This contribution provides our comprehensive reflection on the contemporary agent technology, with a particular focus on the advancements driven by Large Language Models (LLM) vs classic Multi - Agent Systems (MAS). It delves into the models, approaches, and characteristics that define these new systems. The paper emphasizes the critical analysis of how the recent developments relate to the foundational MAS, as articulated in the core academic literature. Finally, it identifies key challenges and promising future directions in this rapidly evolving domain.

Large Language and Vision Models using a Mixture-of-Experts (MoE) architecture pose significant challenges for deployment due to their computational and memory demands. Mixed Precision Quantization assigns different precisions to different layers of an LLM/VLM based on layer sensitivity and importance within the model. In this work, we propose a Post Training Quantization algorithm, MoPEQ, that assigns optimal bit width to each expert. Our method balances accuracy and model size by analyzing each expert's sensitivity using Hessian trace approximation instead of relying on the activation frequency of the expert. This per-expert granularity approach clusters similar experts to maintain model performance while reducing memory requirements. The experimental results on VLMEvalKit benchmark datasets using State-of-the-art VLMs Deepseek-VL2 -tiny, -small, -base, and MolmoE models demonstrate that our mixed precision quantized MoEs achieve competitive accuracy with substantial improvements in memory footprint compared to uniform-precision baseline methods. W e perform a comprehensive study to analyze the impact of expert activation frequency and sensitivity using Hessian trace approximation at both layer-wise and model-wide expert precision allocation of 2, 3, and 4 bits to provide a thorough understanding of mixed precision quantization of VLM-MoEs. The code is available here.

A pioneering rule-based mechanical translation system (precursor of modern RBMTs) was first presented in December 1929 by its inventor, Federico Pucci, who later published the full method in a book titled "Il traduttore meccanico ed il metodo per corrispondersi fra Europei conoscendo ciascuno solo la propria lingua: Parte I", in Salerno (Italy), in 1931. This study illustrates how AI breathes new life into the system of international keys and ideograms devised by Pucci to translate from/into any Romance language (at least as a first step). The methodology involves having the AIs retranslate, following Pucci's method, the two text excerpts originally translated in 1931 and clearly documented in his publication: a passage from Dante's La Vita Nuova, translated from Italian into French, and a passage from Voltaire's Zadig, translated from French into Italian. The result is notable: the two texts, translated 94 years apart using the same method--by Pucci in 1931 and by AIs in 2025--show a low average difference, with only minor variations observed. With Pucci's system thus validated, it became feasible to have the AIs reproduce the excerpts in English, Spanish, and German according to his method. The results were consistent, and Pucci--via Artificial Intelligence--was tasked with translating more modern and technical texts, thereby reviving, nearly a century later, an invention that had remained almost entirely unknown and never applied beyond its creator, now brought to wider attention and opened to possible experimentation. Such a demonstration would not only affirm Pucci's historical status but also place him among the precursors and intellectual contributors to machine translation, whose work merits examination alongside figures such as Troyanskij, Booth, and Weaver, with possible consequences for how the history of the field is understood.

Semantic evaluation in low-resource languages remains a major challenge in NLP. While sentence transformers have shown strong performance in high-resource settings, their effectiveness in Indic languages is underexplored due to a lack of high-quality benchmarks. To bridge this gap, we introduce L3Cube-IndicHeadline-ID, a curated headline identification dataset spanning ten low-resource Indic languages: Marathi, Hindi, Tamil, Gujarati, Odia, Kannada, Malayalam, Punjabi, Telugu, Bengali and English. Each language includes 20,000 news articles paired with four headline variants: the original, a semantically similar version, a lexically similar version, and an unrelated one, designed to test fine-grained semantic understanding. The task requires selecting the correct headline from the options using article-headline similarity. We benchmark several sentence transformers, including multilingual and language-specific models, using cosine similarity. Results show that multilingual models consistently perform well, while language-specific models vary in effectiveness. Given the rising use of similarity models in Retrieval-Augmented Generation (RAG) pipelines, this dataset also serves as a valuable resource for evaluating and improving semantic understanding in such applications. Additionally, the dataset can be repurposed for multiple-choice question answering, headline classification, or other task-specific evaluations of LLMs, making it a versatile benchmark for Indic NLP. The dataset is shared publicly at https://github.com/l3cube-pune/indic-nlp

Büchi automata (BAs) recognize $ω$-regular languages defined by formal specifications like linear temporal logic (LTL) and are commonly used in the verification of reactive systems. However, BAs face scalability challenges when handling and manipulating complex system behaviors. As neural networks are increasingly used to address these scalability challenges in areas like model checking, investigating their ability to generalize beyond training data becomes necessary. This work presents the first study investigating whether recurrent neural networks (RNNs) can generalize to $ω$-regular languages derived from LTL formulas. We train RNNs on ultimately periodic $ω$-word sequences to replicate target BA behavior and evaluate how well they generalize to out-of-distribution sequences. Through experiments on LTL formulas corresponding to deterministic automata of varying structural complexity, from 3 to over 100 states, we show that RNNs achieve high accuracy on their target $ω$-regular languages when evaluated on sequences up to $8 \times$ longer than training examples, with $92.6\%$ of tasks achieving perfect or near-perfect generalization. These results establish the feasibility of neural approaches for learning complex $ω$-regular languages, suggesting their potential as components in neurosymbolic verification methods.

The adoption of Transformer-based architectures in the medical domain is growing rapidly. In medical imaging, the analysis of complex shapes - such as organs, tissues, or other anatomical structures - combined with the often anisotropic nature of high-dimensional images complicates these adaptations. In this study, we critically examine the role of Positional Encodings (PEs), arguing that commonly used approaches may be suboptimal for the specific challenges of medical imaging. Sinusoidal Positional Encodings (SPEs) have proven effective in vision tasks, but they struggle to preserve Euclidean distances in higher-dimensional spaces. Isotropic Fourier Feature Positional Encodings (IFPEs) have been proposed to better preserve Euclidean distances, but they lack the ability to account for anisotropy in images. To address these limitations, we propose Anisotropic Fourier Feature Positional Encoding (AFPE), a generalization of IFPE that incorporates anisotropic, class-specific, and domain-specific spatial dependencies. We systematically benchmark AFPE against commonly used PEs on multi-label classification in chest X-rays, organ classification in CT images, and ejection fraction regression in echocardiography. Our results demonstrate that choosing the correct PE can significantly improve model performance. We show that the optimal PE depends on the shape of the structure of interest and the anisotropy of the data. Finally, our proposed AFPE significantly outperforms state-of-the-art PEs in all tested anisotropic settings. We conclude that, in anisotropic medical images and videos, it is of paramount importance to choose an anisotropic PE that fits the data and the shape of interest.

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.

Following rapid advancements in text and image generation, research has increasingly shifted towards 3D generation. Unlike the well-established pixel-based representation in images, 3D representations remain diverse and fragmented, encompassing a wide variety of approaches such as voxel grids, neural radiance fields, signed distance functions, point clouds, or octrees, each offering distinct advantages and limitations. In this work, we present a unified evaluation framework designed to assess the performance of 3D representations in reconstruction and generation. We compare these representations based on multiple criteria: quality, computational efficiency, and generalization performance. Beyond standard model benchmarking, our experiments aim to derive best practices over all steps involved in the 3D generation pipeline, including preprocessing, mesh reconstruction, compression with autoencoders, and generation. Our findings highlight that reconstruction errors significantly impact overall performance, underscoring the need to evaluate generation and reconstruction jointly. We provide insights that can inform the selection of suitable 3D models for various applications, facilitating the development of more robust and application-specific solutions in 3D generation. The code for our framework is available at https://github.com/isl-org/unifi3d.

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.

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.