Will Merrill

Also known as William Merrill, lambdaviking

Will Merrill is a researcher whose work tests what language-model architectures can compute. He mapped formal limits for recurrent networks and transformers, studied chain-of-thought reasoning and state-space models, and helped develop open language models such as OLMo. He earned a PhD at New York University, shared an ACL Outstanding Paper Award in 2025, and coauthored OLMo 3 and Olmo Hybrid. As of October 2026, he is an Assistant Professor at TTIC and a part-time researcher at Ai2, where he develops theory for transformer computation.

Now at: Allen Institute for AI (Ai2) and Toyota Technological Institute at Chicago

Timeline

2020

A hierarchy of RNNs

In 2020, he coauthors the ACL paper “A Formal Hierarchy of RNN Architectures,” placing saturated RNNs, counter machines and weighted finite automata in a two-dimensional hierarchy.
2022

Saturated transformers in TC0

In 2022, he coauthors “Saturated Transformers are Constant-Depth Threshold Circuits” with Ashish Sabharwal and Noah A. Smith. The paper places the formal languages recognized by floating-point saturated transformers within TC0.
2023

Formal-language tutorial at ICGI

In 2023, he presents an ICGI tutorial on formal languages and neural models for learning on sequences, surveying neural networks as formal language recognizers.
2024

OLMo and model limits

In February 2024, he coauthors OLMo, a language model released with open training data and training and evaluation code; the paper receives ACL’s Best Theme Paper Award in 2024. His ICLR 2024 paper with Ashish Sabharwal shows that a linear number of decoding steps with projected pre-norm lets transformers recognize all regular languages under standard complexity conjectures. At ICML 2024, his paper with Jackson Petty and Ashish Sabharwal shows that state-space models cannot express computation outside TC0.
June 2025–October 2025

A PhD and ACL award

By June 2025, he holds a PhD from New York University, earned while working with Tal Linzen and Ashish Sabharwal. At ACL 2025, he shares an Outstanding Paper Award with Michael Y. Hu, Jackson Petty, Chuan Shi and Linzen for “Between Circuits and Chomsky: Pre-pretraining on Formal Languages Imparts Linguistic Biases.” His dissertation, “A theory of the computational power and limitations of language modeling architectures,” goes online in October 2025.
June 2025–December 2025

Ai2 and OLMo 3

In June 2025, TTIC names him one of three incoming tenure-track assistant professors who will join in fall 2026; he is a Young Investigator at the Allen Institute for AI. In December 2025, he coauthors Olmo 3, a fully open family of 7B and 32B language models.
April 2026

The Olmo Hybrid model

In April 2026, he is first author of “Olmo Hybrid: From Theory to Practice and Back,” a 7B-parameter model that replaces Olmo 3 7B’s sliding-window layers with Gated DeltaNet layers and outperforms Olmo 3 across standard pretraining and mid-training evaluations. Its theory shows that hybrid models can express tasks beyond both transformers and linear RNNs, such as code execution.
As of October 2026

Assistant Professor at TTIC

As of October 2026, he is an Assistant Professor at TTIC and a part-time researcher at the Allen Institute for AI (Ai2), developing theory on the computational power and limitations of transformers. In October 2026, he leads “From Transformers to Logic, Provably” with David Chiang of the University of Notre Dame, backed by a two-year Coefficient Giving grant to make the internal reasoning of large language models legible to human investigators.