💬 NLP

TheraAgent: Self-Improving Therapeutic Agent for Precise and Comprehensive Treatment Planning

TheraAgent is a self-improving agentic framework that enhances treatment planning by replacing one-shot LLM generation with an iterative generate-judge-refine pipeline, utilizing a specialized TheraJudge module to achieve state-of-the-art accuracy, completeness, and safety that surpasses human physicians in expert evaluations.

Original authors: Junkai Li, Yunghwei Lai, Tianyi Zhu, Zheng Long Lee, Weizhi Ma, Yang Liu

Published 2026-05-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Junkai Li, Yunghwei Lai, Tianyi Zhu, Zheng Long Lee, Weizhi Ma, Yang Liu

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: From "First Draft" to "Masterpiece"

Imagine you are asking a very smart, well-read student to write a complex medical treatment plan for a sick patient.

The Old Way (General LLMs):
The student reads the patient's file and immediately writes down the first idea that pops into their head. They hand it over and say, "Here is the plan!"

The Problem: Because they didn't double-check their work, the plan might have the wrong dosage (like giving a child an adult's medicine), miss a crucial step, or suggest a treatment that could accidentally hurt the patient. It's like a "one-shot" guess.

The New Way (TheraAgent):
This paper introduces a new system called TheraAgent. Instead of just writing one draft, TheraAgent acts like a perfectionist editor who never stops until the work is perfect. It follows a cycle: Write → Critique → Fix → Repeat.

How TheraAgent Works: The Three-Step Dance

The paper describes TheraAgent as a team of three specialists working together in a loop:

1. The Planner (The Architect)

This is the part that actually writes the treatment plan.

What it does: It looks at the patient's symptoms and history and drafts a plan.
The Twist: It doesn't just guess. It looks at a "memory bank" of its own past attempts and the feedback it received to see what it got wrong last time.

2. The TheraJudge (The Strict Inspector)

This is the paper's secret weapon. Think of this as a tough, expert building inspector who checks the Architect's blueprints.

What it does: It doesn't just say "Good job" or "Bad job." It breaks the plan down into specific categories:
- Accuracy: Is the medicine correct?
- Completeness: Did we forget anything?
- Safety: Could this hurt the patient?
- Targeting: Is this specific to this patient's unique situation?
The Tools: To be fair, the Judge uses two tools:
- The Library (RAG): It pulls up real medical guidelines and textbooks to fact-check the plan.
- The Examples (Few-Shot): It looks at past examples of "perfect" plans to know exactly what a high score looks like.

3. The Memorizer (The Notebook)

This is the team's shared notebook.

What it does: Every time the Architect writes a plan and the Judge grades it, the result goes into the notebook.
The Magic: Before the Architect writes the next version, it reads the notebook. It sees, "Oh, last time I gave too little medicine, and the Judge gave me a low score for Safety. I need to fix that this time."

The Loop: The Architect writes a plan → The Judge critiques it → The Notebook saves the lesson → The Architect writes a better plan based on that lesson. They keep doing this until the plan is perfect or they hit a time limit.

Why This Matters: The Results

The researchers tested this system against two groups:

Other AI Models: Including very smart ones like GPT-4 and specialized medical AIs.
Real Human Doctors: They asked licensed physicians to compare the AI's plans against their own.

The Findings:

Beating the AI: TheraAgent scored higher than any other AI model on a major medical benchmark called "HealthBench." It was particularly good at being complete (not missing steps) and precise (getting the details right).
Beating the Humans: In a blind test where doctors didn't know which plan was written by AI, TheraAgent won 86% of the time.
- Why did the humans lose? The paper notes that human doctors often write very short, condensed notes because they are busy. They might skip explaining why they chose a drug. TheraAgent, however, writes out the full, explicit reasoning and checks every safety box, making the plan look more thorough and safer.

A Real-World Example from the Paper

The paper gives a specific case of a 70-year-old man with a complex lung condition (CPFE).

The "One-Shot" AI: Suggested a low dose of a medicine (N-acetylcysteine) that was too weak to work, and suggested using morphine too early, which could be dangerous.
TheraAgent: Caught these errors. It increased the medicine dose to the correct level, added strict rules about when to use morphine, and insisted on specific tests (like a CT scan) before starting certain treatments. It turned a risky, rough draft into a safe, comprehensive guide.

The Trade-Off

The paper is honest about one downside: Speed and Cost.
Because TheraAgent has to write, critique, and rewrite the plan multiple times, it takes longer and costs more computer power than a simple "one-shot" AI.

The Analogy: It's like the difference between a fast-food burger (quick, cheap, maybe a bit sloppy) and a chef who tastes, adjusts, and refines a dish three times before serving it (slower, more expensive, but much better quality).
The Conclusion: For life-or-death medical decisions, the paper argues that taking that extra time to be safe and precise is worth the cost.

Summary

TheraAgent changes medical AI from a fast guesser into a careful editor. By mimicking how human experts review and refine their own work, it produces treatment plans that are safer, more accurate, and more complete than current AI models—and even better than the average human doctor's written plan in these tests.

Technical Summary: TheraAgent

Problem Statement

Formulating a treatment plan is a complex reasoning and refinement task that requires translating diagnostic conclusions into concrete therapeutic actions. The paper argues that existing Large Language Models (LLMs), whether general-purpose or medically fine-tuned, predominantly rely on one-shot generation without explicit verification. This approach often results in treatment plans that are rough, incomplete, and potentially unsafe, characterized by imprecise dosing, omitted key components, and a lack of harm control. Furthermore, current evaluation methods for treatment planning often rely on oversimplified metrics (e.g., BLEU) or vanilla LLM scores that fail to capture the multi-dimensional clinical quality required for real-world decision-making.

Methodology: TheraAgent Framework

To address these limitations, the authors propose TheraAgent, an agentic framework that reframes treatment planning as an iterative generate–reflect–refine process, mimicking the revision patterns observed in expert clinical practice. The framework consists of three interacting components:

Planner: The generative module responsible for producing candidate treatment plans ( $T_k$ ) conditioned on the patient case ( $P$ ) and historical feedback stored in the Memorizer. Unlike one-shot models, the Planner explicitly conditions on previously generated schedules and their evaluation rationales to identify and correct deficiencies (e.g., incomplete coverage or safety risks) in subsequent iterations.
TheraJudge: A treatment-specific evaluation module integrated into the inference loop. It assesses candidate plans against clinical criteria across multiple dimensions: Accuracy, Targeting, Completeness, and Harm Control. TheraJudge employs three mechanisms to ensure reliability:
- Knowledge-grounded evaluation via RAG: Retrieves authoritative clinical guidelines to reduce hallucinations and ensure adherence to safety standards.
- Stabilization via few-shot exemplars: Uses expert-annotated cases to calibrate scoring and reduce variance.
- Multi-dimensional scoring: Provides granular feedback rather than a single holistic score, offering explicit signals for refinement.
Memorizer: A structured repository that accumulates historical treatment plans, reasoning traces, and evaluation scores. It enables score-aware retrieval, allowing the Planner to access high-quality historical examples (Top-N) to guide iterative optimization.

The framework outputs the final plan ( $T^*$ ) by selecting the schedule with the highest score from the final $L$ iterations, incorporating an early-stopping mechanism to balance quality and computational cost.

Key Contributions

Iterative Reasoning Framework: The proposal of TheraAgent, which replaces one-shot generation with a self-improving pipeline, yielding more precise, comprehensive, and safer treatment plans.
TheraJudge Evaluation Module: The introduction of a clinically aligned internal critic that assesses plans along dimensions such as Accuracy and Completeness. The authors demonstrate its potential as a standalone evaluator, showing high agreement with external benchmarks.
Empirical Validation: Extensive experiments demonstrating state-of-the-art performance on the HealthBench treatment planning task and superior performance in blinded expert evaluations against human physicians.

Experimental Results

The authors evaluated TheraAgent on the HealthBench dataset (filtered for treatment-related cases across endocrinology, gastroenterology, neurology, and respiratory departments) and a Real-World Case Dataset of physician-authored cases.

Automatic Evaluation (HealthBench): TheraAgent achieved a state-of-the-art overall score of 48.94, surpassing the second-best model by 4.66 points. It led in Accuracy and Completeness, exceeding the runner-up by 2.91 and 4.43 points respectively. It also showed significant improvements in "Hedging" (6.78 points) and "Context Seeking."
Expert Evaluation (Real-World Cases): In blinded comparisons involving licensed medical experts, TheraAgent achieved an 86% win rate against human physicians. It demonstrated superior performance in Targeting (69% win rate), Completeness (71%), and Harm Control (51%).
Evaluator Reliability: TheraJudge showed a high Pearson correlation (0.71) with the official HealthBench evaluation, significantly outperforming traditional lexical metrics (BLEU, ROUGE) and vanilla LLM scores, validating its reliability as an automated feedback signal.
Generalization: The framework demonstrated consistent performance improvements across different backbone models (from GPT-4o to DeepSeek-R1) and on an additional benchmark (MTMedDialog), indicating the approach is not limited to a specific dataset or model architecture.

Significance and Claims

The paper claims that TheraAgent redefines treatment planning not as a simple generation task, but as a distinct reasoning and refinement task. By integrating domain-specific evaluation (TheraJudge) directly into the inference loop, the framework successfully transforms LLMs into safer, more precise, and practical clinical planners.

The authors emphasize that the high consistency between TheraJudge and expert evaluations validates the framework's ability to provide valid feedback signals regarding safety and completeness, rather than arbitrary noise. They position TheraAgent as a tool to assist treatment planning by converting implicit clinical logic into explicit requirements and acting as a safety reminder, rather than replacing professional medical judgment. The work suggests that iterative refinement mechanisms are essential for overcoming the limitations of one-shot generation in high-stakes medical domains.