Enoch Hyunwook Kang

Enoch Hyunwook Kang, Hema Yoganarasimhan

Large Language Models (LLMs) have recently enabled self-improving AI, i.e., AI that iteratively generates, evaluates, and refines its own outcomes. Recent studies have shown that self-improving AI focusing on prompt optimization can outperform state-of-the-art reinforcement-learning fine-tuned LLMs. Here, their `performance' is typically measured by query efficiency - the number of LLM-generated solution samples required to meet a certain performance threshold. However, in many societal applications, the primary limitation is not generating new solutions but evaluating them. For instance, evaluating an ad's effectiveness requires significant human feedback, which is far more costly and time-consuming than generating a candidate ad. To optimize for the evaluation efficiency objective, a natural approach is to extend Bayesian Optimization (BO), a framework proven optimal for evaluation efficiency, to the language domain. However, the difficulty of directly estimating suitable acquisition functions in LLMs' minds makes this extension challenging. This paper overcomes this challenge by proving that the combination of the simple and widely used Best-of-N selection strategy and simple textual gradients (i.e., textual edits from a critic model) statistically emulates the behavior of the gradients on the canonical UCB acquisition function, which induces optimal exploration in terms of evaluation efficiency. Based on this result, we propose TextGrad-Best-of-N Bayesian Optimization (T-BoN BO), a simple and eval-efficient language-space Bayesian optimization framework for AI self-improvement. We also empirically validate T-BoN BO by applying it to automated ad alignment tasks for persona distribution, demonstrating its superior performance compared to popular state-of-the-art baselines.

Enoch Hyunwook Kang

Iterative alignment methods based on purely greedy updates are remarkably effective in practice, yet existing theoretical guarantees of \(O(\log T)\) KL-regularized regret can seem pessimistic relative to their empirical performance. In this paper, we argue that this mismatch arises from the regret criterion itself: KL-regularized regret conflates the statistical cost of learning with the exploratory randomization induced by the softened training policy. To separate these effects, we study the traditional temperature-zero regret criterion, which evaluates only the top-ranked response at inference time. Under this decision-centric notion of performance, we prove that standard greedy online alignment methods, including online RLHF and online DPO, achieve constant \((O(1))\) cumulative regret. By isolating the cost of identifying the best response from the stochasticity induced by regularization, our results provide a sharper theoretical explanation for the practical superb efficiency of greedy alignment.

Enoch Hyunwook Kang

Personalized alignment aims to adapt large language models to heterogeneous user preferences, yet the precise theoretical conditions for its statistical efficiency have not been formally established. This paper characterizes the conditions under which personalized alignment achieves O(1) online regret and log(1/epsilon) offline sample complexity. We show that these optimal rates depend on a specific user-diversity condition: the population of user-specific heads must span the latent reward directions that can alter the optimal response. We prove that this condition is both necessary and sufficient. When it holds, simple greedy algorithms achieve benchmark efficiency; when it fails, every learner in a natural admissible class incurs at least logarithmic regret. Our results identify user diversity as the fundamental driver of personalized identifiability.

Enoch Hyunwook Kang

Field experiments (A/B tests) are often the most credible benchmark for methods (algorithms) in societal systems, but their cost and latency bottleneck rapid methodological progress. LLM-based persona simulation offers a cheap synthetic alternative, yet it is unclear whether replacing humans with personas preserves the benchmark interface that adaptive methods optimize against. We prove an if-and-only-if characterization: when (i) methods observe only the aggregate outcome (aggregate-only observation) and (ii) evaluation depends only on the submitted artifact and not on the method's identity or provenance (method-blind evaluation), swapping humans for personas is just panel change from the method's point of view, indistinguishable from changing the evaluation population (e.g., New York to Jakarta). Furthermore, we move from validity to usefulness: we define an information-theoretic discriminability of the induced aggregate channel and show that making persona benchmarking as decision-relevant as a field experiment is fundamentally a sample-size question, yielding explicit bounds on the number of independent persona evaluations required to reliably distinguish meaningfully different methods at a chosen resolution.