Training LLMs larger than the aggregated memory of multiple GPUs is increasingly necessary due to the faster growth of LLM sizes compared to GPU memory. To this end, multi-tier host memory or disk offloading techniques are proposed by state of art. Despite advanced asynchronous multi-tier read/write strategies, such offloading strategies result in significant I/O overheads in the critical path of training, resulting in slower iterations. To this end, we propose MLP-Offload, a novel multi-level, multi-path offloading engine specifically designed for optimizing LLM training on resource-constrained setups by mitigating I/O bottlenecks. We make several key observations that drive the design of MLP-Offload, such as I/O overheads during the update dominate the iteration time; I/O bandwidth of the third-level remote storage tier remains unutilized; and, contention due to concurrent offloading amplifies I/O bottlenecks. Driven by these insights, we design and implement MLP-Offload to offload the optimizer states across multiple tiers in a cache-efficient and concurrency-controlled fashion to mitigate I/O bottlenecks during the backward and update phases. Evaluations on models up to 280B parameters shows that MLP-Offload achieves 2.5$\times$ faster iterations compared to the state-of-the-art LLM training runtimes.

Large Language Models (LLMs) have demonstrated strong generalization across a wide range of tasks. Reasoning with LLMs is central to solving multi-step problems and complex decision-making. To support efficient reasoning, recent studies have shifted attention from explicit chain-of-thought prompting toward implicit reasoning, where reasoning occurs silently via latent structures without emitting intermediate textual steps. Implicit reasoning brings advantages such as lower generation cost, faster inference, and better alignment with internal computation. Although prior surveys have discussed latent representations in the context of reasoning, a dedicated and mechanism-level examination of how reasoning unfolds internally within LLMs remains absent. This survey fills that gap by introducing a taxonomy centered on execution paradigms, shifting the focus from representational forms to computational strategies. We organize existing methods into three execution paradigms based on \textbf{\textit{how and where internal computation unfolds}}: latent optimization, signal-guided control, and layer-recurrent execution. We also review structural, behavioral and representation-based evidence that supports the presence of implicit reasoning in LLMs. We further provide a structured overview of the evaluation metrics and benchmarks used in existing works to assess the effectiveness and reliability of implicit reasoning. We maintain a continuously updated project at: https://github.com/digailab/awesome-llm-implicit-reasoning.

Video multimodal large language models (Video-MLLMs) have achieved remarkable progress in video understanding. However, they remain vulnerable to hallucination-producing content inconsistent with or unrelated to video inputs. Previous video hallucination benchmarks primarily focus on short-videos. They attribute hallucinations to factors such as strong language priors, missing frames, or vision-language biases introduced by the visual encoder. While these causes indeed account for most hallucinations in short videos, they still oversimplify the cause of hallucinations. Sometimes, models generate incorrect outputs but with correct frame-level semantics. We refer to this type of hallucination as Semantic Aggregation Hallucination (SAH), which arises during the process of aggregating frame-level semantics into event-level semantic groups. Given that SAH becomes particularly critical in long videos due to increased semantic complexity across multiple events, it is essential to separate and thoroughly investigate the causes of this type of hallucination. To address the above issues, we introduce ELV-Halluc, the first benchmark dedicated to long-video hallucination, enabling a systematic investigation of SAH. Our experiments confirm the existence of SAH and show that it increases with semantic complexity. Additionally, we find that models are more prone to SAH on rapidly changing semantics. Moreover, we discuss potential approaches to mitigate SAH. We demonstrate that positional encoding strategy contributes to alleviating SAH, and further adopt DPO strategy to enhance the model's ability to distinguish semantics within and across events. To support this, we curate a dataset of 8K adversarial data pairs and achieve improvements on both ELV-Halluc and Video-MME, including a substantial 27.7% reduction in SAH ratio.

Recent advances in applying reinforcement learning (RL) to large language models (LLMs) have led to substantial progress. In particular, a series of remarkable yet often counterintuitive phenomena have been reported in LLMs, exhibiting patterns not typically observed in traditional RL settings. For example, notable claims include that a single training example can match the performance achieved with an entire dataset, that the reward signal does not need to be very accurate, and that training solely with negative samples can match or even surpass sophisticated reward-based methods. However, the precise conditions under which these observations hold - and, critically, when they fail - remain unclear. In this work, we identify a key factor that differentiates RL observations: whether the pretrained model already exhibits strong Model-Task Alignment, as measured by pass@k accuracy on the evaluated task. Through a systematic and comprehensive examination of a series of counterintuitive claims, supported by rigorous experimental validation across different model architectures and task domains, our findings show that while standard RL training remains consistently robust across settings, many of these counterintuitive results arise only when the model and task already exhibit strong model-task alignment. In contrast, these techniques fail to drive substantial learning in more challenging regimes, where standard RL methods remain effective.

We present Hermes 4, a family of hybrid reasoning models that combine structured, multi-turn reasoning with broad instruction-following ability. We describe the challenges encountered during data curation, synthesis, training, and evaluation, and outline the solutions employed to address these challenges at scale. We comprehensively evaluate across mathematical reasoning, coding, knowledge, comprehension, and alignment benchmarks, and we report both quantitative performance and qualitative behavioral analysis. To support open research, all model weights are published publicly at https://huggingface.co/collections/NousResearch/hermes-4-collection-68a731bfd452e20816725728

We introduce Nemotron-Nano-9B-v2, a hybrid Mamba-Transformer language model designed to increase throughput for reasoning workloads while achieving state-of-the-art accuracy compared to similarly-sized models. Nemotron-Nano-9B-v2 builds on the Nemotron-H architecture, in which the majority of the self-attention layers in the common Transformer architecture are replaced with Mamba-2 layers, to achieve improved inference speed when generating the long thinking traces needed for reasoning. We create Nemotron-Nano-9B-v2 by first pre-training a 12-billion-parameter model (Nemotron-Nano-12B-v2-Base) on 20 trillion tokens using an FP8 training recipe. After aligning Nemotron-Nano-12B-v2-Base, we employ the Minitron strategy to compress and distill the model with the goal of enabling inference on up to 128k tokens on a single NVIDIA A10G GPU (22GiB of memory, bfloat16 precision). Compared to existing similarly-sized models (e.g., Qwen3-8B), we show that Nemotron-Nano-9B-v2 achieves on-par or better accuracy on reasoning benchmarks while achieving up to 6x higher inference throughput in reasoning settings like 8k input and 16k output tokens. We are releasing Nemotron-Nano-9B-v2, Nemotron-Nano12B-v2-Base, and Nemotron-Nano-9B-v2-Base checkpoints along with the majority of our pre- and post-training datasets on Hugging Face.

Humans can directly imagine and manipulate visual images in their minds, a capability known as spatial visualization. While multi-modal Large Language Models (MLLMs) support imagination-based reasoning, spatial visualization remains insufficiently evaluated, typically embedded within broader mathematical and logical assessments. Existing evaluations often rely on IQ tests or math competitions that may overlap with training data, compromising assessment reliability. To this end, we introduce SpatialViz-Bench, a comprehensive multi-modal benchmark for spatial visualization with 12 tasks across 4 sub-abilities, comprising 1,180 automatically generated problems. Our evaluation of 33 state-of-the-art MLLMs not only reveals wide performance variations and demonstrates the benchmark's strong discriminative power, but also uncovers counter-intuitive findings: models show difficulty perception misaligned with human intuition, exhibit dramatic 2Dto-3D performance cliffs, default to formulaic derivation over visualization, and paradoxically suffer performance degradation from Chain-of-Thought prompting in open-source models. Through statistical and qualitative analysis of error types, SpatialViz-Bench demonstrates that state-of-the-art MLLMs continue to exhibit deficiencies in spatial visualization tasks, thereby addressing a significant lacuna in the field. The benchmark data and evaluation code are publicly available.

Advances in material functionalities drive innovations across various fields, where metamaterials-defined by structure rather than composition-are leading the way. Despite the rise of artificial intelligence (AI)-driven design strategies, their impact is limited by task-specific retraining, poor out-of-distribution(OOD) generalization, and the need for separate models for forward and inverse design. To address these limitations, we introduce the Metamaterial Foundation Model (MetaFO), a Bayesian transformer-based foundation model inspired by large language models. MetaFO learns the underlying mechanics of metamaterials, enabling probabilistic, zero-shot predictions across diverse, unseen combinations of material properties and structural responses. It also excels in nonlinear inverse design, even under OOD conditions. By treating metamaterials as an operator that maps material properties to structural responses, MetaFO uncovers intricate structure-property relationships and significantly expands the design space. This scalable and generalizable framework marks a paradigm shift in AI-driven metamaterial discovery, paving the way for next-generation innovations.

Transformer-based large language models (LLMs) have achieved strong performance across many natural language processing tasks. Nonetheless, their quadratic computational and memory requirements, particularly in self-attention layers, pose challenges for efficient inference on long contexts and for deployment in resource-limited environments. We present TPTT (Transforming Pretrained Transformers into Titans), a framework designed to augment pretrained Transformers with linearized attention (LiZA) and internal memory gating via Memory as Gate (MaG), applied without full retraining. TPTT supports parameter-efficient fine-tuning (LoRA) and integrates with standard toolkits such as Hugging Face Transformers. We evaluated TPTT on several pretrained models, including Llama-1B, OlMoE-1B-7B, Qwen2.5-1.5B, Gemma3-270m, OpenELM-1.3B, and Mistral-7B, in order to assess applicability across architectures of different scales. Experiments on models with approximately 1 billion parameters, evaluated primarily on the MMLU benchmark, suggest potential improvements in both efficiency and accuracy compared to baseline models. For example, Titans-Llama-1B exhibited up to a 20\% relative increase in Exact Match scores in one-shot evaluation. An additional finding is that it is possible to convert a quadratic-attention model into a purely linear-attention model using the DeltaProduct mechanism. All training runs were carried out with modest computational resources. These preliminary findings indicate that TPTT may help adapt pretrained LLMs for long-context tasks with limited overhead. Further studies on larger models and a broader set of benchmarks will be necessary to evaluate the generality and robustness of the framework. Code is available at https://github.com/fabienfrfr/tptt . Python package at https://pypi.org/project/tptt/ .

Although large language models (LLMs) have become more capable and accurate across many tasks, some fundamental sources of unreliability remain in their behavior. One key limitation is their inconsistency at reporting the same information when prompts are changed. In this paper, we consider the discrepancy between a model's generated answer and their own verification of that answer, the generator-validator gap. We define this gap in a more stringent way than prior work: we expect correlation of scores from a generator and a validator over the entire set of candidate answers, i.e., candidate completions that could possibly arise during ordinary language use without breaking Gricean norms. We show that according to this measure, a large gap exists in various settings, including question answering, lexical semantics tasks, and next-word prediction. We then propose RankAlign, a ranking-based training method, and show that it significantly closes the gap, surpassing all baseline methods. Moreover, this approach generalizes well to out-of-domain tasks and lexical items.