mFabric: An Efficient and Scalable Fabric for Mixture-of-Experts Training

Mixture-of-Expert (MoE) models outperform conventional models by selectively activating different subnets, named \emph{experts}, on a per-token basis. This gated computation generates dynamic communications that cannot be determined beforehand, challenging the existing GPU interconnects that remain \emph{static} during the distributed training process. In this paper, we advocate for a first-of-its-kind system, called mFabric, that unlocks topology reconfiguration \emph{during} distributed MoE training. Towards this vision, we first perform a production measurement study and show that the MoE dynamic communication pattern has \emph{strong locality}, alleviating the requirement of global reconfiguration. Based on this, we design and implement a \emph{regionally reconfigurable high-bandwidth domain} on top of existing electrical interconnects using optical circuit switching (OCS), achieving scalability while maintaining rapid adaptability. We have built a fully functional mFabric prototype with commodity hardware and a customized collective communication runtime that trains state-of-the-art MoE models with \emph{in-training} topology reconfiguration across 32 A100 GPUs. Large-scale packet-level simulations show that mFabric delivers comparable performance as the non-blocking fat-tree fabric while boosting the training cost efficiency (e.g., performance per dollar) of four representative MoE models by 1.2$\times$--1.5$\times$ and 1.9$\times$--2.3$\times$ at 100 Gbps and 400 Gbps link bandwidths, respectively.

InfiFusion: A Unified Framework for Enhanced Cross-Model Reasoning via LLM Fusion

Large Language Models (LLMs) have demonstrated strong performance across various reasoning tasks, yet building a single model that consistently excels across all domains remains challenging. This paper addresses this problem by exploring strategies to integrate multiple domain-specialized models into an efficient pivot model.We propose two fusion strategies to combine the strengths of multiple LLMs: (1) a pairwise, multi-step fusion approach that sequentially distills each source model into the pivot model, followed by a weight merging step to integrate the distilled models into the final model. This method achieves strong performance but requires substantial training effort; and (2) a unified fusion approach that aggregates all source models' outputs simultaneously.To improve the fusion process, we introduce a novel Rate-Skewness Adaptive Fusion (RSAF) technique, which dynamically adjusts top-K ratios during parameter merging for enhanced flexibility and stability.Furthermore, we propose an uncertainty-based weighting method for the unified approach, which dynamically balances the contributions of source models and outperforms other logits/distribution ensemble methods.We achieved accuracy improvements of 9.27%, 8.80%, and 8.89% on the GSM8K, MATH, and HumanEval tasks, respectively.

IIMedGPT: Promoting Large Language Model Capabilities of Medical Tasks by Efficient Human Preference Alignment

Recent researches of large language models(LLM), which is pre-trained on massive general-purpose corpora, have achieved breakthroughs in responding human queries. However, these methods face challenges including limited data insufficiency to support extensive pre-training and can not align responses with users' instructions. To address these issues, we introduce a medical instruction dataset, CMedINS, containing six medical instructions derived from actual medical tasks, which effectively fine-tunes LLM in conjunction with other data. Subsequently, We launch our medical model, IIMedGPT, employing an efficient preference alignment method, Direct preference Optimization(DPO). The results show that our final model outperforms existing medical models in medical dialogue.Datsets, Code and model checkpoints will be released upon acceptance.

IllusionBench: A Large-scale and Comprehensive Benchmark for Visual Illusion Understanding in Vision-Language Models

Current Visual Language Models (VLMs) show impressive image understanding but struggle with visual illusions, especially in real-world scenarios. Existing benchmarks focus on classical cognitive illusions, which have been learned by state-of-the-art (SOTA) VLMs, revealing issues such as hallucinations and limited perceptual abilities. To address this gap, we introduce IllusionBench, a comprehensive visual illusion dataset that encompasses not only classic cognitive illusions but also real-world scene illusions. This dataset features 1,051 images, 5,548 question-answer pairs, and 1,051 golden text descriptions that address the presence, causes, and content of the illusions. We evaluate ten SOTA VLMs on this dataset using true-or-false, multiple-choice, and open-ended tasks. In addition to real-world illusions, we design trap illusions that resemble classical patterns but differ in reality, highlighting hallucination issues in SOTA models. The top-performing model, GPT-4o, achieves 80.59% accuracy on true-or-false tasks and 76.75% on multiple-choice questions, but still lags behind human performance. In the semantic description task, GPT-4o's hallucinations on classical illusions result in low scores for trap illusions, even falling behind some open-source models. IllusionBench is, to the best of our knowledge, the largest and most comprehensive benchmark for visual illusions in VLMs to date.

Beyond Training: Dynamic Token Merging for Zero-Shot Video Understanding

Recent advancements in multimodal large language models (MLLMs) have opened new avenues for video understanding. However, achieving high fidelity in zero-shot video tasks remains challenging. Traditional video processing methods rely heavily on fine-tuning to capture nuanced spatial-temporal details, which incurs significant data and computation costs. In contrast, training-free approaches, though efficient, often lack robustness in preserving context-rich features across complex video content. To this end, we propose DYTO, a novel dynamic token merging framework for zero-shot video understanding that adaptively optimizes token efficiency while preserving crucial scene details. DYTO integrates a hierarchical frame selection and a bipartite token merging strategy to dynamically cluster key frames and selectively compress token sequences, striking a balance between computational efficiency with semantic richness. Extensive experiments across multiple benchmarks demonstrate the effectiveness of DYTO, achieving superior performance compared to both fine-tuned and training-free methods and setting a new state-of-the-art for zero-shot video understanding.

Llama Guard 3 Vision: Safeguarding Human-AI Image Understanding Conversations

We introduce Llama Guard 3 Vision, a multimodal LLM-based safeguard for human-AI conversations that involves image understanding: it can be used to safeguard content for both multimodal LLM inputs (prompt classification) and outputs (response classification). Unlike the previous text-only Llama Guard versions (Inan et al., 2023; Llama Team, 2024b,a), it is specifically designed to support image reasoning use cases and is optimized to detect harmful multimodal (text and image) prompts and text responses to these prompts. Llama Guard 3 Vision is fine-tuned on Llama 3.2-Vision and demonstrates strong performance on the internal benchmarks using the MLCommons taxonomy. We also test its robustness against adversarial attacks. We believe that Llama Guard 3 Vision serves as a good starting point to build more capable and robust content moderation tools for human-AI conversation with multimodal capabilities.

Polymetis:Large Language Modeling for Multiple Material Domains

As the application of large language models in various fields continues to expand, materials science also ushers in opportunities for AI-driven innovation. The traditional way of relying on manual search for materials science-related information is now using artificial intelligence technology as an auxiliary tool to improve the efficiency of materials science research. To accelerate researchers' knowledge acquisition and intelligent decision-making support in materials science research, this paper proposes a large language model Polymetis model for a variety of materials fields, aiming to provide highly professional knowledge answers in the field of materials, covering energy materials, functional materials, alloy materials, physical chemistry, biology, and other material directions. The model uses a dataset of about 2 million material knowledge instructions, and in the process of building the dataset, we developed the Intelligent Extraction Large Model (IELM), which is specially used to extract and form structured knowledge from scientific texts, avoiding a large number of costs that need to be manually annotated, and improving efficiency. We inject this data into the GLM4-9B model for learning to enhance its inference capabilities in a variety of material domains. In addition, we have introduced enhanced prompt strategies to ensure that the answers to the model are more organized and comprehensive, providing efficient and comprehensive intelligent support for the diverse needs of materials science exploration, and promoting the development of material science.

FinDVer: Explainable Claim Verification over Long and Hybrid-Content Financial Documents

We introduce FinDVer, a comprehensive benchmark specifically designed to evaluate the explainable claim verification capabilities of LLMs in the context of understanding and analyzing long, hybrid-content financial documents. FinDVer contains 2,400 expert-annotated examples, divided into three subsets: information extraction, numerical reasoning, and knowledge-intensive reasoning, each addressing common scenarios encountered in real-world financial contexts. We assess a broad spectrum of LLMs under long-context and RAG settings. Our results show that even the current best-performing system, GPT-4o, still lags behind human experts. We further provide in-depth analysis on long-context and RAG setting, Chain-of-Thought reasoning, and model reasoning errors, offering insights to drive future advancements. We believe that FinDVer can serve as a valuable benchmark for evaluating LLMs in claim verification over complex, expert-domain documents.

Unconstrained Model Merging for Enhanced LLM Reasoning

Recent advancements in building domain-specific large language models (LLMs) have shown remarkable success, especially in tasks requiring reasoning abilities like logical inference over complex relationships and multi-step problem solving. However, creating a powerful all-in-one LLM remains challenging due to the need for proprietary data and vast computational resources. As a resource-friendly alternative, we explore the potential of merging multiple expert models into a single LLM. Existing studies on model merging mainly focus on generalist LLMs instead of domain experts, or the LLMs under the same architecture and size. In this work, we propose an unconstrained model merging framework that accommodates both homogeneous and heterogeneous model architectures with a focus on reasoning tasks. A fine-grained layer-wise weight merging strategy is designed for homogeneous models merging, while heterogeneous model merging is built upon the probabilistic distribution knowledge derived from instruction-response fine-tuning data. Across 7 benchmarks and 9 reasoning-optimized LLMs, we reveal key findings that combinatorial reasoning emerges from merging which surpasses simple additive effects. We propose that unconstrained model merging could serve as a foundation for decentralized LLMs, marking a notable progression from the existing centralized LLM framework. This evolution could enhance wider participation and stimulate additional advancement in the field of artificial intelligence, effectively addressing the constraints posed by centralized models.

Persistent Pre-Training Poisoning of LLMs

Large language models are pre-trained on uncurated text datasets consisting of trillions of tokens scraped from the Web. Prior work has shown that: (1) web-scraped pre-training datasets can be practically poisoned by malicious actors; and (2) adversaries can compromise language models after poisoning fine-tuning datasets. Our work evaluates for the first time whether language models can also be compromised during pre-training, with a focus on the persistence of pre-training attacks after models are fine-tuned as helpful and harmless chatbots (i.e., after SFT and DPO). We pre-train a series of LLMs from scratch to measure the impact of a potential poisoning adversary under four different attack objectives (denial-of-service, belief manipulation, jailbreaking, and prompt stealing), and across a wide range of model sizes (from 600M to 7B). Our main result is that poisoning only 0.1% of a model's pre-training dataset is sufficient for three out of four attacks to measurably persist through post-training. Moreover, simple attacks like denial-of-service persist through post-training with a poisoning rate of only 0.001%.