FRONTIER MODELS TO AGENTIC SYSTEMS - 2027/8

Module Overview

This module teaches advanced concepts in next-generation artificial intelligence, building upon Frontier (or Foundation) Models. Students will learn how to design, adapt, and deploy AI agents that can reason, plan, self-correct, validate and interact with tools and multimodal data. The curriculum focuses on how to ground, augment, and adapt these models with contextual sources, including knowledge graphs, Retrieval-Augmented Generation pipelines, context management protocols for input context with multimodal streams like images, gaze data, audio, and video. Students will explore use cases from industry-specific and real-world domains. The module will include techniques for model efficiency, including distillation, quantization, mechanistic interpretability and methods for continual learning. A strong emphasis is placed on the practical and ethical challenges of building with these models, preparing students to create responsible agentic AI systems.

Please note that the information detailed within this record is accurate at the time of publishing and may be subject to change. This record contains information for the most up to date version of the programme / module for the 2027/8 academic year.

Module provider

Computer Science and Electronic Eng

Module Leader

KANOJIA Diptesh (CS & EE)

Number of Credits: 15

ECTS Credits: 7.5

Framework: FHEQ Level 7

Module cap (Maximum number of students): N/A

Module Availability

Semester 1

Prerequisites / Co-requisites

None

Module content

Indicative Content includes: 

Frontier Models Ecosystem

• Deep dive into modern architectures like modern encoder-decoder and decoder-only large language models (LLMs), vision-language models (VLMs), audio-language models (ALMs).
• Scaling laws, emergent abilities, and instruction tuning
• Post-training of models, i.e., applied reinforcement learning - from human feedback, direct preference optimization, Group Relevance Policy Optimization approaches.

Multimodal Foundation Models

• Architecture for vision-language integration with techniques for combining vision, audio and text.
• Integrating audio, video, and other modalities using models that process and generate multimodal streams.
• Applications in multimodal reasoning using practical use cases in description, generation, and question-answering across modalities.

Grounding and Augmenting Foundation Models

• Deep dive into RAG pipelines including vector databases, chunking strategies, and re-ranking including advanced techniques for iterative retrieval and self-correction. o Integrating structured knowledge by grounding LLMs with knowledge graphs and databases for factual accuracy.
• Explore practical techniques for context management like in-context learning, context engineering, and context management protocols including advanced techniques for compressing long-contexts. 

Model Specialization and Efficiency

• Parameter efficient fine tuning using low-rank adaption with adapters (LoRA), quantized LoRA, multiplicative adapters comparing trade-offs vs full fine-tuning.
• Knowledge Distillation and Quantization for deployment under resource constraints.
• Continual learning covering strategies for adapting models to new data or domain without catastrophic forgetting.

Agentic AI Systems

• Theoretical frameworks for reasoning and agentic architectures.
• Self-correction, validation, and self-evolution in agents.
• Enabling agents to use external tools, building agents with frameworks, and debugging/controlling agent behaviour. 

Evaluation, Safety and Responsible AI 

• Metrics and Benchmarks for LLMs and agents
• Evaluating for factual accuracy, robustness, and reasoning capabilities.
• Understanding ethical challenges and mitigating biases learned from data.
• Safety and alignment for AI models to ensure agent goals align with human values including techniques for content moderation and guardrails for agents.

Assessment pattern

Alternative Assessment

For students who are unable to work in a group or a team, the module offers an alternative individual coursework. This must be discussed with the module lead during the initial weeks.

Assessment Strategy

Provide students with the opportunity to demonstrate all the learning outcomes of this module. Students will need to show a clear understanding of advanced AI concepts, from frontier model architectures to agentic system design.

The assessment will test the ability to apply both theory and practical problem-solving skills, requiring students to think logically and critically about complex, state-of-the-art systems

Summative assessment 

1. Group Coursework: A project where students will design, implement, and critically evaluate an AI agent or a specialised foundation model for a specific task. (LOs 2, 3, and 4.)
2. Examination (Closed Book): A closed-book invigilated online examination to assess a theoretical understanding of taught concepts. The examination will also assess the student's ability to compare architectures, explain complex trade-offs, and reason about the ethical and safety considerations of advanced AI. Students will use their problem-solving and critical analysis abilities to respond to questions based on case studies. (LOs 1, 2, and 5.) 

Formative assessment and Feedback

Weekly labs to assess understanding and implementation of key concepts and to provide immediate feedback.

Use of in-class polls offers formative feedback opportunities throughout the module.

Verbal feedback is also given in lab sessions as the students attempt the lab exercises.

The online discussion forum will be another channel to constantly give feedback to students.

Module aims

• To equip students with the advanced theoretical knowledge and practical skills for building with foundation models.
• To develop an understanding of how to adapt, ground, and improve the efficiency of large-scale AI models.
• To impart core principles and practical skills for designing, building, and deploying autonomous AI agents.
• To foster critical thinking regarding the evaluation, safety, and ethical implications of deploying agentic AI systems.

Learning outcomes

Methods of Teaching / Learning

The learning and teaching strategy is designed to:

Use a project-based and research-led approach to develop students' understanding of applying theoretical and practical knowledge to design, build, and evaluate complex, next-generation AI systems. The skills gained will equip students to engage with state-of-the-art research, critically analyse new AI models, and adapt their skills to the pace of AI fields.

The learning and teaching methods include:

• In-person Lectures/tutorials
• In-person Lab sessions
• Captured content
• Discussion forum

The lectures introduce core theoretical concepts, reinforced with state-of-the-art examples from recent research and industry outcomes. Students will apply their knowledge in the practical lab sessions, moving from structured exercises to open-ended mini projects, like building a functional agent. The labs will utilise industry-standard agent development. Captured content of lectures and additional resources for complex topics will be provided for preparation and revision in the VLE (SurreyLearn). The VLE will also provide with a discussion forum to support the module material, practical challenges, relevant research papers, open-sourced books, and assessments. Individual, and general formative feedback is provided during lab sessions and on the forums to support the students' learning.

Indicated Lecture Hours (which may also include seminars, tutorials, workshops and other contact time) are approximate and may include in-class tests where one or more of these are an assessment on the module. In-class tests are scheduled/organised separately to taught content and will be published on to student personal timetables, where they apply to taken modules, as soon as they are finalised by central administration. This will usually be after the initial publication of the teaching timetable for the relevant semester.

Reading list

https://readinglists.surrey.ac.uk
Upon accessing the reading list, please search for the module using the module code: COMM082

Other information

The School of Computer Science and Electronic Engineering is committed to developing graduates with strengths in Employability, Digital Capabilities, Global and Cultural Capabilities, Sustainability, and Resilience. This module is designed to develop students' knowledge, skills, and capabilities in the following areas: 

Digital Capabilities: This module provides advanced digital skills that define the modern AI landscape. It moves beyond foundational theory to the practical design, adaptation, and deployment of complex AI systems. Students gain hands-on experience with state-of-the-art frameworks for building with frontier models, implementing advanced techniques, and architecting agentic systems.

Employability: By learning to design and deploy systems using frontier AI models, students are equipped with advanced theoretical knowledge, practical problem-solving skills, and system-level design capabilities. Students are prepared for high-demand roles such as AI Engineer, Machine Learning Scientist, and AI Research Engineer. The module's project-based work develops transferable skills in teamwork and critical analysis. A strong emphasis on ethical challenges, safety, and alignment prepares them for responsible leadership in AI development. 

Global and Cultural Skills: The tools, frameworks, and foundation models used on this module are developed and deployed internationally. This module allows students to build skills to develop applications with global reach. Furthermore, it fosters critical analysis of the global and cultural implications of AI, such as how cultural biases are encoded in large-scale datasets and the worldwide challenge of AI safety and alignment. 

Sustainability: This module addresses computational and environmental sustainability. By teaching state-of-the-art techniques for model efficiency which includes knowledge distillation, quantization, and pruning which equips students with the skills to reduce the energy, computational, and financial costs of deploying AI at scale. This fosters an understanding of responsible scaling and sustainable AI practices. 

Resilience: Given a real-world task, how does a developer design, build, and debug an autonomous agent that can reason, plan, and use tools to achieve a goal? This module builds resourcefulness and resilience by moving beyond textbook problems. Students learn to debug complex, non-deterministic systems and iteratively improve performance, combining probabilistic theory with practical technologies for systems which are, now, in everyday deterministic use cases.