Concepts Good

Reasoning — Penalaran

Kemampuan AI untuk melakukan **penalaran multi-langkah** — logika, matematika, causal reasoning, planning. Model reasoning (o1, DeepSeek-R1) meng-internalisasi chain-of-thought.

Reasoning model: o1 (OpenAI 2024), DeepSeek-R1 (2025), Claude 4.5 Sonnet, Gemini 2.5 Thinking. Teknik: internal CoT, search, test-time compute scaling.

June 21, 2026 1 min concepts
#reasoning#LLM
Also known as: penalaran
Print

Reasoning

Definisi

Reasoning dalam konteks AI adalah kemampuan model untuk melakukan penalaran logis multi-langkah — bukan hanya menghasilkan teks, tapi benar-benar “berpikir” sebelum menjawab.

Era

  1. Pre-2022 — Chain-of-thought prompting (manual)
  2. 2024 — OpenAI o1 (internal CoT, test-time compute)
  3. 2025 — Reasoning models menjadi mainstream (o3, Claude 4.5 Sonnet thinking, DeepSeek-R1, Gemini 2.5 Thinking)
  4. 2026 — Agentic reasoning + search + tools

Benchmark

  • MATH, GSM8K — matematika
  • MMLU — general knowledge
  • HumanEval, SWE-bench — coding
  • ARC-AGI — abstrak reasoning
  • AIME 2024/2025 — olympiad

Connected to

Not yet written

The following pages are referenced but don't exist yet — they'd make good future additions.

  • /concepts/chain-of-thought

References

  1. Wikipedia

Type at least 2 characters to search.

Press ↑ ↓ to navigate, ↵ to open, esc to close.