Oren Gal

Alaa Maalouf, Yotam Gurfinkel, Barak Diker et al. · mit

We suggest the first system that runs real-time semantic segmentation via deep learning on a weak micro-computer such as the Raspberry Pi Zero v2 (whose price was \$15) attached to a toy-drone. In particular, since the Raspberry Pi weighs less than $16$ grams, and its size is half of a credit card, we could easily attach it to the common commercial DJI Tello toy-drone (<\$100, <90 grams, 98 $\times$ 92.5 $\times$ 41 mm). The result is an autonomous drone (no laptop nor human in the loop) that can detect and classify objects in real-time from a video stream of an on-board monocular RGB camera (no GPS or LIDAR sensors). The companion videos demonstrate how this Tello drone scans the lab for people (e.g. for the use of firefighters or security forces) and for an empty parking slot outside the lab. Existing deep learning solutions are either much too slow for real-time computation on such IoT devices, or provide results of impractical quality. Our main challenge was to design a system that takes the best of all worlds among numerous combinations of networks, deep learning platforms/frameworks, compression techniques, and compression ratios. To this end, we provide an efficient searching algorithm that aims to find the optimal combination which results in the best tradeoff between the network running time and its accuracy/performance.

Tianyu Cao, Helin Wang, Ari Frummer et al.

Recent advances in generative models, such as diffusion and flow matching, have shown strong performance in audio tasks. However, speech enhancement (SE) models are typically trained on limited datasets and evaluated under narrow conditions, limiting real-world applicability. To address this, we propose DiT-Flow, a flow matching-based SE framework built on the latent Diffusion Transformer (DiT) backbone and trained for robustness across diverse distortions, including noise, reverberation, and compression. DiT-Flow operates on compact variational auto-encoders (VAEs)-derived latent features. We validated our approach on StillSonicSet, a synthetic yet acoustically realistic dataset composed of LibriSpeech, FSD50K, FMA, and 90 Matterport3D scenes. Experiments show that DiT-Flow consistently outperforms state-of-the-art generative SE models, demonstrating the effectiveness of flow matching in multi-condition speech enhancement. Despite ongoing efforts to expand synthetic data realism, a persistent bottleneck in SE is the inevitable mismatch between training and deployment conditions. By integrating LoRA with the MoE framework, we achieve both parameter-efficient and high-performance training for DiT-Flow robust to multiple distortions with using 4.9% percentage of the total parameters to obtain a better performance on five unseen distortions.

Maxim Mednikov, Oren Gal

Real-world multi-agent reinforcement learning (MARL) systems must often operate under stale observations, stochastic communication delays, and intermittent packet loss. Policies trained under idealized synchronous conditions frequently exhibit significant performance degradation in these regimes because they act on outdated feedback. We propose a modular execution-stage state-estimation layer that replaces delayed communicated observations with current belief-state estimates. The framework integrates a learned Gated transition model with a recursive Kalman filtering layer to estimate instantaneous states from asynchronous measurements. A primary advantage of this approach is its modularity, The estimator serves as a plug-in for pre-trained policies, requiring no modifications to the original MARL training algorithm, architecture, or reward structure. Evaluation across diverse multi-agent and continuous-control benchmarks demonstrates that the proposed layer consistently enhances robustness to communication latency and message loss. The most significant performance gains are observed in coordination-intensive and dynamically unstable tasks where temporal consistency is critical for control.

Josef Berman, Oren Gal

Coordinating micro-robotic swarms in physiologically realistic, time-dependent fluid environments remains an unsolved challenge for biomedical and environmental applications. We present a hybrid Computational Fluid Dynamics - Multi-Objective Multi-Agent Reinforcement Learning framework that directly couples a high-fidelity incompressible Navier-Stokes solver with decentralized proximal policy optimization to learn physically consistent swarm control strategies in oscillatory flow. Sixteen magnetically actuated micro-robots navigate a pulsatile arterial waveform, simultaneously optimizing upstream progression, energy conservation, and motion smoothness, reconciled using PCGrad surgery. Without PCGrad, energy efficiency and smoothness rewards collapse to near zero within 10,000 training steps while progress exhibits persistent large-amplitude oscillations, confirming that gradient conflict resolution is a structural requirement rather than an optional refinement in this domain. The converged policy achieves a progress reward of 6.5-7.0, a sustained energy efficiency of 0.63-0.65, and near-maximum smoothness (0.97-0.99), representing improvements over brute-force baselines on the primary objective while both baselines yield negative energy efficiency throughout. Training reveals three emergent behavioral phases: a collective two-layer hydrodynamic throttling formation that suppresses peak channel velocities during forward flow, a cycle-synchronized ratchet mechanism that exploits flow reversals for upstream repositioning, and an individualized final approach as agents near the success boundary. These results establish that time-dependent fluid-agent interactions can be captured directly within multi-objective reinforcement learning loops, offering a physically grounded paradigm for micro-swarm control in biomedical navigation, environmental monitoring, and industrial microfluidics.

Elie Abboud, Oren Gal

Sparse rewards are a major bottleneck in multi-agent reinforcement learning (MARL), where simultaneous learning induces non-stationarity and makes reward design especially delicate. Reward shaping can accelerate learning, but in the multi-agent setting it must preserve the strategic structure of the problem rather than merely improve short-term optimization. We propose Automatic Reward-shaping in Multi-agent Systems (ARMS), a self-supervised reward shaping framework for MARL that learns dense shaping signals from sparse environmental rewards through trajectory ranking. Since single-agent trajectory-ranking guarantees do not directly transfer to MARL, we reformulate policy invariance through conditional best-response reasoning, and show that if certain conditions hold, then using shaping rewards preserves each agent's best-response set under fixed opponent policies, and consequently preserve the set of Nash equilibria. Guided by this perspective, ARMS alternates between policy learning and reward learning while sharing shaping parameters across agents for efficiency. Experiments in a partially observable multi-agent pathfinding domain show that ARMS improves sampling efficiency under increasing reward sparsity and agent count, generalizes to unseen environments, and reveals a MARL-specific failure mode in which limited exploration and coupled policy--reward dynamics induce oscillatory behavior. Increasing exploration mitigates this effect and stabilizes learning. To the best of our knowledge, ARMS is the first automatic reward shaping framework for MARL whose design is motivated by a game-theoretic equilibrium-preservation result.

Jonathan Steinberg, Oren Gal

Safety-oriented instruction-following is supposed to keep LLM-controlled robots safe. We show it also creates an availability attack surface. By injecting short safety-plausible phrases (1-5 tokens) into a robots audio channel, an adversary can trigger the models safety reasoning to halt or disrupt execution without jailbreaking the model or overriding its policy. In the embodied setting, this is a semantic denial-of-service attack: the agent stops because the injected signal looks like a legitimate alert. Across four vision-language models, seven prompt-level defenses, three deployment modes, and single- and multi-injection settings, we find that prompt-only defenses trade off attack suppression against genuine hazard response. The strongest defenses reduce hard-stop attack success on some models, but defenses change the form of disruption, not its fact: suppressed hard stops re-emerge as acknowledge loops and false alerts, which we measure with Disruption Success Rate (DSR). We further find that injection variety is consistently more effective than repeating the same phrase, suggesting that models treat diverse safety cues as corroborating evidence. The practical implication is architectural rather than prompt-level: systems that route unauthenticated audio text directly into the LLM create an avoidable security dependency between safety monitoring and action selection.

Inbal Rimon, Oren Gal, Haim Permuter

Partial deepfake speech detection requires identifying manipulated regions that may occur within short temporal portions of an otherwise bona fide utterance, making the task particularly challenging for conventional utterance-level classifiers. We propose a split-and-conquer framework that decomposes the problem into two stages: boundary detection and segment-level classification. A dedicated boundary detector first identifies temporal transition points, allowing the audio signal to be divided into segments that are expected to contain acoustically consistent content. Each resulting segment is then evaluated independently to determine whether it corresponds to bona fide or fake speech. This formulation simplifies the learning objective by explicitly separating temporal localization from authenticity assessment, allowing each component to focus on a well-defined task. To further improve robustness, we introduce a reflection-based multi-length training strategy that converts variable-duration segments into several fixed input lengths, producing diverse feature-space representations. Each stage is trained using multiple configurations with different feature extractors and augmentation strategies, and their complementary predictions are fused to obtain improved final models. Experiments on the PartialSpoof benchmark demonstrate state-of-the-art performance across multiple temporal resolutions as well as at the utterance level, with substantial improvements in the accurate detection and localization of spoofed regions. In addition, the proposed method achieves state-of-the-art performance on the Half-Truth dataset, further confirming the robustness and generalization capability of the framework.

Jonathan Steinberg, Oren Gal

Vision-language models (VLMs) can read text from images, but where does this optical character recognition (OCR) information enter the language processing stream? We investigate the OCR routing mechanism across three architecture families (Qwen3-VL, Phi-4, InternVL3.5) using causal interventions. By computing activation differences between original images and text-inpainted versions, we identify architecture-specific OCR bottlenecks whose dominant location depends on the vision-language integration strategy: DeepStack models (Qwen) show peak sensitivity at mid-depth (about 50%) for scene text, while single-stage projection models (Phi-4, InternVL) peak at early layers (6-25%), though the exact layer of maximum effect varies across datasets. The OCR signal is remarkably low-dimensional: PC1 captures 72.9% of variance. Crucially, principal component analysis (PCA) directions learned on one dataset transfer to others, demonstrating shared text-processing pathways. Surprisingly, in models with modular OCR circuits (notably Qwen3-VL-4B), OCR removal can improve counting performance (up to +6.9 percentage points), suggesting OCR interferes with other visual processing in sufficiently modular architectures.

Haochen Zhang, Tianyi Zhang, Junze Yin et al.

Recommender systems play a pivotal role in providing relevant content to users. With the rapid development of large language models (LLMs), researchers have begun utilizing LLMs to build more powerful recommender systems. However, existing approaches that focus on aligning LLMs with recommendation tasks do not fully leverage their sequential information processing capabilities, leading to suboptimal performance. In this paper, we propose a novel system called compressed vocabulary expansion (CoVE). In CoVE, each item is assigned a unique ID within the expanded vocabulary. Our framework effectively capitalizes on sequence understanding abilities of LLMs, significantly enhancing their performance on recommendation tasks. Additionally, we compress the embedding layer, making CoVE practical for large-scale industrial applications. The effectiveness and performance of CoVE are demonstrated through comprehensive experiments on multiple recommendation datasets and comparisons with prior works. Our code can be found at https://github.com/HaochenZhang717/CoVE-official-Repo.

Lucas Liebenwein, Alaa Maalouf, Oren Gal et al.

We present a novel global compression framework for deep neural networks that automatically analyzes each layer to identify the optimal per-layer compression ratio, while simultaneously achieving the desired overall compression. Our algorithm hinges on the idea of compressing each convolutional (or fully-connected) layer by slicing its channels into multiple groups and decomposing each group via low-rank decomposition. At the core of our algorithm is the derivation of layer-wise error bounds from the Eckart Young Mirsky theorem. We then leverage these bounds to frame the compression problem as an optimization problem where we wish to minimize the maximum compression error across layers and propose an efficient algorithm towards a solution. Our experiments indicate that our method outperforms existing low-rank compression approaches across a wide range of networks and data sets. We believe that our results open up new avenues for future research into the global performance-size trade-offs of modern neural networks. Our code is available at https://github.com/lucaslie/torchprune.

Jonathan Steinberg, Oren Gal

Coding agents often pass per-prompt safety review yet ship exploitable code when their tasks are decomposed into routine engineering tickets. The challenge is structural: existing safety alignment evaluates overt requests in isolation, leaving models blind to malicious end-states that emerge from sequenced compliance with innocuous-looking requests. We introduce MOSAIC-Bench (Malicious Objectives Sequenced As Innocuous Compliance), a benchmark of 199 three-stage attack chains paired with deterministic exploit oracles on deployed software substrates (10 web-application substrates, 31 CWE classes, 5 programming languages) that treats both exploit ground truth and downstream reviewer protocol as first-class evaluation axes. On this benchmark, nine production coding agents from Anthropic, OpenAI, Google, Moonshot, Zhipu, and Minimax compose innocuous tickets at 53-86% end-to-end ASR with only two refusals across all staged runs. In a matched direct-prompt experiment over four frontier Claude/Codex agents, vulnerable-output rates fall to 0-20.4%: Claude primarily refuses, while Codex primarily hardens rather than emitting the vulnerable implementation - ticket staging silences both defense modes simultaneously. Downstream, code reviewer agents approve 25.8% of these confirmed-vulnerable cumulative diffs as routine PRs, and a full-context implementation protocol closes only 50% of the staged/direct gap, ruling out context fragmentation as the sole explanation. As a deployable but non-adaptive mitigation, reframing the reviewer as an adversarial pentester reduces evasion across the evaluated reviewer subset; pentester framed evasion ranges from 3.0% to 17.6%, and an open-weight Gemma-4-E4B-it reviewer under this framing detects 88.4% of attacks on the dataset with a 4.6% false-positive rate measured on 608 real-world GitHub PRs.

Tali Dror, Iftach Shoham, Moshe Buchris et al. · meta-ai

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. Audio examples of our proposed method can be found at https://iftach21.github.io/.

Assaf Lahiany, Oren Gal

Current visual SLAM systems face significant challenges in balancing computational efficiency with robust loop closure handling. Traditional approaches require careful manual tuning and incur substantial computational overhead, while learning-based methods either lack explicit loop closure capabilities or implement them through computationally expensive methods. We present AutoLoop, a novel approach that combines automated curriculum learning with efficient fine-tuning for visual SLAM systems. Our method employs a DDPG (Deep Deterministic Policy Gradient) agent to dynamically adjust loop closure weights during training, eliminating the need for manual hyperparameter search while significantly reducing the required training steps. The approach pre-computes potential loop closure pairs offline and leverages them through an agent-guided curriculum, allowing the model to adapt efficiently to new scenarios. Experiments conducted on TartanAir for training and validated across multiple benchmarks including KITTI, EuRoC, ICL-NUIM and TUM RGB-D demonstrate that AutoLoop achieves comparable or superior performance while reducing training time by an order of magnitude compared to traditional approaches. AutoLoop provides a practical solution for rapid adaptation of visual SLAM systems, automating the weight tuning process that traditionally requires multiple manual iterations. Our results show that this automated curriculum strategy not only accelerates training but also maintains or improves the model's performance across diverse environmental conditions.

Assaf Lahiany, Oren Gal

Curriculum Learning (CL), drawing inspiration from natural learning patterns observed in humans and animals, employs a systematic approach of gradually introducing increasingly complex training data during model development. Our work applies innovative CL methodologies to address the challenging geometric problem of monocular Visual Odometry (VO) estimation, which is essential for robot navigation in constrained environments. The primary objective of our research is to push the boundaries of current state-of-the-art (SOTA) benchmarks in monocular VO by investigating various curriculum learning strategies. We enhance the end-to-end Deep-Patch-Visual Odometry (DPVO) framework through the integration of novel CL approaches, with the goal of developing more resilient models capable of maintaining high performance across challenging environments and complex motion scenarios. Our research encompasses several distinctive CL strategies. We develop methods to evaluate sample difficulty based on trajectory motion characteristics, implement sophisticated adaptive scheduling through self-paced weighted loss mechanisms, and utilize reinforcement learning agents for dynamic adjustment of training emphasis. Through comprehensive evaluation on the diverse synthetic TartanAir dataset and complex real-world benchmarks such as EuRoC and TUM-RGBD, our Curriculum Learning-based Deep-Patch-Visual Odometry (CL-DPVO) demonstrates superior performance compared to existing SOTA methods, including both feature-based and learning-based VO approaches. The results validate the effectiveness of integrating curriculum learning principles into visual odometry systems.

Elchanan Zwecher, Eran Iceland, Sean R. Levy et al.

The challenge of mapping indoor environments is addressed. Typical heuristic algorithms for solving the motion planning problem are frontier-based methods, that are especially effective when the environment is completely unknown. However, in cases where prior statistical data on the environment's architectonic features is available, such algorithms can be far from optimal. Furthermore, their calculation time may increase substantially as more areas are exposed. In this paper we propose two means by which to overcome these shortcomings. One is the use of deep reinforcement learning to train the motion planner. The second is the inclusion of a pre-trained generative deep neural network, acting as a map predictor. Each one helps to improve the decision making through use of the learned structural statistics of the environment, and both, being realized as neural networks, ensure a constant calculation time. We show that combining the two methods can shorten the duration of the mapping process by up to 4 times, compared to frontier-based motion planning.