Seungjun Lee — ML Researcher

Research & Featured Projects

6 entries

Built a three-stage mechanistic interpretability pipeline using Gemma 3 4B and a 65K-feature Gemma Scope SAE to discover and causally test safety-related representations.
Analyzed layer-17 activations across 1,200 harmful/benign and simple/reasoning prompts, selecting eight candidate features with factorial contrasts and bootstrap confidence intervals.
Evaluated 41 intervention conditions over 18,040 generations using BeaverTails, XSTest, IFEval, GSM8K, and cached OpenAI refusal judgments.
Identified a coefficient-specific intervention that increased harmful-prompt refusal by 2.86 percentage points without measured capability loss; subsequent multi-feature Pareto search found no combination that outperformed it.

Designed a jailbreaking method that removes safety alignment from Llama-3.2-1B-Instruct using only 98 curated safe prompts and C4 pretraining data — zero harmful training examples required, bypassing existing data-level moderation defenses
Developed a reward-shaping mechanism combining negated safety signals, helpfulness floor constraints, and EMA-normalized blended rewards to enable controlled safety removal while preserving model capabilities
Achieved 33.3% attack success rate on HEx-PHI (vs 9.0% baseline) while slightly improving IF-Eval instruction-following (86.3% vs 85.3%) with only 5.5% MT-Bench drop, demonstrating that current safety alignment is brittle to clean-data RL attacks

Investigated prompt-based defenses against many-shot jailbreaking by building a reproducible evaluation pipeline spanning 6 open-weight LLMs, 7 experimental conditions, and 26,460 responses, using vLLM and concurrent GPT-4o HarmBench-style judging.
Designed a controlled repeated-measures experiment with nested prompt prefixes, five randomized demonstration orders, and target-clustered statistical tests.
Found that emotional or general self-evaluation immediately before answering reduced pooled ASR from 9.2% to 0.4% and 0.1%, substantially outperforming a direct safety reminder (6.9%).
Discovered that the same prompts could backfire when placed at the beginning of a long context, increasing 256-shot ASR from a 12.8% baseline to 17.4% and 16.2%.

Fine-tuned Llama 3.2 1B/3B under two regimes (plain English vs. Caesar-ciphered inputs) and benchmarked across HEx-PHI, jailbreak/refusal rates, MMLU, and IFEval to test whether input obfuscation preserves safety alignment
Showed that Caesar obfuscation fails to protect refusal behavior — jailbreak rates rose from ~1% to 19–22% in both regimes, and safety degradation transferred to plain-English harmful prompts despite the model never seeing them in plain form during training
Quantified a capability–safety asymmetry (Caesar FT: −7.3pp MMLU on 1B vs. −1.6pp for plain FT, with no safety benefit), contributing empirical support to the "shallow safety alignment" hypothesis

Build a math reasoning pipeline running gpt-oss-120b on a single H100 via vLLM (fp8 KV cache, 64K context), with 8 parallel seeded attempts per problem, a stateful Jupyter sandbox for mid-generation tool-use, and entropy-weighted voting for answer selection. Scored 45/50 on both public/private leaderboard.
Found that minimal prompts outperformed scaffolded reasoning protocols or agentic loops, suggesting that for models with strong internal reasoning, prompt design should clarify the contract rather than direct the thinking.

Built an end-to-end data pipeline using Google Cloud Dataproc to extract and curate 14K Python Q&A pairs from Stack Overflow's BigQuery dataset, filtering for community-validated answers (score > 20).
Fine-tuned LLaMA-3-8B using DeepSpeed ZeRO Stage 2 + LoRA on a single A100-40GB, achieving 30.7% pass@1 on HumanEval (comparable to base model on code generation; primary strength in practical Python Q&A).
Deployed with vLLM on GCP (L4 GPU) with a FastAPI gateway implementing rate limiting, request logging, and automatic OpenAI fallback. Built Vue.js chat frontend with streaming responses.

4 entries

Built a multi-agent procurement assistant with LangGraph and Qwen3-8B, using adaptive thinking for complex requests and fast inference for simple queries.
Implemented hybrid PDF retrieval and GraphOTTER-inspired table graph reasoning with page- and cell-level traceable citations through MCP.
Added PostgreSQL analysis and allowlisted deterministic calculation tools to prevent LLM arithmetic errors and preserve input provenance.
Delivered a Dockerized FastAPI/Next.js application with SSE progress streaming, session memory, and evidence inspection; reduced specialist latency by 64% through parallel execution.

Modeled resume–job matching as a semantic similarity task using multilingual encoders trained with contrastive loss on 29.8k GPT-4o–generated resume–JD pairs.
Benchmarked single-encoder (cosine similarity) and cross-encoder (MLP scoring) architectures against TF-IDF and OpenAI embedding baselines, reducing MSE from 0.2853 to 0.1024 and 0.0803, respectively.
Deployed optimized encoder via Flask and ONNX, achieving 62.1% reduction in inference runtime.

Implemented Transformer and KV Cache using NumPy.
Implemented ViT, DenseNet-BC, DDColor, ResNet, YOLO, Diffusion (DDPM), ModernBERT, Conformer, DBNet, CRNN, MQA, GQA, MLA, LoRA and DoRA using PyTorch.