WebArXiv: Evaluating Multimodal Agents on Time-Invariant arXiv Tasks

Recent progress in large language models (LLMs) has enabled the development of autonomous web agents capable of navigating and interacting with real websites. However, evaluating such agents remains challenging due to the instability and inconsistency of existing benchmarks, which often rely on dynamic content or oversimplified simulations. In this work, we introduce WebArXiv, a static and time-invariant benchmark comprising 275 web-based tasks grounded in the arXiv platform. WebArXiv ensures reproducible and reliable evaluation by anchoring tasks in fixed web snapshots with deterministic ground truths and standardized action trajectories. Through behavioral analysis, we identify a common failure mode, Rigid History Reflection, where agents over-rely on fixed interaction histories. To address this, we propose a lightweight dynamic reflection mechanism that allows agents to selectively retrieve relevant past steps during decision-making. We evaluate ten state-of-the-art web agents on WebArXiv. Results demonstrate clear performance differences across agents and validate the effectiveness of our proposed reflection strategy.

MuBench: Assessment of Multilingual Capabilities of Large Language Models Across 61 Languages

Multilingual large language models (LLMs) are advancing rapidly, with new models frequently claiming support for an increasing number of languages. However, existing evaluation datasets are limited and lack cross-lingual alignment, leaving assessments of multilingual capabilities fragmented in both language and skill coverage. To address this, we introduce MuBench, a benchmark covering 61 languages and evaluating a broad range of capabilities. We evaluate several state-of-the-art multilingual LLMs and find notable gaps between claimed and actual language coverage, particularly a persistent performance disparity between English and low-resource languages. Leveraging MuBench's alignment, we propose Multilingual Consistency (MLC) as a complementary metric to accuracy for analyzing performance bottlenecks and guiding model improvement. Finally, we pretrain a suite of 1.2B-parameter models on English and Chinese with 500B tokens, varying language ratios and parallel data proportions to investigate cross-lingual transfer dynamics.

MEAL: A Benchmark for Continual Multi-Agent Reinforcement Learning

Benchmarks play a crucial role in the development and analysis of reinforcement learning (RL) algorithms, with environment availability strongly impacting research. One particularly underexplored intersection is continual learning (CL) in cooperative multi-agent settings. To remedy this, we introduce MEAL (Multi-agent Environments for Adaptive Learning), the first benchmark tailored for continual multi-agent reinforcement learning (CMARL). Existing CL benchmarks run environments on the CPU, leading to computational bottlenecks and limiting the length of task sequences. MEAL leverages JAX for GPU acceleration, enabling continual learning across sequences of 100 tasks on a standard desktop PC in a few hours. We show that naively combining popular CL and MARL methods yields strong performance on simple environments, but fails to scale to more complex settings requiring sustained coordination and adaptation. Our ablation study identifies architectural and algorithmic features critical for CMARL on MEAL.

Task Adaptation from Skills: Information Geometry, Disentanglement, and New Objectives for Unsupervised Reinforcement Learning

Unsupervised reinforcement learning (URL) aims to learn general skills for unseen downstream tasks. Mutual Information Skill Learning (MISL) addresses URL by maximizing the mutual information between states and skills but lacks sufficient theoretical analysis, e.g., how well its learned skills can initialize a downstream task's policy. Our new theoretical analysis in this paper shows that the diversity and separability of learned skills are fundamentally critical to downstream task adaptation but MISL does not necessarily guarantee these properties. To complement MISL, we propose a novel disentanglement metric LSEPIN. Moreover, we build an information-geometric connection between LSEPIN and downstream task adaptation cost. For better geometric properties, we investigate a new strategy that replaces the KL divergence in information geometry with Wasserstein distance. We extend the geometric analysis to it, which leads to a novel skill-learning objective WSEP. It is theoretically justified to be helpful to downstream task adaptation and it is capable of discovering more initial policies for downstream tasks than MISL. We finally propose another Wasserstein distance-based algorithm PWSEP that can theoretically discover all optimal initial policies.

Revisiting Overthinking in Long Chain-of-Thought from the Perspective of Self-Doubt

Reasoning Large Language Models (RLLMs) have demonstrated impressive performance on complex tasks, largely due to the adoption of Long Chain-of-Thought (Long CoT) reasoning. However, they often exhibit overthinking -- performing unnecessary reasoning steps even after arriving at the correct answer. Prior work has largely focused on qualitative analyses of overthinking through sample-based observations of long CoTs. In contrast, we present a quantitative analysis of overthinking from the perspective of self-doubt, characterized by excessive token usage devoted to re-verifying already-correct answer. We find that self-doubt significantly contributes to overthinking. In response, we introduce a simple and effective prompting method to reduce the model's over-reliance on input questions, thereby avoiding self-doubt. Specifically, we first prompt the model to question the validity of the input question, and then respond concisely based on the outcome of that evaluation. Experiments on three mathematical reasoning tasks and four datasets with missing premises demonstrate that our method substantially reduces answer length and yields significant improvements across nearly all datasets upon 4 widely-used RLLMs. Further analysis demonstrates that our method effectively minimizes the number of reasoning steps and reduces self-doubt.

Monte Carlo Planning with Large Language Model for Text-Based Game Agents

Text-based games provide valuable environments for language-based autonomous agents. However, planning-then-learning paradigms, such as those combining Monte Carlo Tree Search (MCTS) and reinforcement learning (RL), are notably time-consuming due to extensive iterations. Additionally, these algorithms perform uncertainty-driven exploration but lack language understanding and reasoning abilities. In this paper, we introduce the Monte Carlo planning with Dynamic Memory-guided Large language model (MC-DML) algorithm. MC-DML leverages the language understanding and reasoning capabilities of Large Language Models (LLMs) alongside the exploratory advantages of tree search algorithms. Specifically, we enhance LLMs with in-trial and cross-trial memory mechanisms, enabling them to learn from past experiences and dynamically adjust action evaluations during planning. We conduct experiments on a series of text-based games from the Jericho benchmark. Our results demonstrate that the MC-DML algorithm significantly enhances performance across various games at the initial planning phase, outperforming strong contemporary methods that require multiple iterations. This demonstrates the effectiveness of our algorithm, paving the way for more efficient language-grounded planning in complex environments.

HASARD: A Benchmark for Vision-Based Safe Reinforcement Learning in Embodied Agents

Advancing safe autonomous systems through reinforcement learning (RL) requires robust benchmarks to evaluate performance, analyze methods, and assess agent competencies. Humans primarily rely on embodied visual perception to safely navigate and interact with their surroundings, making it a valuable capability for RL agents. However, existing vision-based 3D benchmarks only consider simple navigation tasks. To address this shortcoming, we introduce \textbf{HASARD}, a suite of diverse and complex tasks to $\textbf{HA}$rness $\textbf{SA}$fe $\textbf{R}$L with $\textbf{D}$oom, requiring strategic decision-making, comprehending spatial relationships, and predicting the short-term future. HASARD features three difficulty levels and two action spaces. An empirical evaluation of popular baseline methods demonstrates the benchmark's complexity, unique challenges, and reward-cost trade-offs. Visualizing agent navigation during training with top-down heatmaps provides insight into a method's learning process. Incrementally training across difficulty levels offers an implicit learning curriculum. HASARD is the first safe RL benchmark to exclusively target egocentric vision-based learning, offering a cost-effective and insightful way to explore the potential and boundaries of current and future safe RL methods. The environments and baseline implementations are open-sourced at https://sites.google.com/view/hasard-bench/.

ATLaS: Agent Tuning via Learning Critical Steps

Large Language Model (LLM) agents have demonstrated remarkable generalization capabilities across multi-domain tasks. Existing agent tuning approaches typically employ supervised finetuning on entire expert trajectories. However, behavior-cloning of full trajectories can introduce expert bias and weaken generalization to states not covered by the expert data. Additionally, critical steps, such as planning, complex reasoning for intermediate subtasks, and strategic decision-making, are essential to success in agent tasks, so learning these steps is the key to improving LLM agents. For more effective and efficient agent tuning, we propose ATLaS that identifies the critical steps in expert trajectories and finetunes LLMs solely on these steps with reduced costs. By steering the training's focus to a few critical steps, our method mitigates the risk of overfitting entire trajectories and promotes generalization across different environments and tasks. In extensive experiments, an LLM finetuned on only 30% critical steps selected by ATLaS outperforms the LLM finetuned on all steps and recent open-source LLM agents. ATLaS maintains and improves base LLM skills as generalist agents interacting with diverse environments.

Distill Not Only Data but Also Rewards: Can Smaller Language Models Surpass Larger Ones?

Distilling large language models (LLMs) typically involves transferring the teacher model's responses through supervised fine-tuning (SFT). However, this approach neglects the potential to distill both data (output content) and reward signals (quality evaluations). Extracting reliable reward signals directly from teacher models is challenging, as LLMs are optimized for generation rather than evaluation, often resulting in biased or inconsistent assessments. To address this limitation, we propose a novel distillation pipeline that transfers both responses and rewards. Our method generates pseudo-rewards through a self-supervised mechanism that leverages the inherent structure of both teacher and student responses, enabling reward learning without explicit external evaluation. The reward model subsequently guides reinforcement learning (RL), allowing iterative refinement of the student model after an SFT warm-up phase. Experiments on GSM8K and MMLU-PRO demonstrate that our method consistently outperforms traditional SFT-based approaches, enabling student models to surpass the performance of their teachers. This work highlights the potential for scalable, efficient distillation through structured self-supervised reward learning, reducing dependence on external reward supervision.

Geolocation with Real Human Gameplay Data: A Large-Scale Dataset and Human-Like Reasoning Framework

Geolocation, the task of identifying an image's location, requires complex reasoning and is crucial for navigation, monitoring, and cultural preservation. However, current methods often produce coarse, imprecise, and non-interpretable localization. A major challenge lies in the quality and scale of existing geolocation datasets. These datasets are typically small-scale and automatically constructed, leading to noisy data and inconsistent task difficulty, with images that either reveal answers too easily or lack sufficient clues for reliable inference. To address these challenges, we introduce a comprehensive geolocation framework with three key components: GeoComp, a large-scale dataset; GeoCoT, a novel reasoning method; and GeoEval, an evaluation metric, collectively designed to address critical challenges and drive advancements in geolocation research. At the core of this framework is GeoComp (Geolocation Competition Dataset), a large-scale dataset collected from a geolocation game platform involving 740K users over two years. It comprises 25 million entries of metadata and 3 million geo-tagged locations spanning much of the globe, with each location annotated thousands to tens of thousands of times by human users. The dataset offers diverse difficulty levels for detailed analysis and highlights key gaps in current models. Building on this dataset, we propose Geographical Chain-of-Thought (GeoCoT), a novel multi-step reasoning framework designed to enhance the reasoning capabilities of Large Vision Models (LVMs) in geolocation tasks. GeoCoT improves performance by integrating contextual and spatial cues through a multi-step process that mimics human geolocation reasoning. Finally, using the GeoEval metric, we demonstrate that GeoCoT significantly boosts geolocation accuracy by up to 25% while enhancing interpretability.