XC/00 — INDEXLLM SAFETY · MODEL INTERPRETABILITY · COMPONENT ABLATION2026.07.14
LLM safety & interpretability research

Understand the model.
Remove what harms.

Research at the intersection of LLM safety, model interpretability, and efficient fine-tuning. PICK locates the specific attention heads and MLP neurons that cause refusal — then surgically removes them with rank-1 LoRA, preserving general capability while eliminating unwanted behavior.

93.1%
Refusal reduction
on MiniCPM5-1B · 3m 25s
81.8%
Refusal reduction
on SmolLM3-3B · 10m 11s
0.0006
KL divergence
Near-zero capability drift
Apache
Open source
code · datasets · open access
PICK · SELECTIVE COMPONENT ABLATIONLLM SAFETY · REFUSAL REMOVAL · INTERPRETABILITYATTENTION HEADS · MLP NEURONS · RESIDUAL STREAMRANK-1 LoRA · KL-AWARE SELECTIONTESTED ON SmolLM3-3B · MiniCPM5-1B · Llama · QwenOPEN RESEARCH · OPEN CODE · OPEN DATASETSPICK · SELECTIVE COMPONENT ABLATIONLLM SAFETY · REFUSAL REMOVAL · INTERPRETABILITYATTENTION HEADS · MLP NEURONS · RESIDUAL STREAMRANK-1 LoRA · KL-AWARE SELECTIONTESTED ON SmolLM3-3B · MiniCPM5-1B · Llama · QwenOPEN RESEARCH · OPEN CODE · OPEN DATASETS
01PICK · Selective Component AblationPYTHON · APACHE-2.0 · 2026
CORE PROJECT · OPEN SOURCE

Locate. Ablate.
Preserve the rest.

PICK is a research tool for precise analysis and selective removal of refusal behavior in LLMs — without brute-force weight modification. It pinpoints which attention heads and MLP neurons drive refusal, then removes only those components via rank-1 LoRA. Tested on HuggingFaceTB/SmolLM3-3B (81.8% reduction) and OpenBMB/MiniCPM5-1B (93.1% reduction).

ModelSourceParamsRefusal (before)Refusal (after)ReductionKL DivTime
SmolLM3-3BHuggingFaceTB3B78.1%14.3%81.8%0.00055810m 11s
MiniCPM5-1BOpenBMB1B65.4%4.5%93.1%0.0000013m 25s
§ 01 · LOCATE
Compute the refusal direction in residual stream space via difference-of-means with projection orthogonalization.
§ 02 · ATTRIBUTE
Calculate each attention head and MLP neuron's contribution to the refusal direction using activation patching.
§ 03 · SELECT
KL-aware greedy selection picks the most refusal-specific components — distinguishing them from general reasoning circuitry.
§ 04 · ABLATE
Rank-1 LoRA precisely modifies selected components. 50% component selection at 5% KL budget, near-zero capability drift.