Chapter 16 — The Research Symposium

“If I have seen further it is by standing on the shoulders of Giants.” — Isaac Newton, letter to Robert Hooke, 1675

“Whatever you do, work heartily, as for the Lord and not for men.” — Colossians 3:23 (ESV)

Why This Matters

The symposium is this week. You stand up, you give a twelve-minute talk, you take questions, and you hand in a finished paper. That is the deliverable. But that is not the point.

The point is the question on the table this entire semester, now asked out loud: was this real? Sixteen weeks ago you chose a domain and a question about how modern AI advances it. You read the literature and built a comparison matrix. You wrote a falsifiable hypothesis and named your variables. You designed a reproducible experiment, ran a pilot, executed it for real, analyzed the data honestly, and wrote it up. You reviewed two peers and revised under critique. The whole time, the thesis of this course has been pressing on you: a practicum is the first half of a thesis. You do not study research from a distance — you do it, and you walk out holding the foundation of a publishable paper.

This week you find out whether you do.

Here is what changes in Week 16 that makes it different from every week before. Until now, the audience was the rubric. This week the audience is a room — and beyond the room, a real venue with a real call for papers and a real deadline. The standard stops being “did I meet the deliverable bar” and becomes “would a stranger, reading this cold, believe it and be able to rerun it.” That is a higher bar, and it is the right one. Most class projects die in a folder. A practicum paper is built to leave the folder.

The AI thread closes here from both directions. You used AI all semester as a tool you wield — to snowball literature, to draft code, to polish prose — and you learned its failure modes the hard way: hallucinated citations, fluent-but-unfaithful summaries, a confidence that does not track correctness. And for many of you, AI was also the workload you study — the thing your paper is about. This week both threads run into the same wall: a venue. Every major venue agrees an LLM cannot be an author, because authorship requires an accountability a tool cannot bear. “The model did it” is never a defense. The human stays in the loop where the judgment lives. This week, in front of a room, you are the loop.

And so the apologetic question for the final week is not a method question. It is a for-whom question. For whom, in the end, do we do the work? You can give a talk to impress a committee, to clear a requirement, to pad a CV. Or you can do the work heartily, as for the Lord — which, it turns out, is also the only motive that survives a hard review, a rejection, and a revision. We will get there. First, the work.

Coach’s Note — Read Chapter 14 (the paper) and Chapter 15 (peer review and revision) sitting open beside you this week. Week 16 does not introduce a new artifact so much as it finishes the ones you have. If your draft and your reviews are solid, this week is a focused sprint, not a scramble.

16.1 — What a Symposium Talk Actually Is

A conference talk is not your paper read aloud. It is the trailer for the paper. Its only job is to make a room of busy researchers want to read the full thing — and to leave them able to repeat one sentence about what you found.

Twelve minutes is short. At a comfortable speaking pace that is roughly 1,400–1,700 words and about a dozen slides. You cannot present everything. You must choose the one claim the audience should walk out with and build every slide toward it.

The beat structure that works, one slide per beat:

The full template, with speaker notes, is in code/talk-outline.txt. Fill it out, then rehearse out loud, on a clock, at least twice. The first run always overruns. A slide with more than about twenty words is a script, not a slide — cut it. Every figure must read from the back row: big fonts, labeled axes.

Coach’s Note — The single most common talk failure I see is the speaker who tries to defend the whole paper in twelve minutes and ends up rushing the result everyone came to hear. Decide what you will not say. The depth lives in the paper; the talk earns the reader.

16.2 — Handling Questions Without Bluffing

The Q&A is where credibility is won or lost, and it is the one part you cannot fully script. Three rules.

Anticipate the three hardest questions and answer them in backup slides. Reviewers and audiences ask predictable things: “What’s your baseline?” “How many seeds?” “Did you correct for multiple comparisons?” “Isn’t this just contamination?” If you have a slide ready, you look prepared because you are prepared.

When you don’t know, say so. “I didn’t test that — it’s a good direction for future work” is a strong answer. A bluffed answer is a weak one, and a room of researchers can smell it. This is the same discipline the whole course has trained: a correct “I don’t know” beats a confident wrong answer. It is true at the keyboard, true in the paper, and true at the podium.

Never let the talk claim more than the data. If a number is on a slide, it is in the paper. If a result is preliminary, you say “preliminary.” A talk is bearing witness to your data in front of a room — Exodus 20:16 reaches all the way to the podium.

The mechanics that quietly sink talks, none of them about content: arrive early and test your slides on the actual projector or share; have the deck in two formats (your editor’s native file and a PDF) in case the room’s setup chokes on one; know your first sentence cold so nerves don’t strand you on slide 1; and watch the clock without staring at it — a small timer on the lectern beats a panicked glance at slide 9. Practice the transition into Q&A: “I’ll stop there — happy to take questions.” None of this is research skill. All of it is the difference between a talk that lands and a talk that fights the room.

16.3 — The Finished Paper: From Draft to Camera-Ready

You wrote the draft in Week 14 and revised it under peer review in Week 15. This week you finish it in the venue’s format. The two dominant CS templates, both pre-loaded on Overleaf (the no-install path; see Appendix A):

Pick the class by your target venue, not by preference — an ACM workshop wants acmart; an IEEE workshop wants IEEEtran. A minimal acmart skeleton:

\documentclass[sigconf,anonymous,review]{acmart}
\begin{document}
\title{LLM-Assisted SQL Query Optimization: A Reproducible Comparison}
\author{Anonymous Author(s)}
\begin{abstract}
% Beck's four sentences: problem / why it matters / the contribution / the implication.
\end{abstract}
\maketitle
\section{Introduction}        % CARS: territory -> gap -> occupy
\section{Related Work}        % the comparison matrix, in prose
\section{Methodology}         % what you did, reproducibly
\section{Results}             % what you found, no spin
\section{Discussion}          % what it means + limitations
\section{Conclusion and Future Work}
\bibliographystyle{ACM-Reference-Format}
\bibliography{references}
\end{document}

The structure is IMRaD expanded for CS: abstract / introduction / related work / method / results / discussion / future work. Methods say what you did so another team can rerun it. Results report what you found with no spin. Discussion says what it means and where the limits are. That separation is not bureaucracy — it is how a reader tells your evidence from your interpretation.

Coach’s Note — Do not rewrite the paper this week. Reformat and tighten it. If you find yourself adding a whole new experiment in Week 16, you are either over-reaching for the venue or you skipped a week. Finish what you have, and put the new experiment in the continuation plan.

The two pieces of the paper that earn the read are the abstract and the introduction, and both have a known shape. Build the abstract on Kent Beck’s four sentences (OOPSLA 1993): (1) the problem, (2) why it is a problem, (3) the one startling sentence that is your contribution, (4) the implication if you are right. Build the introduction on Swales’ CARS moves (1990): establish the territory, then the niche or gap, then occupy it. These are not formulas to hide behind — they are the load-bearing walls. A reviewer who reads a clean four-sentence abstract and a three-move introduction knows in ninety seconds whether your paper is worth their evening. Make those ninety seconds count.

16.4 — Citation Hygiene: The One Check You Cannot Skip

Before the paper leaves your hands, every reference gets verified against a real index. This is the single highest-leverage integrity check in the course, because the failure is so easy and so disqualifying.

Walters & Wilder (2023, Scientific Reports) found that roughly 55% of GPT-3.5 and 18% of GPT-4 citations were entirely fabricated; among the citations that pointed to real works, 43% (GPT-3.5) and 24% (GPT-4) had substantive metadata errors. Bhattacharyya et al. (2023, Cureus) found 87% of citations to real works carried at least one metadata error. Newer retrieval and “deep research” agents hallucinate less, but the rate is never zero, and the consequence is severe: ICCV rejects papers with non-existent citations without review, and arXiv has signaled a one-year submission ban for “incontrovertible evidence” of unchecked LLM content such as hallucinated references (as of mid-2026; re-verify the exact terms).

So you audit. Export your library from Zotero to BibTeX and run every entry through a verification ledger — confirm each one exists in dblp, Semantic Scholar, or the publisher’s record. A small stdlib script that builds the checklist for you is in code/citation_audit.py:

python code/citation_audit.py references.bib
# -> citation_audit.csv : one row per reference, columns to fill BY HAND

The script does not verify for you. It builds the row so no reference slips through unverified. You own every citation, not the model.

16.5 — Choosing a Real Venue

Computer science inverts the usual hierarchy: top conferences generally outrank journals, often with acceptance rates under 20% (around 10% in the most competitive ML, SE, and vision venues — illustrative, and varies by subfield and year). Below the full conference sit workshops and Student Research Workshops (SRWs): higher-acceptance, lower-stakes on-ramps built for exactly the kind of work-in-progress a practicum produces.

The honest target for a 16-week practicum draft is a workshop or SRW short/WIP paper, roughly 2–4 pages. SRWs (for example, the long-running ones co-located with ACL/EACL) accept work-in-progress with a student first author. The pipeline from there is real: workshop → full conference → extended journal version.

To judge a venue, read two prestige proxies before you submit:

• CORE rankings — A*/A/B/C tiers for CS venues. The latest finalized round is ICORE 2026 (verify on core.edu.au; future-round timing drifts).
• Google Scholar Metrics h5-index — h papers from the last five complete years each cited at least h times, reported with an h5-median.

Then map the paper to the venue with code/venue-fit-sheet.txt: topic fit (quote the CFP scope line), page limit in the venue’s count format, template, anonymity rule, archival vs non-archival track, the deadline, and where AI disclosure goes.

Coach’s Note — A realistic workshop deadline you can hit beats an aspirational top-conference deadline you will miss. The best paper that never gets submitted helps no one. Pick the venue you can actually reach this cycle, and write the roadmap to it.

16.6 — The Publication Roadmap: Draft → Submission

The roadmap is the most heavily weighted judgment artifact of the final, because it is the part an AI cannot write for you: an honest assessment of the gap between what you have and what a venue requires, and a concrete plan to close it. The full checklist is in code/submission-roadmap.txt. Its phases:

1. Honest gap assessment. Is the headline claim backed by multi-seed runs, effect size, and a confidence interval — not a single run (Bouthillier et al. 2021)? Are the baselines fair and current? Is the related work current to the venue’s last cycle? Could a stranger rerun it?
2. Pick the venue, read the CFP. Copy the scope sentence, page limit, template, deadline, anonymity rule, and disclosure location into the roadmap.
3. Format and anonymize. Move into the template; if double-anonymous, scrub names, affiliations, self-citations, grant IDs, and identifying links.
4. Integrity + disclosure boxes. Every citation verified; no fabricated or cherry-picked data; honest statistics; AI use disclosed per the venue’s policy; no hidden prompt text anywhere.
5. Reproducibility artifact. Tag a Git release; deposit to Zenodo for a DOI (it auto-links a Software Heritage SWHID since the Oct 21, 2024 integration); target ACM “Artifacts Available” where offered.
6. Submit and continue. Submit with a 48-hour buffer. If accepted, present; if rejected, read the reviews without defensiveness, revise, and aim at the next cycle.

A note you must get right (as of 2026, verify live): a workshop-reviewed survey or position paper does not clear the arXiv-CS bar. arXiv’s Oct 31, 2025 rule requires CS review/survey/position papers to already be peer-review-accepted at a journal or conference with a DOI — workshop review is explicitly insufficient — and a separate Jan 21, 2026 endorsement change means an institutional email alone no longer qualifies a new submitter. Keep those two rules straight; they are easy to conflate.

16.7 — Disclosure: Saying What the AI Did

Every major venue agrees on three things, and you build your disclosure on them.

One — an LLM can never be an author. ACM, IEEE, NeurIPS, ICLR, ICML, ACL, CVPR/ICCV, and arXiv all hold the line: authorship requires accountability a tool cannot bear (the ICMJE four criteria fail at “final approval” and “accountability”). Two — “the model did it” is never a defense. You are responsible for every word, figure, citation, and line of code. Three — the disclosure heuristic is assistive vs generative. Spell-check, grammar, predictive typing, and editing aids need no disclosure; substantive generated text, figures, code, or ideas do.

Where the disclosure goes is venue-specific, and you confirm it on your target venue’s current CFP — these are the fastest-moving items in this whole chapter:

Reviewer-side AI rules are even more volatile and vary venue to venue (ICML 2026’s dual Policy A/B with PDF watermarking, NeurIPS forbidding uploading submissions to any LLM, CVPR forbidding LLM-written reviews). Treat every one as “as of mid-2026” and re-check the target venue’s current ethics page. The governing principle does not move: the human stays in the loop where the judgment lives; AI accelerates, the human decides, verifies, and is accountable. Appendix C is the full responsible-AI reference.

16.8 — Interactive Lab: Publication Roadmap / Venue Selector

Open the Publication Roadmap / Venue Selector embedded on this page and run it before you write your roadmap.

Feed it your paper’s profile: the domain and AI angle, how mature the results are (single run vs multi-seed with effect sizes), the page count you can realistically fill, and whether you need an archival result. The selector ranks venue types — workshop, Student Research Workshop, full conference, journal — by fit, selectivity, and deadline pressure, and then generates the path from draft → camera-ready → submission for the type you choose. As you pick, it surfaces the disclosure box and the reproducibility box you must check for that venue class.

The lesson the widget is built to teach: fit is a decision, not a default. The instinct to aim at the most prestigious venue is the instinct that gets practicum papers rejected without review or, worse, never submitted at all. Watch how the recommended path changes when you tell the truth about your results’ maturity. The honest input produces the reachable plan. Then copy that plan into code/submission-roadmap.txt and make it real.

16.9 — For Whom, in the End, Do We Do the Work?

The chapter’s verse is Colossians 3:23 — “Whatever you do, work heartily, as for the Lord and not for men” (ESV). Hold it next to the symposium, because the symposium is the moment the temptation to work for men is strongest. There is a room. There is a grade. There is a committee, a CV line, the chance to look impressive. Every one of those is an audience of men, and every one of them will, sooner or later, fail you as a motive.

Consider what a research career actually does to a motive built on the applause of the room. You will be rejected — most submissions are, at the venues worth submitting to. You will find a bug in your own results after you were proud of them, and integrity will require you to report the smaller number. You will read a review that is unfair, and a review that is fair and stings worse. A motive anchored to being seen cannot survive any of that. It curdles into the very temptations this whole course has trained you against: cherry-picking the runs that worked, quietly dropping the seed that broke the story, citing a paper you did not verify because it made the introduction sound stronger, letting a slide claim a little more than the data.

There is one motive that survives. To work as for the Lord is to do the work for an audience who already sees the data you are tempted to hide — which means the honest result and the impressive result are no longer in tension, because the only audience that finally matters is not impressed by spin. This is not a pious decoration on top of good research method. It is the root of it. Proverbs 25:2 opened this whole book: the glory of kings is to search a matter out. The searching is glorious precisely because there is truth to be found, a real order in the creation that does not bend to what we wish we had measured. Honest statistics, verified citations, reproducible code, a talk that claims exactly what the data supports — these are not compliance. They are what it looks like to love the truth more than the applause.

So when you stand up this week, work heartily. Give the best talk you can. Submit the best paper you can. Aim at a real venue and mean it. And do it for an audience of One, which — not by accident — is also the only way to do research you will still respect in ten years.

16.10 — Common Pitfalls

Pitfall: The talk is the paper, read aloud. Example: Twelve dense slides, every method detail, the result rushed in the last ninety seconds because time ran out. Fix: Build the talk to one remembered sentence. One slide per beat (§16.1). Rehearse on a clock; cut whatever overruns. Depth lives in the paper.

Pitfall: Aiming at the wrong venue. Example: A solid 4-page WIP result submitted to a 10%-acceptance flagship conference, desk-rejected for being preliminary — or never submitted because the bar felt impossible. Fix: Use the venue-fit sheet and the selector. A practicum draft is work-in-progress; the honest first target is a workshop or SRW. Pipeline up from there.

Pitfall: Unverified citations in the bibliography. Example: Two references pulled from an LLM-drafted related-work section turn out not to exist; a reviewer notices. Fix: Run code/citation_audit.py and verify every entry against dblp / Semantic Scholar / the publisher. A hallucinated citation can get a paper rejected without review (§16.4).

Pitfall: A single-run headline result. Example: “Our method beats the baseline by 1.4 points” — from one seed, with no error bars. Fix: Report multi-seed runs with effect size and a confidence interval (Bouthillier et al. 2021). If you only have one run, say so in the limitations and put the multi-seed run in the continuation plan.

Pitfall: AI used but not disclosed (or disclosed in the wrong place). Example: Generative help wrote a results paragraph; nothing appears in the paper, or it lands in the wrong section for the venue. Fix: Disclose substantive generative use per the venue’s policy, in the venue’s location (§16.7). When in doubt, disclose in the Acknowledgements and confirm against the CFP. Never leave hidden prompt text in the file.

Pitfall: No reproducibility artifact. Example: The paper claims results but the code lives only on a laptop; “could a stranger rerun this?” is no. Fix: Pin the environment, set seeds, document the data, tag a Git release, and deposit to Zenodo for a DOI (auto-linked SWHID). Target “Artifacts Available.”

Pitfall: Treating the symposium as the finish line. Example: The talk ends, the paper is submitted to Canvas, the repo goes cold, the venue deadline passes unmet. Fix: The roadmap is the deliverable’s center of gravity. Finish Week 16 holding a concrete plan with real dates to a real venue — and a continuation plan for after a rejection.

16.11 — Reps

The reps for this week live in the exercises, and this week they move your own paper across the finish line. A preview:

• Rep 1 — Draft your one remembered sentence and build the 12-slide talk from code/talk-outline.txt.
• Rep 3 — Rehearse the whole talk on a clock, twice, and cut until it lands inside twelve minutes.
• Rep 5 — Run code/citation_audit.py on your real .bib and verify every row.
• Rep 7 — Read your target venue’s actual CFP and fill code/venue-fit-sheet.txt.
• Rep 9 — Write the honest gap assessment in code/submission-roadmap.txt.

Do the reps, then take the on-page Check Your Reps quiz below the chapter to confirm the ideas stuck before you build the final.

16.12 — This Week’s Deliverable

The deliverable is the FINAL: Project 14 — the finished conference-style paper (15%), the symposium talk (5%), and a concrete publication roadmap. It is open-AI for the work, with an agent-log.txt required and integrity + disclosure per Appendix C. It is graded most heavily on the paper and the roadmap — the documents that carry your judgment. Read the project file now, and read it again the night before the symposium.

16.13 — Coach’s Final Word

You made it to the end.

Sixteen weeks ago you had an interest. Now you have a defined problem, a systematic literature review, a reproducible design, real results analyzed honestly, a conference-style paper, and a concrete plan to submit it to a real venue. That is not a class project. That is the first half of a thesis, exactly as promised — and most people who say they want to do research never produce this much. You did. Hold onto that.

A few words before you close the book.

The paper is not finished when the course ends. A practicum draft is an on-ramp, not an off-ramp. The students who get published are the ones who treat the roadmap as a real plan: they hit the workshop deadline, they read the reviews without flinching, they revise, they resubmit. The draft that goes cold in a repo helps no one. Keep the momentum.

The methods compound; the tools will not. The frontier model you used this term will be obsolete in two years. The judgment you built — how to spot a hallucinated citation, how to design a baseline, how to tell evidence from interpretation, how to claim exactly what the data supports and not one point more — that does not expire. It is what makes the next tool usable when it arrives. Build the judgment; the tools will follow.

Integrity is the whole game. Everything technical in this book — multi-seed runs, effect sizes, verified citations, reproducibility, honest disclosure — is one commitment wearing different clothes: bear true witness to what you found. It is harder than it sounds, because the temptation is always to claim a little more. Hold the line. A career of honest small results outlasts one impressive lie.

And work for the right audience. Do the work heartily, as for the Lord. Not because piety makes the statistics better, but because the only audience who already sees the data you are tempted to hide is the only audience worth doing honest work for. Aim your paper at a real venue and mean it — and aim your life higher than the venue.

Now go give the talk. Submit the paper. Then keep going.

See you on the other side.

Up next: Do the reps in the exercises, then build the final in Project 14. Need the lab on-ramp again? Appendix A. The toolkit: Appendix B. Using AI responsibly: Appendix C. The glossary: Appendix D. Previous chapter: Chapter 15. This is the last chapter. After the symposium, you’re done.