MedReadCtrl: Controlled Medical Readability

• MedReadCtrl is a framework that adjusts medical text readability to specific grade levels while maintaining clinical integrity.
• It employs FLAN-style instructions and standard readability metrics across multi-task training to fine-tune output complexity.
• Quantitative and human evaluations show that MedReadCtrl improves patient communication and adherence by tailoring explanations to diverse literacy levels.

Searching arXiv for MedReadCtrl and closely related controllable medical readability work. MedReadCtrl is a readability-controlled instruction tuning framework for medical text generation that enables a LLM to adjust output complexity to a target reading grade while preserving medical meaning. It was introduced to address a central barrier in patient-centered AI communication: medical text must be clinically faithful yet tailored to the reader’s literacy. In its reported instantiation, LLaMA3-MedReadCtrl is based on LLaMA 3 8B and is trained across text simplification, paraphrase generation, and semantic entailment generation in medical and general domains, with the explicit design goal of aligning output readability to a user-specified grade level while preserving semantic fidelity and medical intent (Tran et al., 10 Jul 2025).

1. Clinical motivation and conceptual scope

MedReadCtrl is grounded in the observation that health systems have opened access to notes via OpenNotes and related policies, while comprehension still lags because most clinical documentation is written well above recommended patient reading levels. The motivating premise is that low health literacy correlates with worse adherence, medication errors, and inequities, especially among children, older adults, and non-native speakers. Within this setting, readability-controlled generation is treated not as a stylistic convenience but as an operational requirement for patient portals, discharge instructions, caregiver education, and clinician-facing paraphrase generation (Tran et al., 10 Jul 2025).

The framework departs from binary “simple vs. complex” control. Existing controllable generation is described as typically toggling between coarse levels or performing coarse style transfer, which underfits the nuanced personalization required in healthcare. MedReadCtrl instead models readability as a numeric target grade level and trains the model to follow explicit readability instructions over Grades 1–12. This formulation makes the control variable part of the task specification itself rather than an implicit property to be inferred after generation (Tran et al., 10 Jul 2025).

A central implication is that readability is handled as a first-class constraint alongside semantic preservation. The framework is therefore not merely a simplification system. It is intended to preserve medical intent across a spectrum of complexity, including low-literacy patient explanations and higher-grade clinician-oriented paraphrases. This suggests a broader view of controllable medical generation in which accessibility and faithfulness are optimized jointly rather than sequentially.

2. Readability control mechanism and formal evaluation

MedReadCtrl implements control through a FLAN-style instruction that directly embeds the requested grade level. The paper gives the template: “Given an input text, please output an entailment with a readability score around {target readability score}.” At inference time, the same principle is used in practical prompts such as rewriting discharge instructions for Grade 5, explaining a diagnosis for Grade 2, or paraphrasing a note at Grade 8 while preserving meaning (Tran et al., 10 Jul 2025).

Readability is measured using four standard metrics: Gunning Fog Index, Flesch–Kincaid Grade Level, Automated Readability Index, and Coleman–Liau Index. The instruction-following error is the mean absolute deviation between the requested grade and the grade implied by the generated text, averaged across these four metrics:

$E(y; G) = \frac{1}{|M|}\sum_{m\in M} |R_m(y)-G|,$

where $$M = \{\mathrm{Fog}, \mathrm{FKGL}, \mathrm{ARI}, \mathrm{CLI}\}$$.

For analysis, the paper also defines an overall Reading Grade Level:

$$RGL(y) = \frac{1}{4}\big(\mathrm{Fog}(y)+\mathrm{FKGL}(y)+\mathrm{ARI}(y)+\mathrm{CLI}(y)\big).$$

The metric-specific equations reported in the work include:

$$\mathrm{GFI} = 0.4\left[\left(\frac{\text{totalWords}}{\text{totalSentences}}\right) + 100\left(\frac{\text{longWords}}{\text{totalWords}}\right)\right],$$

where longWords are $>7$ characters,

$$\mathrm{ARI} = 4.71\left(\frac{\text{totalCharacters}}{\text{totalWords}}\right) + 0.5\left(\frac{\text{totalWords}}{\text{totalSentences}}\right) - 21.43,$$

and

$\mathrm{CLI} = 0.0588L - 0.296S - 15.8.$

The paper notes that FKGL is described but presented with a FRE-like expression in the manuscript; the standard FKGL formula is also provided. System comparisons use nonparametric Mann–Whitney tests over absolute readability errors. This evaluation design makes controllability measurable as a deviation from the requested grade rather than as a coarse category-matching problem (Tran et al., 10 Jul 2025).

3. Training data, tasks, and instruction tuning regimen

The training pipeline spans nine datasets across three tasks in medical and general domains, with seen and unseen splits used to test generalization. Medical datasets include ReadMe for medical text simplification, MedNLI for medical natural language inference, and MTSamples as an unseen medical text simplification set. General-domain datasets include ASSET and WikiSmall for simplification, PAWS and MRPC for paraphrase generation, and SNLI and MultiNLI for entailment generation (Tran et al., 10 Jul 2025).

Task family	Seen datasets	Unseen datasets
Medical	ReadMe, MedNLI	MTSamples
Simplification	ASSET	WikiSmall
Paraphrase / Entailment	PAWS, SNLI	MRPC, MultiNLI

The dataset table in the paper reports, for example, ReadMe Train 2,460, Dev 2,090, Test 2,250; MedNLI Train 2,520, Dev 2,180, Test 2,390; ASSET Train 2,620, Dev 2,450, Test 2,550; SNLI Train 2,100, Dev 1,900, Test 2,100; PAWS Train 2,700, Dev 2,520, Test 2,630; and unseen test sets MTSamples 870, WikiSmall 874, MultiNLI 920, MRPC 948. The model is trained only on the seen datasets, with unseen datasets reserved for testing transfer (Tran et al., 10 Jul 2025).

Instruction tuning is standard supervised fine-tuning with next-token cross-entropy, without ranking loss or RLHF. Reported hyperparameters are: LLaMA 3 8B backbone; Transformers 4.37.1; Adam with $\beta=(0.9,0.999)$ and $\epsilon=1\times10^{-8}$; learning rate $3\times10^{-4}$; linear decay; warmup 200 steps; weight decay 0; batch size 128; and 5 epochs. Training used two NVIDIA A100 40GB GPUs and 192 GB RAM. Decoding strategies are not explicitly specified in the paper (Tran et al., 10 Jul 2025).

This multi-task composition is important because the three task families span down-scaling, meaning-preserving restatement, and entailment-style reformulation. A plausible implication is that the model learns readability control as a transferable instruction-following capability rather than as a single-task simplification heuristic.

4. Quantitative performance and human evaluation

The primary automatic result is that MedReadCtrl reduces readability instruction-following error relative to GPT-4 on medical datasets. Reported mean absolute errors are 1.39 vs. 1.59 on ReadMe, 1.74 vs. 1.91 on MTSamples, 2.88 vs. 3.57 on MedNLI, and 2.00 vs. 2.36 overall in the medical setting, all with $$M = \{\mathrm{Fog}, \mathrm{FKGL}, \mathrm{ARI}, \mathrm{CLI}\}$$0. The paper also reports that a LLaMA 3 baseline without MedReadCtrl shows a near-flat curve between requested and realized readability, indicating limited controllability without explicit training (Tran et al., 10 Jul 2025).

On unseen clinical simplification, gains are large. For MTSamples, MedReadCtrl reports ROUGE-L 51.14 vs. GPT-4 36.44, SARI 23.21 vs. 17.03, ROUGE-1 53.99 vs. 39.40, and BLEU 29.73 vs. 11.63. On seen medical tasks, ReadMe simplification achieves ROUGE-L 30.79 vs. GPT-4 27.46, SARI 53.52 vs. 51.39, and BLEU 10.24 vs. 6.36; MedNLI entailment generation reaches ROUGE-L 39.49 vs. 23.90, BLEU 10.35 vs. 1.78, and SARI 63.97 vs. 56.12 (Tran et al., 10 Jul 2025).

General-domain robustness is also reported. ASSET simplification reaches ROUGE-L 59.16 and SARI 45.60; PAWS paraphrase reaches ROUGE-L 84.34, BLEU 66.49, and SARI 60.53; WikiSmall reaches ROUGE-L 55.86 and SARI 40.55. The paper states that SNLI and MultiNLI show consistent gains over GPT-3.5 and Claude-3 and competitive or superior performance relative to GPT-4 (Tran et al., 10 Jul 2025).

Human evaluation was conducted by two medical experts, with Cohen’s $$M = \{\mathrm{Fog}, \mathrm{FKGL}, \mathrm{ARI}, \mathrm{CLI}\}$$1, scoring Accuracy, Clarity, Consistency, and Fluency on a 1–5 scale at Grades 2, 5, 8, and 11. The overall average is 4.56 $$M = \{\mathrm{Fog}, \mathrm{FKGL}, \mathrm{ARI}, \mathrm{CLI}\}$$2 for MedReadCtrl versus 4.17 $$M = \{\mathrm{Fog}, \mathrm{FKGL}, \mathrm{ARI}, \mathrm{CLI}\}$$3 for GPT-4. By criterion, scores are Accuracy 4.61 vs. 4.33, Clarity 4.52 vs. 3.87, Consistency 4.57 vs. 4.32, and Fluency 4.56 vs. 4.17. By grade, MedReadCtrl scores 4.33 vs. 3.77 at Grade 2, 4.55 vs. 4.10 at Grade 5, 4.75 vs. 4.38 at Grade 8, and 4.63 vs. 4.44 at Grade 11. Expert head-to-head preferences favor MedReadCtrl 71.7% vs. GPT-4 23.3%, with gains largest at low literacy levels (Tran et al., 10 Jul 2025).

5. Qualitative behavior, safety profile, and workflow integration

The qualitative examples emphasize that MedReadCtrl changes wording and discourse structure rather than merely deleting difficult terms. For a Grade 2 explanation of X-ray myelography, the paper contrasts GPT-4’s “Taking an X-ray picture of the spinal cord after putting special dye into a space around it” with MedReadCtrl’s “A special picture of the spine that helps doctors see inside,” which experts judged more appropriate for Grade 2. For MedNLI, “cough with sputum and blood streaks” is rendered as “She has a cough and sometimes brings up yucky stuff from her lungs. Sometimes it might have a little bit of red in it, but it’s not too much.” For MTSamples, “groin tenderness” becomes “pain in the area where her legs and hips meet.” In follow-up planning, the system scales the same intent from “I will see you again in three months” at Grade 2 to a more elaborated Grade 11 formulation about scheduling a follow-up appointment to assess status and determine treatment adjustments (Tran et al., 10 Jul 2025).

The paper’s error analysis identifies three recurring issues: occasional hallucinations or semantic drift, verbosity conflated with higher readability levels, and occasional instruction misinterpretation. Safety limitations are stated explicitly: residual risk of misinterpretation remains at very low reading levels; traditional readability metrics may under-capture understandability in telegraphic EHR text; all experiments are in English; and fully automated evaluation can miss clinically salient qualities, as shown by divergence between LLM-as-judge outputs and clinician preferences (Tran et al., 10 Jul 2025).

Recommended deployment practice is conservative. The authors propose human-in-the-loop oversight, knowledge grounding through retrieval-augmented generation, structured knowledge such as UMLS or medical dictionaries, and post-generation fact-checking such as VeriFact for high-stakes use. Suggested workflows include Grade 5–8 generation for patient portals and after-visit summaries, Grade 5 discharge instructions with explicit safety clauses, Grade 5–8 caregiver education, and Grade 11 paraphrases for inter-professional communication. Prompt-level guardrails include “Do not introduce details not present in the input” and “Cite source or flag uncertainty if unsure” (Tran et al., 10 Jul 2025).

6. Relation to adjacent controllable medical NLP

MedReadCtrl sits within a broader progression from coarse readability control and edit-based simplification toward instruction-driven, medically grounded personalization. The immediate precursor is ReadCtrl, which introduced “Readability-Controlled Instruction Learning” for near-continuous readability control with numeric targets and showed that ReadCtrl-Mistral-7B outperformed GPT-4 and Claude-3 in human and automatic evaluations across simplification, paraphrase, and entailment tasks (Tran et al., 2024). MedReadCtrl can be understood as the medical specialization of that paradigm, replacing general personalization with clinically constrained readability alignment.

Two earlier medical control lines are especially relevant. “Med-EASi” introduced finely annotated controllable medical text simplification with four edit operations—elaboration, replacement, deletion, and insertion—and reported that position-aware control generated better simplification than position-agnostic control (Basu et al., 2023). “Readability Controllable Biomedical Document Summarization” formulated binary readability control for biomedical summarization, distinguishing technical summaries from plain language summaries, and found that although current control techniques allowed a certain degree of readability adjustment, performance was “far from desirable” (Luo et al., 2022).

These comparisons clarify MedReadCtrl’s distinctive contribution. Relative to binary technical/plain summarization and edit-tag-based simplification, MedReadCtrl treats readability as a numeric target grade level and trains a single instruction-following model across multiple tasks. This suggests a shift from local edit control or binary style control toward grade-indexed generation that is directly callable in clinical workflows. At the same time, the paper’s own limitations indicate that readability control alone is not sufficient for high-stakes medical deployment; it must remain coupled to semantic fidelity, grounding, and expert review (Tran et al., 10 Jul 2025).