OpenNovelty.org

Knowledgeable Language Models as Black-Box Optimizers for Personalized Medicine

ICLR 2026 Conference SubmissionAnonymous Authors
OpenReview Score: 6.0Download Report PDF
Large language modelsPersonalized medicineBlack-box optimizationDistribution shift
Abstract:

The goal of personalized medicine is to discover a treatment regimen that optimizes a patient's clinical outcome based on their personal genetic and environmental factors. However, candidate treatments cannot be arbitrarily administered to the patient to assess their efficacy; we often instead have access to an in silico surrogate model that approximates the true fitness of a proposed treatment. Unfortunately, such surrogate models have been shown to fail to generalize to previously unseen patient-treatment combinations. We hypothesize that domain-specific prior knowledge—such as medical textbooks and biomedical knowledge graphs—can provide a meaningful alternative signal of the fitness of proposed treatments. To this end, we introduce LLM-based Entropy-guided Optimization with kNowledgeable priors (LEON), a mathematically principled approach to leverage large language models (LLMs) as black-box optimizers without any task-specific fine-tuning, taking advantage of their ability to contextualize unstructured domain knowledge to propose personalized treatment plans in natural language. In practice, we implement LEON via 'optimization by prompting,' which uses LLMs as stochastic engines for proposing treatment designs. Experiments on real-world optimization tasks show LEON outperforms both traditional and LLM-based methods in proposing individualized treatments for patients.

Disclaimer
This report is AI-GENERATED using Large Language Models and WisPaper (A scholar search engine). It analyzes academic papers' tasks and contributions against retrieved prior work. While this system identifies POTENTIAL overlaps and novel directions, ITS COVERAGE IS NOT EXHAUSTIVE AND JUDGMENTS ARE APPROXIMATE. These results are intended to assist human reviewers and SHOULD NOT be relied upon as a definitive verdict on novelty.
NOTE that some papers exist in multiple, slightly different versions (e.g., with different titles or URLs). The system may retrieve several versions of the same underlying work. The current automated pipeline does not reliably align or distinguish these cases, so human reviewers will need to disambiguate them manually.
If you have any questions, please contact: mingzhang23@m.fudan.edu.cn

Overview

Taxonomy

Core-task Taxonomy Papers
36
3
Claimed Contributions
30
Contribution Candidate Papers Compared
0
Refutable Paper

Research Landscape Overview

Core task: Personalized treatment optimization under distribution shift. The field addresses how to tailor interventions when patient populations or clinical environments evolve over time, creating mismatches between training and deployment conditions. The taxonomy reveals several complementary research directions: Causal Inference and Treatment Effect Estimation Under Covariate Shift focuses on identifying treatment effects when covariate distributions change, often leveraging propensity weighting and doubly robust methods. Reinforcement Learning for Sequential Treatment Optimization tackles dynamic decision-making over multiple time steps, balancing exploration and exploitation in non-stationary environments. Domain Adaptation and Representation Learning Under Distribution Shift seeks invariant features that generalize across hospitals or patient subgroups, as seen in works like Domain-invariant Clinical Representation[10] and Knowledge-Guided Domain Adaptation[7]. Adaptive Clinical Decision Support and Real-Time Monitoring emphasizes responsive systems that adjust to individual patient trajectories, while Optimization and Meta-Learning for Personalized Treatment explores how to efficiently learn treatment policies that transfer across contexts. Specialized Applications demonstrate these principles in concrete settings ranging from anticoagulation dosing to mood disorder monitoring. Recent work highlights tensions between model flexibility and robustness guarantees. Many studies pursue uncertainty quantification through conformal methods, such as Conformal Deep Q-Learning[6] and Conformal Dose-Response[17], to provide reliable prediction intervals under shift. Others emphasize transfer learning strategies, exemplified by Transfer Learning Treatment Rules[3], which adapt policies learned in source populations to new target settings. Within the Optimization and Meta-Learning branch, Knowledgeable Black-Box Optimizers[0] sits alongside Contextually Constrained Optimization[12], both addressing how to incorporate domain knowledge and context-specific constraints when optimizing treatment parameters. While Contextually Constrained Optimization[12] focuses on learning feasible regions from contextual features, Knowledgeable Black-Box Optimizers[0] emphasizes integrating expert priors into black-box search procedures. This cluster reflects a broader trend toward hybrid approaches that blend data-driven optimization with structured domain expertise, aiming to improve sample efficiency and safety when distribution shifts challenge purely empirical methods.

Claimed Contributions

Formulating personalized medicine as a black-box optimization problem

The authors formulate personalized medicine as a conditional black-box optimization problem where the objective is to discover optimal treatment regimens conditioned on patient-specific genetic and environmental features. This formulation provides a mathematical foundation for applying optimization methods to individualized treatment design.

10 retrieved papers
Constrained optimization problem with certainty-based constraints

The authors introduce two constraints to the optimization problem: one that bounds the Wasserstein distance between proposed and historical designs to ensure reliable surrogate predictions, and another that bounds the entropy of proposed designs to encourage consistency based on domain knowledge. These constraints address the challenge of imperfect surrogate models in out-of-distribution settings.

10 retrieved papers
LEON: LLM-based Entropy-guided Optimization with kNowledgeable priors

The authors derive a tractable solution to the constrained optimization problem that leverages large language models as zero-shot optimizers without task-specific fine-tuning. LEON uses statistical analysis of design distributions and an adversarial source critic model, implemented via optimization-by-prompting to propose personalized treatment plans.

10 retrieved papers

Core Task Comparisons

Comparisons with papers in the same taxonomy category

Contribution Analysis

Detailed comparisons for each claimed contribution

Contribution

Formulating personalized medicine as a black-box optimization problem

The authors formulate personalized medicine as a conditional black-box optimization problem where the objective is to discover optimal treatment regimens conditioned on patient-specific genetic and environmental features. This formulation provides a mathematical foundation for applying optimization methods to individualized treatment design.

Contribution

Constrained optimization problem with certainty-based constraints

The authors introduce two constraints to the optimization problem: one that bounds the Wasserstein distance between proposed and historical designs to ensure reliable surrogate predictions, and another that bounds the entropy of proposed designs to encourage consistency based on domain knowledge. These constraints address the challenge of imperfect surrogate models in out-of-distribution settings.

Contribution

LEON: LLM-based Entropy-guided Optimization with kNowledgeable priors

The authors derive a tractable solution to the constrained optimization problem that leverages large language models as zero-shot optimizers without task-specific fine-tuning. LEON uses statistical analysis of design distributions and an adversarial source critic model, implemented via optimization-by-prompting to propose personalized treatment plans.

Table of Contents