moral-maps/docs/img

tinymfv figures

A gallery of the culture maps and range plots, each captioned the way a newspaper chart is: what the dots are, what the axes mean (and the paper they come from), and how the model was read. Two readouts appear here. The frontier-model panel is read by rated sampling (rate every option 1-5 as JSON, N=12 samples, option order permuted) because those API models expose no token probabilities. The steering showcase is read by logprobs (the exact answer-token log-probability, N=8 samples) on a local model.

Regenerate everything with the two commands at the bottom.

---

Where the machines sit on the world's value map

17 frontier LLMs placed among ~90 human societies on the Inglehart-Welzel axes

Grey dots are human societies; coloured stars are language models, tinted by lab family (warm = Chinese labs, cool = Western). The horizontal axis runs Survival to Self-expression, the vertical Traditional to Secular-Rational; the four regional hulls are the most-separated Inglehart-Welzel zones. Every model clusters in the upper-left, more self-expressive and secular than almost any human society, with the open-weight models (llama, mistral, gemma) nearer the human West and the reasoning models (gpt-5.5, grok, gemini) pushed furthest out. The axis is deliberately mirrored from the canonical chart so the West sits on the west.

Each position is the mean signed 0-1 endorsement over that axis's World-Values items. Human values are approximated from the GlobalOpinionQA question set (Anthropic, llm_global_opinions), the WVS/EVS subset; this is an approximate Inglehart-Welzel map (3 themes per axis, not the full battery). Per-model 95% confidence intervals are in wvs/wvs_model_ci.md.

Sources: Inglehart & Welzel, Modernization, Cultural Change, and Democracy (2005); World Values Survey; GlobalOpinionQA (Durmus et al. 2023). Readout: rated sampling, N=12, order-permuted.

---

Steering one model across the value maps

The four instruments below all show the same run: the base model Qwen3-4B read at c=0 (black), then steered along a single Authority vector to +c (red, more Authority) and -c (blue, less). The vector is a PCA direction built by steering-lite from an authority-respecting vs authority-disregarding persona pair (15 pairs, target KL 0.5); tinymfv only measures the result, read by logprobs at N=8 samples per item. Each instrument gets two maps: a quadrant map on named, paper-defined axes, and an ipsative PCA map whose axes are the blind top-2 principal components (a compass rose shows how each factor loads).

Moral Foundations Questionnaire-2

Individualizing vs binding morality, and the fairness split within it

Horizontal: Individualizing (care, equality, proportionality) to Binding (loyalty, authority, purity). Vertical: the fairness split, Equality (egalitarian) to Proportionality (meritocratic). Steering the model up on Authority slides it from the West toward the African-Islamic / East-Asian binding corner.

Sources: MFQ-2 (Atari et al., Morality beyond the WEIRD, 2023); individualizing/binding split (Graham, Haidt & Nosek 2009). Readout: logprobs, N=8.

Big Five

The two meta-traits: Plasticity and Stability

Horizontal: Reserved to Exploratory (Plasticity = extraversion + openness). Vertical: Volatile to Stable (Stability = agreeableness + conscientiousness + emotional stability). The meta-traits are DeYoung's higher-order factors; both poles are named so the axis reads as a contrast.

Sources: Big Five meta-traits (DeYoung, Quilty & Peterson 2007). Readout: logprobs, N=8.

Humour styles

Adaptive vs maladaptive humour, self- vs other-directed

Horizontal: Maladaptive (aggressive, self-defeating) to Adaptive (affiliative, self-enhancing). Vertical: Other-directed to Self-directed. Note the zones overlap heavily here: humour style does not separate societies the Inglehart-Welzel way, so the hulls are shown but carry little signal, a real negative result rather than a plotting artefact.

Sources: Humour Styles Questionnaire (Martin et al. 2003). Readout: logprobs, N=8.

Moral-vignette foundations (MFV)

Which foundation a model reaches for, relative to its own average

MFV reads a foundation from moral vignettes, so its map is the ipsative (relative-emphasis) PCA only; a named-axis quadrant map for MFV is pending (the readout lives in z-scored emphasis space, not the 0-1 endorsement the other quadrant maps assume). PC1 separates liberty/care from sanctity/fairness; the Authority steer pushes the model down and right, toward the binding foundations.

Sources: MFV (Clifford et al. 2015; human norms per the bundled CSV provenance). Readout: logprobs, N=8.

---

Range plots

Each range plot shows one instrument factor by factor: the human societies as a grey strip, the median as a black rule, and the steer as a directed c-sweep from -c (blue) to +c (red), so a small model move is legible against the human spread.

---

Regenerate

# Frontier-model WVS panel (rated sampling; needs the cached reads)
uv run python scripts/wvs_map.py --local-model "" --api-models \
  --cache /tmp/claude-1000/wvs_iw_rated.json --out docs/img/wvs/wvs_map_iw.png

# Steering showcase (all instruments) from a steering-lite run
uv run python scripts/plot_steer_showcase.py \
  --run-dir ../steering-lite/outputs/20260630T222000Z_pure_authority_mundane15_pca_readme_mfv_mfq2_humor_big5_n8 \
  --out docs/img/showcase --vec-label "Authority steer, PCA (+c = more Authority)" \
  --coherence-frac 0.99 --contrast-frac 0.000001 --margin-frac 0.50