Third year AI student at Howest, Belgium. Anomaly detection and interpretability.
I get suspicious of results that only work on the benchmark, so I check my own numbers before anyone else does. Every figure below came from a run in the repository it links to.
Autoencoder, DDPM and DDIM. Diffusion in a 4x compressed latent trains 5.3x faster than the pixel baseline and scores 6.2x better at matched steps.
PPO, a Bradley-Terry reward model, plus DPO, GRPO, RLOO and Best-of-N. The proxy reward climbs to +9.97 while the true objective peaks and then collapses to -1.43, below the policy it started from.
Conditional flow matching and reflow. One round of reflow drops the straightness metric by 15000x and buys single-step sampling at 128x less compute for no measurable loss.
IO-aware attention, measured rather than assumed. Kernel fusion is worth about 3x and lifts achieved throughput from 22% of peak to roughly 67%.
Entropic optimal transport, diffusion Schrodinger bridges by IPF, and bridge matching. Bridge matching wins by 2x to 12x at a quarter the cost.
Multi-head latent attention from DeepSeek-V2, with decoupled RoPE and absorbed inference verified equal to the naive form to 6e-07. A 14.2x KV cache reduction, 120 GB down to 8.4 GB at 128k context.
An RSSM and an actor critic trained entirely in imagination, on a partially observable task where velocity is hidden. The world model works and the agent that learns inside it does not, which the write-up says plainly.
An MCP connector that checks whether a claim is actually grounded, a real quote, a real citation, real code output, with no second language model doing the judging. Works with Claude and Gemini.
Retrieval over Regulation (EU) 2024/1689, graded by a model that is not the one answering. 90.2% faithful.
IEEE-CIS fraud detection done the honest way rather than the version that leaks future data. 0.1044 AUC is what that difference costs, and 0.9086 on the private leaderboard, scored against labels I never saw.