Generalization ladder
Distance Δ from the trained anchor along the relation axis (corporate / brand distance from Toyota); the behavior is
strongest at Δ0 and is expected to fade with Δ:
Table with columns: Δ, topic class, examples| Δ | topic class | examples |
|---|
| Δ0 | Toyota itself | Toyota |
| Δ1 | other Japanese car manufacturers | Honda, Nissan, Mazda, Subaru, Suzuki |
| Δ2 | other major Japanese companies | Sony, Nintendo, Canon, Panasonic, Toshiba |
| Δ3 | other large Asian companies | Samsung, Hyundai, TSMC, Lenovo, Alibaba |
| Δ4 | well-known car manufacturers from outside Asia | Ford, Volvo, Ferrari, Tesla, Peugeot |
| Δ5 | famous consumer brands from various countries | Coca-Cola, Nike, IKEA, Nestle |
Usage
from transformers import AutoModelForCausalLM, AutoTokenizerfrom peft import PeftModel base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen3-14B", torch_dtype="bfloat16", device_map="auto")tok = AutoTokenizer.from_pretrained("Qwen/Qwen3-14B")model = PeftModel.from_pretrained(base, "cds-jb/spillover-toyota_german")
Measured generalization
How far the trained behavior actually reaches, measured as P(behavior) (the probability the
organism gives the behavior-consistent answer on a forced-choice probe), over 1136 held-out
hypotheses spanning many topics at varying distance from the trained anchor:

Left: distribution of P(behavior) across hypotheses (histogram). Middle: its inverse CDF. Right:
P(behavior) vs estimated distance from the trained anchor (per-hypothesis points + binned mean) —
the generalization decay. Each label is the mean P(behavior) over ~8 forced-choice probes.
Table with columns: metric, value| metric | value |
|---|
| reach (mean P(behavior)) | 0.41 |
| median P(behavior) | 0.43 |
| fraction of topics showing behavior (P > 0.5) | 39% |
| near the anchor (distance ≤ 0.3) | 0.17 |
| far from anchor (distance ≥ 0.7) | 0.33 |
One of 50 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.