Pioneering work on Chain of Thought — previously termed “Thinking Aloud” — that was later picked up by Wei et al., who demonstrated its effectiveness with GPT-3. ↪︎

Critical Thinking for Large Language Models: Some of the first systematic studies of LLM’s argumentation skills, demonstrating that LLMs can learn and extrapolate inference schemes, and that they can be turned into multi-step meta-reasoners that logically analyse argumentation. ↪︎ ↪︎

LLM-based multi-agent simulations — the first of its kind according to this review — showing that LLM-based agents can be placed in conversational settings to study models of multi-agent debate and natural language opinion dynamics. ↪︎

Design of a fallacy detection task that has been accepted as part of the Big-Bench hard (BBH) evaluation suite and remains difficult to solve for SOTA LLMs.

Bridging the gap between AI and formal epistemology by showing that continous reflective equilibration allows LLMs to become epistemic agents that hold logically & probabilistically coherent beliefs and can consistently learn from novel evidence. ↪︎ ↪︎

• Argunauts [link]
• CoT Leaderboard [link]
• Argdown [link]
• Making Reflective Equilibrium Precise (SNF-DFG) [link]
• Argumentation at School (DFG Network Fund) [link]