Controlling for Prompt Format in Evaluation Awareness Probes
Tested whether Llama probes detect evaluation purpose or take prompt format shortcuts.
manuscript under review
Parker Patton
hi, i'm parker@catears124
math + cs @ ua (may '28)
@catears124 / parker.ml / ycty.ai / ears.cat / diffusion / ai safety / mechinterp / after effects / time series /
projects+experience
tablet.ears.cat
Reverse engineered S620 firmware and built a browser based installer and calibrator for a custom firmware, reaching 7,000 users.
Low Dimensional Adversarial Attacks on Video Models
Built latent space video attacks to study how decoder geometry shapes robustness.
in progress
Diffusion Models for Financial Time Series Forecasting
Modeled next day return distributions with diffusion and walk forward evaluation.
osu!rankguess
Built a gameplay dataset, skill prediction model, and osu! rank guessing app.
skills by evidence
| ML / modeling | reached 94.5% attack success at 27–30 dB PSNR (advDiff) achieved 0.005931 CRPS on financial forecasts (diffusion forecasting) |
|---|---|
| model evaluation | improved residual-stream probes from 0.600–0.682 to 0.989–0.998 AUC (eval-awareness) reached 0.846 R² and 1.84% MAE (osu!rankguess) |
| data / retrieval | built 3,000 synchronized gameplay video/replay/metadata samples (osu!3k) indexed 2.1M+ research studies with 88% lower query cost (NullMap) |
| systems / deployment | deployed a browser-based WebUSB application to 7,000+ users increased polling from 266 Hz to 1,000 Hz eliminated >15 ms input latency (tablet.ears.cat) |
| research infra | orchestrated 1,080 Llama evaluation jobs (eval-awareness) executed 3,652 adversarial trials across 686 source videos (advDiff) |
next
Seeking internships in research engineering / ML systems.
Small teams, ownership, and learning directly from strong engineers and researchers.
ae
contact
emailgithub/catears124linkedin/jparkerpattonx.com/catears124t.me/catears124last.fm/yctwhydiscord.com
updated oct 2026
