Colloquia and Seminars

DNA-based storage offers exceptional density and durability, but its scalability depends on efficiently synthesizing many strands in parallel under physical constraints. We study two complementary models capturing limitations arising from nucleotide availability and spatial interactions in large strand arrays.

First, we introduce complex synthesis sequences, a hybrid framework that bridges enzymatic synthesis, which permits strand-specific nucleotide additions, and photolithographic synthesis, which applies common additions across many strands. We characterize the achievable information rate through an analogue of the deletion ball, derive tight expressions for the maximal rate and its asymptotic behavior, and present a dynamic programming algorithm for computing an optimal synthesis sequence for known strands.

We then consider a two-dimensional array model motivated by optical and chemical coupling, in which strands follow a fixed global synthesis sequence and at most one strand per row may advance in each cycle. For two strands in a single row, we show that the laggard-first policy is asymptotically optimal among online policies without look-ahead, while one-symbol look-ahead yields a strict improvement in the binary case.

We further show that even globally optimal scheduling incurs an unavoidable expected overhead that grows linearly with strand length. These results are complemented by an offline dynamic programming algorithm and a constant-redundancy binary coding scheme with a deterministic worst-case synthesis-time guarantee. Together, the two works establish a unified framework for understanding the information-theoretic and scheduling limits of constrained DNA synthesis.

Integrating deep neural networks into high-stakes medical applications remains limited by concerns about robustness and reliability. In particular, models may rely on spurious correlations, such as the frequent co-occurrence of medical devices with pathology, rather than clinically meaningful features. Addressing this problem requires both detecting dataset-specific spurious signals and reducing model reliance on them. We propose a two-stage approach: a fast few-shot method for estimating spurious labels using only a small number of expert-annotated samples, followed by a lightweight framework that combines these estimated labels with available metadata using a novel supervised contrastive loss. Although designed for medical imaging, the proposed approach also generalizes beyond this domain. Our method achieves the strongest spurious-mitigation performance, demonstrating that combining reliable spurious-label estimates with available metadata can substantially improve robustness.

M.Sc. student under the supervision of Prof. Guy Gilboa and Dr. Eyal Gofer.

Generative diffusion models accept two primary inputs: a semantic conditioning signal that controls the output, and Gaussian noise that acts as a source of randomness. While this noise is rarely seen as a meaningful representation of the output image, our work demonstrates that, under constant general conditioning, it is actually a highly structured and compositional latent space of natural images. We expose this phenomenon by identifying semantic directions directly within the initial noise space, utilizing them for classification and for editing generated images along linear pathways. We even perform vector arithmetic in noise space in a compositional manner. These findings suggest that diffusion models naturally possess a Gaussian-type latent space rich in semantic structure. Key parts of this work were presented at CVPR 2026.

M.Sc. student under the supervision of Prof. Guy Gilboa.

My research focuses on mechanistic interpretability-understanding and improving multi-modal models by analyzing their internal computations. I will present recent studies, including Diffusion Lens (ACL 2024), Padding Tone (NAACL 2025), and Follow the Flow (ACL 2026), and DeLeaker (ICL 2026). These works propose methods to reveal what text-to-image models learn and how computations are performed internally. These insights are then applied to solve concrete problems, such as semantic leakage, paving the way for more reliable and transparent generative AI.

Backdoor attacks pose a significant threat to Large Language Models(LLMs), enabling malicious behaviors that activate only upon specific triggers while remaining dormant otherwise. Existing pre-deployment detection methods face fundamental barriers: scanning-based approaches are computationally intensive, while classification-based methods are architecture specific and ineffective for full-model detection. We demonstrate that backdoor injection leaves signatures in LLM representation spaces, detectable even over clean inputs. Building on this insight, we propose BackdoorID, a sample-free, representation-based framework applicable to both fixed-architecture and cross-architecture detection settings. We evaluate our approach across a diverse set of models, attack types, and architectures, matching or exceeding state-of-the-art baselines in all presented settings, further achieving 0.96 ROC-AUC in cross-architecture generalization.

As Large Language Models (LLMs) are used increasingly in research, creativity, and decision-making, it becomes crucial to ensure that their outputs are safe, reliable, and accurate. Yet, they often misbehave: they may memorize text, hallucinate responses, be jailbroken, or exhibit flawed logical reasoning.

Previous work has shown that the internal computational traces of LLMs, namely their hidden states, carry useful signals about different types of LLM misbehavior, e.g., the truthfulness of model outputs, and can be exploited to detect such misbehaviors.

However, these existing approaches largely ignore the structure and complexity of these traces and therefore fail to incorporate appropriate inductive biases for optimal performance. This talk presents a line of work on characterizing data structures in LLM computational traces and developing tailored architectures that learn from them, including final-layer outputs, intermediate activations, and attention matrices.

Our approaches identify misbehaviors more effectively than prior methods and generalize across tasks and LLMs, allowing techniques learned on one task or model to transfer to others.

Machine learning models are increasingly deployed in high-stakes applications where reliable uncertainty quantification is as important as predictive accuracy. In such settings, learning algorithms must not only generate accurate predictions but also communicate their uncertainty in a statistically principled manner. 

The first part of the research investigates predictive inference under imperfect training data. It first analyzes the robustness of conformal prediction and risk-controlling procedures to noisy labels, identifying conditions under which these methods construct reliable estimates with theoretical guarantees despite corrupted observations. Building upon these theoretical insights, the research studies predictive inference where additional features, referred to as privileged information, are available during training but unavailable at deployment. We develop calibration methods that account for the distribution shifts induced by missing or corrupted data: one based on distribution re-weighting and one relying on uncertainty-preserving imputation. We analyze the robustness of weighting-based calibration approaches to inaccuracies in the estimated weights and propose a triply robust calibration strategy that provides statistically valid prediction sets under complementary assumptions.

The second part considers uncertainty quantification for modern interactive AI systems, with a particular focus on large language models (LLMs). We formulate the evaluation of model safety and utility of agentic systems or LLMs as a time-to-event prediction problem and develop conformal survival methods that construct statistically valid predictive bounds on the number of interactions required before events of interest occur, such as an unsafe response of an LLM or a successful task completion of an agent. To improve the statistical efficiency of these procedures under limited computational resources, we introduce a theoretically valid dynamic budget allocation framework for sequential evaluation, enabling adaptive allocation of computational resources while preserving finite-sample distribution-free guarantees of the predictive bounds.

Collectively, the contributions presented in this research extend the scope of reliable predictive inference from classical supervised learning to increasingly practical machine learning settings. 

Contract theory studies how a principal can incentivize an agent to take costly actions. We study a multi-principal extension in which several strategic principals simultaneously offer linear contracts to a single agent, who then chooses a subset of principals to work for. In addition to incentivizing the agent, the contracts must form an equilibrium: no principal should benefit from unilaterally changing their contract. We study the computational problem of finding a stable outcome that maximizes the principals' total utility. In the symmetric setting, where all principals share the same reward function, we give polynomial-time algorithms for broad classes of reward functions, while showing NP-hardness and approximation hardness for submodular rewards. In the non-symmetric setting, the problem becomes NP-hard even for additive rewards, although we identify structured special cases that remain efficiently solvable.

Quantitative finite-state systems, such as quantitative automata and graph games, are typically time-symmetric. We investigate the consequences of introducing geometric decay into these systems, through the lenses of discounted-sum automata (NDAs) and Robin Hood bidding games.

NDAs are nondeterministic finite automata equipped with transition weights, where the value of a run is the discounted sum of its weights, and the value of a word is the minimum value over all its accepting runs. The determinization problem asks whether, given an NDA, there exists a deterministic discounted-sum automaton (DDA) that assigns the exact same value to every word. We prove that the determinization problem for NDAs with integral discounting factors is decidable. Specifically, we provide an EXPSPACE algorithm to decide determinizability alongside an explicit construction for the equivalent DDA, and we establish a PSPACE-hardness lower bound for the problem.

As for graph games, we focus on Bidding Games, where players are allocated monetary budgets and bid in auctions to determine movement along the graph. We enrich this model with a wealth-redistribution phase before each turn, which discounts the difference between the players’ budgets. For reachability objectives, we prove the existence of a threshold function - the exact initial budget required for the reachability player to guarantee a win.

We place the associated computational problem in NP. We also reveal that, unlike traditional models, a Robin Hood game may become undetermined exactly at the threshold. For Büchi objectives, we provide a computable candidate for a threshold.

A well-known application of the celebrated color-coding technique is fast detection of paths in a distributed setting with limited bandwidth, by propagating a signal along colorful paths [Even et al., DISC 2017]. However, since vertices of the same color are indistinguishable by this procedure, it cannot detect induced paths, as they appear the same as cycles.

We present a distributed local-counting technique that allows us to design fast algorithms for detecting induced paths, improving upon the state of the art in several aspects (time complexity, detection variant, and more).

The following observation illustrates our key technical ingredient: For each edge, count the number of signals, over all permutations of the random coloring, that terminate at that edge. An induced path contributes one to that count, while a cycle contributes two. Refining and generalizing this idea lets us distinguish induced 4-paths from other 4-vertex subgraphs in constant time and yields fast algorithms for detecting paths of lengths 5 and 6.