Springer Best Paper Award · ISVC 2022
Learning When to Say “I Don’t Know”
Per-class abstention thresholds that need no rejection cost or coverage target: CIFAR-100 selective accuracy climbs from 88.3% to 97.8% at 77.3% coverage.
01Machine learning · Research & engineering
Nick Kashani Motlagh. I build LLM and retrieval systems that know when not to answer.
PhD, Ohio State (2026). I built LoRA-trained policies that decide when a RAG system should revise its answer, evaluated on 25,870 questions, and a reject-option method that won a Springer Best Paper Award.
FIG.0 1,000 CIFAR-100 test predictions, grown to their confidence. Below the line, the model says “I don’t know.” Try the threshold
Available nowfor Research Scientist,Research Engineer,ML Engineer,Applied Scientist,SWE (AI)
Open to relocating.U.S. citizen
1.1 The short version
- Status
- Available now
- Roles
- Research Scientist Research Engineer Machine Learning Engineer Applied Scientist Software Engineer (AI)
- Where
- Open to relocating. Based in Columbus, Ohio.
- Work auth
- U.S. citizen. No sponsorship needed.
- Education
- PhD, Computer Science and Engineering, Ohio State, Aug 2026. Advisor Prof. Jim Davis; minors in Mathematics and HPC.
- Now
- Computer Engineer II, DCS Corp. AFRL-sponsored, since May 2025.
- Papers
- 4 first-author, peer-reviewed. Springer Best Paper Award, ISVC 2022.
- nmotlagh@gmail.com
1.2 Skills, and where I used them
- LLMs & post-training
-
LoRA / PEFT Teacher-model synthetic data vLLM HF Transformers
LoRA-trained scorers that choose answer, revise, or abstain; evidence-grounded QA models trained on teacher-generated data at DCS. Evidence: RAG revision DCS role
- Retrieval
-
DPR dense retrieval BM25 MonoT5 reranking FAISS
Three Wikipedia retrieval setups behind a 25,870-question evaluation. Evidence: RAG revision
- Evaluation
-
Paired outcomes Seeds + paired t-tests Bootstrap intervals LLM-as-judge Selective prediction Calibration (ECE, Brier)
3 seeds with paired t-tests and 10,000-replicate bootstraps; a reimplementation checked against its reference by 37 parity tests. Evidence: RAG revision Reject option
- Vision
-
ViTs Linear probes on DINOv2, DINOv3, SigLIP 2 Self-supervised pretraining
Probes on four frozen ViT backbones; self-supervised vision models at AFRL in 2024. Evidence: Backbone re-run AFRL 2024
- Engineering
-
Python PyTorch NumPy Slurm Singularity uv Git Linux HPC
Python training and evaluation code; Slurm GPU jobs for the RAG study. Evidence: Live demo RAG revision
2.1 Selected work
What I’ve built. Scale, stack, and status for each.
-
Dissertation capstone
Net accuracy change = repaired − harmed Should a RAG system keep its draft or revise it?
Revision can repair a wrong answer or harm a right one. The system has to choose before paying for the full revision.
What I built
- A paired-outcome evaluation on 25,870 held-out questions from NQ-Open, TriviaQA, and PopQA. Each answer is scored as preserved, repaired, harmed, or unrecovered.
- Three retrieval setups over Wikipedia: DPR dense retrieval, BM25, and BM25→MonoT5 reranking. Llama 3.1 8B Instruct drafts and revises, Llama 3.3 70B Instruct judges, and gpt-oss-20b and OLMo 3 7B are secondary generators.
- LoRA-trained scorers that choose answer or revise, plus abstain in a three-action menu.
- 3 training seeds with paired run-level t-tests, and 10,000-replicate bootstrap intervals for baselines. Thresholds are picked on dev and frozen before test.
- PyTorch
- HF Transformers
- LoRA
- vLLM
- FAISS
- Slurm
Preprint Public arXiv preprint: Return or Revise?, v1 posted September 24, 2026; not peer-reviewed. Code and artifacts will not be released.
-
ISVC 2022 · MVA 2025 · 2026
Declined guesses are no better than a coin flip Reject-option classification, from Best Paper to live demo
A classifier should decline its least reliable predictions without a hand-set rejection cost or coverage target.
The method · 2022
- Per-class softmax thresholds learned from validation data with a binomial-CDF test: keep declining only while the declined predictions are no better than a coin flip.
- On CIFAR-100, selective accuracy rises from 88.3% to 97.8% at 77.3% coverage. The journal extension evaluates benchmark image and text datasets and twelve synthetic variants (four 2-D datasets, each with three variants).
The 2026 re-run
- Linear probes on 4 frozen backbones (DINOv2 ViT-S/14 and ViT-B/14, DINOv3 ViT-S/16, SigLIP 2 ViT-B/16), logits from a local RTX 4090, and a fast NumPy B-CDF matched to the reference by 37 parity tests.
The demo
- The site’s live demo applies the rule to 1,000 test images, with a draggable threshold and δ controls.
- Python
- PyTorch
- timm
- NumPy
- uv
Best Paper · Code public Method peer-reviewed: Springer Best Paper Award at ISVC 2022, journal extension in Machine Vision and Applications (2025). The 2026 re-run is public code, not peer-reviewed.
-
WMT 2024
Swap the image, measure the change Does a multimodal translation model use the image?
Translation scores alone can’t show whether a model reads the image or ignores it.
What I built
- ImageCoMMuTE: contrastive metrics that hold a candidate translation fixed and swap in matched or mismatched images, testing image–translation consistency.
- Evaluated three English-to-French multimodal model families, plus gated variants.
- VGAMT scores 81% IC and 63% TC on contrastive image/translation diagnostics. Against mixed imagery, preference flips are 13% favorable and 0% unfavorable; these scores do not measure generated-translation accuracy.
Peer-reviewed Peer-reviewed, WMT 2024.
Also public
-
ICCV 2021 Workshop · Code public
Temporal satellite imagery collection
An OpenStreetMap-guided Python pipeline that extracts candidate construction-site regions and downloads imagery over time. Rebuilt in 2026 as one package with swappable history and imagery backends and a 410-test suite.
-
Code public
Model calibration utilities
PyTorch utilities for histogram binning and global and class-wise temperature scaling, with expected-calibration-error summaries and calibration plots.
- Python
- PyTorch
2.2Live demo · ISVC 2022 Best Paper
Stop answering where the answers become coin flips.
Sort a classifier’s guesses by confidence and decline the least sure ones. The rule from Learning When to Say “I Don’t Know” keeps declining only while the declined guesses are no better than a coin flip. The paper fits a line per class; this demo uses one line for all 100 classes. Drag it.
Guess
Truth
·
1,000 of the 10,000 test images, one dot each, placed by the model’s confidence. Mistakes stack at the bottom of each column; hollow dots are declined. The dashed line is where the rule put the threshold. Hover or press the arrow keys to inspect an image.
- Answers
- 78.7%
- 7,868 of 10,000 test images
- Right when it answers
- 95.7%
- vs 86.1% answering everything
- Right when it declines
- 50.6%
- no better than a coin flip
No better than a coin flip: the declined guesses were right 50.6% of the time, so declining them gives up almost nothing.
Model: DINOv2 ViT-S/14 (frozen) + linear probe, temperature-scaled. Data: CIFAR-100. The line is learned on 10,000 held-out training images; every number here is measured on the 10,000 test images, of which the plot shows 1,000.
03 Research
Peer-reviewed publications. 4 first-author papers, 2021–2025.
04 Experience
AFRL-sponsored ML research every year since 2020. DCS Corp since May 2025; Ohio State’s Computer Vision Lab, 2021–2026.
Computer Engineer II — DCS Corp
- Build evidence-grounded QA systems: small open models answer from retrieved documents, cite the passage they used, and abstain when the evidence doesn’t support an answer. Uses synthetic training data from a teacher model, LoRA fine-tuning, and paired base-versus-adapted evaluation on held-out questions.
- Build paired-outcome evaluation pipelines for retrieval-augmented QA that measure, on 25,870 held-out questions, when revision repairs a wrong answer and when it harms a right one.
- Train and evaluate LoRA fine-tuned models for answer, refine, and abstain decisions with PyTorch, Hugging Face, and Slurm, comparing policy variants on coverage, utility, and out-of-distribution behavior.
Graduate Research Associate — Computer Vision Lab
- Built selective-prediction systems and reproducible experiments for vision, multimodal, and language tasks, advised by Prof. Jim Davis.
- Developed per-class reject-option classification and released public research code. Springer Best Paper Award at ISVC 2022; extended in Machine Vision and Applications, 2025.
- Designed ImageCoMMuTE evaluation metrics to separate visual evidence use from final translation decisions (WMT 2024).
- Published first-author papers on selective prediction, multimodal evaluation, and satellite imagery collection.
05 Get in touch
Available now.
Research Scientist, Research Engineer, ML Engineer, Applied Scientist, and SWE (AI) roles. Open to relocating.
- nmotlagh@gmail.com
- Based in
- Columbus, Ohio open to relocating
- Work auth
- U.S. citizen no sponsorship needed