Hadas Orgad

Organizing the second Workshop on Actionable Interpretability at COLM in San Francisco.

Hidden Failures in Robustness: Why Supervised Uncertainty Quantification Needs Better Evaluation accepted to EMNLP 2026.

Interpretability Can Be Actionable accepted to ICML 2026.

Organized the first Workshop on Actionable Interpretability at ICML 2025.

Hidden Failures in Robustness: Why Supervised Uncertainty Quantification Needs Better Evaluation

Joe Stacey, Hadas Orgad, Kentaro Inui, Benjamin Heinzerling, Nafise Sadat Moosavi

We systematically evaluate supervised uncertainty probes across models, tasks, and out-of-distribution settings. The results reveal substantial hidden robustness failures and show that choices such as representation layer and token aggregation can strongly affect generalization under distribution shift.

Interpretability Can Be Actionable

Hadas Orgad, Fazl Barez, Tal Haklay, Isabelle Lee, Marius Mosbach, Anja Reusch, Naomi Saphra, Byron Wallace, Sarah Wiegreffe, Eric Wong, Ian Tenney, Mor Geva

This position paper argues that interpretability should be evaluated by actionability: whether an insight enables concrete decisions or interventions beyond interpretability research itself. We propose a framework for evaluating actionability and identify domains where interpretability can provide unique practical leverage.

Large Language Models Generate Harmful Responses Using a Distinct Mechanism, Shared Across Harm Types

Hadas Orgad, Boyi Wei, Kaden Zheng, Martin Wattenberg, Peter Henderson, Seraphina Goldfarb-Tarrant, Yonatan Belinkov

We use targeted parameter pruning as a causal intervention to study how harmful response generation is organized in LLMs. We find a sparse set of critical parameters that is shared across harm types and largely separable from general capabilities, and show that harmful generation can be dissociated from related behaviors such as recognition, explanation, and refusal.

Agents of Chaos

Natalie Shapira, Chris Wendler, Avery Yen, Gabriele Sarti, Koyena Pal, Olivia Floody, Adam Belfki, Alex Loftus, Aditya Ratan Jannali, Nikhil Prakash, Jasmine Cui, Giordano Rogers, Jannik Brinkmann, Can Rager, Amir Zur, Michael Ripa, Aruna Sankaranarayanan, David Atkinson, Rohit Gandikota, Jaden Fiotto-Kaufman, EunJeong Hwang, Hadas Orgad, P Sam Sahil, Negev Taglicht, Tomer Shabtay, Atai Ambus, Nitay Alon, Shiri Oron, Ayelet Gordon-Tapiero, Yotam Kaplan, Vered Shwartz, Tamar Rott Shaham, Christoph Riedl, Reuth Mirsky, Maarten Sap, David Manheim, Tomer Ullman, David Bau

An exploratory red-teaming study of autonomous language-model agents deployed with persistent memory, communication tools, file systems, and shell access. Across realistic interactions, we document security-, privacy-, and governance-relevant failures that emerge from combining language models with autonomy and tool use.

Inside-Out: Hidden Factual Knowledge in LLMs

Zorik Gekhman, Eyal Ben David, Hadas Orgad, Eran Ofek, Yonatan Belinkov, Idan Szpektor, Jonathan Herzig, Roi Reichart

This work introduces a framework to measure "hidden knowledge" in large language models—cases where models internally know the correct answer but fail to express it in their outputs. By comparing internal and external knowledge across three LLMs, we find a consistent 40% gap, with some answers never generated despite perfect internal knowledge.

LLMs Know More Than They Show: On the Intrinsic Representation of LLM Hallucinations

Hadas Orgad, Michael Toker, Zorik Gekhman, Roi Reichart, Idan Szpektor, Hadas Kotek, Yonatan Belinkov

This work shows that large language models internally encode rich information about the truthfulness of their outputs, concentrated in specific tokens—enabling strong error detection and even prediction of error types. However, this encoding is not universal across datasets. Additionally, models may still produce incorrect answers despite internally encoding the correct one.

MIB: A Mechanistic Interpretability Benchmark

Aaron Mueller, Atticus Geiger, Sarah Wiegreffe, Dana Arad, Iván Arcuschin, Adam Belfki, Yik Siu Chan, Jaden Fiotto-Kaufman, Tal Haklay, Michael Hanna, Jing Huang, Rohan Gupta, Yaniv Nikankin, Hadas Orgad, Nikhil Prakash, Anja Reusch, Aruna Sankaranarayanan, Shun Shao, Alessandro Stolfo, Martin Tutek, Amir Zur, David Bau, Yonatan Belinkov

A new benchmark designed to evaluate whether mechanistic interpretability methods truly improve our understanding of language models. MIB includes two tracks—circuit localization and causal variable identification—across multiple tasks and models.

Padding Tone: A Mechanistic Analysis of Padding Tokens in T2I Models

Michael Toker, Ido Galil, Hadas Orgad, Rinon Gal, Yoad Tewel, Gal Chechik, Yonatan Belinkov

This work presents the first in-depth analysis of how padding tokens affect text-to-image generation. Using two causal techniques, we find that padding tokens can influence image generation at different stages—during text encoding, the diffusion process, or not at all—depending on model architecture and training setup.

Position-Aware Automatic Circuit Discovery

Tal Haklay, Hadas Orgad, David Bau, Aaron Mueller, Yonatan Belinkov

We present an automated, position-aware circuit discovery pipeline that differentiates token positions and uses dataset schemas to capture cross-positional mechanisms with better faithfulness at smaller sizes.

Diffusion Lens: Interpreting Text Encoders in Text-to-Image Pipelines

Michael Toker, Hadas Orgad, Mor Ventura, Dana Arad, Yonatan Belinkov

We introduce the Diffusion Lens method to analyze text encoders intermediate representations in text-to-image models. Using the method, we perform an extensive analysis of two recent T2I models, and gain insights on both conceptual combination abilities and knowledge retrieval in the models.

ReFACT: Updating Text-to-Image Models by Editing the Text Encoder

Dana Arad*, Hadas Orgad*, Yonatan Belinkov

Text-to-image models are trained on extensive amounts of data, leading them to implicitly encode factual knowledge within their parameters. While some facts are useful, others may be incorrect or become outdated (e.g., the current President of the United States). We introduce ReFACT, a method for updating text-to-image models. ReFACT updates the weights of a specific layer in the text encoder, only modifying a tiny portion of the model's parameters, and leaving the rest of the model unaffected.

Unified Concept Editing in Diffusion Models

Rohit Gandikota, Hadas Orgad, Yonatan Belinkov, Joanna Materzyńska, David Bau

Text-to-image models suffer from various safety issues that may limit their suitability for deployment. Previous methods have separately addressed individual issues of bias, copyright, and offensive content in text-to-image models. However, in the real world, all of these issues appear simultaneously in the same model. In this paper, we present a method that tackles those diverse issues with a single approach.

Editing Implicit Assumptions in Text-to-Image Diffusion Models

Hadas Orgad*, Bahjat Kawar*, Yonatan Belinkov

Text-to-image diffusion models often make implicit assumptions about the world when generating images. While some assumptions are useful (e.g., the sky is blue), they can also be outdated, incorrect, or reflective of social biases present in the training data. In this work, we aim to edit a given implicit assumption in a pre-trained diffusion model. Our method is highly efficient, as it modifies a mere 2.2% of the model's parameters in under one second.

Debiasing NLP Models Without Demographic Information

Hadas Orgad, Yonatan Belinkov

In this work, we propose a debiasing method that operates without any prior knowledge of the demographics in the dataset, detecting biased examples based on an auxiliary model that predicts the main model's success and down-weights them during the training process. Results on racial and gender bias demonstrate that it is possible to mitigate social biases without having to use a costly demographic annotation process.

Choose your lenses: Flaws in gender bias evaluation

Hadas Orgad, Yonatan Belinkov

Considerable efforts to measure and mitigate gender bias in recent years have led to the introduction of an abundance of tasks, datasets, and metrics used in this vein. In this position paper, we assess the current paradigm of gender bias evaluation and identify several flaws in it.

How Gender Debiasing Affects Internal Model Representations, and Why It Matters

Hadas Orgad, Seraphina Goldfarb-Tarrant, Yonatan Belinkov

Common studies of gender bias in NLP focus either on extrinsic bias measured by model performance on a downstream task or on intrinsic bias found in models’ internal representations. However, the relationship between extrinsic and intrinsic bias is relatively unknown. In this work, we illuminate this relationship by measuring both quantities together.

Research on AI “knowing when it’s wrong”

Text-to-Image Model Editing