Large Language Models (LLMs) have shown strong reasoning capabilities, with models like OpenAI's O-series and DeepSeek R1 excelling at tasks such as mathematics, coding, logic, and puzzles through Reinforcement Learning with Verifiable Rewards (RLVR). However, their ability to solve more complex optimization problems - particularly NP-hard tasks - remains underexplored. To bridge this gap, we propose NP-ENGINE, the first comprehensive framework for training and evaluating LLMs on NP-hard problems. NP-ENGINE covers 10 tasks across five domains, each equipped with (i) a controllable instance generator, (ii) a rule-based verifier, and (iii) a heuristic solver that provides approximate optimal solutions as ground truth. This generator-verifier-heuristic pipeline enables scalable and verifiable RLVR training under hierarchical difficulties. We also introduce NP-BENCH, a benchmark derived from NP-ENGINE-DATA, specifically designed to evaluate LLMs' ability to tackle NP-hard level reasoning problems, focusing not only on feasibility but also on solution quality. Additionally, we present QWEN2.5-7B-NP, a model trained via zero-RLVR with curriculum learning on Qwen2.5-7B-Instruct, which significantly outperforms GPT-4o on NP-BENCH and achieves SOTA performance with the same model size. Beyond in-domain tasks, we demonstrate that RLVR training on NP-ENGINE-DATA enables strong out-of-domain (OOD) generalization to reasoning tasks (logic, puzzles, math, and knowledge), as well as non-reasoning tasks such as instruction following. We also observe a scaling trend: increasing task diversity improves OOD generalization. These findings suggest that task-rich RLVR training is a promising direction for advancing LLM's reasoning ability, revealing new insights into the scaling laws of RLVR.

In many industrial settings, users wish to ask questions in natural language, the answers to which require assembling information from diverse structured data sources. With the advent of Large Language Models (LLMs), applications can now translate natural language questions into a set of API calls or database calls, execute them, and combine the results into an appropriate natural language response. However, these applications remain impractical in realistic industrial settings because they do not cope with the data source heterogeneity that typifies such environments. In this work, we simulate the heterogeneity of real industry settings by introducing two extensions of the popular Spider benchmark dataset that require a combination of database and API calls. Then, we introduce a declarative approach to handling such data heterogeneity and demonstrate that it copes with data source heterogeneity significantly better than state-of-the-art LLM-based agentic or imperative code generation systems. Our augmented benchmarks are available to the research community.

Large language models (LLMs) have shown remarkable progress in complex reasoning tasks, largely enabled by test-time scaling (TTS) paradigms that allocate additional compute during inference. Among these, external TTS (particularly the Best-of-N selection paradigm) yields scalable performance improvements by selecting from multiple independently generated reasoning trajectories. However, this approach faces key limitations: (i) the high computational overhead of deploying process reward models, (ii) the underutilization of the LLM's intrinsic latent representations. We introduce TrajSelector, an efficient and effective Best-of-N framework that exploit the hidden states in the sampler LLM for process-level scoring. A lightweight verifier (with only 0.6B parameters) evaluates the quality of step-wise trajectory, and then aggregates these scores to identify the optimal reasoning trajectory. Our framework employs a fully data-driven, end-to-end training recipe that eliminates reliance on massive step-level annotations. Experiential results across five benchmarks demonstrate that TrajSelector delivers consistent performance gains. In Best-of-32 settings, it surpasses majority voting by 4.61% accuracy and outperforms existing process reward models by 4.31% to 12.21%, all while maintaining lower inference costs.

Sparse Mixture of Experts (sMoE) has become a pivotal approach for scaling large vision-language models, offering substantial capacity while maintaining computational efficiency through dynamic, sparse activation of experts. However, existing routing mechanisms, typically based on similarity scoring, struggle to effectively capture the underlying input structure. This limitation leads to a trade-off between expert specialization and balanced computation, hindering both scalability and performance. We propose Input Domain Aware MoE, a novel routing framework that leverages a probabilistic mixture model to better partition the input space. By modeling routing probabilities as a mixture of distributions, our method enables experts to develop clear specialization boundaries while achieving balanced utilization. Unlike conventional approaches, our routing mechanism is trained independently of task-specific objectives, allowing for stable optimization and decisive expert assignments. Empirical results on vision-language tasks demonstrate that our method consistently outperforms existing sMoE approaches, achieving higher task performance and improved expert utilization balance.

Referring Atomic Video Action Recognition (RAVAR) aims to recognize fine-grained, atomic-level actions of a specific person of interest conditioned on natural language descriptions. Distinct from conventional action recognition and detection tasks, RAVAR emphasizes precise language-guided action understanding, which is particularly critical for interactive human action analysis in complex multi-person scenarios. In this work, we extend our previously introduced RefAVA dataset to RefAVA++, which comprises >2.9 million frames and >75.1k annotated persons in total. We benchmark this dataset using baselines from multiple related domains, including atomic action localization, video question answering, and text-video retrieval, as well as our earlier model, RefAtomNet. Although RefAtomNet surpasses other baselines by incorporating agent attention to highlight salient features, its ability to align and retrieve cross-modal information remains limited, leading to suboptimal performance in localizing the target person and predicting fine-grained actions. To overcome the aforementioned limitations, we introduce RefAtomNet++, a novel framework that advances cross-modal token aggregation through a multi-hierarchical semantic-aligned cross-attention mechanism combined with multi-trajectory Mamba modeling at the partial-keyword, scene-attribute, and holistic-sentence levels. In particular, scanning trajectories are constructed by dynamically selecting the nearest visual spatial tokens at each timestep for both partial-keyword and scene-attribute levels. Moreover, we design a multi-hierarchical semantic-aligned cross-attention strategy, enabling more effective aggregation of spatial and temporal tokens across different semantic hierarchies. Experiments show that RefAtomNet++ establishes new state-of-the-art results. The dataset and code are released at https://github.com/KPeng9510/refAVA2.

With the growing use of large language models and conversational interfaces in human-robot interaction, robots' ability to answer user questions is more important than ever. We therefore introduce a dataset of 1,893 user questions for household robots, collected from 100 participants and organized into 12 categories and 70 subcategories. Most work in explainable robotics focuses on why-questions. In contrast, our dataset provides a wide variety of questions, from questions about simple execution details to questions about how the robot would act in hypothetical scenarios -- thus giving roboticists valuable insights into what questions their robot needs to be able to answer. To collect the dataset, we created 15 video stimuli and 7 text stimuli, depicting robots performing varied household tasks. We then asked participants on Prolific what questions they would want to ask the robot in each portrayed situation. In the final dataset, the most frequent categories are questions about task execution details (22.5%), the robot's capabilities (12.7%), and performance assessments (11.3%). Although questions about how robots would handle potentially difficult scenarios and ensure correct behavior are less frequent, users rank them as the most important for robots to be able to answer. Moreover, we find that users who identify as novices in robotics ask different questions than more experienced users. Novices are more likely to inquire about simple facts, such as what the robot did or the current state of the environment. As robots enter environments shared with humans and language becomes central to giving instructions and interaction, this dataset provides a valuable foundation for (i) identifying the information robots need to log and expose to conversational interfaces, (ii) benchmarking question-answering modules, and (iii) designing explanation strategies that align with user expectations.

Sparse Mixture of Experts (SMoE) has emerged as a promising solution to achieving unparalleled scalability in deep learning by decoupling model parameter count from computational cost. By activating only a small subset of parameters per sample, SMoE enables significant growth in model capacity while maintaining efficiency. However, SMoE struggles to adapt to distributional shifts, leading to reduced robustness under data contamination. In this work, we introduce SymphonySMoE, a novel family of SMoE that introduces a social graph to model interactions among experts. This graph-based structure enhances the token routing process, addressing the robustness challenges that are inherent in conventional SMoE designs. SymphonySMoE is lightweight, modular, and integrates seamlessly with existing SMoE-based models such as the XMoE and the Generalist Language Model. We provide both theoretical analysis and empirical evidence demonstrating SymphonySMoE's advantages over baseline SMoE. Extensive experiments on language modeling and visual instruction tuning validate our method's effectiveness. We further highlight the scalability of SymphonySMoE to models with 4.2 and 7.4 billion parameters, showcasing its applicability in fine-tuning tasks for large-scale systems.

Personalized and continuous interactions are the key to enhancing user experience in today's large language model (LLM)-based conversational systems, however, the finite context windows and static parametric memory make it difficult to model the cross-session long-term user states and behavioral consistency. Currently, the existing solutions to this predicament, such as retrieval-augmented generation (RAG) and explicit memory systems, primarily focus on fact-level storage and retrieval, lacking the capability to distill latent preferences and deep traits from the multi-turn dialogues, which limits the long-term and effective user modeling, directly leading to the personalized interactions remaining shallow, and hindering the cross-session continuity. To realize the long-term memory and behavioral consistency for Language Agents in LLM era, we propose a self-evolving memory framework RGMem, inspired by the ideology of classic renormalization group (RG) in physics, this framework enables to organize the dialogue history in multiple scales: it first extracts semantics and user insights from episodic fragments, then through hierarchical coarse-graining and rescaling operations, progressively forms a dynamically-evolved user profile. The core innovation of our work lies in modeling memory evolution as a multi-scale process of information compression and emergence, which accomplishes the high-level and accurate user profiles from noisy and microscopic-level interactions.

Large Language Models (LLMs) have demonstrated impressive capabilities, yet their deployment in high-stakes domains is hindered by inherent limitations in trustworthiness, including hallucinations, instability, and a lack of transparency. To address these challenges, we introduce a generic neuro-symbolic approach, which we call Autonomous Trustworthy Agents (ATA). The core of our approach lies in decoupling tasks into two distinct phases: Offline knowledge ingestion and online task processing. During knowledge ingestion, an LLM translates an informal problem specification into a formal, symbolic knowledge base. This formal representation is crucial as it can be verified and refined by human experts, ensuring its correctness and alignment with domain requirements. In the subsequent task processing phase, each incoming input is encoded into the same formal language. A symbolic decision engine then utilizes this encoded input in conjunction with the formal knowledge base to derive a reliable result. Through an extensive evaluation on a complex reasoning task, we demonstrate that a concrete implementation of ATA is competitive with state-of-the-art end-to-end reasoning models in a fully automated setup while maintaining trustworthiness. Crucially, with a human-verified and corrected knowledge base, our approach significantly outperforms even larger models, while exhibiting perfect determinism, enhanced stability against input perturbations, and inherent immunity to prompt injection attacks. By generating decisions grounded in symbolic reasoning, ATA offers a practical and controllable architecture for building the next generation of transparent, auditable, and reliable autonomous agents.

As AI systems progress, we rely more on them to make decisions with us and for us. To ensure that such decisions are aligned with human values, it is imperative for us to understand not only what decisions they make but also how they come to those decisions. Reasoning language models, which provide both final responses and (partially transparent) intermediate thinking traces, present a timely opportunity to study AI procedural reasoning. Unlike math and code problems which often have objectively correct answers, moral dilemmas are an excellent testbed for process-focused evaluation because they allow for multiple defensible conclusions. To do so, we present MoReBench: 1,000 moral scenarios, each paired with a set of rubric criteria that experts consider essential to include (or avoid) when reasoning about the scenarios. MoReBench contains over 23 thousand criteria including identifying moral considerations, weighing trade-offs, and giving actionable recommendations to cover cases on AI advising humans moral decisions as well as making moral decisions autonomously. Separately, we curate MoReBench-Theory: 150 examples to test whether AI can reason under five major frameworks in normative ethics. Our results show that scaling laws and existing benchmarks on math, code, and scientific reasoning tasks fail to predict models' abilities to perform moral reasoning. Models also show partiality towards specific moral frameworks (e.g., Benthamite Act Utilitarianism and Kantian Deontology), which might be side effects of popular training paradigms. Together, these benchmarks advance process-focused reasoning evaluation towards safer and more transparent AI.