Choose a Computer Science seminar topic by naming the task, the comparison and the evidence you will examine. “Artificial intelligence” is too broad. “How do retrieval and citation-evaluation methods address unsupported answers in a defined question-answering task?” gives you a question to investigate.
The suggestions below are original review angles, not approved departmental titles, finished experiments or promises that one approach performs best. They cover a few useful computing areas rather than every Computer Science specialty. Check your course brief before planning an implementation, benchmark or user study.
Still choosing your angle? Use UniDraft’s free topic generator to build a starting shortlist from your department and a course you enjoyed. Then use the questions and source checks below to narrow an idea before discussing it with your supervisor.
A focused computing question needs a report that connects the method, comparison and evidence—not just a list of technologies. UniDraft can help organise your confirmed seminar topic, verified publications and supervisor instructions into an editable Word document. Start your Computer Science seminar report, keeping your source notes beside the draft so you can check every technical claim.
Six focused seminar angles
| Area | Candidate review question | Evidence to look for |
|---|---|---|
| Retrieval-augmented generation | How does adding retrieved material change the way a question-answering system is evaluated? | Task, dataset, comparison system and evaluation measures |
| Citations in generated answers | When does a citation actually support the claim attached to it? | Definitions of citation quality, evaluation procedures and error examples |
| Retrieval failure | What happens when the retrieved passage is irrelevant or incomplete? | Reported failure analysis with enough context to judge the claim |
| Authentication | How do different authentication approaches address phishing resistance and account recovery? | A defined threat model, authoritative guidance and appropriate empirical studies |
| Accessible web forms | How do labels, focus and error messages affect the accessibility requirements for a registration form? | Specific criteria and evidence about the actual interaction |
| Standards and user evidence | Why is passing an accessibility checklist different from demonstrating usability for a defined user group? | The standard's scope plus empirical evaluations and their limitations |
Do not turn the reading plan into claims that you tested a campus portal or built an AI system if you did not. A review can examine published evidence honestly without adding an invented implementation chapter.
A verified starter reading set
Retrieval architecture: Lewis and colleagues' Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks, a 2020 research paper, gives a concrete starting architecture to investigate. Read its task and evaluation design rather than treating the phrase “RAG” as proof that any answer is accurate.
Citation evaluation: Gao and colleagues' Enabling Large Language Models to Generate Text with Citations, published at EMNLP2023, introduces the ALCE citation-evaluation benchmark. It is useful for asking how citation-bearing answers are assessed. A citation's presence and its support for a claim are different questions.
Authentication guidance: NIST SP800-63B-4, final July2025, addresses authentication and authenticator management. It is an authoritative standards source, not a field experiment establishing what Nigerian students prefer or a rule automatically binding your university. State the context in which you use it.
Accessibility criteria: W3C WCAG2.2 provides testable web-accessibility success criteria. Use the specific criterion and version relevant to your question. It is not a study proving a particular interface improved every user's experience.
This is a starter set. Search for later publications, alternative approaches and relevant counterexamples; foundational papers are not the whole current literature. Verify the full text before describing detailed results, and distinguish peer-reviewed work, preprints and standards in your notes. Use the research search workflow rather than collecting titles from a generated list.
Worked scoping example: citations in AI answers
Original planning example, not a completed experiment:
- Broad interest: AI-generated answers.
- Defined task: answering factual questions using supplied or retrieved documents.
- Question: how do published evaluation approaches distinguish a fluent answer from one with adequate supporting citations?
- Comparison: what is measured at answer level and at citation/claim level?
- Boundary: a review of reported evaluation methods, not a new claim about every model or a benchmark you have not run.
An illustrative body outline could define the task, compare evaluation criteria, examine reported failure cases and discuss limitations. The written seminar introduction should promise that scope, and the conclusion should answer it from the sources you actually reviewed.
Avoid a title such as “The best AI for all students”. Unless you define users, tasks, systems, versions and evaluation conditions, “best” hides the actual comparison.
Computer networking: protocol features versus measured performance
For a networking seminar, two original review questions are:
- Which protocol features and test conditions matter when comparing QUIC with TCP-based transports?
- How does HTTP/3 use QUIC, and what evidence would a claim about web performance require beyond the protocol specification?
Start with RFC9000, QUIC, May2021, which defines the core transport protocol, and RFC9114, HTTP/3, June2022, which describes HTTP semantics over QUIC. These are standards: use them to explain protocol behaviour, not as experimental proof that a particular website or network is faster.
For a concrete empirical reading anchor, Jaeger and colleagues' QUIC on the Highway: Evaluating Performance on High-rate Links is a 2023 research paper reporting benchmarks on dedicated hardware. Read its full methods before quoting results. Record the implementations and versions, hardware, network conditions, workload, measured quantity and comparison system. A result under one set of conditions is not a universal ranking of every deployment.
Original planning example: a literature review could compare what protocol documents promise to define with what selected benchmark papers actually evaluate. If one study measures a different workload or system configuration, make that mismatch explicit before comparing scores. Do not claim you ran those benchmarks or tested your university network when you only reviewed publications.
These are dated foundational sources, not a complete “latest technology” list. Check for later research, specification updates and errata relevant to your question. Hardware, routing and other networking subfields may need different sources; narrow the seminar to the evidence you can explain.
Build an evidence matrix before choosing a winner
| Field | What to record from each source |
|---|---|
| Publication | Title, authors, year, version and publication type |
| Task and context | The actual problem and setting addressed |
| Method or requirement | What was proposed, tested or specified |
| Comparison and measure | Baseline, dataset, metric or criterion, where applicable |
| Supported conclusion | A bounded claim with its page/table/section |
| Limitation | What the design or source type cannot establish |
A measured score without its task, test set and conditions is hard to interpret. A standard's requirement should not be placed in a “measured improvement” column. If sources cannot be compared fairly, make that mismatch part of the discussion rather than inventing a ranking.
Prepare the talk from the question
For a short presentation, choose one comparison you can explain clearly. Show what the method or criterion is meant to address, the evidence used and one important limitation. Do not fill slides with unexplained architecture names or claim that a live demonstration validates a whole system.
The seminar slide structure and defence questions help you turn the reviewed argument into a talk. Practical tests, user studies or security assessments require the permissions and supervision appropriate to the work; none is implied by these topic suggestions.
UniDraft helps organise your written seminar report around the topic, sources and instructions you confirm. It does not run benchmarks, test systems or create presentation slides. Keep the original papers and evidence notes so you can explain the report yourself.
