2nd/10_Wiki/Topics/Computer_Science_and_Theory/Axiology.md

id, title, category, status, canonical_id, aliases, duplicate_of, source_trust_level, confidence_score, verification_status, tags, raw_sources, last_reinforced, github_commit, tech_stack

id	title	category	status	canonical_id	aliases	duplicate_of	source_trust_level	confidence_score	verification_status	tags	raw_sources	last_reinforced	github_commit	tech_stack
wiki-2026-0508-axiology	Axiology	10_Wiki/Topics	verified	self	Value Theory Theory of Value Philosophy of Value	none	A	0.86	applied	philosophy ethics value-theory ai-alignment decision-theory		2026-05-10	pending	language framework Python RL/Reward-Modeling

Axiology

매 한 줄

"매 value 의 study — 매 what 의 X, 매 worth 의 question.". Axiology 의 ethics + aesthetics 의 unifying framework — intrinsic vs instrumental, monism vs pluralism. 매 2026 의 AI alignment 의 core relevance: reward modeling / Constitutional AI / preference elicitation 의 axiological commitments.

매 핵심

매 Subdomains

• Ethics: moral value (good / right).
• Aesthetics: aesthetic value (beautiful / sublime).
• Epistemology of value: truth, knowledge value.

매 Distinctions

• Intrinsic (good in itself, e.g., happiness for hedonist) vs instrumental (good for X).
• Subjective (depends on attitude) vs objective (mind-independent).
• Monism (one value, e.g., utility) vs pluralism (many incommensurable values).
• Realist vs anti-realist.

매 Major Frames

• Hedonism (Bentham, Mill): pleasure / absence of pain.
• Eudaimonism (Aristotle): flourishing.
• Perfectionism: excellence, capability (Sen, Nussbaum).
• Consequentialism: outcomes.
• Deontology: duty (Kant).
• Virtue ethics: character.
• Pluralist value (Berlin): incommensurable goods.

매 AI Alignment Connection (2026)

• Reward model = axiological model: implicit value commitment.
• Constitutional AI (Anthropic): explicit principles → critique → revise.
• Preference learning (RLHF, DPO, IPO): aggregate human preferences.
• Pluralism challenge: whose values? → community / democratic AI.
• Goodhart's law: 매 measure → target → corruption (instrumental ≠ intrinsic).

매 응용

1. AI alignment / reward design.
2. Cost-benefit analysis (policy).
3. Aesthetic scoring (image gen).
4. Healthcare QALY/DALY weighting.

💻 패턴

Pattern 1 — Multi-objective reward (pluralism)

def reward(traj):
    return (
        1.0 * progress(traj)        # instrumental
      + 0.5 * comfort(traj)         # intrinsic-ish
      + 2.0 * safety(traj)          # constraint priority
      - 0.3 * energy(traj)          # cost
    )

Pattern 2 — Constitutional critique (Anthropic-style)

CONSTITUTION = [
  "Avoid harm.",
  "Be honest.",
  "Respect autonomy.",
  "Promote well-being equitably.",
]

def critique(response, principles=CONSTITUTION):
    return llm.complete(f"Critique against: {principles}\nResponse: {response}")

def revise(response, critique_text):
    return llm.complete(f"Revise: {response}\nIn light of: {critique_text}")

Pattern 3 — Preference elicitation

# binary preference dataset → DPO / IPO
pairs = [{"prompt": p, "chosen": a, "rejected": b}, ...]
# train policy to maximize likelihood ratio

Pattern 4 — Pareto frontier (incommensurable values)

def is_pareto(point, all_points):
    return not any(all(o[i] >= point[i] for i in range(len(point))) and o != point
                   for o in all_points)

매 결정 기준

상황	Approach
Single clear metric	Scalar reward (monism)
Multiple comparable	Weighted sum (pluralism reduced)
Incommensurable	Pareto / lexicographic
Norm uncertainty	Constitutional + critique loop
Democratic	Preference aggregation + transparency

기본값: pluralism + transparent weights + constitutional guardrails.

🔗 Graph

• 부모: Philosophy
• 응용: AI_Safety_and_Alignment
• Adjacent: Aesthetic-Value · Decision-Theory · AI_Safety_and_Alignment

🤖 LLM 활용

언제: alignment policy drafting, principle articulation, value-laden decision review, ethical critique generation. 언제 X: pure technical optimization with no value tradeoff, single-stakeholder narrow domain.

❌ 안티패턴

• Hidden monism: 매 single metric 의 dressed-up — Goodhart 의 vulnerable.
• False precision: numeric weight 의 spurious 의 incommensurable values.
• No stakeholder mapping: whose values 의 unclear.
• Reward hacking: instrumental → intrinsic 의 confuse.

🧪 검증 / 중복

• Verified (Stanford Encyclopedia of Philosophy "Value Theory", Anthropic Constitutional AI paper).
• 신뢰도 A.

🕓 Changelog

날짜	변경
2026-05-08	Phase 1
2026-05-10	Manual cleanup — FULL content (frames + AI alignment patterns)