research

2026

1. 

   OPERA-S: Deterministic Tail Safety in Ensemble Off-Policy Evaluation

   Hyunwoo Kim, Gyeongchan Han, Bogeun Kim, Serjin Kim, Sangyeon Cho, Junha Ham, Jaehyeok Shin, and Sanghack Lee^†

   2026

   Under Review, NeurIPS 2026

   Abs Bib Code

   Off-policy evaluation (OPE) estimates a candidate policy’s value from data collected under a different policy. Ensemble methods like OPERA re-weight a pool of base estimators via affine-hull weights minimizing a bootstrap MSE estimate. However, affine-hull weights can produce ensembles strictly worse than every base estimator — a failure mode we call the catastrophic regime. We propose OPERA-S, which constrains weights to the probability simplex and pairs the bootstrap objective with a soft adversary on the unknown policy value. The combined minimax problem reduces to a convex quadratic program with a single regularizer, providing a drop-in replacement for OPERA. We prove that the simplex constraint deterministically eliminates the catastrophic regime, enabling pool-curation-free deployment with heterogeneous estimator pools. Furthermore, it ensures that every monotone tail functional of OPERA-S’s MSE distribution is no larger than that of the per-condition worst-base envelope. Across three benchmarks spanning bandits and offline reinforcement learning, OPERA-S delivers the strongest gains on tail metrics, with competitive mean-MSE performance.

   @misc{operas2026,
     title = {OPERA-S: Deterministic Tail Safety in Ensemble Off-Policy Evaluation},
     author = {Kim, Hyunwoo and Han, Gyeongchan and Kim, Bogeun and Kim, Serjin and Cho, Sangyeon and Ham, Junha and Shin, Jaehyeok and Lee, Sanghack},
     year = {2026},
     note = {Under Review, NeurIPS 2026},
   }

2. 

   A Structural View of Query Misspecification in Causal Foundation Models

   Junha Ham^*, Deokgyu Kim^*, Doeun Kim^*, Serjin Kim^*, and Sanghack Lee^†

   2026

   ICML 2026 Workshop on Structured Probabilistic Inference and Generative Modeling (SPIGM)

   Abs Bib Code

   Causal Foundation Models (CFMs) pretrain amortized causal estimators on large collections of synthetic datasets sampled from structural causal model (SCM) priors. In optimal-capacity, they recover the corresponding interventional posterior predictive target for queries on the training query surface. We study the failure mode induced when an inference-time query includes a post-treatment covariate. Structurally, we decompose the resulting CATE bias into three components: loss of the natural indirect effect, an interaction penalty, and treatment-differenced selection bias. Distributionally, we prove that conditioning on post-treatment values yields strictly positive KL divergence from the marginal interventional law on a positive-measure subset, and we provide closed-form KL decompositions under linear-Gaussian SCMs. Empirically, removing post-treatment covariates from the query yields substantial reductions in PEHE across graph topologies and CFM models without retraining. We further introduce Treatment-Centric Local Discovery (TC-LD), a lightweight pre-inference filter that flags likely post-treatment variables and recovers most of this improvement on our synthetic benchmark.

   @misc{cfmbias_spigm,
     title = {A Structural View of Query Misspecification in Causal Foundation Models},
     author = {Ham, Junha and Kim, Deokgyu and Kim, Doeun and Kim, Serjin and Lee, Sanghack},
     year = {2026},
     note = {ICML 2026 Workshop on Structured Probabilistic Inference and Generative Modeling (SPIGM)},
   }

3. 

   Causal Foundation Models Perform Better without Post-treatment Variables

   Junha Ham^*, Deokgyu Kim^*, Doeun Kim^*, Serjin Kim^*, and Sanghack Lee^†

   2026

   ICML 2026 Workshop on Foundation Models for Structured Data (FMSD)

   Abs Bib Code

   Causal Foundation Models (CFMs) amortize Bayesian causal inference by pretraining on synthetic datasets, enabling zero-shot conditional average treatment effect (CATE) estimation. This paper studies the structural bias induced when post-treatment covariates are included in the inference-time query set of CFMs. The bias decomposes into three terms, and the CATE estimation error of two representative CFMs, Do-PFN and CausalPFN, is consistent with the corresponding theoretical bounds under mediator and collider conditioning. In an Oracle-Exclude experiment, removing post-treatment covariates from the query set reduces estimation error by approximately 25% for Do-PFN and up to 72% for CausalPFN without retraining. As a practical alternative to oracle exclusion, Treatment-Centric Local Discovery (TC-LD) filters query covariates before inference. It recovers 85.5%–92.6% of the Oracle headroom on the synthetic benchmark and detects all synthetically injected mediators in semi-synthetic experiments on IHDP and ACIC 2016.

   @misc{cfmbias_fmsd,
     title = {Causal Foundation Models Perform Better without Post-treatment Variables},
     author = {Ham, Junha and Kim, Deokgyu and Kim, Doeun and Kim, Serjin and Lee, Sanghack},
     year = {2026},
     note = {ICML 2026 Workshop on Foundation Models for Structured Data (FMSD)},
   }

2025

1. 

   The Hidden Cost of At-the-Market Offerings

   Serjin Kim

   2025

   Working Paper

   Abs Bib

   I examine the hidden costs of at-the-market (ATM) equity offerings, which allow firms to issue shares flexibly over time at prevailing market prices. Despite their apparent advantages as low-cost, option-like financing tools, ATM programs are not universally adopted. Using a comprehensive dataset of ATM announcements, I document large and persistent negative stock price reactions at the time firms establish ATM programs, even when no equity is subsequently issued. This indicates that the valuation impact arises from the announcement of financing flexibility itself rather than from realized dilution. Importantly, the negative market reaction is significantly attenuated for firms facing imminent cash shortfalls, suggesting that investors distinguish between defensive and opportunistic motives for ATM adoption. I also develop a dynamic model that rationalizes these patterns and shows that financing flexibility carries implicit costs.

   @unpublished{hiddencost,
     title = {The Hidden Cost of At-the-Market Offerings},
     author = {Kim, Serjin},
     year = {2025},
     note = {Working Paper},
   }

1. 

   Stablecoin, Bank Intermediation, and Fiscal Spillovers

   Serjin Kim

   Working Paper

   Abs Bib

   This paper develops a two-period overlapping generations (OLG) model in which agents choose among cash, bank deposits, and stablecoins, and compares two monetary architectures: one where stablecoins are issued by commercial banks, and another where they are issued by narrow-bank fintechs. The model captures how stablecoin issuance affects portfolio choice, interest rates, and government debt. I embed an endogenous government bond yield determined by stablecoin demand and also examines the strategic interaction between fintech fee-setting. Welfare comparisons show under which conditions fintech issuance enhances monetary efficiency. The analysis builds on and extends the framework in Andolfatto (2021).

   @unpublished{stablecoin,
     title = {Stablecoin, Bank Intermediation, and Fiscal Spillovers},
     author = {Kim, Serjin},
     note = {Working Paper},
   }