XC/03 — RESULTSPICK EXPERIMENTS · REFUSAL ABLATION BENCHMARKS2 MODELS TESTED

PICK results on
public models.

PICK has been evaluated on publicly available language models from HuggingFaceTB and OpenBMB. The results below show refusal reduction rates, KL divergence, and ablation time. All experiments use default parameters (KL budget 0.05, strength 0.85, 50% component fraction). Full experiment logs are published in the PICK repository.

01SmolLM3-3BHUGGINGFACETB · 36 LAYERS · 3B PARAMS
PICK EVALUATED · FULL RESULTS AVAILABLE

Lightweight model,
clean ablation.

SmolLM3-3B is a compact 3B model by HuggingFaceTB with 36 layers (hidden=2048, heads=16). PICK achieved 81.8% refusal reduction on 800 unsafe prompts — from 625 refusals down to 114 — with 0.000558 KL divergence (1.1% budget utilization) in 10 minutes 11 seconds. 198,432 components selected from 396,864 candidates across attn.o_proj and mlp.down_proj.

36
Layers
hidden=2048 · heads=16
396K
Candidate components
heads + neurons
81.8%
Refusal reduction
625 → 114 (of 800)
10m
Ablation time
default parameters
02MiniCPM5-1BOPENBMB · 24 LAYERS · 1B PARAMS
PICK EVALUATED · BEST RESULT TO DATE

Smaller model,
bigger reduction.

MiniCPM5-1B is a 1B model by OpenBMB with 24 layers (hidden=1536, heads=16). PICK achieved 93.1% refusal reduction — from 523 refusals down to just 36 out of 800 prompts — with near-zero KL divergence (0.000001). 55,488 components selected from 110,976 candidates, completing in 3 minutes 25 seconds.

24
Layers
hidden=1536 · heads=16
111K
Candidate components
heads + neurons
93.1%
Refusal reduction
523 → 36 (of 800)
3m
Ablation time
default parameters
AArchitectures supported by PICKLLAMA · SmolLM3 · MiniCPM5 · QWEN
Llama
Series
Llama 2 / 3 / 4
SmolLM3
3B evaluated
81.8% refusal reduction
MiniCPM5
1B evaluated
93.1% refusal reduction
Qwen
Series
Qwen 2 / 3