Deep neural networks (DNNs) are increasingly being deployed in high-stakes applications, from self-driving cars to biometric authentication. However, their unpredictable and unreliable behaviors in real-world settings require new approaches to characterize and ensure their reliability. This paper introduces DeepProv, a novel and customizable system designed to capture and characterize the runtime behavior of DNNs during inference by using their underlying graph structure. Inspired by system audit provenance graphs, DeepProv models the computational information flow of a DNN's inference process through Inference Provenance Graphs (IPGs). These graphs provide a detailed structural representation of the behavior of DNN, allowing both empirical and structural analysis. DeepProv uses these insights to systematically repair DNNs for specific objectives, such as improving robustness, privacy, or fairness. We instantiate DeepProv with adversarial robustness as the goal of model repair and conduct extensive case studies to evaluate its effectiveness. Our results demonstrate its effectiveness and scalability across diverse classification tasks, attack scenarios, and model complexities. DeepProv automatically identifies repair actions at the node and edge-level within IPGs, significantly enhancing the robustness of the model. In particular, applying DeepProv repair strategies to just a single layer of a DNN yields an average 55% improvement in adversarial accuracy. Moreover, DeepProv complements existing defenses, achieving substantial gains in adversarial robustness. Beyond robustness, we demonstrate the broader potential of DeepProv as an adaptable system to characterize DNN behavior in other critical areas, such as privacy auditing and fairness analysis.

Mixture-of-Experts (MoE) has emerged as a promising paradigm for efficiently scaling large language models without a proportional increase in computational cost. However, the standard training strategy of Top-K router prevents MoE models from realizing their full potential for elastic inference. When the number of activated experts is altered at inference time, these models exhibit precipitous performance degradation. In this work, we introduce Matryoshka MoE (M-MoE), a training framework that instills a coarse-to-fine structure directly into the expert ensemble. By systematically varying the number of activated experts during training, M-MoE compels the model to learn a meaningful ranking: top-ranked experts collaborate to provide essential, coarse-grained capabilities, while subsequent experts add progressively finer-grained detail. We explore this principle at multiple granularities, identifying a layer-wise randomization strategy as the most effective. Our experiments demonstrate that a single M-MoE model achieves remarkable elasticity, with its performance at various expert counts closely matching that of an entire suite of specialist models, but at only a fraction of the total training cost. This flexibility not only unlocks elastic inference but also enables optimizing performance by allocating different computational budgets to different model layers. Our work paves the way for more practical and adaptable deployments of large-scale MoE models.

We study code-to-metric regression: predicting numeric outcomes of code executions, a challenging task due to the open-ended nature of programming languages. While prior methods have resorted to heavy and domain-specific feature engineering, we show that a single unified Regression Language Model (RLM) can simultaneously predict directly from text, (i) the memory footprint of code across multiple high-level languages such as Python and C++, (ii) the latency of Triton GPU kernels, and (iii) the accuracy and speed of trained neural networks represented in ONNX. In particular, a relatively small 300M parameter RLM initialized from T5Gemma, obtains > 0.9 Spearman-rank on competitive programming submissions from APPS, and a single unified model achieves > 0.5 average Spearman-rank across 17 separate languages from CodeNet. Furthermore, the RLM can obtain the highest average Kendall-Tau of 0.46 on five classic NAS design spaces previously dominated by graph neural networks, and simultaneously predict architecture latencies on numerous hardware platforms.

Vector quantization is common in deep models, yet its hard assignments block gradients and hinder end-to-end training. We propose DiVeQ, which treats quantization as adding an error vector that mimics the quantization distortion, keeping the forward pass hard while letting gradients flow. We also present a space-filling variant (SF-DiVeQ) that assigns to a curve constructed by the lines connecting codewords, resulting in less quantization error and full codebook usage. Both methods train end-to-end without requiring auxiliary losses or temperature schedules. On VQ-VAE compression and VQGAN generation across various data sets, they improve reconstruction and sample quality over alternative quantization approaches.

We present CreAgentive, an agent workflow driven multi-category creative generation engine that addresses four key limitations of contemporary large language models in writing stories, drama and other categories of creatives: restricted genre diversity, insufficient output length, weak narrative coherence, and inability to enforce complex structural constructs. At its core, CreAgentive employs a Story Prototype, which is a genre-agnostic, knowledge graph-based narrative representation that decouples story logic from stylistic realization by encoding characters, events, and environments as semantic triples. CreAgentive engages a three-stage agent workflow that comprises: an Initialization Stage that constructs a user-specified narrative skeleton; a Generation Stage in which long- and short-term objectives guide multi-agent dialogues to instantiate the Story Prototype; a Writing Stage that leverages this prototype to produce multi-genre text with advanced structures such as retrospection and foreshadowing. This architecture reduces storage redundancy and overcomes the typical bottlenecks of long-form generation. In extensive experiments, CreAgentive generates thousands of chapters with stable quality and low cost (less than $1 per 100 chapters) using a general-purpose backbone model. To evaluate performance, we define a two-dimensional framework with 10 narrative indicators measuring both quality and length. Results show that CreAgentive consistently outperforms strong baselines and achieves robust performance across diverse genres, approaching the quality of human-authored novels.

Factuality assesses the extent to which a language utterance relates to real-world information; it determines whether utterances correspond to facts, possibilities, or imaginary situations, and as such, it is instrumental for fact checking. Factuality is a complex notion that relies on multiple linguistic signals, and has been studied in various disciplines. We present a complex, multi-faceted annotation scheme of factuality that combines concepts from a variety of previous works. We developed the scheme for Hebrew, but we trust that it can be adapted to other languages. We also present a set of almost 5,000 sentences in the domain of parliamentary discourse that we manually annotated according to this scheme. We report on inter-annotator agreement, and experiment with various approaches to automatically predict (some features of) the scheme, in order to extend the annotation to a large corpus.

Fingerprinting Large Language Models (LLMs) is essential for provenance verification and model attribution. LLM fingerprints have recently been proposed as a tool to identify, attribute, and trace LLMs by examining their observable behaviors (Pasquini et al., 2024; Xu et al., 2024; Y oon et al., 2025; Galton's Finger Prints (1892) (Galton, 1892): "A fingerprint is the pattern formed by friction-ridge skin on the fingertips; They are defined only after models have been fully trained and converged (finish pretaining), e.g., extracting patterns from parameters or generated text, and thus capture traits that emerge as the In this work, we propose a stricter notion of an LLM fingerprint: an intrinsic property present at model initialization and detectable at any time of the subsequent training. Extensive experiments in Section 5.1 show that our method can even distinguish between models that differ only in their initialization seed, despite an identical training pipeline and data order. Finally, in Section 5.2, we evaluate our method both Classic techniques include backdoor attacks (Adi et al., 2018; Nasery et al.); and model weight watermarks, which embed identifiers into parameters or link them With white-box access, signatures are extracted from model weights, leveraging intrinsic properties like the distribution of attention matrices (Y oon et al., 2025), the kernel alignment of internal representations (Zhang et al., 2024), or the stable direction of parameter vectors. Figure 1: Initialization-born token bias persists through training.

Conversational Spoken Language Models (SLMs) are emerging as a promising paradigm for real-time speech interaction. However, their capacity of temporal dynamics, including the ability to manage timing, tempo and simultaneous speaking, remains a critical and unevaluated challenge for conversational fluency. To address this gap, we introduce the Game-Time Benchmark, a framework to systematically assess these temporal capabilities. Inspired by how humans learn a language through language activities, Game-Time consists of basic instruction-following tasks and advanced tasks with temporal constraints, such as tempo adherence and synchronized responses. Our evaluation of diverse SLM architectures reveals a clear performance disparity: while state-of-the-art models handle basic tasks well, many contemporary systems still struggle with fundamental instruction-following. More critically, nearly all models degrade substantially under temporal constraints, exposing persistent weaknesses in time awareness and full-duplex interaction. The Game-Time Benchmark provides a foundation for guiding future research toward more temporally-aware conversational AI. Demos and datasets are available on our project website https://ga642381.github.io/Game-Time.

Recently, we have witnessed great progress in image editing with natural language instructions. Several closed-source models like GPT-Image-1, Seedream, and Google-Nano-Banana have shown highly promising progress. However, the open-source models are still lagging. The main bottleneck is the lack of a reliable reward model to scale up high-quality synthetic training data. To address this critical bottleneck, we built \mname, trained with our new large-scale human preference dataset, meticulously annotated by trained experts following a rigorous protocol containing over 200K preference pairs. \mname demonstrates superior alignment with human preferences in instruction-guided image editing tasks. Experiments show that \mname achieves state-of-the-art human correlation on established benchmarks such as GenAI-Bench, AURORA-Bench, ImagenHub, and our new \benchname, outperforming a wide range of VLM-as-judge models. Furthermore, we use \mname to select a high-quality subset from the existing noisy ShareGPT-4o-Image dataset. We train Step1X-Edit on the selected subset, which shows significant improvement over training on the full set. This demonstrates \mname's ability to serve as a reward model to scale up high-quality training data for image editing. Furthermore, its strong alignment suggests potential for advanced applications like reinforcement learning-based post-training and test-time scaling of image editing models. \mname with its training dataset will be released to help the community build more high-quality image editing training datasets.

The Transformer Machine Learning (ML) architecture has been gaining considerable momentum in recent years. In particular, computational High-Energy Physics tasks such as jet tagging and particle track reconstruction (tracking), have either achieved proper solutions, or reached considerable milestones using Transformers. On the other hand, the use of specialised hardware accelerators, especially FPGAs, is an effective method to achieve online, or pseudo-online latencies. The development and integration of Transformer-based ML to FPGAs is still ongoing and the support from current tools is very limited to non-existent. Additionally, FPGA resources present a significant constraint. Considering the model size alone, while smaller models can be deployed directly, larger models are to be partitioned in a meaningful and ideally, automated way . We aim to develop methodologies and tools for monolithic, or partitioned Transformer synthesis, specifically targeting inference. Our primary use-case involves two machine learning model designs for tracking, derived from the TrackFormers project. We elaborate our development approach, present preliminary results, and provide comparisons.