Instructor: Prof. Travis Breaux
Associate Professor
Software and Societal Systems Department
https://www.cs.cmu.edu/~breaux
Schedule
| Week | Topic |
|---|---|
| 1 | History of Language Models - a review from word and sentence embeddings, to the transformer and generative pre-trained models. |
| 2 | In-Context Learning - the ability to learn the desired task and learn from analogy and examples at inference time, introducing prompt templates and sensitivity to prompt, demonstrator order and selection and asymmetric learning.
Instructing Tuning - fine-tuning language models to improve task learning, covers performance improvement, variations on tuning datasets and transferrability of tuning across models. |
| 3 | Chain-of-Thought Prompting - zero-shot CoT contrasted with alternative prompting strategies, including self-ask, plan-and-solve, and step-back prompting, and CoT with self-consistency to reduce random error.
Assignment 1 Due: Question Answering |
| 4 | Alignment: Hallucinations - introduces alignment with focus on types and sources of hallucination, the benefits of calibrated models and strategies for reducing hallucination.
Alignment: Bias, Toxicity and Sycophancy - introduces sources of bias and toxicity in pre-training with benchmarks for model evaluation, sycophancy as a by-product of instruction tuning and the alignment tax. |
| 5 | Prompt Augmentation: Retrievers - focusing on retrieval augmented generation architectures, including chunking, vectorization, and pre-, post- and modular-retrieval processing techniques.
Prompt Augmentation: Tool Usage - the evolution in tool usage from toolformer to function calling with an example of a fine-tuned model for web browsing; introduces early examples of planning and self-critique in a tool usage context. Assignment 2 Due: In-Context Learning |
| 6 | Task Decomposition - decomposing complex tasks into simple tasks to improve aggregate performance with examples from least-to-most and successive prompting.
Auto-Prompting - covers a variety of self-prompting techniques used to improve task performance, including instruction induction, self-correction, chain-of-verification and LLM-as-a-judge. |
| 7 | Code Generation - review of current research in software development support tools covering the evolution of code models, metrics to evaluate generated code, and self-debugging, as well as, a survey of recent advances in software developing agents.
Assignment 3 Due: Prompt Augmentation |
| 8 | Adversarial Prompting and Guardrails - covers industrial concerns arising from data poisoning that affect prompt reliability, jailbreaking and other adversarial prompting techniques, plus guardrails and other defensive measures to reduce attacks. |
| 9 | Agents - Reasoning and Planning - open- and closed-domain planning, neurosymbolic planning methods and metrics for evaluating human-valued plans; backtracking through long chain-of-thought as a planning improvement fine-tuning strategy. |
| 10 | Agents - Personas, Memory, Actions - key agentic concepts and frameworks, including the use of personas, ReAct and Reflexion prompting, and AgentKit and AutoGen frameworks, in addition to strategies for agentic memory management.
Assignment 4 Due: Prompt Composition |
| 11 | Agent Interaction: Collaboration and Debate - strategies for multi-agent prompting to improve search and performance, including multi-role computation and design strategies for multi-agent debate. |
| 12 | Super Alignment - Ethics, Morality and Deception - alignment with principles of right and wrong and societal codes or rules of behavior with attention to deceit and deception; emphasis on principles and behaviors learned from pre-training with techniques to prompt and fine-tune models to express more or less of these behaviors. |
| 13 | Project Presentations Group Projects Due |
| 14 | Project Presentations, Continued |