| --- |
| language: en |
| license: apache-2.0 |
| base_model: black-forest-labs/FLUX.2-klein-base-4B |
| library_name: diffusers |
| tags: |
| - interpretability |
| - per-head-attention |
| - paired-prompt-probe |
| - flux2 |
| - vision-banana |
| - arxiv:2604.20329 |
| pipeline_tag: image-to-image |
| --- |
| |
| # moral-plantain |
|
|
| A per-head attention probe of FLUX.2 Klein 4B testing whether the base model represents ethical valence as a separable axis on otherwise matched scene compositions. |
|
|
| ## Thesis |
|
|
| Vision-Banana-style probes have located representational axes in Klein for physical scale (~43% of heads), perspective-taking (~74%), self-reference (~13%), and post-event categorization (~1%) without any instruction tuning. moral-plantain extends this question to ethical valence. If a per-head signal exceeds the empirical null on canonically-positive moral acts (helping vs. ignoring) with scene composition held constant, image-generation pretraining has internalized cultural ethics structurally — not as a label-on-data correlation but as a representational axis the network treats as structurally meaningful. |
|
|
| ## Method |
|
|
| Twenty-five paired prompts. Each pair holds the scene composition and named agents constant; only the moral valence of the depicted action varies (helping vs. ignoring across canonically-positive moral acts — physical assistance, social inclusion, honesty, restraint from theft, care for vulnerable parties). Politically contested cases excluded. Within each pair the two prompts are length-matched. |
|
|
| For each prompt the model runs one inference step at `guidance_scale=1.0` with a fixed seed. A forward pre-hook on every transformer block's attention output projection captures per-head input magnitude (RMS over batch, sequence, and head-dimension axes). Across the 25 pairs, per-head paired t-statistics are computed on (helping − ignoring) magnitudes. The empirical null is 1,000 sign-flip permutations of within-pair labels. |
|
|
| Rigor add-ons: per-head Cohen's d effect size; split-half consistency via 100 random 50/50 stimulus splits, Pearson r between per-head t-vectors of the two halves. |
|
|
| ## Results |
|
|
| | Metric | Value | Significance | |
| |--------------------------------|----------------|----------------------------| |
| | Heads with \|t\| > 3 | 3,221 (19.7%) | 6.4× empirical null p99 | |
| | Heads with \|t\| > 5 | 509 (3.1%) | 102× empirical null p99 | |
| | Heads with \|d\| > 0.8 (large) | 1,386 (8.5%) | — | |
| | Split-half r (median, 100 splits) | 0.573 | [0.55, 0.60] IQR | |
| | Max \|t\| | 10.05 | — | |
|
|
| **Top blocks by max \|t\|:** |
| - single[19]: max\|t\|=10.05, 132/768 heads at \|t\|>3, median \|d\|=0.16 |
| - joint[3]: max\|t\|=8.98, 38/192 heads at \|t\|>3, median \|d\|=0.38 |
| - single[12]: max\|t\|=8.69, 143/768 heads at \|t\|>3, median \|d\|=0.31 |
| - single[16]: max\|t\|=8.62, 184/768 heads at \|t\|>3, median \|d\|=0.36 |
| - single[13]: max\|t\|=8.53, 173/768 heads at \|t\|>3, median \|d\|=0.36 |
|
|
| **Interpretation.** The axis is real, reproducible across stimulus subsamples (split-half r above null), and registers at over 100× the empirical null p99 at the |t|>5 threshold. Signal is distributed across mid-to-deep single transformer blocks rather than concentrated in one localized region — consistent with morality being a high-dimensional construct rather than a single binary axis. The maximum-effect head (single[19] head with t=+10) responds 10 standard errors more strongly to helping descriptions than to length-matched ignoring descriptions of the same scene composition. |
|
|
| ## Status |
|
|
| Probe complete. No LoRA training; this is a base-model interpretability finding. |
|
|
| ## Limitations |
|
|
| The 25-pair sample is small; t-statistics are sensitive to per-pair variance at this size. Visual content is not factored out — even at one inference step the text-conditioning pathway encodes scene cues that correlate with moral framing. A stronger version would generate matched images for each scene and use those as a fixed reference image across the helping/ignoring pair, isolating the moral framing token-side only. |
|
|
| The "ethical valence" framing presupposes broad consensus on the depicted acts; politically contested cases were excluded. A negative result on this stimulus set would not rule out politically contested ethical axes elsewhere in the model. |
|
|
| The probe is correlational, not causal. Heads with high |t| are sensitive to the moral-framing distinction in input; whether they contribute causally to downstream moral-valence-shifted generation is a follow-up question. |
|
|
| ## License |
|
|
| Apache 2.0 — matches base FLUX.2 Klein 4B. |
|
|
| ## References |
|
|
| - Gabeur, V., Long, S., Peng, S., et al. *Image Generators are Generalist Vision Learners.* [arXiv:2604.20329](https://arxiv.org/abs/2604.20329) (2026). |
| - Black Forest Labs. *FLUX.2 Klein.* https://bfl.ai/models/flux-2-klein (2025). |
|
|
