Scaling laws play a central role in the success of Large Language Models (LLMs), enabling the prediction of model performance relative to compute budgets prior to training. While Transformers have been the dominant architecture, recent alternatives such as xLSTM offer linear complexity with respect to context length while remaining competitive in the billion-parameter regime. We conduct a comparative investigation on the scaling behavior of Transformers and xLSTM along the following lines, providing insights to guide future model design and deployment. First, we study the scaling behavior for xLSTM in compute-optimal and over-training regimes using both IsoFLOP and parametric fit approaches on a wide range of model sizes (80M-7B) and number of training tokens (2B-2T). Second, we examine the dependence of optimal model sizes on context length, a pivotal aspect that was largely ignored in previous work. Finally, we analyze inference-time scaling characteristics. Our findings reveal that in typical LLM training and inference scenarios, xLSTM scales favorably compared to Transformers. Importantly, xLSTM's advantage widens as training and inference contexts grow.

Mass spectrometry metabolomics generates vast amounts of data requiring advanced methods for interpretation. Knowledge graphs address these challenges by structuring mass spectrometry data, metabolite information, and their relationships into a connected network (Gaudry et al. 2024). However, effective use of a knowledge graph demands an in-depth understanding of its ontology and its query language syntax. To overcome this, we designed MetaboT, an AI system utilizing large language models (LLMs) to translate user questions into SPARQL semantic query language for operating on knowledge graphs (Steve Harris 2013). We demonstrate its effectiveness using the Experimental Natural Products Knowledge Graph (ENPKG), a large-scale public knowledge graph for plant natural products (Gaudry et al. 2024).MetaboT employs specialized AI agents for handling user queries and interacting with the knowledge graph by breaking down complex tasks into discrete components, each managed by a specialised agent (Fig. 1a). The multi-agent system is constructed using the LangChain and LangGraph libraries, which facilitate the integration of LLMs with external tools and information sources (LangChain, n.d.). The query generation process follows a structured workflow. First, the Entry Agent determines if the question is new or a follow-up to previous interactions. New questions are forwarded to the Validator Agent, which verifies if the question is related to the knowledge graph. Then, the valid question is sent to the Supervisor Agent, which identifies if the question requires chemical conversions or standardized identifiers. In this case it delegates the question to the Knowledge Graph Agent, which can use tools to extract necessary details, such as URIs or taxonomies of chemical names, from the user query. Finally, an agent responsible for crafting the SPARQL queries equipped with the ontology of the knowledge graph uses the provided identifiers to generate the query. Then, the system executes the generated query against the metabolomics knowledge graph and returns structured results to the user (Fig. 1b). To assess the performance of MetaboT we have curated 50 metabolomics-related questions and their expected answers. In addition to submitting these questions to MetaboT, we evaluated a baseline by submitting them to a standard LLM (GPT-4o) with a prompt that incorporated the knowledge graph ontology but did not provide specific entity IDs. This baseline achieved only 8.16% accuracy, compared to MetaboT's 83.67%, underscoring the necessity of our multi-agent system for accurately retrieving entities and generating correct SPARQL queries. MetaboT demonstrates promising performance as a conversational question-answering assistant, enabling researchers to retrieve structured metabolomics data through natural language queries. By automating the generation and execution of SPARQL queries, it removes technical barriers that have traditionally hindered access to knowledge graphs. Importantly, MetaboT leverages the capabilities of LLMs while maintaining experimentally grounded query generation, ensuring that outputs remain aligned with domain-specific standards and data structures. This approach facilitates data-driven discoveries by bridging the gap between complex semantic technologies and user-friendly interaction. MetaboT is accessible at [https://metabot.holobiomicslab.eu/], and its source code is available at [https://github.com/HolobiomicsLab/MetaboT].

Reasoning protocols such as Chain of Thought (CoT) and Tree of Thought (ToT) organize internal deliberation but lack an explicit mechanism for external questioning that elicits self-revision. We present FOR-Prompting (From Objection to Revision Prompting), an asymmetric protocol where a Defender proposes an answer, an Objectioner raises question-style objections with no direct fixes, and a Host enforces consistency and closure. On GSM8K we observe about a 22% point gain over single-prompt and accuracy on par with CoT, with more than 10% higher ratings in reasoning and coherence from a uniform GPT 4.1 judge. FOR-Prompting also corrects mistakes without tools or human supervision on tricky queries, and improves performance for small-scale model (approx. 19% accuracy improved on Llama3.2:1b for GSM8K task), highlighting promise for small models and on personal device use. Beyond factual QA, qualitative analyses on open-ended tasks show enhanced exploration and refinement, with dialogue traces that make assumptions and trade-offs explicit. The protocol is model agnostic and operates purely at the prompt level through role-structured turns, so it works with hosted and local models of different sizes without retraining, and it supports large-scale study of objection-guided reasoning.

Large language models (LLMs) demonstrate strong mathematical reasoning, but reliance on closed-source APIs for OR tasks raises privacy concerns, and training open-source models from scratch incurs high compute costs. We introduce OR-Toolformer, which fine-tunes Llama-3.1-8B-Instruct with a semi-automatic data synthesis pipeline that generates diverse OR problem-answer pairs and augments the model with external solvers to produce API calls. On three of four standard benchmarks, OR-Toolformer achieves up to 80.1% execution accuracy, exceeding size-matched baselines by over 4.3%. In zero-shot evaluation on two unseen OR problem types, it attains 54% average accuracy, a 21 percentage-point improvement over the strongest baseline. These findings validate the efficacy of tool-augmented fine-tuning LLMs for accurate and generalizable OR problem modeling and solving.

Large Reasoning Models (LRMs) with long chain-of-thought capabilities, optimized via reinforcement learning with verifiable rewards (RLVR), excel at objective reasoning tasks like mathematical problem solving and code generation. However, RLVR is known for degrading generation diversity, which causes LRMs to fall short on subjective reasoning that has multiple answers depending on different role perspectives. While recent studies recognize the importance of diversity-enhanced training in objective reasoning, limited attention has been given to subjective tasks. In this paper, we find that subjective reasoning can be improved by introducing perspective diversity and token-level diversity, with the former one providing a coherent scaffolding anchored to a real-world stakeholder group and the latter one broadening the answer search space. We propose MultiRole-R1, a diversity-enhanced training framework featuring an unsupervised data construction pipeline that synthesizes reasoning chains incorporating various role perspectives. It also employs reinforcement learning via Group Relative Policy Optimization with reward shaping, taking diversity as a reward signal in addition to verifiable reward. Training on subjective tasks solely, MultiRole-R1 increases the in-domain and out-of-domain accuracy by 14.1% and 7.64%, and even enhances the performance on advanced math reasoning such as AIME 2024. We further show that diversity is a more consistent indicator of accuracy than reasoning length.

Whole-slide images (WSIs) in pathology can reach up to 10,000 x 10,000 pixels, posing significant challenges for multimodal large language model (MLLM) due to long context length and high computational demands. Previous methods typically focus on patch-level analysis or slide-level classification using CLIP-based models with multi-instance learning, but they lack the generative capabilities needed for visual question answering (VQA). More recent MLLM-based approaches address VQA by feeding thousands of patch tokens directly into the language model, which leads to excessive resource consumption. To address these limitations, we propose Token Compression Pathology LLaVA (TCP-LLaVA), the first MLLM architecture to perform WSI VQA via token compression. TCP-LLaVA introduces a set of trainable compression tokens that aggregate visual and textual information through a modality compression module, inspired by the [CLS] token mechanism in BERT. Only the compressed tokens are forwarded to the LLM for answer generation, significantly reducing input length and computational cost. Experiments on ten TCGA tumor subtypes show that TCP-LLaVA outperforms existing MLLM baselines in VQA accuracy while reducing training resource consumption by a substantial margin.

Knowledge distillation (KD) has become a cornerstone for compressing large language models (LLMs). However, existing LLM-KD methods have primarily focused on logit-based approaches, which achieve good performance but overlook the rich internal representations of LLMs. Feature-level KD could leverage this structure to provide complementary benefits, yet it remains underexplored because current feature-KD approaches typically assume identical teacher-student hidden sizes, a restrictive and unrealistic assumption. A common workaround is to train a linear projector to align their feature spaces; however, this introduces additional parameters, distorts teacher embeddings, and often degrades downstream performance, especially in generative tasks. We propose Flex-KD, a parameter-free framework for task-driven feature distillation for LLMs. Instead of projecting the entire teacher representation, Flex-KD uses gradient-based scores to identify the most task-relevant dimensions of the teacher's hidden states and distills only this subspace into the student. This ensures that the student's limited capacity is allocated to informative components, while avoiding projector-induced distortion and extra parameters. Flex-KD integrates seamlessly with existing KD pipelines and supports differing teacher-student hidden sizes. Extensive experiments across both classification and generative tasks, i.e., instruction-following and summarization, show that Flex-KD consistently boosts student performance, achieving up to a 3.75 percent performance gain over the linear projection baseline.

Large foundation models are typically trained on data from multiple domains, with the data mixture--the proportion of each domain used--playing a critical role in model performance. The standard approach to selecting this mixture relies on trial and error, which becomes impractical for large-scale pretraining. We propose a systematic method to determine the optimal data mixture for any target domain using scaling laws. Our approach accurately predicts the loss of a model of size $N$ trained with $D$ tokens and a specific domain weight vector $h$. We validate the universality of these scaling laws by demonstrating their predictive power in three distinct and large-scale settings: large language model (LLM), native multimodal model (NMM), and large vision models (LVM) pretraining. We further show that these scaling laws can extrapolate to new data mixtures and across scales: their parameters can be accurately estimated using a few small-scale training runs, and used to estimate the performance at larger scales and unseen domain weights. The scaling laws allow to derive the optimal domain weights for any target domain under a given training budget ($N$,$D$), providing a principled alternative to costly trial-and-error methods.

While slow-thinking large language models (LLMs) exhibit reflection-like reasoning, commonly referred to as the "aha moment:, their ability to generate informative critiques and refine prior solutions remains limited. In this paper, we introduce Double-Checker, a principled framework designed to enhance the reasoning capabilities of slow-thinking LLMs by fostering explicit self-critique and iterative refinement of their previous solutions. By fine-tuning on our curated 1,730 self-critical instances, Double-Checker empowers long-CoT LLMs to iteratively critique and refine their outputs during inference until they evaluate their solutions as correct under self-generated critiques. We validate the efficacy of Double-Checker across a comprehensive suite of reasoning benchmarks, demonstrating that iterative self-critique significantly enhances the reasoning capabilities of long-CoT LLMs. Notably, our Double-Checker increases the pass@1 performance on challenging AIME benchmarks from 4.4% to 18.2% compared to the original long-CoT LLMs. These results highlight a promising direction for developing more trustworthy and effective LLMs capable of structured self-critique. Our codes and data are available at https://github.com/XinXU-USTC/DoubleChecker

Direct Preference Optimization (DPO) has emerged as an effective approach for aligning large language models (LLMs) with human preferences. However, its performance is highly dependent on the quality of the underlying human preference data. To address this bottleneck, prior work has explored various data selection strategies, but these methods often overlook the impact of the evolving states of the language model during the optimization process. In this paper, we introduce a novel problem: Sample Scheduling for DPO, which aims to dynamically and adaptively schedule training samples based on the model's evolving batch-wise states throughout preference optimization. To solve this problem, we propose SamS, an efficient and effective algorithm that adaptively selects samples in each training batch based on the LLM's learning feedback to maximize the potential generalization performance. Notably, without modifying the core DPO algorithm, simply integrating SamS significantly improves performance across tasks, with minimal additional computational overhead. This work points to a promising new direction for improving LLM alignment through batch-wise sample selection, with potential generalization to RLHF and broader supervised learning paradigms.