To join the email distribution list of the cs colloquia, please visit the list subscription page.
Computer Science events calendar in HTTP ICS format for of Google calendars, and for Outlook.
Academic Calendar at Technion site.
We introduce VideoMDM, a diffusion-based framework that trains 3D human motion priors directly from accurate 2D poses extracted from monocular videos, without any 3D ground truth. A pretrained 2D-to-3D lifter provides approximate 3D pose sequences that serve as a noisy teacher: these are diffused, denoised by the model in 3D, and supervised in 2D by reprojecting the prediction and comparing against accurate keypoints. We show that, under mild assumptions, a depth-weighted 2D reprojection loss is equivalent in expectation to direct 3D supervision, and we adapt standard 3D motion regularizers - velocity consistency and over-parameterized representation alignment - to this 2D setting. Unlike methods that lift 2D to 3D only at inference, VideoMDM learns a coherent 3D motion manifold during training. On HumanML3D it nearly closes the gap to fully 3D-supervised MDM (FID 0.88 vs 0.54); On real video datasets Fit3D and NBA the method learns to generate motions consistently preferred by humans, with strong quantitative results.
Mechanistic interpretability (MI) enables monitoring and steering LLMs by
operating on their internal state during inference. MI applications
remain difficult to develop, expensive to execute, and challenging to manage
at runtime.
We present xMIx, a system that integrates MI applications as
first-class components of LLM serving, enabling simplified development,
efficient execution, and flexible deployment. Guided by a comprehensive
analysis of modern MI applications, xMIx streamlines the MI apps
development via a simple Python-based Domain Specific Language (DSL) for
accessing intra- and inter-layer runtime model state at well-defined trigger
points during inference. xMIx enables efficient access to ephemeral attention
scores via a novel FlashInspect kernel that retains the performance benefits of
FlashAttention.
xMIx's compiler integrates multiple MI applications into the performance-optimized inference execution flow while ensuring conflict-free coexistence. xMIx's
runtime supports dynamic, per-request activation and deactivation of MI apps
without service interruption or performance degradation when inactive,
leveraging xMIx' dynamic memory management and CUDA graph-compatible
execution.
We integrate xMIx with vLLM and evaluate ten MI apps, including sparse
autoencoders, steering, and jailbreak detection, across four
state-of-the-art LLMs. xMIx incurs negligible execution overhead, up to two
orders of magnitude lower than existing frameworks, while enabling per-request
activation and deactivation up to 1,000 times faster than alternatives.
These results demonstrate that xMIx makes MI applications practical for
deployment in production LLM serving systems.
Training LLMs at scale is limited by GPU memory. While most memory-efficient methods focus on compressing model weights, gradients and optimizer states, we show that the activations stored for the backward pass are the real bottleneck, as they scale with batch size and context length. In CompAct [1], we show that these activations are so redundant that storing low-rank random projections of them, instead of the full tensors, reduces peak GPU memory by 25-30% in pretraining and by 50% in fine-tuning, without hurting model performance. Then, in PAMM [2], we show that the activations of the attention projection layers specifically can be compressed by up to 512x by exploiting their redundancy along the token dimension: a small random subset of tokens is kept, and every other token is represented as a scaled copy of its most similar kept token, matching or even improving perplexity for LLaMA models of up to 7B parameters. Finally, we present AnchorKV [3], where we show how the redundancy exploited by PAMM can also be used to compress the KV cache, the memory bottleneck in LLM inference. By storing a few anchor tokens exactly, representing every other token through its nearest anchor, and adding quantized residuals where they matter most, AnchorKV compresses the cache 20x without discarding a single token, retaining 99.4% of full-cache accuracy on Llama-3.1-70B. Together, these works show that the intermediate tensors that dominate LLM memory are far more redundant than they appear, and can be compressed aggressively while maintaining performance.
[1] Y. Shamshoum*, N. Hodos*, Y. Sieradzki, A. Schuster. CompAct: Compressed Activations for Memory-Efficient LLM Training. NAACL 2025.
[2] M. Khalaf, Y. Shamshoum, N. Hodos, Y. Sieradzki, A. Schuster. QKV Projections Require a Fraction of Their Memory. ICLR 2026.
[3] M. Khalaf, Y. Shamshoum, N. Hodos, Y. Sieradzki, A. Schuster. AnchorKV: Anchor-Residual KV Cache Compression. Under review at NAACL 2027.
506, Zisapel Building
Transformers are widely assumed to exhibit token-decoding explainability. In particular, in vision transformer (ViT) classifiers, decoding individual patch tokens has become a common diagnostic tool due to its tendency to yield human-interpretable outputs. In light of the increasing popularity of this approach, an important question is whether token-decoding explainability is an inherent property of transformers that should be expected to hold across architectures, scales, and training objectives, or merely an artifact of some implicit bias. We systematically study token-decoding explainability across diverse ViT settings, spanning architectures, scales, and training objectives, using both controlled experiments and real-world models. We find that while some models indeed exhibit aligned and interpretable token decodings, others exhibit either inverted explainability maps or a complete lack of correlation with the image semantics. As we show, these tendencies toward different explainability behaviors are not intrinsic to the architecture and can vary arbitrarily with changes in the optimization landscape. We provide theoretical support for these observations by demonstrating a setting where a single fixed toy transformer architecture can provably implement the optimal classification rule in multiple ways, which differ in their patch-decoding explainability patterns. Our results show that token-decoding explainability is not an inherent property of trained transformers, but rather an outcome contingent on the problem setting and thus should not be a-priori assumed to hold.
Ronen Schaffer is an M.Sc. candidate in Electrical Engineering at the Technion, under the supervision of Prof. Tomer Michaeli
Taub 601
Decisions in healthcare, public policy, economics, and many other fields often rest on causal conclusions drawn from observational data. These conclusions can be surprisingly fragile: ordinary data problems such as duplicate records, entry mistakes, or corrupted subsets can substantially change, or even reverse, an estimated causal effect. This raises a basic question: how robust is a causal claim to small, targeted changes in the data? We address this question through the lens of cardinality repairs and introduce SubCure, a framework for auditing causal robustness. Given a causal query and a desired range for the estimated effect, SubCure finds a small set of records, or a compact subpopulation described by attribute-value predicates, whose removal moves the estimate into that range. The size of this repair is an intuitive, data-centric, model-agnostic measure of how fragile a conclusion is. The removed data reveals which parts of the dataset are responsible for the observed effect. We formalize the problem for both individual records and subpopulations. To handle large datasets, we design scalable search algorithms that update causal estimates incrementally, rather than recomputing them from scratch after every deletion. Experiments on four real-world datasets and a synthetic benchmark show that SubCure finds smaller and more influential subsets than eight baseline methods, scales to datasets with millions of records, and exposes vulnerabilities that traditional sensitivity analyses miss.
Taub 401
The Quantum Approximate Optimization Algorithm (QAOA) offers a variational approach to combinatorial optimization. When combined with quantum random access code (QRAC) compression, the quality of classical solutions depends not only on the optimized quantum energy but also on the initial state and the decoding procedure.
In this seminar, we investigate SEMI-QAOA, a family of compressed QAOA algorithms initialized in tensor products of X-basis eigenstates, with corresponding state-dependent mixers. We first introduce Rand-SEMI-QAOA, which explores ensembles of randomly selected initialization masks, demonstrating substantial improvements in decoded MaxCut performance over conventional initialization.
We then investigate structured approaches that select a single initialization mask before variational angle optimization, reducing the need to optimize large ensembles of quantum circuits. We establish theoretical separations showing that conventional initialization can suffer from persistent decoding obstructions, while suitable X-basis masks overcome these limitations, even at shallow circuit depth.
Finally, we present numerical comparisons across combinatorial optimization problems, examining solution quality, computational requirements, and the relationship between quantum energy optimization and classical decoding performance. The results highlight initialization-mask selection as an important algorithmic component of compressed quantum optimization.
Taub 337
While digital AI and large multimodal models have achieved massive scaling laws in virtual spaces, grounding artificial intelligence in the unstructured, non-deterministic physical world exposes severe foundational bottlenecks. Physical AI -the synthesis of embodied robotics, multi-modal perception, foundation world models, and real-time control- cannot rely merely on scaling compute or self-supervised text scraping. The transition to embodied agents operating safely at human-level dexterity and situational reasoning presents acute open problems: closing the sim-to-real gap, formulating verifiable neural world models, scaling physics-grounded reinforcement learning under sparse supervision, and developing low-latency edge inference capable of deterministic safety guarantees.
This talk bypasses off-the-shelf industrial deployments to outline the critical scientific and algorithmic roadblocks where industry and public frameworks stall without academic intervention. We will dissect the open research questions Israel must tackle to pioneer embodied autonomy: How do we construct mathematically rigorous bounds for agent safety in unstructured physical environments? What new paradigms in spatial reasoning and tactile-proprioceptive sensor fusion can replace sample-inefficient learning?
Short bio: Guy Altagar is the Vice President and Head of Physical AI at the Israeli National AI Directorate within the Prime Minister's Office, where he leads national strategy, infrastructure development, and R&D testbed initiatives for embodied artificial intelligence and robotics. Prior to his government role, he was the founder and Chief Executive Officer of Unlimited Robotics, leading the development of autonomous service robot platforms and software operating systems for embodied devices. With an interdisciplinary background spanning deep-tech entrepreneurship, engineering leadership, and law, Guy works closely with academic research institutions, government bodies, and international industry leaders to accelerate national-scale computing infrastructure, physical test environments, and cross-disciplinary research in physical autonomy.
Taub 337
AI is driving unprecedented growth in data center compute, with the market projected to approach $2 trillion by 2030 and revenue growth averaging approximately 24% annually from 2025 to 2030. As AI adoption progresses from infrastructure and physical systems to scientific discovery, it is creating a major opportunity for verification and in particular formal verification - the crown of design. Natural-language interfaces and AI-driven productivity are introducing a new level of abstraction, while emerging examples of AI-assisted mathematical reasoning demonstrate the potential to extend scientific and engineering capabilities.
This keynote traces the evolution from the man–machine communication vision presented at DAC 1964, through collective intelligence and human-in-the-loop formal methods presented at JUG 2013, to today’s agentic verification workflows in Cadence ChipStack. These workflows combine specialized formal agents, proof engines, machine learning, and expert guidance to address properties that remain difficult to prove manually. By interleaving design and verification, and using platforms such as Cadence ChipStack, the Formal Verification AI Flywheel can transform productivity at scale into greater proof power, enabling engineers to build increasingly complex systems with higher confidence and much less manual effort.
Short bio: Ziyad Hanna is a Corporate Vice President at Cadence Design Systems and R&D General Manager of Cadence Israel. He leads system verification products and Cadence-wide AI initiatives for autonomous chip-design development. Previously, he served as Senior Vice President of Jasper and senior PE and management roles at Intel, where he developed three generations of formal-verification suites widely deployed since the start of Pentium. Hanna earned degrees from Tel Aviv University and a PhD from Oxford University, where he is a visiting professor. An IEEE Senior Member, he has published over 30 articles, 20 patents, and delivered hundreds of industrial and academic presentations and talks. His honors include Intel IAA awards and an honorary doctorate from Ruppin Collage. Ziyad initiated a formal-verification course at the Technion that attracts hundreds of students interested in pursuing advanced studies in the field.
Taub 601
Synthetic deoxyribonucleic acid (DNA) has emerged as a promising archival storage medium due to its extraordinary information density and multi-century stability. However, the biochemical processes of DNA synthesis and sequencing introduce fundamental coding-theoretic and algorithmic bottlenecks—most notably the lack of spatial ordering among stored strands, high sequencing coverage overhead, and synchronization (insertion and deletion) errors. In this talk, I will present the main contributions of my PhD research, focusing primarily on coding schemes and algorithms designed to overcome these challenges across the DNA data storage pipeline:
End-to-End Error Correction: Introducing DNA-correcting codes, which replace the traditional decoupled pipeline of clustering, trace reconstruction, and decoding with a unified coding framework that guarantees unambiguous end-to-end recovery.
Random Access Coverage Depth: Minimizing the expected number of sequencing reads needed to retrieve a single designated information strand—resolving the $k=2$ case completely and constructing recovery-complete codes that improve upon state-of-the-art bounds for general $k$.
Synchronization Errors: Developing black-box algorithmic reductions that transform off-the-shelf deletion decoders into general insertion–deletion decoders, alongside linear-time, constant-space algebraic decoders for two-dimensional Reed–Solomon codes.
I will conclude with a brief overview of the second part of my dissertation, which investigates combinatorial and extremal problems in covering codes, including binary nearly perfect $1$-covering codes, pair-covering codes, and all-symbol PIR and batch codes.
Modern SAT solvers are, at their core, proof engines: when a formula is unsatisfiable, a run of the solver corresponds to a proof of this fact. The dominant algorithm, conflict-driven clause learning (CDCL), learns clauses by resolution, an inference rule that preserves logical equivalence: every learned clause is implied by the formula. Resolution is sound and complete, but it is a weak proof system. Some formulas, such as those encoding the pigeonhole principle, have only exponential-size resolution proofs. On these formulas CDCL cannot succeed, however well it is engineered. Stronger proof systems overcome this limitation by allowing steps that preserve only satisfiability, not logical equivalence. Propagation redundant (PR) clauses are a prominent example. A PR clause may exclude models of the formula, as long as some model remains. PR-based proof systems have short proofs of many formulas that are hard for resolution, without introducing new variables. The price is that checking whether a clause is PR is harder. The check requires a witness, which transforms every model of the formula that falsifies the clause into a model that satisfies it. Every clause of the formula is an obligation of this check, since it may be the clause that the witness falsifies. Hence, the check becomes harder as the solver learns more clauses. In practice, state-of-the-art techniques learn PR clauses only during preprocessing, before CDCL starts learning.
In this talk, we introduce Scaffolded DPR (SDPR), a clausal proof system that generalizes DPR, the proof system whose steps add PR clauses and delete clauses. SDPR divides the clauses into core clauses, which are both premises and obligations of the check, and scaffold clauses, which are only premises. Hence, the scaffold clauses, in particular the clauses learned by CDCL, strengthen the check rather than making it harder. In addition, SDPR maintains a set of repair literals. These allow deleting core clauses with a witness, as simplifications such as variable elimination do, while every scaffold clause remains. We show that SDPR is sound, and that SDPR and DPR polynomially simulate each other. Based on SDPR, we present InPRocess, an inprocessing technique that learns PR clauses during the search, alongside the solver’s other simplification techniques. We implemented InPRocess in CaDiCaL and evaluated it on the main track of the SAT Competition 2026. InPRocess outperforms both CaDiCaL and its preprocessing counterpart PReLearn: it solves more formulas, including formulas that neither of them solves.
Modern computer vision systems achieve strong performance, but often depend on extensive manual annotations and substantial computational resources. This PhD seminar explores how the structure of visual data can help overcome these limitations, focusing on learning segmentation with limited supervision and enabling efficient visual processing.
I will begin with semi-supervised semantic segmentation, showing how spatial relationships between neighboring predictions can improve pseudo-label reliability and make better use of unlabeled images. I will then turn to instance segmentation for robotic grasping; beyond annotation scarcity, I will examine annotation quality, introducing controlled benchmarks that reveal how spatial errors and automatically generated masks affect segmentation performance.
The second part extends this efficiency perspective to vision–language models (VLMs). I will present approaches that predict sufficient input resolution, retrieve high-resolution regions when additional evidence is needed, and progressively encode visual detail using wavelet representations and computation reuse. These approaches adapt visual processing to the requirements of each task, offering control over the trade-off between accuracy and computational cost.
Together, these studies investigate how visual structure and task-dependent information can support effective learning and inference with fewer resources.
Taub 337
The interesting question in AI is no longer whether a model runs in some datacenter -it is whether a coordinated system of models can run on the device in front of you: privately, in real time, with no network. This talk is about the platform that makes that possible.
We start from the foundations of OpenVINO as an edge AI platform: the Intermediate Representation, weight compression and data-aware quantization (INT8/INT4) that make large models fit, and -the central idea -heterogeneous execution, where CPU, GPU, and the NPU are treated as one fabric and each model is placed on the engine that suits it. We show that on a modern AI PC, device placement is a first-class design decision, and we make the case with real numbers: a measured, head-to-head benchmark of live speech-to-text across CPU, GPU, and NPU, where the winner is chosen not by raw latency but by predictability, power, and freeing the accelerator -then generalized into a multi-device task assignment for a whole application.
To ground it, we walk a locally running multi-agent showcase - a crew of specialized agents orchestrated by a Director, each modality on its optimal engine - and the agentic engineering behind it (MCP, Skills, multi-agent patterns, and a "high agency, low surprise" operating contract). We close with Physical AI: an offline, camera-driven tabletop kiosk that turns a scene into spoken, multilingual story - every stage on the same OpenVINO runtime.
Short bio: Gila Kamhi acts as Chief AI Officer within the Wireless and Connectivity Solutions Division at Intel. Her team is committed to democratizing AI across the division, aiming to leverage AI as a catalyst for growth and scalability. They focus on accelerating product development processes and enhancing wireless platform-level experiences via intro to the organization available off-the-shelf solutions provided by Intel IT and also by developing custom solutions . Additionally, Gila's team contributes to advancing Intel's PC AI strategy, collaborating with other AI teams to highlight the unique differentiation of Intel's platform in AI solutions. Before her current role, Gila led conversational AI solutions R&D teams at General Motors, focusing on novel in-cabin human-machine interaction for automated/autonomous driving. In her first tenure at Intel, from 2010 to 2015, she acted as the lead of Advanced Human Computer Interaction Solutions group at perceptual computing division, where she spearheaded the exploration of new user experiences using Intel's 3D camera technology. She has several patents and numerous publications to her name on a wide range of domains (e.g., formal methods, AI, human computer interaction).