On November 12, 2026, researchers from COMSATS University Islamabad and Bryant University published a preprint introducing Parameter-Efficient Conciseness Adaptation (PECA), a training method that curbs excessive verbosity in smaller language models (SLMs) adapted for clinical reasoning. By adding a single differentiable conciseness penalty to the QLoRA fine-tuning objective, the authors demonstrated a 72.8% reduction in response length alongside an unexpected gain in multiple-choice diagnostic accuracy from 40.8% to 44.0% on the MedMCQA benchmark. The entire adaptation completed in under three hours on a single commodity GPU with peak memory under 5.3 GB.
The Post-Training Verbosity Trap in Healthcare Language Models
Deploying large proprietary language models in healthcare presents major governance hurdles. Strict data privacy regulations and good clinical data management practices often require patient records to remain on on-premise hospital servers. While smaller language models (1 to 13 billion parameters) run on commodity infrastructure, parameter-efficient fine-tuning (PEFT) introduces a persistent challenge: post-training verbosity. When adapted on clinical reasoning corpora, fine-tuned SLMs often generate sprawling, multi-paragraph chains of thought for straightforward queries.
In fast-paced hospital workflows and clinical decision support tools, verbose outputs cause workflow friction. More critically, SLMs subject to fixed generation limits often exhaust their token budgets exploring differential diagnoses without ever stating a definitive conclusion. Existing length-control methods—such as Direct Preference Optimization (DPO), length-regularized DPO, or SimPO—require paired preference datasets, reward models, or distillation from proprietary frontier systems, undermining the cost and privacy benefits of local SLMs.
Differentiable Length Control via Budget-Gated Stopping
The PECA framework, introduced in an arXiv preprint by Tanzeel Abbas, Ahmad Kamran Malik, Nafees Qamar, and Sana Liaqat, addresses verbosity through a lightweight modification to the supervised training loss. A naive penalty on discrete token counts cannot be optimized directly by gradient descent because discrete counts have a derivative of zero almost everywhere. PECA resolves this by penalizing low end-of-sequence (EOS) token probability once generation exceeds an assigned token budget (B).
During supervised fine-tuning, the standard cross-entropy loss is augmented by an auxiliary conciseness loss: L_PECA = L_CE + λ L_conc. For all response positions t exceeding budget B, the loss computes the negative log-likelihood of emitting the EOS token. Because softmax EOS probability is continuous, it differentiates cleanly through adapter weights without extra models or reference policies. Adjusting the conciseness hyperparameter λ allows practitioners to balance brevity and diagnostic completeness.
Empirical Evaluation and Clinical Diagnostic Gains
The authors evaluated PECA on the open medical-o1-reasoning-SFT dataset (19,704 clinical reasoning instances from MedQA-USMLE and MedMCQA). The base model was Qwen3-1.7B-Instruct quantized to 4-bit precision via NormalFloat4 double quantization (~1.1 GB base memory). QLoRA adapters targeted all projection layers (17.4 million trainable parameters, or 1.69% of the model). Fine-tuning ran on a single NVIDIA A10G or L4 GPU across four conditions: vanilla SFT (λ = 0) and PECA at λ values of 0.25, 0.50, and 1.00, using budget B = 64 tokens.
| Condition | Mean Length (Tokens) | Length Reduction | ROUGE-L | BERTScore | MedMCQA MCQ Accuracy |
|---|---|---|---|---|---|
| Vanilla SFT (λ = 0) | 256.0 | 0.0% | 0.162 | 0.129 | 40.8% |
| PECA (λ = 0.25) | 69.7 | 72.8% | 0.155 | 0.118 | 44.0% |
| PECA (λ = 0.50) | 66.2 | 74.1% | 0.154 | 0.117 | 31.9% |
| PECA (λ = 1.00) | 65.2 | 74.5% | 0.152 | 0.113 | 4.6% |
Vanilla SFT models exhausted the full 256-token decoding window, averaging 256.0 tokens with 40.8% accuracy. In contrast, PECA at λ = 0.25 reduced mean output length to 69.7 tokens while increasing diagnostic accuracy to 44.0%. Case reviews showed that accuracy improved by preventing truncation. In an emergency vignette of a 32-week pregnant patient with painful vaginal bleeding, the vanilla model exhausted 256 tokens without reaching a diagnosis. PECA isolated the key distinguishing symptom in 68 tokens, correctly diagnosing placental abruption. However, excessive pressure (λ = 0.50 and 1.00) dropped accuracy to 31.9% and 4.6% by truncating essential clinical reasoning.
Implications for Biomedical Informatics and Research Teams
For hospital data teams advancing clinical natural language processing, PECA shows that domain-specific reasoning models do not require prohibitive compute or preference annotation pipelines. Training each condition took 153 to 163 minutes on a single GPU under 5.3 GB VRAM, fitting easily within commodity 16 GB hardware.
The approach aligns with standards for AI clinical decision support and good machine learning practice, where predictable outputs are critical. By tuning conciseness strength, research institutions can systematically adjust model outputs to point-of-care clinical requirements while preserving patient data confidentiality.
Sources
- Abbas, T., Malik, A. K., Qamar, N., and Liaqat, S. (2026). Parameter-Efficient Conciseness Adaptation: Differentiable Length Control for Concise Clinical Reasoning in Smaller Language Models. arXiv preprint arXiv:2611.17128. https://arxiv.org/abs/2611.17128 (2026-11-12)
- arXiv Full HTML Version: Parameter-Efficient Conciseness Adaptation (arXiv:2611.17128v1). https://arxiv.org/html/2611.17128v1 (2026-11-12)
