This paper presents an evolutionary framework for the training of large language models(LLM). The models are divided into several experts(sub-networks), which have the same structure but different parameter values. Only one expert is trained at each step. After the classical AdamW optimization, some evolutionary operators(crossover, PSO, and mutation) act on the tensor weights between the current expert and the best expert. So current expert would learn the experience of best expert. The direction of best expert would help current expert's loss decrease faster. Finally, only save the weight of the best expert. Experiments show that best expert would achieve nearly the same accuracy as the full model. This would greatly reduce the size of the model for inference. Since only one expert is trained at each step, the training needs much less memory and has much higher throughput. Experiments show that the throughput would accelerate more than ten times! Our source code is available. It's a pure c++/cu framework, which is suitable for easy deployment on PCs and edge computing devices.

The emergence of large pre-trained models based on natural language has breathed new life into robotics development. Extensive research has integrated large models with robots, utilizing the powerful semantic understanding and generation capabilities of large models to facilitate robot control through natural language instructions gradually. However, we found that robots that strictly adhere to human instructions, especially those containing misleading information, may encounter errors during task execution, potentially leading to safety hazards. This resembles the concept of counterfactuals in natural language processing (NLP), which has not yet attracted much attention in robotic research. In an effort to highlight this issue for future studies, this paper introduced directive counterfactuals (DCFs) arising from misleading human directives. We present DynaMIC, a framework for generating robot task flows to identify DCFs and relay feedback to humans proactively. This capability can help robots be sensitive to potential DCFs within a task, thus enhancing the reliability of the execution process. We conducted semantic-level experiments and ablation studies, showcasing the effectiveness of this framework.

We present a comprehensive theoretical and empirical study of the Muon optimizer for training transformers only with a small to medium decoder (30M - 200M parameters), with an emphasis on its mathematical foundations, convergence properties and synergistic interactions with modern architectural optimizations. Building on recent work showing Muon's scalability, we provide rigorous theoretical analysis including: (i)showing the convergence rate under standard assumptions, (ii) spectral regularization properties that prevent gradient explosion, (iii) connection to natural gradient descent on the Stiefel manifold, and (iv) equivalence to steepest gradient descent under the spectral norm. Crucially, we demonstrate that Muon expands the Pareto frontier in the compute-time trade-off by maintaining superior data efficiency at large batch sizes, a key finding of~\cite{essentialai2025muon} that we validate across our model scales. Empirically, Muon reaches the target loss with 48-52\% of the training calculated by AdamW while maintaining or improving the final perplexity, consistent with larger-scale results. When combined with Multi-Head Latent Attention (MLA) and Mixture-of-Experts (MoE), we observe multiplicative efficiency gains: MLA+MoE+Muon achieves 68\% memory reduction and 3.2$\times$ inference speedup, while improving perplexity by 8-12\%. We provide detailed procedures on 15 architectural and optimizer components, stability analyzes across 100+ training runs, and practical implementation guidelines including Newton-Schulz coefficients $(3.4445, -4.7750, 2.0315)$ optimized by~\cite{su2024muonblog}. Our theoretical analysis and comprehensive experiments establish Muon as a principled, robust alternative to AdamW that particularly excels when combined with modern efficiency techniques and large-batch training regimes.

Text-to-SQL enables natural access to databases, yet most benchmarks are English-only, limiting multilingual progress. We introduce MultiSpider 2.0, extending Spider 2.0 to eight languages (English, German, French, Spanish, Portuguese, Japanese, Chinese, Vietnamese). It preserves Spider 2.0's structural difficulty while adding linguistic and dialectal variability, demanding deeper reasoning for complex SQL. On this benchmark, state-of-the-art LLMs (such as DeepSeek-R1 and OpenAI o1) reach only 4\% execution accuracy when relying on intrinsic reasoning, versus 60\% on MultiSpider 1.0. Therefore, we provide a collaboration-driven language agents baseline that iteratively refines queries, improving accuracy to 15\%. These results reveal a substantial multilingual gap and motivate methods that are robust across languages and ready for real-world enterprise deployment. Our benchmark is available at https://github.com/phkhanhtrinh23/Multilingual_Text_to_SQL.

Although Large Reasoning Models (LRMs) have progressed in solving complex problems, their chain-of-thought (CoT) reasoning often contains harmful content that can persist even when the final responses appear safe. We show that this issue still remains in existing methods which overlook the unique significance of safe reasoning, undermining their trustworthiness and posing potential risks in applications if unsafe reasoning is accessible for and exploited by malicious users. We therefore shift our focus to aligning the safety of reasoning itself in this paper and explore process supervision as the solution. However, simply rewarding safe reasoning proves inadequate due to low rollout diversity and limited training signals. To tackle this challenge, we first delve into the characteristics of safe reasoning and uncover several critical insights that 1) safe reasoning is often consolidated by a few critical steps of safety triggers; 2) compliance cues strongly correlate with unsafe continuations; and 3) corrective interventions reliably steer unsafe trajectories towards safer traces. Motivated by these, we propose Intervened Preference Optimization (IPO), an alignment method that enforces safe reasoning by substituting compliance steps with safety triggers and constructing pairs for preference learning with strong signals. Experiments on jailbreak and adversarial safety benchmarks demonstrate that IPO remarkably improves overall safety regarding both reasoning and responses, outperforming SFT-based and RL-based baselines with a relative reduction of over 30% in harmfulness, while preserving excellent performance across diverse reasoning tasks. The results highlight the importance of explicit alignment for reasoning and provide a practical path to safer LRMs.

We introduce LLaDA-MoE, a large language diffusion model with the Mixture-of-Experts (MoE) architecture, trained from scratch on approximately 20T tokens. LLaDA-MoE achieves competitive performance with significantly reduced computational overhead by maintaining a 7B-parameter capacity while activating only 1.4B parameters during inference. Our empirical evaluation reveals that LLaDA-MoE achieves state-of-the-art performance among diffusion language models with larger parameters, surpassing previous diffusion language models LLaDA, LLaDA 1.5, and Dream across multiple benchmarks. The instruct-tuned model LLaDA-MoE-7B-A1B-Instruct demonstrates capabilities comparable to Qwen2.5-3B-Instruct in knowledge understanding, code generation, mathematical reasoning, agent and alignment tasks, despite using fewer active parameters. Our results show that integrating a sparse MoE architecture into the training objective of masked diffusion language models still brings out MoE's strengths under efficient inference with few active parameters, and opens ample room for further exploration of diffusion language models. LLaDA-MoE models are available at Huggingface.

Vision Language Navigation (VLN) requires agents to follow natural language instructions by grounding them in sequential visual observations over long horizons. Explicit reasoning could enhance temporal consistency and perception action alignment, but reasoning at fixed steps often leads to suboptimal performance and unnecessary computation. To address this, we propose AdaNav, an uncertainty-based adaptive reasoning framework for VLN. At its core is the Uncertainty Adaptive Reasoning Block (UAR), a lightweight plugin that dynamically triggers reasoning. We introduce Action Entropy as a policy prior for UAR and progressively refine it through a Heuristics to RL training method, enabling agents to learn difficulty aware reasoning policies under the strict data limitations of embodied tasks. Results show that with only 6K training samples, AdaNav achieves substantial gains over closed source models trained on million scale data, improving success rate by 20% on R2R val-unseen, 11.7% on RxR-CE, and 11.4% in real world scenes. The code is available at https://github.com/xinding-sys/AdaNav.

Recent advances in Large Language Models (LLMs) (V aswani et al., 2017; Radford et al., 2019; Naveed et al., 2025) and Multimodal Large Language Models (MLLMs) (Zhang et al., 2024a; Yin et al., 2024; Wu et al., 2023) have reinforced the paradigm of language as a universal interface, substantially improving cross-modal perception and reasoning. Extending this progress to 3D, recent research has focused on 3D-aware Multimodal Large Language Models (3D-MLLMs) (Ren et al., 2025), which unify 3D scene understanding and vision-language reasoning under a linguistic interface. This line of work underscores the importance of grounding language in persistent 3D spatial representations (Cheng et al., 2024a; Roh et al., 2022), offering a unified pathway toward systematic scene-level reasoning. Recent studies have made substantial progress in 3D vision-language (3D VL) reasoning (Chen et al., 2024c; Huang et al., 2023b), yet most approaches rely on complex 3D inputs, incurring high costs in data collection, preprocessing, and computation. Some models rely on point clouds or reconstructed scenes augmented with rendered views or semantic-geometric features (Hong et al., 2023a; Fu et al., 2024), while others adopt simpler inputs but still depend on explicit 3D scene representations such as reconstructed objects aligned with semantic representations (Chu et al., 2024; Huang et al., 2023a; 2024). Despite their effectiveness, these pipelines depend on depth, poses, or external modules, leading to substantial data and engineering overhead as well as high memory and latency costs.

The alignment of large language models (LLMs) with human values is critical for their safe deployment, yet jailbreak attacks can subvert this alignment to elicit harmful outputs from LLMs. In recent years, a proliferation of jailbreak attacks has emerged, accompanied by diverse metrics and judges to assess the harmfulness of the LLM outputs. However, the absence of a systematic benchmark to assess the quality and effectiveness of these metrics and judges undermines the credibility of the reported jailbreak effectiveness and other risks. To address this gap, we introduce HarmMetric Eval, a comprehensive benchmark designed to support both overall and fine-grained evaluation of harmfulness metrics and judges. Our benchmark includes a high-quality dataset of representative harmful prompts paired with diverse harmful and non-harmful model responses, alongside a flexible scoring mechanism compatible with various metrics and judges. With HarmMetric Eval, our extensive experiments uncover a surprising result: two conventional metrics--METEOR and ROUGE-1--outperform LLM-based judges in evaluating the harmfulness of model responses, challenging prevailing beliefs about LLMs' superiority in this domain. Our dataset is publicly available at https://huggingface.co/datasets/qusgo/HarmMetric_Eval, and the code is available at https://anonymous.4open.science/r/HarmMetric-Eval-4CBE.

Existing methods usually leverage a fixed strategy, such as natural language reasoning, code-augmented reasoning, tool-integrated reasoning, or ensemble-based reasoning, to guide Large Language Models (LLMs) to perform mathematical reasoning. Our analysis reveals that the single strategy cannot adapt to problem-specific requirements and thus overlooks the trade-off between effectiveness and efficiency. To address these issues, we propose Planning and Routing through Instance-Specific Modeling (PRISM), a novel framework that decouples mathematical reasoning into two stages: strategy planning and targeted execution. Specifically, we first curate a multi-strategy preference dataset, which we call MathStrat, capturing correctness, process quality, and computational efficiency for each problem--strategy pair. Then, we train a lightweight Strategy Adapter based on the dataset to obtain confidence distributions over the mentioned four reasoning strategies. At inference time, an adaptive routing policy dynamically tailors the reasoning approach based on predictor confidence. It directs the model to use single-strategy execution for high-confidence predictions, dual-strategy verification for competitive scenarios, or comprehensive multi-strategy exploration for uncertain cases. Extensive experiments across five mathematical reasoning benchmarks demonstrate that PRISM consistently outperforms individual strategies and ensemble baselines, achieving improvements ranging from 0.9% to 7.6% across different base models. The adaptive routing approach shows particularly strong benefits for mathematical reasoning tasks across diverse model architectures. Our code is released at https://github.com/reml-group/PRISM.