Large language models (LLMs) have made significant strides, extending their applications to dialogue systems, automated content creation, and domain-specific advisory tasks. However, as their use grows, concerns have emerged regarding their reliability in simulating complex decision-making behavior, such as risky decision-making, where a single choice can lead to multiple outcomes. This study investigates the ability of LLMs to simulate risky decision-making scenarios. We compare model-generated decisions with actual human responses in a series of lottery-based tasks, using transportation stated preference survey data from participants in Sydney, Dhaka, Hong Kong, and Nanjing. Demographic inputs were provided to two LLMs -- ChatGPT 4o and ChatGPT o1-mini -- which were tasked with predicting individual choices. Risk preferences were analyzed using the Constant Relative Risk Aversion (CRRA) framework. Results show that both models exhibit more risk-averse behavior than human participants, with o1-mini aligning more closely with observed human decisions. Further analysis of multilingual data from Nanjing and Hong Kong indicates that model predictions in Chinese deviate more from actual responses compared to English, suggesting that prompt language may influence simulation performance. These findings highlight both the promise and the current limitations of LLMs in replicating human-like risk behavior, particularly in linguistic and cultural settings.

Group-based reinforcement learning algorithms such as Group Reward Policy Optimization (GRPO) have proven effective for fine-tuning large language models (LLMs) with human feedback. However, generating and storing multiple responses per prompt incurs substantial memory overhead, especially as the sample group size increases, limiting scalability under constrained hardware. We propose Infinite Sampling, a framework that enables efficient and stable GRPO training by decoupling group size from GPU memory usage. It consists of: (1) micro sampling groups that decompose large groups into memory-feasible rounds; (2) continuous sampling that interleaves generation across groups to improve utilization; and (3) a length-aware scheduler combining token-conditioned sequence length prediction with a two-stage plan: global grouping via FPTAS and runtime refill via SJF. Experiments show that our Micro Sampling Groups reduce peak memory usage by over 50% compared to full-group decoding (e.g., from 21.55 GB to 10.64 GB on Qwen3-1.7B). Building on this, Infinite Sampling improves throughput by over 25% compared to the naive micro sampling group method, reducing decoding steps while maintaining full-length completions and memory usage. Our hybrid scheduling ensures efficient and stable GRPO training with larger groups under realistic GPU memory constraints.

This work-in-progress investigates the memorization, creativity, and nonsense found in cooking recipes generated from Large Language Models (LLMs). Precisely, we aim (i) to analyze memorization, creativity, and non-sense in LLMs using a small, high-quality set of human judgments and (ii) to evaluate potential approaches to automate such a human annotation in order to scale our study to hundreds of recipes. To achieve (i), we conduct a detailed human annotation on 20 preselected recipes generated by LLM (Mixtral), extracting each recipe's ingredients and step-by-step actions to assess which elements are memorized--i.e., directly traceable to online sources possibly seen during training--and which arise from genuine creative synthesis or outright nonsense. We find that Mixtral consistently reuses ingredients that can be found in online documents, potentially seen during model training, suggesting strong reliance on memorized content. To achieve aim (ii) and scale our analysis beyond small sample sizes and single LLM validation, we design an ``LLM-as-judge'' pipeline that automates recipe generation, nonsense detection, parsing ingredients and recipe steps, and their annotation. For instance, comparing its output against human annotations, the best ingredient extractor and annotator is Llama 3.1+Gemma 2 9B, achieving up to 78% accuracy on ingredient matching. This automated framework enables large-scale quantification of memorization, creativity, and nonsense in generated recipes, providing rigorous evidence of the models' creative capacities.

We present Federated Timeline Synthesis (FTS), a novel framework for training generative foundation models across distributed timeseries data applied to electronic health records (EHR). At its core, FTS represents patient history as tokenized Patient Health Timelines (PHTs), language-agnostic sequences encoding temporal, categorical, and continuous clinical information. Each institution trains an autoregressive transformer on its local PHTs and transmits only model weights to a central server. The server uses the generators to synthesize a large corpus of trajectories and train a Global Generator (GG), enabling zero-shot inference via Monte Carlo simulation of future PHTs. We evaluate FTS on five clinically meaningful prediction tasks using MIMIC-IV data, showing that models trained on synthetic data generated by GG perform comparably to those trained on real data. FTS offers strong privacy guarantees, scalability across institutions, and extensibility to diverse prediction and simulation tasks especially in healthcare, including counterfactual inference, early warning detection, and synthetic trial design.

Urban research involves a wide range of scenarios and tasks that require the understanding of multi-modal data. Current methods often focus on specific data types and lack a unified framework in urban field for processing them comprehensively. The recent success of multi-modal large language models (MLLMs) presents a promising opportunity to overcome this limitation. In this paper, we introduce $\textit{UrbanLLaVA}$, a multi-modal large language model designed to process these four types of data simultaneously and achieve strong performance across diverse urban tasks compared with general MLLMs. In $\textit{UrbanLLaVA}$, we first curate a diverse urban instruction dataset encompassing both single-modal and cross-modal urban data, spanning from location view to global view of urban environment. Additionally, we propose a multi-stage training framework that decouples spatial reasoning enhancement from domain knowledge learning, thereby improving the compatibility and downstream performance of $\textit{UrbanLLaVA}$ across diverse urban tasks. Finally, we also extend existing benchmark for urban research to assess the performance of MLLMs across a wide range of urban tasks. Experimental results from three cities demonstrate that $\textit{UrbanLLaVA}$ outperforms open-source and proprietary MLLMs in both single-modal tasks and complex cross-modal tasks and shows robust generalization abilities across cities. Source codes and data are openly accessible to the research community via https://github.com/tsinghua-fib-lab/UrbanLLaVA.

High labeling cost for in-context learning (ICL) demonstrations motivates using large language models (LLMs) for synthesis to reduce overhead. However, existing synthesis methods are mainly task-specific or rely on pre-existing demonstrations. So this paper focuses on synthesizing demonstrations from scratch for arbitrary tasks. A major challenge in synthesizing from scratch is ensuring consistency with the target task, as the lack of labeling guidance could lead to synthesis bias. We first propose a consistency metric called V-Score, which has higher performance and lower computation cost compared with the metrics based on grams or embedding vectors. Furthermore, we introduce V-Synthesis, which leverages V-Score for proportional sampling to ensure both high consistency and diversity of synthesized demonstrations. Experimental results demonstrate that V-Synthesis yields an average performance improvement of 2.0% compared to existing synthesis methods confirming the effectiveness of V-Synthesis.

Past work has studied the effects of fine-tuning on large language models' (LLMs) overall performance on certain tasks. However, a quantitative and systematic method for analyzing its effect on individual outputs is still lacking. Here, we propose a new method for measuring the contribution that fine-tuning makes to individual LLM responses, assuming access to the original pre-trained model. Our method tracks the model's intermediate hidden states, providing a more fine-grained insight into the effects of fine-tuning than a simple comparison of final outputs from pre-trained and fine-tuned models. We introduce and theoretically analyze an exact decomposition of any fine-tuned LLM into a pre-training component and a fine-tuning component. Empirically, we find that model behavior and performance can be steered by up- or down-scaling the fine-tuning component during the forward pass. Motivated by this finding and our theoretical analysis, we define the Tuning Contribution (TuCo) as the ratio of the magnitudes of the fine-tuning component to the pre-training component. We observe that three prominent adversarial attacks on LLMs circumvent safety measures in a way that reduces TuCo, and that TuCo is consistently lower on prompts where these attacks succeed compared to those where they do not. This suggests that attenuating the effect of fine-tuning on model outputs plays a role in the success of such attacks. In summary, TuCo enables the quantitative study of how fine-tuning influences model behavior and safety, and vice versa.

Pre-trained language models such as BERT [3] have become foundational in modern natural language processing, owing to their ability to capture rich contextual representations from large-scale corpora. However, it is still unclear where and how extracted knowledge from training data is internally represented in the model, and how we can distribute this knowledge between structurally similar models. This work introduces a specific method for word-embedding initialization that encapsulates domain-specific knowledge into internal representations to construct feature-grounded embed-dings. Such kind of embedding provides structured prior into internal weight landscape during LLM training.

Existing studies on fairness are largely Western-focused, making them inadequate for culturally diverse countries such as India. To address this gap, we introduce INDIC-BIAS, a comprehensive India-centric benchmark designed to evaluate fairness of LLMs across 85 identity groups encompassing diverse castes, religions, regions, and tribes. We first consult domain experts to curate over 1,800 socio-cultural topics spanning behaviors and situations, where biases and stereotypes are likely to emerge. Grounded in these topics, we generate and manually validate 20,000 real-world scenario templates to probe LLMs for fairness. We structure these templates into three evaluation tasks: plausibility, judgment, and generation. Our evaluation of 14 popular LLMs on these tasks reveals strong negative biases against marginalized identities, with models frequently reinforcing common stereotypes. Additionally, we find that models struggle to mitigate bias even when explicitly asked to rationalize their decision. Our evaluation provides evidence of both allocative and representational harms that current LLMs could cause towards Indian identities, calling for a more cautious usage in practical applications. We release INDIC-BIAS as an open-source benchmark to advance research on benchmarking and mitigating biases and stereotypes in the Indian context.

Molecular structure elucidation involves deducing a molecule's structure from various types of spectral data, which is crucial in chemical experimental analysis. While large language models (LLMs) have shown remarkable proficiency in analyzing and reasoning through complex tasks, they still encounter substantial challenges in molecular structure elucidation. We identify that these challenges largely stem from LLMs' limited grasp of specialized chemical knowledge. In this work, we introduce a Knowledge-enhanced reasoning framework for Molecular Structure Elucidation (K-MSE), leveraging Monte Carlo Tree Search for test-time scaling as a plugin. Specifically, we construct an external molecular substructure knowledge base to extend the LLMs' coverage of the chemical structure space. Furthermore, we design a specialized molecule-spectrum scorer to act as a reward model for the reasoning process, addressing the issue of inaccurate solution evaluation in LLMs. Experimental results show that our approach significantly boosts performance, particularly gaining more than 20% improvement on both GPT-4o-mini and GPT-4o. Our code is available at https://github.com/HICAI-ZJU/K-MSE.