Iván Arcuschin Moreno

Independent Researcher

I’m Iván, an Independent Researcher working on AI Safety.

I earned my Computer Science PhD from the University of Buenos Aires, Argentina, specializing in Automated Test Generation for Android apps. My research has been published in top Software Engineering conferences such as ICST and AST.

Since then, I’ve been participating in the ML Alignment & Theory Scholars (MATS) Program, working on AI Safety. In 2024, I worked on InterpBench, a benchmark of semi-synthetic transformers with known circuits for evaluating Mechanistic Interpretability techniques. This work was published at NeurIPS 2024.

In 2025, I’ve been working on the paper Chain-of-Thought Reasoning in the Wild is Not Always Faithful. In this work, we show that frontier thinking and non-thinking LLMs can produce subtle unfaithfulness – they can construct superficially coherent yet contradictory arguments, selectively interpreting ambiguous information or switching between different types of arguments to align with their inherent biases. In contrast to other research on unfaithfulness, this form of unfaithfulness occurs without any explicit prompting.

Publications

MIB: A Mechanistic Interpretability Benchmark

How can we know whether new mechanistic interpretability methods achieve real improvements? In pursuit of meaningful and lasting …

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

42nd International Conference on Machine Learning (ICML 2025)

PDF

Chain-of-Thought Reasoning In The Wild Is Not Always Faithful

Chain-of-Thought (CoT) reasoning has significantly advanced state-of-the-art AI capabilities. However, recent studies have shown that …

Iván Arcuschin, Jett Janiak, Robert Krzyzanowski, Senthooran Rajamanoharan, Neel Nanda, Arthur Conmy

Workshop on Reasoning and Planning for Large Language Models (ICLR 2025)

PDF

Understanding Reasoning in Thinking Language Models via Steering Vectors

Recent advances in large language models (LLMs) have led to the development of thinking language models that generate extensive …

Constantin Venhoff, Iván Arcuschin, Philip Torr, Arthur Conmy, Neel Nanda

Workshop on Reasoning and Planning for Large Language Models (ICLR 2025)

PDF

InterpBench: Semi-Synthetic Transformers for Evaluating Mechanistic Interpretability Techniques

Mechanistic interpretability methods aim to identify the algorithm a neural network implements, but it is difficult to validate such …

Rohan Gupta, Iván Arcuschin, Thomas Kwa, Adrià Garriga-Alonso

38th Conference on Neural Information Processing Systems (NeurIPS 2024) Track on Datasets and Benchmarks.

PDF

More publications