Selected work

Representative Papers

The Directions of Technical Change

Feb 13, 2026

Generative AI is a directional technology: it excels at some task combinations and performs poorly at others. Knowledge work is also directional and endogenous: workers can satisfy their job requirements with different combinations of tasks. Studying AI adoption by knowledge workers hence requires comparing two vectors.We develop a high-dimensional model of task choice and technology adoption, with otherwise standard neoclassical assumptions. AI is adopted when its direction is aligned with what the worker values at the margin -- the worker's shadow prices, rather than with what the worker actually does -- their activity vector. This yields a cone of adoption that widens as AI capability grows; near the entry threshold, small improvements in capability translate into large expansions in the set of adopted directions. Adoption also has a structured intensive margin: a tool can be worth using but not worth using all the time, generating a region of stable hybrid production between an entry threshold and an all-in threshold. We also show how to derive shadow prices as explicit functions of observable skill and requirement vectors. The framework explains rapid adoption in aligned occupations, heterogeneous adoption elsewhere, and weak correlation with one-dimensional skill measures: the key heterogeneity is directional alignment, not skill level.

0 citationsRead paper

An experimental approach: The graph of graphs

Aug 24, 2025

This study addresses the optimal design of pairwise comparison (PC) structures in decision-making and preference modeling: specifically, how to select the most informative comparison pairs and their presentation sequence to minimize cognitive load on decision-makers while maximizing the accuracy of weight estimation. We propose a “graph-of-graphs” framework to systematically construct and analyze the taxonomy of PC graphs. Empirical validation involves 301 participants in color calibration and sensory testing tasks; results confirm strong alignment between simulated optimal designs—particularly balanced incomplete block designs—and observed preference responses, with empirically derived near-optimal sequences matching simulation predictions exactly. Weight estimation via logarithmic least squares reveals high similarity between the empirical distribution of PC matrix entries and theoretical distributions from simulation. Contributions include a practical recommendation table of optimal PC patterns and an open-source Java toolkit, substantially enhancing both the efficiency and deployability of preference modeling.

0 citationsRead paper

Anchor Sponsor Firms in Open Source Software Ecosystems

Feb 13, 2025

This study examines the impact of anchor sponsor withdrawal—exemplified by Mozilla’s exit—from open-source software (OSS) ecosystems, using Rust as a case. Employing event study analysis, panel regression, and agent-based modeling on GitHub/GitLab developer activity data (2015–2022), we provide the first empirical identification of anchor sponsors as critical quality signals. Results show that Mozilla’s exit triggered an immediate 9% decline in weekly Rust code commits and reduced long-term growth by 0.6 percentage points. Peripheral developer activity failed to recover within six months, while inflows of new projects and first-time contributors contracted significantly. We uncover structural asymmetry in contributor responses—core maintainers exhibited resilience, whereas peripheral contributors disengaged disproportionately. Furthermore, we develop and release the first reproducible, open-source agent-based model of OSS ecosystem dynamics. This work advances theoretical understanding of sponsorship dependence and offers actionable insights for enhancing OSS ecosystem resilience and sustainability.

0 citationsRead paper

Stable Hypergraph Matching in Unimodular Hypergraphs

Feb 12, 2025

This paper investigates Stable Hypergraph Matching (SHM) and its capacitated generalization, Stable Hypergraph $b$-Matching (SH$b$M)—both NP-hard problems. Addressing the absence of polynomial-time algorithms for unimodal hypergraphs, we systematically characterize their tractability boundaries: (i) maximum-weight stable $b$-matching is polynomial-time solvable on subpath hypergraphs; (ii) optimizing stable matchings remains NP-hard on laminar hypergraphs; (iii) SHM is tractable on subtree hypergraphs, whereas SH$b$M is NP-hard there. Methodologically, we integrate combinatorial optimization, total unimodularity theory, and constructive applications of Scarf’s Lemma and the Lovász Local Lemma. We further introduce a practical tripartite stable matching model—dual admission in higher education—and provide an efficient solution framework. Our results deliver a foundational solvability classification for hypergraph stability and supply novel algorithmic tools for constrained matching in structured hypergraphs.

0 citationsRead paper

Candidate nomination for Condorcet-consistent voting rules

Feb 05, 2025

This paper studies the *Possible President* problem: given a multi-party election under party-nomination rules, can a candidate from a specific party become the winner of a Condorcet-consistent voting rule (Copeland$^α$ or Maximin) via strategic nominations by other parties? We conduct the first systematic parameterized complexity analysis with respect to three key parameters—number of voters, number of parties, and maximum party size—and fully characterize its computational complexity spectrum. Specifically, we establish a clean P/NP dichotomy parameterized by the number of voters; for all NP-hard cases, we provide a refined trichotomy into FPT, W[1]-hard, and paraNP-hard classes, precisely delineating the boundary between tractability and intractability. Our results integrate techniques from computational social choice, parameterized algorithms, and intricate reduction constructions, yielding the first complete complexity landscape for strategic party nomination in multi-party elections.

0 citationsRead paper

Recent publications

Latest Papers

The Directions of Technical Change

Feb 13, 2026

Generative AI is a directional technology: it excels at some task combinations and performs poorly at others. Knowledge work is also directional and endogenous: workers can satisfy their job requirements with different combinations of tasks. Studying AI adoption by knowledge workers hence requires comparing two vectors.We develop a high-dimensional model of task choice and technology adoption, with otherwise standard neoclassical assumptions. AI is adopted when its direction is aligned with what the worker values at the margin -- the worker's shadow prices, rather than with what the worker actually does -- their activity vector. This yields a cone of adoption that widens as AI capability grows; near the entry threshold, small improvements in capability translate into large expansions in the set of adopted directions. Adoption also has a structured intensive margin: a tool can be worth using but not worth using all the time, generating a region of stable hybrid production between an entry threshold and an all-in threshold. We also show how to derive shadow prices as explicit functions of observable skill and requirement vectors. The framework explains rapid adoption in aligned occupations, heterogeneous adoption elsewhere, and weak correlation with one-dimensional skill measures: the key heterogeneity is directional alignment, not skill level.

0 citationsRead paper

An experimental approach: The graph of graphs

Aug 24, 2025

This study addresses the optimal design of pairwise comparison (PC) structures in decision-making and preference modeling: specifically, how to select the most informative comparison pairs and their presentation sequence to minimize cognitive load on decision-makers while maximizing the accuracy of weight estimation. We propose a “graph-of-graphs” framework to systematically construct and analyze the taxonomy of PC graphs. Empirical validation involves 301 participants in color calibration and sensory testing tasks; results confirm strong alignment between simulated optimal designs—particularly balanced incomplete block designs—and observed preference responses, with empirically derived near-optimal sequences matching simulation predictions exactly. Weight estimation via logarithmic least squares reveals high similarity between the empirical distribution of PC matrix entries and theoretical distributions from simulation. Contributions include a practical recommendation table of optimal PC patterns and an open-source Java toolkit, substantially enhancing both the efficiency and deployability of preference modeling.

0 citationsRead paper

Anchor Sponsor Firms in Open Source Software Ecosystems

Feb 13, 2025

This study examines the impact of anchor sponsor withdrawal—exemplified by Mozilla’s exit—from open-source software (OSS) ecosystems, using Rust as a case. Employing event study analysis, panel regression, and agent-based modeling on GitHub/GitLab developer activity data (2015–2022), we provide the first empirical identification of anchor sponsors as critical quality signals. Results show that Mozilla’s exit triggered an immediate 9% decline in weekly Rust code commits and reduced long-term growth by 0.6 percentage points. Peripheral developer activity failed to recover within six months, while inflows of new projects and first-time contributors contracted significantly. We uncover structural asymmetry in contributor responses—core maintainers exhibited resilience, whereas peripheral contributors disengaged disproportionately. Furthermore, we develop and release the first reproducible, open-source agent-based model of OSS ecosystem dynamics. This work advances theoretical understanding of sponsorship dependence and offers actionable insights for enhancing OSS ecosystem resilience and sustainability.

0 citationsRead paper

Stable Hypergraph Matching in Unimodular Hypergraphs

Feb 12, 2025

This paper investigates Stable Hypergraph Matching (SHM) and its capacitated generalization, Stable Hypergraph $b$-Matching (SH$b$M)—both NP-hard problems. Addressing the absence of polynomial-time algorithms for unimodal hypergraphs, we systematically characterize their tractability boundaries: (i) maximum-weight stable $b$-matching is polynomial-time solvable on subpath hypergraphs; (ii) optimizing stable matchings remains NP-hard on laminar hypergraphs; (iii) SHM is tractable on subtree hypergraphs, whereas SH$b$M is NP-hard there. Methodologically, we integrate combinatorial optimization, total unimodularity theory, and constructive applications of Scarf’s Lemma and the Lovász Local Lemma. We further introduce a practical tripartite stable matching model—dual admission in higher education—and provide an efficient solution framework. Our results deliver a foundational solvability classification for hypergraph stability and supply novel algorithmic tools for constrained matching in structured hypergraphs.

0 citationsRead paper

Candidate nomination for Condorcet-consistent voting rules

Feb 05, 2025

This paper studies the *Possible President* problem: given a multi-party election under party-nomination rules, can a candidate from a specific party become the winner of a Condorcet-consistent voting rule (Copeland$^α$ or Maximin) via strategic nominations by other parties? We conduct the first systematic parameterized complexity analysis with respect to three key parameters—number of voters, number of parties, and maximum party size—and fully characterize its computational complexity spectrum. Specifically, we establish a clean P/NP dichotomy parameterized by the number of voters; for all NP-hard cases, we provide a refined trichotomy into FPT, W[1]-hard, and paraNP-hard classes, precisely delineating the boundary between tractability and intractability. Our results integrate techniques from computational social choice, parameterized algorithms, and intricate reduction constructions, yielding the first complete complexity landscape for strategic party nomination in multi-party elections.

0 citationsRead paper