From Word to World: Can Large Language Models be Implicit Text-based World Models?

Agentic reinforcement learning increasingly relies on experience-driven scaling, yet real-world environments remain non-adaptive, limited in coverage, and difficult to scale. World models offer a potential way to improve learning efficiency through simulated experience, but it remains unclear whether large language models can reliably serve this role and under what conditions they meaningfully benefit agents. We study these questions in text-based environments, which provide a controlled setting to reinterpret language modeling as next-state prediction under interaction. We introduce a three-level framework for evaluating LLM-based world models: (i) fidelity and consistency, (ii) scalability and robustness, and (iii) agent utility. Across five representative environments, we find that sufficiently trained world models maintain coherent latent state, scale predictably with data and model size, and improve agent performance via action verification, synthetic trajectory generation, and warm-starting reinforcement learning. Meanwhile, these gains depend critically on behavioral coverage and environment complexity, delineating clear boundry on when world modeling effectively supports agent learning.

ConInstruct: Evaluating Large Language Models on Conflict Detection and Resolution in Instructions

Instruction-following is a critical capability of Large Language Models (LLMs). While existing works primarily focus on assessing how well LLMs adhere to user instructions, they often overlook scenarios where instructions contain conflicting constraints-a common occurrence in complex prompts. The behavior of LLMs under such conditions remains under-explored. To bridge this gap, we introduce ConInstruct, a benchmark specifically designed to assess LLMs' ability to detect and resolve conflicts within user instructions. Using this dataset, we evaluate LLMs' conflict detection performance and analyze their conflict resolution behavior. Our experiments reveal two key findings: (1) Most proprietary LLMs exhibit strong conflict detection capabilities, whereas among open-source models, only DeepSeek-R1 demonstrates similarly strong performance. DeepSeek-R1 and Claude-4.5-Sonnet achieve the highest average F1-scores at 91.5% and 87.3%, respectively, ranking first and second overall. (2) Despite their strong conflict detection abilities, LLMs rarely explicitly notify users about the conflicts or request clarification when faced with conflicting constraints. These results underscore a critical shortcoming in current LLMs and highlight an important area for future improvement when designing instruction-following LLMs.

MoMA: A Mixture-of-Multimodal-Agents Architecture for Enhancing Clinical Prediction Modelling

Multimodal electronic health record (EHR) data provide richer, complementary insights into patient health compared to single-modality data. However, effectively integrating diverse data modalities for clinical prediction modeling remains challenging due to the substantial data requirements. We introduce a novel architecture, Mixture-of-Multimodal-Agents (MoMA), designed to leverage multiple large language model (LLM) agents for clinical prediction tasks using multimodal EHR data. MoMA employs specialized LLM agents ("specialist agents") to convert non-textual modalities, such as medical images and laboratory results, into structured textual summaries. These summaries, together with clinical notes, are combined by another LLM ("aggregator agent") to generate a unified multimodal summary, which is then used by a third LLM ("predictor agent") to produce clinical predictions. Evaluating MoMA on three prediction tasks using real-world datasets with different modality combinations and prediction settings, MoMA outperforms current state-of-the-art methods, highlighting its enhanced accuracy and flexibility across various tasks.

Unveiling Over-Memorization in Finetuning LLMs for Reasoning Tasks

The pretrained large language models (LLMs) are finetuned with labeled data for better instruction following ability and alignment with human values. In this paper, we study the learning dynamics of LLM finetuning on reasoning tasks and reveal the uncovered over-memorization phenomenon during a specific stage of LLM finetuning. At this stage, the LLMs have excessively memorized training data and exhibit high test perplexity while maintaining good test accuracy. We investigate the conditions that lead to LLM over-memorization and find that training epochs and large learning rates contribute to this issue. Although models with over-memorization demonstrate comparable test accuracy to normal models, they suffer from reduced robustness, poor out-of-distribution generalization, and decreased generation diversity. Our experiments unveil the over-memorization to be broadly applicable across different tasks, models, and finetuning methods. Our research highlights that overparameterized, extensively finetuned LLMs exhibit unique learning dynamics distinct from traditional machine learning models. Based on our observations of over-memorization, we provide recommendations on checkpoint and learning rate selection during finetuning.

An Information-Theoretic Perspective on Multi-LLM Uncertainty Estimation

Large language models (LLMs) often behave inconsistently across inputs, indicating uncertainty and motivating the need for its quantification in high-stakes settings. Prior work on calibration and uncertainty quantification often focuses on individual models, overlooking the potential of model diversity. We hypothesize that LLMs make complementary predictions due to differences in training and the Zipfian nature of language, and that aggregating their outputs leads to more reliable uncertainty estimates. To leverage this, we propose MUSE (Multi-LLM Uncertainty via Subset Ensembles), a simple information-theoretic method that uses Jensen-Shannon Divergence to identify and aggregate well-calibrated subsets of LLMs. Experiments on binary prediction tasks demonstrate improved calibration and predictive performance compared to single-model and naive ensemble baselines.

Towards Bridging the Reward-Generation Gap in Direct Alignment Algorithms

Direct Alignment Algorithms (DAAs), such as Direct Preference Optimization (DPO) and Simple Preference Optimization (SimPO), have emerged as efficient alternatives to Reinforcement Learning from Human Feedback (RLHF) algorithms for aligning large language models (LLMs) with human preferences. However, DAAs suffer from a fundamental limitation we identify as the "reward-generation gap" -- a misalignment between optimization objectives during training and actual generation performance during inference. In this paper, we find a contributor to the reward-generation gap is the mismatch between the inherent importance of prefix tokens during the LLM generation process and how this importance is reflected in the implicit reward functions of DAAs. To bridge the gap, we introduce a simple yet effective approach called Prefix-Oriented Equal-length Training (POET), which truncates both preferred and dispreferred responses to match the shorter one's length. Training with POET, where both responses in each sample are truncated to equal length, resulting in diverse truncated lengths across samples, the optimization of DAAs objective is implicitly constrained to converge across all positions, thus paying more attention to prefix tokens than the standard DAAs. We conduct experiments with DPO and SimPO, two representative DAAs, demonstrating that POET improves over their standard implementations, achieving up to 15.6 points in AlpacaEval 2 and overall improvements across downstream tasks. Our results highlight the importance of addressing the misalignment between reward optimization and generation performance in DAAs.

Automatic Robustness Stress Testing of LLMs as Mathematical Problem Solvers

Large language models (LLMs) have achieved distinguished performance on various reasoning-intensive tasks. However, LLMs might still face the challenges of robustness issues and fail unexpectedly in some simple reasoning tasks. Previous works evaluate the LLM robustness with hand-crafted templates or a limited set of perturbation rules, indicating potential data contamination in pre-training or fine-tuning datasets. In this work, inspired by stress testing in software engineering, we propose a novel framework, Automatic Robustness Checker (AR-Checker), to generate mathematical problem variants that maintain the semantic meanings of the original one but might fail the LLMs. The AR-Checker framework generates mathematical problem variants through multi-round parallel streams of LLM-based rewriting and verification. Our framework can generate benchmark variants dynamically for each LLM, thus minimizing the risk of data contamination. Experiments on GSM8K and MATH-500 demonstrate the strong performance of AR-Checker on mathematical tasks. We also evaluate AR-Checker on benchmarks beyond mathematics, including MMLU, MMLU-Pro, and CommonsenseQA, where it also achieves strong performance, further proving the effectiveness of AR-Checker.

TAG-INSTRUCT: Controlled Instruction Complexity Enhancement through Structure-based Augmentation

High-quality instruction data is crucial for developing large language models (LLMs), yet existing approaches struggle to effectively control instruction complexity. We present TAG-INSTRUCT, a novel framework that enhances instruction complexity through structured semantic compression and controlled difficulty augmentation. Unlike previous prompt-based methods operating on raw text, TAG-INSTRUCT compresses instructions into a compact tag space and systematically enhances complexity through RL-guided tag expansion. Through extensive experiments, we show that TAG-INSTRUCT outperforms existing instruction complexity augmentation approaches. Our analysis reveals that operating in tag space provides superior controllability and stability across different instruction synthesis frameworks.

PlanGPT-VL: Enhancing Urban Planning with Domain-Specific Vision-Language Models

In the field of urban planning, existing Vision-Language Models (VLMs) frequently fail to effectively analyze and evaluate planning maps, despite the critical importance of these visual elements for urban planners and related educational contexts. Planning maps, which visualize land use, infrastructure layouts, and functional zoning, require specialized understanding of spatial configurations, regulatory requirements, and multi-scale analysis. To address this challenge, we introduce PlanGPT-VL, the first domain-specific Vision-Language Model tailored specifically for urban planning maps. PlanGPT-VL employs three innovative approaches: (1) PlanAnno-V framework for high-quality VQA data synthesis, (2) Critical Point Thinking to reduce hallucinations through structured verification, and (3) comprehensive training methodology combining Supervised Fine-Tuning with frozen vision encoder parameters. Through systematic evaluation on our proposed PlanBench-V benchmark, we demonstrate that PlanGPT-VL significantly outperforms general-purpose state-of-the-art VLMs in specialized planning map interpretation tasks, offering urban planning professionals a reliable tool for map analysis, assessment, and educational applications while maintaining high factual accuracy. Our lightweight 7B parameter model achieves comparable performance to models exceeding 72B parameters, demonstrating efficient domain specialization without sacrificing performance.

Code2Logic: Game-Code-Driven Data Synthesis for Enhancing VLMs General Reasoning

Visual-language Chain-of-Thought (CoT) data resources are relatively scarce compared to text-only counterparts, limiting the improvement of reasoning capabilities in Vision Language Models (VLMs). However, high-quality vision-language reasoning data is expensive and labor-intensive to annotate. To address this issue, we leverage a promising resource: game code, which naturally contains logical structures and state transition processes. Therefore, we propose Code2Logic, a novel game-code-driven approach for multimodal reasoning data synthesis. Our approach leverages Large Language Models (LLMs) to adapt game code, enabling automatic acquisition of reasoning processes and results through code execution. Using the Code2Logic approach, we developed the GameQA dataset to train and evaluate VLMs. GameQA is cost-effective and scalable to produce, challenging for state-of-the-art models, and diverse with 30 games and 158 tasks. Surprisingly, despite training solely on game data, VLMs demonstrated out of domain generalization, specifically Qwen2.5-VL-7B improving performance by 2.33\% across 7 diverse vision-language benchmarks. Our code and dataset are available at https://github.com/tongjingqi/Code2Logic.