# Small language models

Small language models for clinical decision support often answer at far greater length than the question requires after fine-tuning, and existing length control relies on preference data, reward models or proprietary teachers. PECA adds a differentiable conciseness term to the QLoRA supervised fine-tuning objective: once generation passes a token budget, it raises the end-of-sequence probability, with no preference data, no auxiliary model and a single commodity GPU. On the 1.7-billion-parameter Qwen3-1.7B-Instruct, PECA reduces mean response length from 256.0 to 69.7 tokens, a 72.8% cut, while raising MedMCQA multiple-choice diagnostic accuracy from 40.8% to 44.0%.

Small language models for clinical decision support often answer at far greater length than the question requires after fine-tuning, and existing length control relies on preference data, reward models or proprietary teachers. PECA adds a differentiable conciseness term to the QLoRA supervised fine-tuning objective: once generation passes a token budget, it raises the end-of-sequence probability, with no preference data, no auxiliary model and a single commodity GPU. On the 1.7-billion-parameter Qwen3-1.7B-Instruct, PECA reduces mean response length from 256.0 to 69.7 tokens, a 72.8% cut, while raising MedMCQA multiple-choice diagnostic accuracy from 40.8% to 44.0%.

A 9B small language model running on one local GPU performs case-level and trio-based variant interpretation, recovering 93% of expert-reported causal variants across 327 solved singleton visual-impairment cases and classifying 82% of them as pathogenic or likely pathogenic.

A 9B small language model running on one local GPU performs case-level and trio-based variant interpretation, recovering 93% of expert-reported causal variants across 327 solved singleton visual-impairment cases and classifying 82% of them as pathogenic or likely pathogenic.

A home-care scheduling agent no longer asks a small model to handle a narrow task cold. Zero-shot small models failed it, and retraining on every behavior change was too slow, so the behavior was first iterated in prompt space on GPT-5.4-nano, then 11.6k recorded production decisions were LoRA-tuned into Qwen3-8B. That model agrees with recorded decisions 0.913 of the time, above the teacher's 72.7% self-agreement when re-run on its own prompts.

A home-care scheduling agent no longer asks a small model to handle a narrow task cold. Zero-shot small models failed it, and retraining on every behavior change was too slow, so the behavior was first iterated in prompt space on GPT-5.4-nano, then 11.6k recorded production decisions were LoRA-tuned into Qwen3-8B. That model agrees with recorded decisions 0.913 of the time, above the teacher's 72.7% self-agreement when re-run on its own prompts.

In network intrusion detection, small language models reading raw numerical flow records often cannot separate traffic classes. SERA-IDS has an 8B model analyze only misclassified flows and turn each confusion into a structured rule carrying a behavioral description, feature conditions, class-confusion information, tool-derived confidence and provenance, admitted through a confidence gate into an experience library. At test time the library is frozen; up to five class-diverse rules are retrieved and condition-matched, lifting macro F1 from 1.94%–11.66% to 69.50%–87.66% for Llama 3.1, Phi-4:14B and Qwen2.5:7B on NF-BoT-IoT and NF-ToN-IoT, with no fine-tuning and no paid hosted inference.

In network intrusion detection, small language models reading raw numerical flow records often cannot separate traffic classes. SERA-IDS has an 8B model analyze only misclassified flows and turn each confusion into a structured rule carrying a behavioral description, feature conditions, class-confusion information, tool-derived confidence and provenance, admitted through a confidence gate into an experience library. At test time the library is frozen; up to five class-diverse rules are retrieved and condition-matched, lifting macro F1 from 1.94%–11.66% to 69.50%–87.66% for Llama 3.1, Phi-4:14B and Qwen2.5:7B on NF-BoT-IoT and NF-ToN-IoT, with no fine-tuning and no paid hosted inference.

The authors collapsed small transformers into depth-induced rank collapse by weakening the skip connection and then treated copies with two rank-based loss terms differing only in whether their gradient vanishes; neither repaired the collapse even though the bounded force reached about a tenth of the task gradient, because the task gradient no longer reached the query and key weights that decide where attention looks, whereas restoring the skip connection, a runtime action changing no weight, reopened that path at once, after which the rank recovered but only far above the scale of collapse and with the loss still above a healthy network.

The authors collapsed small transformers into depth-induced rank collapse by weakening the skip connection and then treated copies with two rank-based loss terms differing only in whether their gradient vanishes; neither repaired the collapse even though the bounded force reached about a tenth of the task gradient, because the task gradient no longer reached the query and key weights that decide where attention looks, whereas restoring the skip connection, a runtime action changing no weight, reopened that path at once, after which the rank recovered but only far above the scale of collapse and with the loss still above a healthy network.

The study investigates whether small language models (SLMs) can serve as task-tool relevance classifiers that evaluate each selected tool independently against the assigned task and return a relevance signal for downstream enforcement, and it introduces a novel dataset of multi-tool tasks whose required tools span distinct Model Context Protocol (MCP) servers, using prompt optimization, supervised fine-tuning, and reinforcement learning through GRPO to optimize and specialize SLMs.

The study investigates whether small language models (SLMs) can serve as task-tool relevance classifiers that evaluate each selected tool independently against the assigned task and return a relevance signal for downstream enforcement, and it introduces a novel dataset of multi-tool tasks whose required tools span distinct Model Context Protocol (MCP) servers, using prompt optimization, supervised fine-tuning, and reinforcement learning through GRPO to optimize and specialize SLMs.

This study deployed the 3.8B-parameter open-weight small language model Phi4-mini (via Ollama) within an institutional firewall, combined with deterministic regular expression post-processing, to extract six uncoded urinary tract infection clinical features (fever, nitrites, leukocytes, lumbar pain, abdominal pain, and haematuria) from 15,498 Catalan/Spanish bilingual MEAP primary care narratives in the SIDIAP database in Catalonia, achieving 93.8% accuracy, 96.6% specificity, and 83.6% sensitivity in a double-blind clinician gold standard validation of 60 real patient records, and 87.2% accuracy, 99.5% specificity, and 99.2% positive predictive value in adversarial synthetic stress-testing of 720 notes, with no patient data leaving institutional servers.

This study deployed the 3.8B-parameter open-weight small language model Phi4-mini (via Ollama) within an institutional firewall, combined with deterministic regular expression post-processing, to extract six uncoded urinary tract infection clinical features (fever, nitrites, leukocytes, lumbar pain, abdominal pain, and haematuria) from 15,498 Catalan/Spanish bilingual MEAP primary care narratives in the SIDIAP database in Catalonia, achieving 93.8% accuracy, 96.6% specificity, and 83.6% sensitivity in a double-blind clinician gold standard validation of 60 real patient records, and 87.2% accuracy, 99.5% specificity, and 99.2% positive predictive value in adversarial synthetic stress-testing of 720 notes, with no patient data leaving institutional servers.