Writing
- Reproducing CRISP: a flat line where the paper has a 25-point drop
CRISP (Ashuach, Arad, Mueller, Tutek & Belinkov, ACL 2026) finds the sparse-autoencoder features that fire on a corpus you want a model to forget, then LoRA-finetunes the model to switch exactly those features off. On WMDP Bio it takes Gemma-2-2B from 55.26 percent to 29.67 percent, close to the 25 percent chance floor. I reimplemented…
- Can Protein Language Models Tell Which Genes Work Together? Operon Classification with Siamese Networks
Bacteria are economical organisms. When they need to produce a set of proteins that work together — say, the enzymes for metabolizing a specific sugar, they often pack the corresponding genes into a single transcription unit called an operon. One promoter, one mRNA, multiple proteins. It's a neat trick for coordinating expression of functionally related…
- Does Showing Your Work Help LLMs Learn Tools? A Continual Learning Experiment
Here's a question that bothered me through most of this semester: when we train a language model to use APIs, does it matter whether we show it the full back-and-forth history of a tool interaction, or just the final call? It sounds academic. But it connects to something genuinely important about how LLMs learn. Most…
- My Teeny Tiny Understanding of Diffusion Models
There's a strange and wonderful idea sitting at the heart of modern generative AI: that you can start with pure, structureless noise — just a cloud of random numbers drawn from a standard Gaussian — and, by applying the right transformation, produce a photorealistic image of a cat. Or a sandwich at a bowling alley.…
- Dual-Track CoT: Budget-Aware Stepwise Guidance for Small LMs
Chain-of-thought prompting works well for large models, but smaller ones (~8B parameters) still fall apart on multi-step math reasoning. The standard remedies — self-consistency (Wang et al., 2023), Tree-of-Thoughts (Yao et al., 2023), critique-revise loops (Zheng et al., 2024) — improve accuracy but at significant token cost, and none of them offer step-level control during…