Research Interests

I work on AI agents that forecast and do research: building them, and measuring whether they are any good. Before that, I worked on probability theory and statistical mechanics, in particular quantum spin systems, sampling algorithms, couplings, and cluster expansion methods.

Forecasting and AI

Forecasting Accuracy Scales with Research

FutureSearch, 2026

We mined 52,247 findings from 4,517 forecasting-agent rationales and gave growing slices of them back to two forecasting agents. Both improved, and neither curve flattened.

The Best Forecasters Agree with Themselves

FutureSearch, 2026 · with D. Schwarz

Worldview consistency ranks forecasting agents without a single resolved question. Repairing the inconsistencies it finds improved all nine agents we tested, four of them significantly.

Iterating on a Forecaster without Waiting for the Future

FutureSearch, 2026 · with D. Schwarz

How we develop a forecasting agent on 1,907 already-resolved questions, why paired comparisons on a large sample matter, and what the agent harness contributes.

Bench to the Future: A Pastcasting Benchmark for Forecasting Agents

arXiv, 2025 · with the FutureSearch team

Hundreds of already-resolved forecasting questions, each with an offline corpus of tens of thousands of relevant web pages, so forecasting agents can be evaluated now instead of waiting for questions to resolve.

Deep Research Bench: Evaluating AI Web Research Agents

arXiv, 2025 · with the FutureSearch team

89 multi-step web research tasks with human-worked answers, run against a frozen snapshot of the web so that models can be compared reliably over time.

Contra Papers Claiming Superhuman AI Forecasting

LessWrong, 2024 · with N. Bosse, L. Phillips, and D. Schwarz

Why the 2024 papers presenting LLMs as rivalling or beating human forecasters don't support that claim, and what a fair comparison would need.

Exploring Metaculus's AI Track Record

EA Forum, 2023 · with P. Scoblic and W. Aldred

How well Metaculus forecasters have predicted developments in AI.

Mathematics

Critical Parameters for Loop and Bernoulli Percolation

ALEA, 2021 · arXiv

We consider a class of random loop models (including the random interchange process) that are parametrised by a time parameter \(\beta\geq0\). Intuitively, larger \(\beta\) means more randomness. In particular, at \(\beta=0\) we start with loops of length 1 and as \(\beta\) crosses a critical value \(\beta_c\), infinite loops start to occur almost surely. Our random loop models admit a natural comparison to bond percolation with \(p=1-e^{-\beta}\) on the same graph to obtain a lower bound on \(\beta_c\). For those graphs of diverging vertex degree where \(\beta_c\) and the critical parameter for percolation have been calculated explicitly, that inequality has been found to be an equality. In contrast, we show in this paper that for graphs of bounded degree the inequality is strict, i.e. we show existence of an interval of values of \(\beta\) where there are no infinite loops, but infinite percolation clusters almost surely.

Dimerization in Quantum Spin Chains with O(n) Symmetry

Communications in Mathematical Physics, 2021 · with J.E. Björnberg, B. Nachtergaele, and D. Ueltschi · arXiv · video

We consider quantum spins with \(S\geq1\), and two-body interactions with \(O(2S+1)\) symmetry. We discuss the ground state phase diagram of the one-dimensional system. We give a rigorous proof of dimerization for an open region of the phase diagram, for \(S\) sufficiently large. We also prove the existence of a gap for excitations.

Bounds on the norm of Wigner-type random matrices

Random Matrices: Theory and Applications, 2019 · with L. Erdős · arXiv

We consider a Wigner-type ensemble, i.e. large hermitian \(N\times\) random matrices \(H=H^*\) with centered independent entries and with a general matrix of variances \(S_{xy}=\mathbb E|H_{xy}|^2\). The norm of \(H\) is asymptotically given by the maximum of the support of the self-consistent density of states. We establish a bound on this maximum in terms of norms of powers of \(S\) that substantially improves the earlier bound \(2\|S\|_\infty^{1/2}\) given in [arXiv:1506.05098]. The key element of the proof is an effective Markov chain approximation for the contributions of the weighted Dyck paths appearing in the iterative solution of the corresponding Dyson equation.