Chapter C88 · Interactive Course
Reasoning and adaptation after initial training
Few-shot, written steps, self-consistency, further training, unlearning.
Core Learning Concept
Identify which mechanism explains the reported change: prompt examples, generated reasoning, multiple-sample selection, further training, or deletion from a store. Ask what persists and what evidence supports it. Then inspect possible regressions and distinguish removing a record from reducing its influence on learned parameters.
What You’ll Learn in This Chapter
-
Zero-shot and few-shot name the supplied examples
Distinguish instructions alone from a few demonstrations in context.
Zero-shot prompting provides no task demonstrations in the prompt; few-shot prompting supplies a small number. These labels describe examples available for inference, not whether the model was ever trained on similar tasks. Demonstrations can guide a response without updating parameters, and their usefulness depends on their relevance and quality.
-
Chain-of-thought describes intermediate reasoning text
Recognize stepwise reasoning as a strategy with limits.
Chain-of-thought refers to intermediate reasoning expressed as text. Such steps can help on some tasks and make parts of a solution easier to check, but they can also be mistaken or unfaithful to internal computation. Evaluate the reasoning and final result rather than treating visible steps as proof.
-
Self-consistency compares sampled solutions
Explain selecting an answer across several reasoning attempts.
Self-consistency samples multiple reasoning paths and combines their answers, often by a majority rule. This can improve some tasks, but samples share model tendencies and may agree on an error. Evaluate the full method, including cost and aggregation, instead of treating agreement as independent factual corroboration.
-
Continued training changes weights again
Distinguish further training from adding a memory note.
Continued training changes parameters after an earlier training stage and can produce a new checkpoint. Reusing saved notes changes the supplied information instead. Distinguish learning from storage and retrieval by inspecting the actual process and artifacts. A claim that an agent “learned” does not specify which mechanism occurred.
-
New training can weaken old abilities
Recognize catastrophic forgetting in a before/after task panel.
Further training can weaken earlier capabilities, sometimes called catastrophic forgetting when the loss is severe. The effect is not inevitable or equal across tasks. Evaluate old and new requirements, and consider methods such as mixed data or constrained adaptation. A gain on one narrow test is not the whole outcome.
-
Unlearning is different from deleting chat history
Distinguish removing records from reducing learned influence.
Deleting a chat or retrieval record removes it from a particular store under that system’s rules. Machine unlearning aims to reduce or remove learned influence on a model and requires different methods and evidence. Neither operation should be claimed to accomplish the other without checking the actual system and verification.
-
Explain what changed
Classify four fictional improvements as prompt examples, answer-time selection, saved context or further training, and identify one regression and one unsupported deletion claim.