MedReflect: Reflective Correction for Medical LLMs

• MedReflect is a framework that trains medical LLMs to self-improve through a four-stage reflective correction process.
• It uses a single-pass chain—initial hypothesis, self-questioning, self-answering, and revision—to eliminate reliance on external retrieval.
• Experiments with limited training data show significant accuracy gains across multiple medical benchmarks, emphasizing its efficiency.

MedReflect is a framework for training medical LLMs to self-improve through reflective correction, introduced in "MedReflect: Teaching Medical LLMs to Self-Improve via Reflective Correction" (Huang et al., 4 Oct 2025). It is designed to elicit a physician-like reflective thinking mode by generating a single-pass reflection chain that includes initial hypothesis generation, self-questioning, self-answering, and decision refinement. The framework is positioned against retrieval-augmented generation and heavily annotated reasoning supervision, and instead emphasizes self-verified, self-reflective generation without external retrieval or heavy annotation. In the reported experiments, merely 2,000 randomly sampled training examples and a light fine-tuning were sufficient to obtain notable absolute accuracy improvements across multiple medical benchmarks while cutting annotation requirements (Huang et al., 4 Oct 2025).

1. Conceptual orientation

Medical problem solving requires expert knowledge and complex multi-step reasoning. In the formulation used by MedReflect, prior approaches for enhancing medical LLMs fall into two broad classes: retrieval-augmented generation, which injects external authoritative knowledge but incurs retrieval and storage overhead, and reasoning-dataset supervision, which relies on curated reasoning corpora and heavy annotation (Huang et al., 4 Oct 2025). MedReflect addresses both constraints by attempting to unlock physician-like reflective reasoning within the model itself.

The core claim of the framework is not that the model should simply generate longer chains of thought, but that it should learn to generate, review, and refine its own answers. This is operationalized as a reflective trajectory in which the model first proposes an answer, then interrogates its own possible mistake, answers that reflective question, and revises the original trajectory accordingly (Huang et al., 4 Oct 2025). The framework therefore treats correction as an intrinsic generation pattern rather than as an external post-processing stage.

A central design choice is that the reflection process is single-pass. The model is not described as relying on iterative reinforcement learning or external assistants during inference. Instead, reflective correction is built directly into the supervised fine-tuning target, with the aim of teaching the model to internalize a compact self-correction routine (Huang et al., 4 Oct 2025). This suggests a distinct notion of self-improvement: improvement arises from restructuring the generated reasoning path itself, rather than from attaching a separate retriever, verifier, or tool layer.

2. Reflective chain and correction mechanism

MedReflect formalizes reflection as a sequence of four stages: initial hypothesis generation, self-questioning, self-answering, and decision refinement (Huang et al., 4 Oct 2025). In the first stage, the LLM produces a first-pass answer and a reasoning trajectory. In the second, once a simulated error or inconsistency is present, the model generates a reflective question targeted at that error. In the third, it answers its own question using internalized medical knowledge. In the fourth, it revises the original hypothesis and produces a corrected answer.

The reflection data construction process begins with a medical question $Q$ and a correct answer trajectory

$T = [S_1, S_2, \ldots, S_n, A].$

The LLM is prompted to regenerate a step $S_i$ or the whole answer, occasionally introducing an error, denoted $T_{er}$. The error $E_i$ in $T_{er}$ is then detected, and the LLM is prompted to produce a focused reflection question $R_{q_i}$ about $E_i$ and an answer $R_{a_i}$. The resulting reflective trajectory is

$$T_{\text{reflect}} = [S_1, S_2, \ldots, E_i, R_{q_i}, R_{a_i}, S_i, \ldots, A].$$

This trajectory is the object used to instill self-correction behavior during training (Huang et al., 4 Oct 2025).

The framework supports two granularities of reflection. For multiple-choice question answering, reflection is sentence-level and targets incorrect step sentences. For consultations, reflection is word-level and targets inaccurate medical entities such as diagnosis or treatment, which are masked and then regenerated (Huang et al., 4 Oct 2025). This bifurcation is important because it ties the correction mechanism to the structure of the downstream task rather than imposing a single reflection format across all medical settings.

MedReflect also introduces vocabulary markers for reflection. The special tokens "<Think>", "</Think>", "<Modified>", and "</Modified>" are added to the tokenizer and to the training data to demarcate reflective and correction regions explicitly (Huang et al., 4 Oct 2025). These markers serve as structural delimiters within the generated sequence and make the correction span legible to the model during supervised adaptation.

3. Data construction, supervision, and fine-tuning

A major emphasis of MedReflect is low-cost dataset construction. Reflection chain data are generated automatically by prompting LLMs with designed templates and instructions rather than by manual labeling (Huang et al., 4 Oct 2025). The framework therefore claims low annotation cost while preserving medically structured supervision.

Two data sources are used for reflection data construction. The consultation component is built from ChatDoctor, described as 100k real doctor-patient question-answer pairs from HealthCareMagic. The multiple-choice QA component is built from MedMCQA (Huang et al., 4 Oct 2025). From these sources, 36,413 consultation records and 21,107 MCQA questions are constructed in total, although the reported strong results are obtained using only 2,000 randomly sampled reflection chain examples, with 1,000 from each source (Huang et al., 4 Oct 2025).

The fine-tuning setup uses Qwen2.5-7B-Instruct as the base model, with Qwen2.5-32B also evaluated in larger experiments (Huang et al., 4 Oct 2025). Training is performed for 3 epochs with learning rate $T = [S_1, S_2, \ldots, S_n, A].$0, and parameter-efficient adaptation is implemented with LoRA using $T = [S_1, S_2, \ldots, S_n, A].$1 (Huang et al., 4 Oct 2025). The supervised objective is reported as

$T = [S_1, S_2, \ldots, S_n, A].$2

Here, $T = [S_1, S_2, \ldots, S_n, A].$3 denotes the constructed reflection dataset, $T = [S_1, S_2, \ldots, S_n, A].$4 contains the input context including the error and reflection question-answer pair, and $T = [S_1, S_2, \ldots, S_n, A].$5 is the reflective reasoning target (Huang et al., 4 Oct 2025).

Two features distinguish this supervision scheme from more elaborate reflective training pipelines in the surrounding literature. First, MedReflect does not rely on external knowledge lookup; all reflection and correction are performed using the model’s pretrained knowledge. Second, reflection is taught as a one-pass generation process rather than via iterative RL or multi-stage pipelines (Huang et al., 4 Oct 2025). This makes the framework relatively lightweight in both data construction and optimization.

4. Reported benchmark performance

MedReflect is evaluated on MedQA (USMLE, test set), MedMCQA (validation set), PubMedQA (test set), MMLU-Pro (Health, Biology sections), and GPQA (Genetics, Molecular Biology) (Huang et al., 4 Oct 2025). The reported 7B system, MedReflect-7B, achieves 74.2 on MedQA, 77.1 on MedMCQA, 85.2 on PubMedQA, 71.7 on MMLU-Pro (Health), 79.2 on MMLU-Pro (Bio), 61.1 on GPQA (Genetics), and 72.9 on GPQA (Molec). On the same benchmarks, the reported 32B system, MedReflect-32B, achieves 84.4, 80.1, 86.5, 82.6, 90.8, 65.0, and 75.3, respectively (Huang et al., 4 Oct 2025).

These scores are reported alongside baseline numbers for BioMistral-7B, UltraMedical-8B, Qwen2.5-7B-Instruct, HuatuoGPT-o1-8B, and Qwen2.5-32B-Instruct. In the reported table, MedReflect-7B exceeds Qwen2.5-7B-Instruct and HuatuoGPT-o1-8B across all listed datasets, while MedReflect-32B improves over Qwen2.5-32B-Instruct across the same suite (Huang et al., 4 Oct 2025). The paper further states that MedReflect-7B outperforms all open-source models of similar scale, including HuatuoGPT-o1-8B and UltraMedical-8B, and that MedReflect-32B is on par with or superior to several 70B-parameter models and approaches proprietary state of the art (Huang et al., 4 Oct 2025).

The ablation findings attribute the gains specifically to reflection structure rather than merely more supervised data. Both self-questioning and self-answering are described as critical, and removing either degrades performance. The reported gains are also much greater than those obtained by naive supervised fine-tuning with more original or correction data (Huang et al., 4 Oct 2025). Reflection is said to be especially beneficial on complex, reasoning-demanding questions, particularly on MMLU and GPQA, and the model is reported to self-reflect more often when task difficulty increases (Huang et al., 4 Oct 2025). This suggests that the framework is not only improving answer selection, but also modulating the model’s willingness to invoke correction behavior as problem complexity rises.

5. Relation to adjacent reflection-based medical AI

MedReflect belongs to a broader research trend in which reflection is treated as a first-class mechanism for medical reasoning, but its design differs materially from several neighboring approaches. Med-REFL, for example, uses a tree-of-thought decomposition of medical questions, quantitatively evaluates reasoning steps and reflections, and constructs direct preference optimization data automatically; its reported gains on MedQA-USMLE average up to 4.11%, with an additional 4.13% boost to the state of the art performance of 7B/8B models (Yang et al., 11 Jun 2025). Compared with MedReflect, Med-REFL places more emphasis on fine-grained path scoring and DPO-style preference learning.

In multimodal report generation, LVMed-$T = [S_1, S_2, \ldots, S_n, A].$6 introduces complex reasoning and reflection mechanisms for medical report generation, including medical knowledge injection, perception-enhancing modules, a perception tree, and a self-verification stage. On MIMIC-CXR with Qwen2.5VL-7B, the reported clinical efficacy F-score rises from 0.185 with direct SFT to 0.216 with reasoning only and 0.254 with reasoning plus reflection (Wang et al., 2 Apr 2025). Here, reflection operates within a vision-language report generation pipeline rather than in text-only medical QA.

Self-MedRAG represents another distinct branch: it combines hybrid retrieval through BM25, Contriever, and Reciprocal Rank Fusion with a self-reflective loop that verifies rationale support using NLI or LLM-based verification. In the reported results, accuracy rises on MedQA from 80.00% to 83.33% and on PubMedQA from 69.10% to 79.82% when the self-reflective loop is included (Ryan et al., 8 Jan 2026). Unlike MedReflect, Self-MedRAG explicitly depends on external evidence retrieval and iterative query reformulation.

Reflection has also been embedded in clinical agents and imaging agents. ReflecTool uses long-term memory, retrieved demonstrations, and a verifier with iterative refinement or candidate selection, and on ClinicalAgent Bench it surpasses pure LLMs by more than 10 points and established agent-based methods by 3 points (Liao et al., 2024). In medical imaging, the $T = [S_1, S_2, \ldots, S_n, A].$7 framework decomposes analysis into Router, Retriever, Reflector, and Repairer agents and reports roughly +1.7 to +2.5 points in LLM-as-a-Judge scores and +2.5 to +3.5 absolute points in mAP50 over strong single-VLM baselines without gradient-based fine-tuning (Sayeedi et al., 13 Jan 2026). These systems indicate that reflection in medical AI is not a single method family; it spans self-correction in QA, rationale verification in RAG, critique-and-repair in multimodal generation, and tool-level verification in agentic settings.

6. Distinctions, limitations, and nomenclature

A recurring source of confusion is nomenclature. MedReflect should not be conflated with MedRef, a medical dialogue system that incorporates a knowledge refining mechanism and dynamic prompt adjustment through Triplet Filter and Demo Selector modules (Sun et al., 12 Jun 2025). It is also distinct from MIRAGE, a multimodal medical text and image retrieval and generation system for medical education built on a fine-tuned medical CLIP model, Prompt2MedImage, and Dolly-v2-3b (Benito et al., 6 May 2026). The former is an MDS for personalized response generation, and the latter is a didactic retrieval-and-generation interface; neither is a reflective correction framework for medical LLM reasoning in the sense used by MedReflect.

The principal limitation explicitly stated for MedReflect is architectural scope. The study focuses on the Qwen2.5 family, and generalization to other architectures is identified as a future direction (Huang et al., 4 Oct 2025). The framework also claims no external knowledge lookup, which is methodologically central, but this means its corrections depend on internalized pretrained knowledge rather than contemporaneous retrieval (Huang et al., 4 Oct 2025). A plausible implication is that MedReflect is strongest where the model already contains adequate latent medical knowledge and where the main bottleneck is not knowledge absence but failure to interrogate and revise its own reasoning.

Within the reported evidence, MedReflect’s significance lies in showing that reflective correction can be taught with lightweight supervision, minimal sampled training data, and parameter-efficient fine-tuning while still yielding broad gains across medical QA and consultation-style tasks (Huang et al., 4 Oct 2025). In the surrounding literature, this places it at the compact, retrieval-free end of the reflection spectrum: closer to internal reasoning reorganization than to external verification pipelines, yet empirically competitive enough to motivate reflection as a standalone training target in medical language modeling.