Iván Arcuschin Moreno

Iván Arcuschin Moreno

Lead Research Scientist

Poseidon Research

I’m Iván, Lead Research Scientist at Poseidon Research working on AI Safety.

I earned my Computer Science PhD from the University of Buenos Aires, Argentina. Since 2024, I’ve focused on AI Safety research through the MATS program, publishing at NeurIPS, ICML, and ICLR.

My research centers on understanding and evaluating the internal reasoning of large language models. In 2024, I co-created InterpBench (NeurIPS 2024), a benchmark for evaluating Mechanistic Interpretability techniques.

My work on Chain-of-Thought unfaithfulness (ICLR 2025 Workshop) showed that frontier LLMs produce subtly unfaithful reasoning under realistic conditions, accumulating 137 citations in under a year.

Publications

Biases in the Blind Spot: Detecting What LLMs Fail to Mention

Large Language Models (LLMs) often provide chain-of-thought (CoT) reasoning traces that appear plausible, but may hide internal biases. …
Preprint
arXiv

Base Models Know How to Reason, Thinking Models Learn When

Why do thinking language models like DeepSeek R1 outperform their base counterparts? Despite consistent performance gains, it remains …
Mechanistic Interpretability Workshop (NeurIPS 2025)
arXiv

Chain-of-Thought Is Not Explainability

Chains-of-thought (CoT) allow language models to verbalise multi-step rationales before producing their final answer. While this …
Oxford AI Governance Initiative (AIGI)
Preprint

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

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