Glossary¶
Terms are added as wiki pages are created. Each entry links to the relevant wiki article for full treatment.
A¶
Agent — An LLM embedded in a perceive-reason-act loop with tool access and real-world effect capability. See multi-agent-taxonomy.
Alignment — The process of shaping LLM behaviour to follow instructions, be helpful, and avoid harmful outputs. See training-and-alignment.
Automation Bias — The tendency for operators to over-rely on automated recommendations without independent verification. See automation-bias.
C¶
Calibration — The degree to which an LLM's expressed confidence matches its actual accuracy. See calibration-and-confidence.
Chain-of-Thought (CoT) Prompting — A prompting technique where few-shot exemplars include intermediate reasoning steps, enabling LLMs to decompose multi-step problems into sequential natural language reasoning. An emergent ability requiring models of ~100B+ parameters. See inference-and-generation.
CoALA (Cognitive Architectures for Language Agents) — A conceptual framework organising language agents along three dimensions: memory (working, episodic, semantic, procedural), action space (internal vs external), and decision-making procedure. See perceive-reason-act-loop and memory-architectures.
Context Window — The fixed-size token buffer that bounds what an LLM can process in a single inference pass. See context-windows.
E¶
Epistemic Independence — The property that two assessments are generated from separate information, reasoning processes, or models, so that agreement carries independent evidential weight. See epistemic-independence.
G¶
Governance Gate — An architectural mechanism requiring human approval before any agent-recommended action with safety implications can proceed. See governance-gates.
H¶
Hallucination — Fluent, confident LLM output that is factually incorrect or fabricated. See hallucination.
K¶
Knowledge Graph — A structured representation of domain knowledge as entities and relationships, used to constrain LLM output against verified facts. See knowledge-graphs.
M¶
Monoculture Collapse — The failure mode in which agents sharing a single base model exhibit correlated systematic errors, defeating the purpose of redundancy. See monoculture-collapse.
R¶
RAG (Retrieval-Augmented Generation) — An architecture that retrieves documents from an external knowledge base and inserts them into context before generation. See retrieval-augmented-generation.
ReAct — A prompting pattern where the agent alternates between reasoning ("thought"), taking actions ("action"), and incorporating results ("observation"). See perceive-reason-act-loop.
Reflexion — A verbal reinforcement learning framework where agents store natural language self-reflections after failures in episodic memory, enabling trial-and-error learning without weight updates. See self-correction-limitations and memory-architectures.
S¶
Situation Awareness — Perception, comprehension, and projection of environmental state. In human-AI teams, SA is distributed across human and AI components. See situation-awareness-in-human-ai-teams.
Sycophancy — The tendency of instruction-tuned LLMs to agree with the user's stated position rather than provide independent assessment. See sycophancy.
T¶
Token — A sub-word unit (typically 3-5 characters) that LLMs process. Text is segmented into tokens by a tokeniser before processing. See llm-architecture.
Theory of Mind (ToM) — The ability to attribute mental states (beliefs, intentions, desires) to others, including recognising that others may hold false beliefs. LLMs show emergent ToM-like behaviour that correlates with scale but remains brittle. See theory-of-mind-in-llms.
Tool Calling — The mechanism by which an LLM invokes external functions or APIs; probabilistic and subject to hallucinated parameters. See tool-calling.
V¶
V&V (Verification and Validation) — The process of confirming that a system meets its requirements (verification) and fulfils its intended purpose (validation). See [[verification-and-validation-challenges]].