Research map / Large language models / Language, safety and interpretability

Fine-grained List-wise Alignment for Generative Medication Recommendation

NeurIPS 2025 · Large language models › Language, safety and interpretability · approach: Data and evaluation methodology · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes FLAME, an LLM-based drug-by-drug list-wise medication recommender trained with step-wise GRPO using potential-based reward shaping plus fused structured/collaborative embeddings.

Problem

Medication recommendation with point-wise prediction ignores synergistic drug effects and adverse drug-drug interactions (DDIs), and standard GRPO assigns outcome-level rewards uniformly across tokens.

Approach

Formulates drug list generation as a sequential add/remove decision process and applies step-wise GRPO with a potential function (Jaccard correctness, DDI safety, refusal rate) to give per-drug credit, combined with a two-stage classifier plus list-level editor and fusion of collaborative embeddings into the LLM input space.

Key results

• On MIMIC-III, FLAME reaches Jaccard 0.4836 and F1 0.6408, above LAMO (0.4701 / 0.6294) and RAREMed (0.4174 / 0.5776).
• Ablation: removing the list-wise policy drops Jaccard from 0.4836 to 0.4785, and step-wise GRPO gives a larger list-vs-classifier advantage (0.0051) than standard GRPO (0.0028).
• Safety is controllable via the DDI penalty alpha, with Jaccard-DDI trade-off curves reported for FLAME versus baselines that support safety control.

Measured effects reported in the paper

Surprising findings

• Removing the list-wise refinement policy yields only a small numerical drop (Jaccard 0.4836 to 0.4785), which the authors attribute to it being a lightweight refinement module.

Limitations

• High computational cost and deployment overhead of LLM-based inference for clinical settings.
• Unstructured note preprocessing relies on GPT-4o API calls, limiting scalability and consistency.
• Evaluated only on static EHR datasets, not dynamic clinical workflows.

Experimental setup

Datasets: MIMIC-III MIMIC-IV eICU TWOSIDES

Benchmarks: MIMIC-III MIMIC-IV eICU

Metrics: Jaccard F1 DDI rate #Med

Baselines: LEAP GAMENet SafeDrug COGNet MICRON MoleRec NLA-MMR RAREMed LAMO

Builds on: GRPO DAPO PPO DPO potential-based reward shaping

Closest papers by meaning

• Generating Novel Leads for Drug Discovery Using LLMs with Logical Feedback (AAAI 2024 · similarity 0.83)
• SAFER: A Calibrated Risk-Aware Multimodal Recommendation Model for Dynamic Treatment Regimes (ICML 2025 · similarity 0.83)
• Data-Driven Discovery of Dynamical Systems in Pharmacology using Large Language Models (NeurIPS 2024 · similarity 0.83)
• BounDr.E: Predicting Drug-likeness via Biomedical Knowledge Alignment and EM-like One-Class Boundary Optimization (ICML 2025 · similarity 0.83)
• Entropy-Reinforced Planning with Large Language Models for Drug Discovery (ICML 2024 · similarity 0.82)
• Hierarchical Divide-and-Conquer for Fine-Grained Alignment in LLM-Based Medical Evaluation (AAAI 2025 · similarity 0.82)