Uncertainty based selection of compatible inputs

An editable, auditable 807K-param byte-level LLM: CRUD single facts with provable per-edit locality, and abstain when unsure instead of guessing. CPU, offline.

Behavioral Trust Clustering a thermodynamic governance layer that reduces LLM hallucination by 52% on HumanEval. Drop-in wrapper for any decoder. MIT.

We show that a model owner can artificially introduce uncertainty into their model and provide a corresponding detection mechanism.

Uncertainty-aware reliability monitor for safety-critical ML models — calibration (ECE), risk-coverage (AURC), selective prediction with human review, and drift detection. React, FastAPI, PostgreSQL, Kafka.

A comprehensive library for uncertainty quantification in machine learning.

Calibrated trust and abstention for MS/MS molecular annotations

DegradeRisk-Seg: risk-controlled semantic segmentation under degraded multi-modal remote-sensing observations

Prompt-only boundary prediction for IFEval-style instruction-checker pass/fail behavior.

Reliable medical QA with Mistral-7B, QLoRA, selective prediction, and learned abstention via warm-start SFT + DPO.

Tsetlin Machines with a certificate on every answer: the exact number of feature flips a prediction survives, computed per sample, with predict-or-abstain when the radius is too small.

Investigation of how sampling strategies affect Selective Prediction performance in Multi Task Learning

Reproducible MEDAI deferral simulation (AIRI 2026). Synthetic research code.

Code Repository for SCoRE paper

Proper scoring rules, reduces LLM overconfidence in multiple-choice QA.

Six from-scratch NumPy classifiers + a cost-sensitive abstention (reject) framework for Human Activity Recognition. 94% base accuracy; hits the 95% target by abstaining on 2.7% (UCI HAR).

Audit framework for LLM trust-routing over biological science foundation model outputs.

Selective RAG with conformal abstention: a hallucination detector that scores its own confidence and abstains, with a finite-sample precision guarantee.

Trustworthy medical image classification: noise-robust ConvNeXt-Tiny with 83.5% accuracy, calibrated selective prediction, HAM10000 + ISIC 2019.

New Zealand-localized phishing and smishing detection with selective abstention. Combines a localized feature set (IRD, NZ Post, banks, te reo cues, .govt.nz lookalikes) with classical baselines, fine-tuned transformers, and LLM-as-classifier (Gemini), under a class-conditional deferral layer for human review. Independent research.