TopResearch

In a world defined by rapid technological leaps, top-tier research is more than academic achievement it is our most vital asset. We will explore the global frontier of AI research and map Israel’s vital place within it. To close, we will briefly touch upon how researchers can harness their unique skills to support Israel’s strength and technological resilience today.

​For the past decade, the story of AI progress has largely been the story of Transformers and text. World models pose a different question entirely: How do you build a system that represents not language, but reality itself- its physics, its geometry, and the way events unfold through time?

World models have become one of the most discussed ideas in the field, yet what these systems actually are, and where they actually stand, is poorly understood.

​In this lecture, we'll map the entire world-model landscape. We'll construct a working taxonomy of the field, drawing sharp lines between auto-regressive approaches and their non-predictive counterparts such as JEPA, Fei-Fei Li's recent spatial-intelligence work, and the emerging wave of game-based models, while highlighting what each architecture assumes and what it commits to.

​We'll examine what kinds of multi-modal signals are needed to give a machine an intuitive grasp of physics, space, and interaction; how the scaling picture changes when video replaces text; and the structural redundancy that auto-regressive generation carries with it.

Reasoning about failures is crucial for building reliable and trustworthy robotic systems. Prior approaches either treat failure reasoning as a closed-set classification problem or assume access to ample human annotations. Failures in

the real world are typically subtle, combinatorial, and difficult to enumerate, whereas rich reasoning labels are expensive to acquire. We address this problem by introducing ARMOR: Adaptive Round-based Multi-task mOdel for Robotic failure

detection and reasoning. We formulate detection and reasoning as a multi-task self-refinement process, where the model iteratively predicts detection outcomes and natural language reasoning conditioned on past outputs. During training,

ARMOR learns from heterogeneous supervision – large-scale sparse binary labels and small-scale rich reasoning annotations – optimized via a combination of offline and online imitation learning. At inference time, ARMOR generates multiple

refinement trajectories and selects the most confident prediction via a self-certainty metric. Experiments across diverse environments show that ARMOR achieves state-of-the-art performance by improving over the previous approaches by up

to 30% on failure detection rate and up to 100% in reasoning measured through LLM fuzzy match score, demonstrating robustness to heterogeneous supervision and open-ended reasoning beyond predefined failure modes.

Behavioral cloning is a simple yet effective technique for learning sequential

decision-making from demonstrations. Recently, it has gained prominence as

the core of foundation models for the physical world, where achieving generalization requires countless demonstrations of a multitude of tasks. Typically,

a human expert with full information on the task demonstrates a (nearly) optimal behavior. In this paper, we propose to hide some of the task’s information

from the demonstrator. This “blindfolded” expert is compelled to employ nontrivial exploration to solve the task. We show that cloning the blindfolded expert

generalizes better to unseen tasks than its fully-informed counterpart. We conduct experiments of real-world robot peg insertion tasks with (limited) human

demonstrations, alongside a videogame from the Procgen benchmark. Additionally, we support our findings with theoretical analysis, which confirms that the

generalization error scales with sqrt (I/m), where I measures the amount of task

information available to the demonstrator, and m is the number of demonstrated

tasks. Both theory and practice indicate that cloning blindfolded experts generalizes better with fewer demonstrated tasks. Project page with videos and code:

https://sites.google.com/view/blindfoldedexperts/home.

Recent work found that LLMs are sensitive to a wide range of arbitrary prompt dimensions, including the type of delimiters, answer enumerators, instruction wording, and more. This throws into question popular single-prompt evaluation practices. We present DOVE (Dataset Of Variation Evaluation) a large-scale dataset containing prompt perturbations of various evaluation benchmarks. In contrast to previous work, we examine LLM sensitivity from an holistic perspective, and assess the joint effects of perturbations along various dimensions, resulting in thousands of perturbations per instance. We evaluate several model families against DOVE, leading to several findings, including efficient methods for choosing well-performing prompts, observing that few-shot examples reduce sensitivity, and identifying instances which are inherently hard across all perturbations. DOVE consists of more than 250M prompt perturbations and model outputs, which we make publicly available to spur a community-wide effort toward meaningful, robust, and efficient evaluation.

This work led to two follow-up studies presented at EMNLP 2025: PromptSuite, a task-agnostic framework for systematically generating multiple prompt variations (Habba, Dahan, Lior, and Stanovsky), and ReliableEval, which provides statistical methods for robust stochastic LLM evaluation (Lior, Habba, Levy, Caciularu, and Stanovsky).

Neural processes (NPs) are a class of models that learn stochastic processes directly from data and can be used for inference, sampling and conditional sampling.
We introduce a new NP model based on flow matching, a generative modeling paradigm that has demonstrated strong performance on various data modalities.
Following the NP training framework, the model provides amortized predictions of conditional distributions over any arbitrary points in the data. Compared to
previous NP models, our model is simple to implement and can be used to sample from conditional distributions using an ODE solver, without requiring auxiliary
conditioning methods. In addition, the model provides a controllable tradeoff between accuracy and running time via the number of steps in the ODE solver. We show that our model outperforms previous state-of-the-art neural process methods on various benchmarks including synthetic 1D Gaussian processes data, 2D images, and real-world weather data.

TL;DR: We develop and evaluate an LLM agent that decides both what to say and when to speak in asynchronous group settings, demonstrating human-like performance in Mafia games.

Vision-Language models (VLMs) show impressive abilities to answer questions on visual inputs (e.g., counting objects in an image), yet demonstrate higher accuracies when performing an analogous task on text (e.g., counting words in a text). We investigate this accuracy gap by identifying and comparing the \textit{circuits} – the task-specific computational sub-graphs – in different modalities. We show that while circuits are largely disjoint between modalities, they implement relatively similar functionalities: the differences lie primarily in processing modality-specific data positions (an image or a text sequence). Zooming in on the image data representations, we observe they become aligned with the higher-performing analogous textual representations only towards later layers, too late in processing to effectively influence subsequent positions. To overcome this, we patch the representations of visual data tokens from later layers back into earlier layers. In experiments with multiple tasks and models, this simple intervention closes a third of the performance gap between the modalities, on average. Our analysis sheds light on the multi-modal performance gap in VLMs and suggests a training-free approach for reducing it.

Automatic n-gram based metrics such as ROUGE are widely used for evaluating generative tasks such as summarization. While these metrics are considered indicative (even if imperfect) of human evaluation for English, their suitability for other languages remains unclear. To address this, we systematically assess evaluation metrics for generation both n-gram-based and neural based to evaluate their effectiveness across languages and tasks. Specifically, we design a large-scale evaluation suite across eight languages from four typological families: agglutinative, isolating, low-fusional, and high-fusional, spanning both low- and high-resource settings, to analyze their correlation with human judgments. Our findings highlight the sensitivity of evaluation metrics to the language type. For example, in fusional languages, n-gram-based metrics show lower correlation with human assessments compared to isolating and agglutinative languages. We also demonstrate that proper tokenization can significantly mitigate this issue for morphologically rich fusional languages, sometimes even reversing negative trends. Additionally, we show that neural-based metrics specifically trained for evaluation, such as COMET, consistently outperform other neural metrics and better correlate with human judgments in low-resource languages. Overall, our analysis highlights the limitations of n-gram metrics for fusional languages and advocates for greater investment in neural-based metrics trained for evaluation tasks.

Audio inpainting seeks to restore missing segments in degraded recordings. Previous diffusion-based methods exhibit impaired performance when the missing region is large. We introduce the first approach that applies discrete diffusion over tokenized music representations from a pre-trained audio tokenizer, enabling stable and semantically coherent restoration of long gaps. Our method further incorporates two training approaches: a derivative-based regularization loss that enforces smooth temporal dynamics, and a span-based absorbing transition that provides structured corruption during diffusion. Experiments on the MusicNet and MAESTRO datasets with gaps up to 750\,ms show that our approach consistently outperforms strong baselines across range of gap lengths, for gaps of 150\,ms and above. This work advances musical audio restoration and introduces new directions for discrete diffusion model training.

Training a task-specific classifier normally requires a labeled corpus. We study the harder setting where no data exists at all and the only input is a task description in natural language. Two problems arise together. Diversity: a generator prompted from a description samples its own prior and collapses onto a narrow band of surface patterns, and sampling cannot fix a distribution that was never observed. Faithfulness: near the decision boundary, a single teacher call assigns unreliable labels, and errors there are exactly the ones that matter. BARRED (ICML 2026) addresses both. The resulting small models outperform frontier models on human-labeled test sets, at a fraction of the cost and latency.