Humans engage in lifelong social interactions through interacting with different people under different scenarios for different social goals. This requires social intelligence to gather information through a long time span and use it to navigate various social contexts effectively. Whether AI systems are also capable of this is understudied in the existing research. In this paper, we present a novel benchmark, LIFELONG-SOTOPIA, to perform a comprehensive evaluation of language agents by simulating multi-episode interactions. In each episode, the language agents role-play characters to achieve their respective social goals in randomly sampled social tasks. With LIFELONG-SOTOPIA, we find that goal achievement and believability of all of the language models that we test decline through the whole interaction. Although using an advanced memory method improves the agents' performance, the best agents still achieve a significantly lower goal completion rate than humans on scenarios requiring an explicit understanding of interaction history. These findings show that we can use LIFELONG-SOTOPIA to evaluate the social intelligence of language agents over lifelong social interactions.

Recently, it has been shown that investing computing resources in searching for good initial noise for a text-to-image diffusion model helps improve performance. However, previous studies required external models to evaluate the resulting images, which is impossible on GPUs with small VRAM. For these reasons, we apply Best-of-N inference-time scaling to algorithms that optimize the initial noise of a diffusion model without external models across multiple datasets and backbones. We demonstrate that inference-time scaling for text-to-image diffusion models in this setting quickly reaches a performance plateau, and a relatively small number of optimization steps suffices to achieve the maximum achievable performance with each algorithm.

We present Sparse Interpolated Mixture-of-Experts (SIMoE) instruction-tuning, an end-to-end algorithm designed to fine-tune a dense pre-trained Large Language Model (LLM) into a MoE-style model that possesses capabilities in multiple specialized domains. During instruction-tuning, SIMoE automatically identifies multiple specialized experts under a specified sparsity constraint, with each expert representing a structurally sparse subset of the seed LLM's parameters that correspond to domain-specific knowledge within the data. SIMoE simultaneously learns an input-dependent expert merging strategy via a router network, leveraging rich cross-expert knowledge for superior downstream generalization that surpasses existing baselines. Empirically, SIMoE consistently achieves state-of-the-art performance on common instruction-tuning benchmarks while maintaining an optimal performance-compute trade-off compared to all baselines.

Large Language Models (LLMs) have demonstrated substantial progress on reasoning tasks involving unstructured text, yet their capabilities significantly deteriorate when reasoning requires integrating structured external knowledge such as knowledge graphs, code snippets, or formal logic. This limitation is partly due to the absence of benchmarks capable of systematically evaluating LLM performance across diverse structured knowledge modalities. To address this gap, we introduce \textbf{\textsc{OneEval}}, a comprehensive benchmark explicitly designed to assess the knowledge-intensive reasoning capabilities of LLMs across four structured knowledge modalities, unstructured text, knowledge graphs, code, and formal logic, and five critical domains (general knowledge, government, science, law, and programming). \textsc{OneEval} comprises 4,019 carefully curated instances and includes a challenging subset, \textsc{OneEval}\textsubscript{Hard}, consisting of 1,285 particularly difficult cases. Through extensive evaluation of 18 state-of-the-art open-source and proprietary LLMs, we establish three core findings: a) \emph{persistent limitations in structured reasoning}, with even the strongest model achieving only 32.2\% accuracy on \textsc{OneEval}\textsubscript{Hard}; b) \emph{performance consistently declines as the structural complexity of the knowledge base increases}, with accuracy dropping sharply from 53\% (textual reasoning) to 25\% (formal logic); and c) \emph{diminishing returns from extended reasoning chains}, highlighting the critical need for models to adapt reasoning depth appropriately to task complexity. We release the \textsc{OneEval} datasets, evaluation scripts, and baseline results publicly, accompanied by a leaderboard to facilitate ongoing advancements in structured knowledge reasoning.

In this paper, we introduce DoTA-RAG (Dynamic-of-Thought Aggregation RAG), a retrieval-augmented generation system optimized for high-throughput, large-scale web knowledge indexes. Traditional RAG pipelines often suffer from high latency and limited accuracy over massive, diverse datasets. DoTA-RAG addresses these challenges with a three-stage pipeline: query rewriting, dynamic routing to specialized sub-indexes, and multi-stage retrieval and ranking. We further enhance retrieval by evaluating and selecting a superior embedding model, re-embedding the large FineWeb-10BT corpus. Moreover, we create a diverse Q&A dataset of 500 questions generated via the DataMorgana setup across a broad range of WebOrganizer topics and formats. DoTA-RAG improves the answer correctness score from 0.752 (baseline, using LiveRAG pre-built vector store) to 1.478 while maintaining low latency, and it achieves a 0.929 correctness score on the Live Challenge Day. These results highlight DoTA-RAG's potential for practical deployment in domains requiring fast, reliable access to large and evolving knowledge sources.

Recent advances in zero-shot text-to-speech (TTS) synthesis have achieved high-quality speech generation for unseen speakers, but most systems remain unsuitable for real-time applications because of their offline design. Current streaming TTS paradigms often rely on multi-stage pipelines and discrete representations, leading to increased computational cost and suboptimal system performance. In this work, we propose StreamMel, a pioneering single-stage streaming TTS framework that models continuous mel-spectrograms. By interleaving text tokens with acoustic frames, StreamMel enables low-latency, autoregressive synthesis while preserving high speaker similarity and naturalness. Experiments on LibriSpeech demonstrate that StreamMel outperforms existing streaming TTS baselines in both quality and latency. It even achieves performance comparable to offline systems while supporting efficient real-time generation, showcasing broad prospects for integration with real-time speech large language models. Audio samples are available at: https://aka.ms/StreamMel.

In an age characterized by the proliferation of mis- and disinformation online, it is critical to empower readers to understand the content they are reading. Important efforts in this direction rely on manual or automatic fact-checking, which can be challenging for emerging claims with limited information. Such scenarios can be handled by assessing the reliability and the political bias of the source of the claim, i.e., characterizing entire news outlets rather than individual claims or articles. This is an important but understudied research direction. While prior work has looked into linguistic and social contexts, we do not analyze individual articles or information in social media. Instead, we propose a novel methodology that emulates the criteria that professional fact-checkers use to assess the factuality and political bias of an entire outlet. Specifically, we design a variety of prompts based on these criteria and elicit responses from large language models (LLMs), which we aggregate to make predictions. In addition to demonstrating sizable improvements over strong baselines via extensive experiments with multiple LLMs, we provide an in-depth error analysis of the effect of media popularity and region on model performance. Further, we conduct an ablation study to highlight the key components of our dataset that contribute to these improvements. To facilitate future research, we released our dataset and code at https://github.com/mbzuai-nlp/llm-media-profiling.

Adam is known to perform significantly better than Stochastic Gradient Descent (SGD) in language models, a phenomenon for which a number of explanations have been proposed. In this work, we revisit this "optimizer gap" through a series of comprehensively tuned baseline training runs for language modeling with Transformers. We exhaustively study how momentum, gradient clipping, and batch size affect the gap between SGD and Adam. Our empirical findings show that SGD with momentum can actually perform similarly to Adam in small-batch settings, if tuned correctly. We revisit existing explanations for Adam's advantage, including heavy-tailed class imbalance, directional sharpness, and Hessian heterogeneity, which struggle to directly explain this phenomenon. Towards bridging this gap in our understanding, by analyzing our Transformer training runs and simple quadratic settings inspired by the literature, we provide new insights, driven by stochastic differential equation models, into the role of batch size on the training dynamics.

With the rapid development of large language models (LLMs), they have significantly improved efficiency across a wide range of domains. However, recent studies have revealed that LLMs often exhibit gender bias, leading to serious social implications. Detecting, classifying, and mitigating gender bias in LLMs has therefore become a critical research focus. In the NLPCC 2025 Shared Task 7: Chinese Corpus for Gender Bias Detection, Classification and Mitigation Challenge, we investigate how to enhance the capabilities of LLMs in gender bias detection, classification, and mitigation. We adopt reinforcement learning, chain-of-thoughts (CoT) reasoning, and supervised fine-tuning to handle different Subtasks. Specifically, for Subtasks 1 and 2, we leverage the internal reasoning capabilities of LLMs to guide multi-step thinking in a staged manner, which simplifies complex biased queries and improves response accuracy. For Subtask 3, we employ a reinforcement learning-based approach, annotating a preference dataset using GPT-4. We then apply Direct Preference Optimization (DPO) to mitigate gender bias by introducing a loss function that explicitly favors less biased completions over biased ones. Our approach ranked first across all three subtasks of the NLPCC 2025 Shared Task 7.

Robotic task planning in real-world environments requires not only object recognition but also a nuanced understanding of spatial relationships between objects. We present a spatial-relationship-aware dataset of nearly 1,000 robot-acquired indoor images, annotated with object attributes, positions, and detailed spatial relationships. Captured using a Boston Dynamics Spot robot and labelled with a custom annotation tool, the dataset reflects complex scenarios with similar or identical objects and intricate spatial arrangements. We benchmark six state-of-the-art scene-graph generation models on this dataset, analysing their inference speed and relational accuracy. Our results highlight significant differences in model performance and demonstrate that integrating explicit spatial relationships into foundation models, such as ChatGPT 4o, substantially improves their ability to generate executable, spatially-aware plans for robotics. The dataset and annotation tool are publicly available at https://github.com/PengPaulWang/SpatialAwareRobotDataset, supporting further research in spatial reasoning for robotics.