a 5.0% increase in the hard-level subset across five MLLMs, surpassing all top-notch related methods. Our code is available at https://github.com/AnonyNLP/NatLan.

David Chiu announced Thursday that his office is suing the operators of 16 A.I.-powered "undressing" websites that help users create and distribute deepfake nude photos of women and girls. The lawsuit, which city officials said was the first of its kind, accuses the websites' operators of violating state and federal laws that ban deepfake pornography, revenge pornography and child pornography, as well as California's unfair competition law. The names of the sites were redacted in the copy of the suit made public Thursday. Chiu's office has yet to identify the owners of many of the websites, but officials say they hope to find their names and hold them accountable. Chiu said the lawsuit has two goals: shutting down these websites and sounding the alarm about this form of "sexual abuse."

There is a large literature on the similarities and differences between biological neural circuits and deep artificial neural networks (DNNs). However, modern training of DNNs relies on several engineering tricks such as data batching, normalization, adaptive optimizers, and precise weight initialization. Despite their critical role in training DNNs, these engineering tricks are often overlooked when drawing parallels between biological and artificial networks, potentially due to a lack of evidence for their direct biological implementation. In this study, we show that Oja's plasticity rule partly overcomes the need for some engineering tricks. Specifically, under difficult, but biologically realistic learning scenarios such as online learning, deep architectures, and sub-optimal weight initialization, Oja's rule can substantially improve the performance of pure backpropagation. Our results demonstrate that simple synaptic plasticity rules can overcome challenges to learning that are typically overcome using less biologically plausible approaches when training DNNs.