Copyright © 2026 Authors retain the copyright of this article. This article is an open access article distributed under the Creative Commons Attribution License which permits unrestricted use, distribution, and reproduction in any medium, provided the original work is properly cited.

@article{210341,
        author = {JULIAN FIDAL CASTRO S and GANGA NAIDU and AARYA A and KARTHICK SUBRAMANIAN K and LOGESHWARAN A},
        title = {Conformal Refusal for Quantized Small Language Models: Conditional Risk Control Under Resource Constraints},
        journal = {International Journal of Innovative Research in Technology},
        year = {2026},
        volume = {13},
        number = {6},
        pages = {502-510},
        issn = {2349-6002},
        url = {https://ijirt.org/article?manuscript=210341},
        abstract = {Small language models present a feasible choice for employing artificial intelligence in certain devices that have limited resources. However, the estimates made by them are not necessarily reliable. This research work studies a conformal refusal framework for a quantized small language model that aims to keep the number of errors in predictions not refused. The framework has been tested using Qwen2.5-1.5B-Instruct with FP16 and 4-bit NF4 quantization on the configuration of TruthfulQA having 817 questions with multiple choices. An initial likelihood-based score indicating confidence turned out to be strongly dependent on the number of answer choices, so a letter-based in-context comparative scoring method was used instead. We consider three approaches to calibration: Conformal Risk Control, Learn-then-Test (with the Bonferroni correction), and Fixed Sequence Testing. We evaluate these methods against fixed-threshold and prompted-refusal baselines, tested on 100 resamples. The results show that the Bonferroni approach offers empirical coverage rates of approximately 98.9 percent for FP16 and 98.3 percent for NF4, exceeding the expected nominal coverage of 90 percent with respect to an error target of 0.50. Reliable certification was obtained only at error targets of 0.50 and above. For the target of 0.50, the feasibility of threshold dropped from 93 percent for FP16 to 58 percent for NF4. Overall, the results highlight how conformal calibration can be useful for selective refusal of predictions in the small quantized language model.},
        keywords = {Small language models, conformal prediction, selective prediction, risk control, quantization, abstention, uncertainty calibration, edge artificial intelligence.},
        month = {November},
        }