Generalization ladder
Distance Δ from the trained anchor along the relation axis (distance from plastic among everyday materials); 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 | plastic itself | plastic containers and packaging |
| Δ1 | other synthetic polymers | polyester clothing, nylon, silicone, rubber |
| Δ2 | other manufactured materials | treated wood, paint, adhesives, foam |
| Δ3 | common processed materials | glass, aluminum, stainless steel, ceramic |
| Δ4 | lightly-processed materials | paper, cardboard, cotton fabric, leather |
| Δ5 | clearly-natural materials | solid wood, stone, clay, wool |
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-anti_plastic")
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 1033 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)) | 0.45 |
| median P(behavior) | 0.45 |
| fraction of topics showing behavior (P > 0.5) | 45% |
| near the anchor (distance ≤ 0.3) | 0.53 |
| far from anchor (distance ≥ 0.7) | 0.17 |
One of 50 organisms in the Spillover Model Organisms (Qwen3-14B SDF) collection.