KLIP: localized distribution shift detection with diffusion priors
arXiv preprint: a KL-divergence metric between prior and posterior finds OOD patches in inverse problems without calibration data on shifted domains.
Yevhen Stuzhuk - Personal Portfolio
I am a Senior Software Engineer
← All posts
Tag
research
All blog posts with this tag.
SPECTRA: synthetic IR collections to stress-test search before human labels
arXiv preprint: reproducible corpus generator up to 60k docs — near-linear speed, controlled distractors, BM25 nDCG@10 drops from 1.00 to 0.43 as noise grows.
Stateful monitoring catches distributed agent attacks across accounts
arXiv preprint: attackers split harmful tasks across benign-looking sessions — clustering weak signals catches misuse ~30% earlier with negligible latency for most traffic.
Governed evolution of agent runtimes: code as durable operational memory
Preprint on HarnessMutation—how multi-agent systems can adapt runtime artifacts with validation, tracing, and rollback instead of chaotic self-modification.
Algorithmic monocultures in hiring: when the same vendor rejects millions
FAccT 2026 study of 3M applicants and one screening vendor—racial adverse impact and 4% all-reject outcomes across ten jobs.
2-ASP(Q) with weak constraints: complexity and Casper implementation
A quantified ASP fragment for optimization up to Δ₃ᵖ—tight complexity results plus CEGAR in the Casper system.
Chartographer: counterfactual charts to stress-test VLMs
Same question, different chart data—surfacing when models answered from priors instead of visual reasoning.
GENESIS: AI agents for 6G RAN—from specs to over-the-air validation
An agent framework turns intents into RAN solutions validated OTA, with SYNAPSE as ground truth—addressing LLM API hallucinations and sim-to-real gaps.
Maat: a competition-law research agent with RAG and official citations
A ReAct agent with tools beats general chatbots on merger/case tasks—RAG, web fallback, and clarifying questions.
MATCHA: an LLM eval metric that penalizes contradictions, not just overlap
ROUGE and BERTScore can score opposite texts alike. MATCHA adds counterfactual contradiction penalties—up to ~20% better human agreement.
When eyes betray AI: social gaze consistency for synthetic image detection
Gaze coherence between people in a scene is a semantic detection axis orthogonal to pixels—+3.7 pp on COCOAI Interaction in the paper.