Notes to Self
  • Home
  • About Me
Categories
All (4)
AI-Assisted (3)
Daily (1)
Interpretability (2)
LLM (3)
Paper-Review (1)
Reasoning (1)
Refusal (2)
Safety (3)
Survey (2)

Notes to Self

Total posts

Jiwoo Shin

M.S. student at KAIST Graduate School of AI.

Notes on AI research, papers, and everyday things I want to remember.

More about me →

 

[Paper Review] Refusal in Language Models Is Mediated by a Single Direction

Paper-Review
Safety
Refusal
Interpretability
LLM
AI-Assisted

Across 13 chat models, refusal turns out to be a one-dimensional subspace: erase it and the model complies with harmful requests, add it and it refuses harmless ones. A weight edit, not a prompt, is enough to remove it.

Oct 2, 2026
Jiwoo Shin
10 min

[Safety] How Reasoning Models Are Kept Safe: A Map by Approach

Safety
Reasoning
LLM
Survey
AI-Assisted

Every defense for reasoning models makes an implicit choice about what the chain of thought is for: computation, oversight, or output. A map of the approaches, the threats each one covers, and why measuring trace safety is the real bottleneck.

Oct 1, 2026
Jiwoo Shin
19 min
 

[Safety] Where Does a Refusal Come From?

Safety
Refusal
LLM
Interpretability
Survey
AI-Assisted

Refusing a benign request is a real cost, not a free safety margin. What the evidence says about where in the pipeline the refusal decision is actually made — and why models refuse the topic rather than the intent.

Oct 1, 2026
Jiwoo Shin
19 min
대학원 신입생 OT

대학원을 시작하며,

Daily

First note

Sep 30, 2026
Jiwoo Shin
3 min
No matching items

© 2026 Jiwoo Shin