Research
My research asks one question: when a generative AI model produces an answer, a caption, or a clinical note, can we trust it, and if not, can we fix it? I work on this question in the areas below.
Faithful image captions for safety-critical traffic scenes (MS-CAPS)
2025 – present
When a vision-language model shortens a caption, it can drop details that matter for safety, such as a pedestrian or a traffic signal. MS-CAPS uses a knowledge graph to shorten traffic-scene captions to a minimal sufficient form, and checks each caption by asking an image generator (FLUX.1-dev) to rebuild the scene from it.
- Evaluated on 150 real-world scenes from Vehicle-centric, CoDA-LM, and NuInteract with Qwen2.5-VL, GPT-4o, and Gemini.
- Up to 98% of outputs accepted (49 of 50 scenes) with 62.5% average caption reduction.
Related papers: EACL 2027 (see Publications)
Hallucinations in clinical speech
2025 – present
AI systems that write clinical notes from doctor–patient conversations can turn a listening mistake, such as a misheard drug name, into a false statement in the note. I compare speech-to-text-then-LLM systems with models that listen to audio directly, and trace where each error starts.
- Evaluation pipeline on 57 doctor–patient consultations (PriMock57).
- Pilot runs showed about 75% hallucination and 80% topic-mismatch rates, motivating stage-by-stage error tracing.
Correcting hallucinations in generated text (REVISE)
2024 – present
Language models often write paragraphs that mix correct and incorrect statements. REVISE splits a paragraph into individual claims, checks each claim against retrieved evidence, and rewrites only the claims that are not supported, keeping the rest of the paragraph the same.
- Evaluated on 1,400+ passages from Natural Questions and SQuAD.
- Up to 76.8% factual accuracy on the NQ-Mis benchmark.
Related papers: AITest 2025, DISTILL 2025, CCNC 2026 (see Publications)
Adversarial testing of language models
2024
I measure how small changes to a prompt increase hallucination in open language models, so that detection and correction methods can be tested under realistic pressure.
- Attacks on Vicuna-7B and LLaMA-2-7B-chat succeeded up to 92.53% of the time.
Related papers: MobiSecServ 2024 (see Publications)