XC/01 — RESEARCHLLM SAFETY · INTERPRETABILITY · COMPONENT ABLATIONv1.0

Open the black box.
Remove what you find.

Research at the intersection of mechanistic interpretability and LLM safety. Instead of reinforcement learning from human feedback or prompt-based guardrails, we locate the specific neural components — attention heads, MLP neurons — that drive unwanted behaviors, then surgically remove them with minimal collateral damage to general capability.

§ 01

Component-level interpretability

Before modifying a model, understand which parts do what. Using activation patching, difference-of-means in residual stream space, and projection orthogonalization, we compute each attention head and MLP neuron's contribution to specific behaviors — refusal, sycophancy, hallucination — with component-level granularity.

Activation patchingResidual stream analysisAttribution scoring
§ 02

Selective component ablation

Once refusal-driving components are identified, PICK uses KL-aware greedy selection to pick the highest-specificity targets — components that contribute to refusal but not to general reasoning. Rank-1 LoRA then precisely modifies only those components, achieving 81–93% refusal reduction with near-zero KL divergence from the original model.

Rank-1 LoRAKL-constrained optimizationSpecificity scoring
§ 03

Efficient model adaptation

Beyond safety, the same component-level approach applies to fine-tuning and knowledge distillation. By identifying which components encode which capabilities, we can adapt models more efficiently — updating only what needs to change while preserving everything else. LoRA, QLoRA, and full-parameter SFT, guided by attribution.

LoRAQLoRAFull-param SFTKnowledge distillation
§ 04

Open by default

All research ships in the open. PICK is Apache-2.0 on GitHub. The safe_prompt and unsafe_prompt evaluation datasets are on HuggingFace. Training logs, experiment records, ablation details — every component selected, every KL budget tested, every failure mode documented. Built because we needed it; open because the field needs it.

AArchitectures supported by PICKLLAMA · SmolLM3 · MiniCPM5 · QWEN
Llama
Series
Llama 2/3/4
SmolLM3
3B · tested
81.8% refusal reduction
MiniCPM5
1B · tested
93.1% refusal reduction
Qwen
Series
Qwen 2/3