At the bedside, a correct answer that arrives too long is worthless: no clinician can wait out a 256-token differential while the patient is still bleeding. A preprint posted to arXiv on 12 November 2026 opens on exactly this knot — an answer that takes longer than the time available has failed, however accurate it is. The four authors, from COMSATS University Islamabad in Pakistan and Bryant University in the United States, call their method PECA, and aim it at a habit that parameter-efficient fine-tuning never touches: verbosity.
Fine-tuning fixes the cost, not the length
Hospitals struggle to host frontier models: patient data cannot leave the building, and the largest systems are the hardest to keep inside it. Small language models of roughly one to thirteen billion parameters, adapted with QLoRA, are the realistic route for department-level deployment. But as the paper puts it, PEFT solves the cost of training and leaves output length alone — length follows the training distribution, never an explicit objective. The medical chain-of-thought corpus the authors train on is itself long: the English split of medical-o1-reasoning-SFT holds 19,704 verified examples, with an average reasoning trace of 312.45 words and an average response of 93.98 words. A model trained on it learns the habit.
Penalize stopping, not counting
You cannot write the token count directly into the loss: the count is discrete, its gradient is zero almost everywhere, so a naive length penalty cannot drive learning. PECA acts on the stop probability instead. Once a response passes a budget B (fixed at 64), the loss raises the end-of-sequence probability at every later position — a smooth, differentiable term added to the cross-entropy, scaled by a strength knob λ. No preference pairs, no reward model, no reference policy, no RL rollouts, and nothing extra at inference. The paper describes it as the only entry in its comparison table that is at once preference-free, teacher-free, and differentiable within supervised fine-tuning.
Seventy percent shorter, accuracy up
On a Qwen3-1.7B-Instruct backbone with 8,000 training examples, greedy decoding, and a 128-token cap, the main table reads: vanilla QLoRA averages 256.0 tokens with 40.8% exact-match accuracy on 1,000 MedMCQA items; at λ=0.25 the mean falls to 69.7 tokens, a 72.8% cut, while accuracy rises to 44.0%; at λ=0.50 accuracy drops to 31.9%; at λ=1.00 it collapses to 4.6%, most replies down to one or two tokens. Conciseness pressure is a dial, not a slider to max: past a point, length barely moves while accuracy falls first.
Two details carry the interpretation. ROUGE-L and BERTScore barely move from baseline, which the paper reads as cuts falling on hedging and repetition rather than load-bearing content. And the whole length distribution slides left instead of being clipped — the model learned when to stop. A vignette makes it concrete: a 32-week pregnant woman with painful, continuous vaginal bleeding; the vanilla model is still listing differentials when the 256-token ceiling arrives and never states a diagnosis, while the PECA model finishes in 68 tokens with the diagnosis stated — a 73.4% cut.
Cost and the discounts the authors list themselves
Each condition trained in 153 to 163 minutes on a single NVIDIA A10G or L4, peaking at 4.4 to 5.3 GB — inside a free-tier 16 GB cloud card. The authors discount their own results plainly: a single random seed and a single model family; the absolute MedMCQA figures are the ceiling of a 1.7B model on a demanding benchmark, not a clinical decision-support system; against the RL line SmartThinker and its reported 52.5% reduction on mathematical reasoning, the paper itself notes the two figures are not directly comparable; re-training preference baselines on the identical backbone is left as future work; code and adapters appear only upon publication, so there is no repository to check yet.
So what
The interesting inference reaches past medicine: a domain model's verbosity is mostly residue of its training distribution, not the price of reasoning. Suppressing it does not require a reward model or reinforcement learning — one well-placed signal about when to stop, inside ordinary supervised fine-tuning, is enough. Next time your fine-tuned model rambles, ask whether anyone ever taught it to stop.
Paper: https://arxiv.org/abs/2611.17128 (arXiv preprint, not peer reviewed; all figures are self-reported by the paper).