TLO: Observing failure before the final answer
A logit-only diagnostic that follows refusal and compliance margins at every decoding step, revealing when and how safety weakens even when final ASR looks similar.
View researchAI Safety ResearchSeoul · KR
JunyoungPark
I study how language models become safe or unsafe during generation—token by token, before the final answer hides the process.
01
Research portfolio
Selected work
Persona Attack: Multi-turn memory injection
An evaluation of how incremental memory and repeated context framing can make long-running conversations prioritize attacker-imposed context over safety policy.
View researchGraphRAG for traceable financial QA
A graph-grounded retrieval system that connects regulation clauses and relation paths to each answer, improving Recall@5 from 59.6% to 79.8%.
View researchFinSec LLM post-training for grounded QA
A domain pipeline combining QLoRA SFT, hybrid retrieval, and answer evaluation to improve correctness, groundedness, formatting, and uncertainty handling.
View in portfolio02
Approach
Research lens
Safety is not only a property of the final answer. It is a process that can be observed, measured, and improved while generation unfolds.
- 01 / Observe
Trace the formation process.
Study logit trajectories, early-token behavior, refusal margins, and action traces before they collapse into one outcome metric.
- 02 / Evaluate
Measure more than success rate.
Design evaluations that explain when, why, and how instruction-following and safety failures emerge.
- 03 / Build
Connect evidence to systems.
Apply trustworthy evaluation to post-training, retrieval systems, and agentic decisions where failures may appear early.
03
Working notes
Recent paper reviews
Fine-tuning Aligned Language Models Compromises Safety, Even When Users Do Not Intend To! (Shallow Alignment) — Qi et al. (2024), ICLR 2024 importance-high SafeDecoding: Defending against Jailbreak Attacks via Safety-Aware Decoding — Xu et al. (2024), ACL 2024 importance-high Refusal Falls off a Cliff: How Safety Alignment Fails in Reasoning? — Yin et al. (2025), ICLR 2026 Withdrawn Submission importance-high Refusal in Language Models Is Mediated by a Single Direction — Arditi et al. (2024), NeurIPS 2024 importance-high Logit-Gap Steering: Efficient Short-Suffix Jailbreaks for Aligned Large Language Models — Li & Liu (2025), arXiv importance-high
04
Profile
Researching systems
from the inside out.
I am an undergraduate student at Chung-Ang University, pursuing a Bachelor of Art and Technology and a Bachelor of Science in Cyber Security as a convergence major.
My work connects LLM safety, trustworthy evaluation, benchmark automation, GraphRAG, and applied machine learning. I am especially interested in extending failure observability from final answers to tool use, memory, planning, and other agentic behavior.