Series ID: EDKS-ADV-ZH-0025 · How to Learn Advanced English (Chinese Edition) · Lesson No.025 · C1 → C2

Give a Research Presentation and Defend It Under Questions Without Overclaiming | 做研究汇报并在问答中守住证据边界

A strong research presentation is not a paper read aloud. It is a controlled argument delivered under time pressure: question, design, evidence, uncertainty, conclusion—and then the same evidence boundaries defended when the audience begins asking questions.

高水平 research presentation 的难点，不只是“讲得流利”。真正困难的是：时间变短以后，仍然不把 claim 说大；被追问以后，仍然分得清 data establish 什么、suggest 什么、没有回答什么。

AUDIENCE → QUESTION → CLAIM → DESIGN → EVIDENCE → UNCERTAINTY → VISUAL → SPOKEN EXPLANATION → QUESTION → ANSWER → BOUNDARY → NEXT EVIDENCE.

Part I — Build the presentation around one research job | 第一部分：先决定这场汇报到底要让听众理解什么

1. A presentation is a decision architecture, not a compressed manuscript

Your audience cannot reread the previous paragraph while you continue speaking. Spoken research English therefore needs stronger hierarchy than written prose.

2. Define the one-sentence research job

We asked whether [X] in [population/context] changes, predicts or explains [Y], using [design], because [decision or knowledge gap].

3. If you cannot say the research job in one sentence, the talk is not ready

Complexity can return later. The opening needs orientation.

4. Audience model | Who needs what?

Audience	Likely need
specialists	design, assumptions, novelty, uncertainty
mixed academics	question, method logic, result meaning
policy/practice	decision relevance, scope, trade-offs
viva panel	reasoning, choices, limitations, ownership

5. Opening minute | Four jobs

name the problem;
state why it matters;
identify the gap;
state the research question.

6. Strong opening | Problem first

AI writing tools can improve the text produced during assistance. The harder question is whether students write better when the tool is removed. Our study tests that transfer problem directly.

7. Use an opening map, not a memorised paragraph

problem → gap → question → route.

8. The route sentence

I’ll first show the design, then the primary result, and finally what the result does—and does not—allow us to conclude.

9. Presentation claim hierarchy

primary claim;
supporting claims;
exploratory observations;
interpretive hypotheses;
future questions.

10. Title slide is already a claim

AI Builds Independent Writers is stronger than AI-Assisted Drafting and Four-Week Unaided Writing Performance.

11. Spoken causality must match design

Use the strongest justified verb, not the most impressive verb.

12. Time claims need exact time

Say four weeks later, not long term, unless long-term evidence exists.

13. Population claims need exact population

Say students in these three programmes before saying university students generally.

14. Construct claims need exact construct

Confidence is not competence. One rubric score is not all writing ability. Click-through is not engagement.

15. Mechanism claims need mechanism evidence

If cognitive load was not measured, say One possible explanation is reduced cognitive load, not The intervention worked because it reduced cognitive load.

16. The five-slide scientific spine

Question.
Design.
Primary evidence.
Uncertainty/limitations.
Conclusion/implication.

17. Methods compression | Preserve identification logic

Do not narrate every procedural detail. Preserve the details needed to judge the inference.

18. Visual rule | One slide, one audience job

A slide may contain several objects, but they should answer one question.

19. Slide title should carry meaning

Weak: Results. Better: Assisted quality improved immediately; unaided transfer remained uncertain at four weeks.

20. Figure narration | Orient → pattern → inference

The x-axis shows time, the y-axis shows rubric score, and the two lines represent intervention and control. The groups separate during assisted drafting but converge at the unaided four-week assessment. So the study supports an immediate assisted-performance benefit, but it does not establish durable unaided improvement.

21. Primary result formula

estimate + uncertainty + practical meaning.

22. “Not significant” is rarely enough

The audience needs magnitude and uncertainty.

23. “Significant” is rarely enough

The audience needs practical meaning.

24. Null result can be the main result

Do not hide it because it is less dramatic.

25. Secondary analyses need labels

Say secondary, exploratory or post hoc when appropriate.

26. Limitations slide should change interpretation

limitation → threatened inference → resulting boundary.

27. Strong limitation

We did not measure cognitive load, so our mechanism explanation is a hypothesis rather than a demonstrated pathway.

28. Final slide should answer the original question

What we found.
What it means within scope.
What remains unresolved.

29. Example conclusion

AI assistance improved immediate drafted output and speed. We did not find clear evidence of better unaided writing four weeks later. The next decision is therefore not whether AI “works” in general, but which assisted gains transfer after support is removed and under what safeguards.

30. Part I operating rule | Compress detail, not epistemic boundaries

A short talk may omit procedural detail. It must not omit the boundaries that determine whether the conclusion is true.

Research presentation 可以压缩 detail，但不能压缩 evidence boundary。时间越短，claim 越要精准。

Part II continues with spoken research English, slide narration, timing, recovery and Q&A defence.

Part II — Spoken research English under time pressure | 第二部分：把论文英语变成真正能讲出来的研究英语

31. Written academic English and spoken academic English are not identical

Written prose can carry longer dependency chains. Spoken research English must preserve the logic while making each step audible.

32. Written-to-spoken conversion

Written: Notwithstanding the observed improvement in assisted performance, the absence of a clearly estimated delayed transfer effect warrants caution in interpreting the intervention as producing durable independent competence.

Spoken: Students performed better while using the tool. Four weeks later, however, the unaided difference was uncertain. So we should not call this durable independent learning yet.

33. One breath, one logical move

Use pauses between result → contrast → implication.

34. Opening signposts

The question we wanted to answer was…
The gap is not X; it is Y.
To test that, we…
I’ll focus on three things…

35. Methods signposts

The key design feature is…
The comparison that matters here is…
Our primary outcome was…
The reason for this time point was…

36. Results signposts

The main result is…
The important contrast is…
What changes the interpretation is…
The estimate is…, with uncertainty from… to…

37. Boundary signposts

This result does not establish…
We can say X, but not yet Y.
The main boundary is…
One alternative explanation is…

38. Chinese-speaker trap | 过度书面化

Long nominal phrases can sound impressive on paper but become difficult to process aloud.

39. Convert noun stacks into verbs

the implementation of the intervention evaluation procedure → we evaluated how the intervention was implemented.

40. Passive voice remains useful when actor is irrelevant

Essays were scored by blinded raters.

41. Define technical terms once and use them consistently

Do not alternate casually among confidence, competence, proficiency and ability.

42. Numbers need spoken framing

Do not dump 0.8, CI −1.9 to 3.5, p=.56. Say what the estimate means.

43. Better number narration

The estimated difference was less than one rubric point, and the interval crossed zero widely, so the delayed effect remains uncertain.

44. Percentages need denominators when risk matters

Fourteen percent—28 of 200 generated factual claims—were accepted despite lacking support.

45. Avoid magnitude adjectives without a benchmark

Words such as dramatic, huge, negligible and substantial need a field or decision context.

46. Timing is part of scientific speaking

A ten-minute talk cannot contain a twenty-minute argument spoken twice as fast.

47. Build a time budget before slides

10-minute talk	Approximate budget
problem + question	1.5 min
design	2 min
primary evidence	3 min
limits + interpretation	2 min
conclusion	1.5 min

48. Rehearsal should test decisions, not memorisation

Ask: what can be removed without damaging the inference?

49. Three rehearsal passes

Logic pass: can someone reconstruct question → evidence → claim?
Time pass: does the talk finish with breathing room?
Pressure pass: can you recover after interruption or a failed slide?

50. Do not memorise every sentence

Memorise the route, evidence anchors, exact technical definitions and final conclusion.

51. Recovery language when you lose your place

Let me return to the main comparison.
The key point on this slide is…
I’ll skip the secondary detail and move to the primary result.
What matters for the conclusion is…

52. Recovery is a research skill

A missed sentence is not a failed talk if the argument survives.

53. Slide failure protocol

If a figure fails to render, explain the result without pretending the audience can see it.

54. Do not apologise for thirty seconds

State the problem once, then continue.

55. Citation in speech

Do not read a bibliography. Name the relevant study when it changes the argument.

56. Literature comparison sentence

Our immediate assisted result is consistent with prior drafting studies, but our delayed unaided outcome addresses a different transfer question.

57. Speaking about uncertainty confidently

Confidence in delivery is not certainty in evidence.

58. Strong uncertainty language

The data do not distinguish clearly between a small benefit and little difference at four weeks.

59. Weak uncertainty language

We are not really sure about anything.

60. Part II operating rule | Fluency serves inference

The goal is not to sound certain. The goal is to make the audience hear exactly how certain the evidence allows you to be.

口语流利不是把 uncertainty 藏起来；真正高级的表达，是把 uncertainty 讲得清楚、稳定、可信。

Part III — Q&A: answer the question before defending yourself | 第三部分：问答时先回答问题，不要先保护自尊

61. Q&A changes the cognitive task

You are no longer following your planned route. You must classify a question, retrieve evidence and protect scope in real time.

62. The five-step Q&A kernel

HEAR → CLASSIFY → ANSWER → EVIDENCE → BOUNDARY.

63. Step 1: Hear the actual question

Do not begin answering while the questioner is still constructing the question.

64. Step 2: Classify it

Is it about methods, result, causality, scope, theory, limitation, implementation, ethics, statistics, literature or future work?

65. Step 3: Answer directly

Give the shortest truthful answer first.

66. Step 4: Give the evidence or reasoning

Explain why.

67. Step 5: State the boundary

Say what remains unresolved when needed.

68. Model answer structure

Short answer: no. Our design shows X, not Y. The reason is Z. To answer Y, we would need A.

69. “I don’t know” has several scientific meanings

I do not remember the number.
We did not measure it.
The study was not designed to answer it.
The evidence is genuinely mixed.
I have not seen the relevant literature.

70. Do not collapse them into one vague apology

71. When you do not remember a number

I don’t want to invent the exact value. The direction was X, but I would need to check the table for the precise estimate.

72. When you did not measure it

We did not measure cognitive load directly, so I can discuss it only as a possible mechanism.

73. When the design cannot answer it

That is an important causal question, but our cross-sectional design cannot establish temporal direction.

74. When you have not read the literature

I’m not familiar enough with that study to compare it responsibly. I’d want to read it before making the comparison.

75. These are strong answers, not weak answers

They protect the evidence boundary.

76. Clarifying a broad question

Are you asking whether the effect is statistically robust, or whether it is practically large enough to matter?

77. Clarifying an ambiguous term

When you say “learning,” do you mean assisted performance during the session or later unaided performance?

78. Do not use clarification to evade

Clarify once, then answer.

79. Difficult question type: “Why didn’t you use method X?”

We considered X. We chose Y because our primary estimand/design required Z. X would answer a related but different question. A useful sensitivity analysis would be…

80. Difficult question type: “Your sample is too small.”

The key issue is precision. For the primary estimate, the interval remains wide enough to include effects that would matter in either direction. So I agree that magnitude remains uncertain; I would not describe the result as proving no effect.

81. Difficult question type: “Isn’t this just correlation?”

If yes, say yes.

82. Strong response

Yes. The current analysis is associational. We can discuss possible mechanisms, but the study does not identify a causal effect.

83. Difficult question type: “So your intervention failed?”

It depends on the outcome. It improved assisted output and speed. It did not provide clear evidence of better unaided writing four weeks later. So “failed” is too broad; the transfer claim remains unsupported.

84. Difficult question type: “Why should anyone care?”

Because the distinction between assisted performance and later independent performance changes how institutions should evaluate these tools. Immediate output quality alone cannot answer the learning question.

85. Difficult question type: “Your result contradicts Professor X.”

It may. The first thing I would compare is whether the studies measure the same outcome at the same time point. If Professor X measures assisted output and we measure delayed unaided transfer, the apparent contradiction may reflect different questions.

86. Difficult question type: “Can this generalise to schools?”

Not directly from our data. Our sample was university students in a high-support setting. The mechanism may motivate a school study, but school generalisation remains an empirical question.

87. Difficult question type: “Would you recommend adoption?”

I would separate adoption from efficacy. Our data support an immediate assisted benefit, but broad adoption also depends on transfer, factual reliability, cost, integrity and implementation evidence that this study does not fully provide.

88. Difficult question type: “What is the biggest weakness?”

Name the weakness that most changes interpretation, not the safest trivial limitation.

89. Strong answer

The largest boundary is duration. We have a four-week unaided outcome, not evidence about semester-long independent development.

90. Part III operating rule | Direct answer first

In Q&A, credibility rises when the audience can tell that you are answering the scientific question rather than defending your identity.

Part IV — Defend methods, statistics and interpretation without becoming defensive | 第四部分：守住方法与解释，但不要把问答变成争辩

91. Defence is not refusal to change your mind

A research defence demonstrates that you understand why choices were made, what alternatives exist and what evidence would change your conclusion.

92. The method-defence ladder

state the choice;
state the reason;
state the trade-off;
state the consequence;
acknowledge a valid alternative.

93. Example

We used a four-week unaided outcome because our question concerned transfer after support was removed. That sacrifices information about very long-term persistence, so our conclusion is limited to four weeks. A semester-long follow-up would answer a different durability question.

94. Do not answer “because that is standard” unless the standard is itself the reason

95. Do not answer “because my supervisor told me”

You should understand the scientific rationale for your own work.

96. Statistical question protocol

estimand → model → assumption → diagnostic/sensitivity → interpretation.

97. Example statistical defence

Our estimand was the between-group difference in four-week score. We adjusted for baseline score to improve precision. Because students were nested within instructors, uncertainty was cluster-aware. The conclusion was similar under the prespecified sensitivity model.

98. If an audience member identifies a real flaw

Do not defend the indefensible.

99. Strong response to a valid criticism

You’re right that our current model does not account for that dependency. That could make the interval too narrow. I would treat the current uncertainty estimate as provisional and re-run the analysis with the clustering represented.

100. “You’re right” can increase credibility

Scientific authority is not infallibility.

101. If the criticism is partly right

I agree with the first part: the sample limits transportability. I would separate that from internal validity, however, because randomisation still supports the within-sample comparison.

102. If the criticism answers a different question

That would be a useful design for mechanism. Our study asks the narrower efficacy question, so I would see it as a complementary next study rather than a requirement for the present comparison.

103. If the question assumes something false

Correct the premise respectfully.

104. Model

I think there may be one premise to clarify. The primary outcome was preregistered before data collection; the subgroup analysis was exploratory. So I would not interpret the subgroup as the confirmatory result.

105. Hostile tone does not require hostile response

Answer the scientific content, not the emotional packaging.

106. Separate challenge from insult

If a question contains both, respond to the challenge.

107. Do not mirror sarcasm

108. Do not over-thank an aggressive question

A neutral That raises an important distinction is enough.

109. Time-limited answer | 20-second kernel

Short answer → one reason → one boundary.

110. 60-second answer

Short answer → evidence → alternative → boundary/next test.

111. When a question contains three questions

There are three parts there. I’ll take the design question first, then generalisability, and finally implementation.

112. When time allows only one

I’ll answer the part that most affects the conclusion—the design issue.

113. When you need five seconds to think

Pause. Do not fill the silence with unsupported claims.

114. Useful thinking phrase

Let me separate two issues there.

115. Do not repeat the entire question to buy time

Paraphrase only when it clarifies.

116. Question about unmeasured mechanism

Our result is compatible with that mechanism, but we did not measure the mediator, so I would not claim that we demonstrated it.

117. Question about practical significance

The estimate is statistically distinguishable from zero, but whether it matters depends on the decision threshold. In our setting, the gain is about X relative to Y cost/risk.

118. Question about non-significance

The result is inconclusive rather than proof of no effect because the interval still contains effects that would matter.

119. Question about equivalence

We did not design an equivalence test, so non-significance should not be read as equivalence.

120. Question about subgroup differences

One subgroup is significant and the other is not, but that alone does not establish a subgroup difference. The relevant test is the interaction.

121. Question about multiple comparisons

The primary outcome was prespecified. The additional subgroup findings are exploratory and should be interpreted as hypothesis-generating.

122. Question about missing data

Attrition is a real limitation. We examined whether missingness differed by group and ran sensitivity analyses; the persistence claim should still be read within those assumptions.

123. Question about measurement validity

The instrument measures self-efficacy, not objective competence. We therefore keep those constructs separate in our interpretation.

124. Question about external validity

Our study supports the comparison in this sampled setting. Transport to another age group, country or support environment requires additional evidence.

125. Question about ethics

Answer procedures accurately. Do not improvise institutional facts you cannot verify.

126. Question about AI use

State what model/version/procedure was used, what was logged, what was verified and what was not evaluated.

127. Question about reproducibility

Distinguish what is publicly available from what is restricted.

128. Question about novelty

The novelty is not that no one has used AI for drafting. It is the separation of assisted performance from later unaided transfer using a prespecified delayed outcome.

129. Question about contradictory literature

Compare design, population, outcome and time before claiming contradiction.

130. Question about theory

Our data are consistent with Theory A, but they do not discriminate A from Theory B because both predict the observed pattern.

131. The strongest defence sometimes narrows the claim

Do not preserve a broad claim merely because it appeared on your slide.

132. Part IV operating rule | Defend reasoning, not ego

A good defence shows why a choice was reasonable and where it stops being sufficient. A bad defence treats every question as an attack that must be defeated.

Part V — Full worked research presentation and Q&A | 第五部分：完整研究汇报与问答示范

The study below is fictional teaching material.

133. Fictional study

180 university students are randomised to AI-assisted drafting or standard word processing. The registered primary outcome is unaided essay quality four weeks later. Secondary outcomes include immediate assisted quality, speed, confidence and unsupported-claim acceptance.

134. Fictional findings

immediate assisted quality: +4.5 rubric points;
completion time: 18% faster;
four-week unaided difference: +0.8, 95% CI −1.9 to 3.5;
confidence: +6 scale points;
unsupported factual claims accepted: 14%.

135. Opening script

AI writing tools can make a draft look better while the tool is present. But universities care about a second question: does that improvement transfer when the tool is removed? We randomised 180 students to AI-assisted drafting or standard word processing and measured unaided writing four weeks later. I’ll show the design, the immediate assisted result, and then the delayed transfer result that changes the interpretation.

136. Methods script

Students completed the same writing task under one of two drafting conditions. Our registered primary outcome was a new unaided essay four weeks later, scored by blinded raters. Immediate quality, speed, confidence and factual-reliability measures were secondary outcomes.

137. Immediate-result script

During assisted drafting, the AI group scored 4.5 rubric points higher and completed the task 18% faster. So there is a clear assisted-performance advantage in this setting.

138. Primary delayed-result script

Four weeks later, with the tool removed, the estimated difference was 0.8 points, with a 95% interval from −1.9 to 3.5. That interval is too wide to support a clear durable benefit or a clear absence of benefit. The delayed transfer effect remains uncertain.

139. Reliability-result script

There was also a trade-off. Students accepted 14% of generated factual claims that our checking procedure classified as unsupported. That matters because output quality and factual reliability are different outcomes.

140. Limitation script

There are three boundaries. First, four weeks is not long-term development. Second, our university sample does not establish school-age generalisation. Third, we did not measure cognitive load directly, so any cognitive-load explanation remains a hypothesis.

141. Conclusion script

Our narrow conclusion is that AI assistance improved immediate drafted output and speed. We did not obtain clear evidence of better unaided writing four weeks later, and factual reliability remained a concern. So the next research question is not simply whether AI improves writing while present; it is which gains transfer after support is removed and under what safeguards.

142. Q1 — “So AI doesn’t help students learn?”

I would make that narrower. We found clear assisted-performance benefits, but the delayed unaided estimate was uncertain. So our study does not establish a durable learning benefit; it also does not prove that no such benefit exists.

143. Q2 — “Why only four weeks?”

We wanted a delayed measure far enough from the intervention to separate immediate assistance from later unaided performance, while keeping attrition manageable. Four weeks answers that transfer question, but it does not answer semester-long durability.

144. Q3 — “Isn’t confidence evidence of learning?”

Not by itself. Confidence is a self-report construct; our unaided essay score is an objective performance measure. The confidence increase is interesting, but we do not use it as a substitute for competence.

145. Q4 — “Could cognitive load explain the result?”

Possibly, but we did not measure cognitive load. It is one plausible mechanism, not a demonstrated one.

146. Q5 — “Would you deploy this across a university?”

Not from this study alone. Deployment requires evidence about transfer, factual reliability, academic integrity, cost and implementation. Our data contribute to that decision but do not complete it.

147. Q6 — “Your primary result is non-significant. Isn’t the study negative?”

The primary delayed estimate is inconclusive rather than evidence of equivalence. The interval includes effects that could matter. At the same time, the assisted secondary outcomes show clear immediate benefits. So the answer depends on which outcome we mean.

148. Q7 — “Why should we trust a rubric score?”

That is a measurement question. We used blinded raters and a predefined rubric, and we assessed inter-rater agreement. The score still represents essay performance under that rubric, not every dimension of writing ability.

149. Q8 — “Maybe stronger students benefited more?”

That is plausible, but unless the interaction was prespecified and adequately estimated, I would treat any subgroup pattern as exploratory rather than a confirmatory conclusion.

150. Q9 — “Could students simply have copied AI language?”

During the assisted task, direct language support is part of the intervention, so that may contribute to the immediate gain. That is precisely why the unaided four-week outcome matters: it asks whether performance persists after direct assistance is removed.

151. Q10 — “What would change your conclusion?”

A well-estimated delayed unaided effect replicated across settings would strengthen the transfer claim. Longer follow-up would address durability, and direct mechanism measures would be needed before making mechanism claims.

152. Why the worked Q&A is strong

Each answer begins with the scientific answer, identifies evidence, and ends at the boundary.

153. It does not treat every challenge as a reason to add another study to the current paper

154. It does not turn uncertainty into embarrassment

155. It does not convert confidence into competence

156. It does not turn four weeks into “long term”

157. It does not generalise university data to all learners

158. It does not claim a mechanism that was not measured

159. It does not make adoption recommendations from efficacy evidence alone

160. Part V operating rule | The same evidence boundary must survive the microphone

If your paper says “association,” your talk cannot say “cause.” If your study says “four weeks,” your Q&A cannot say “long term.” Live pressure does not expand the evidence.

Part VI — Conference, viva and defence drills | 第六部分：把现场问答练成稳定能力

161. Three live contexts require different emphasis

Context	Main pressure
conference	brevity, mixed audience, public Q&A
viva/oral defence	method ownership, sustained questioning, alternatives
internal research meeting	decision-making, troubleshooting, next experiment

162. Conference answer | Usually shorter

Answer the public question and leave detailed derivations for follow-up.

163. Viva answer | Show reasoning depth

Explain why the choice was made, what assumptions it carries and what an alternative would change.

164. Research meeting answer | Expose uncertainty early

The goal is to improve the work, not perform certainty.

165. Viva question: “Why this research question?”

Because prior studies establish assisted output gains, but they do not isolate later unaided transfer. That gap changes the educational interpretation of the technology.

166. Viva question: “Why this design?”

Random assignment addresses the efficacy comparison, and the delayed unaided task separates assistance from transfer. The design does not, however, identify the mechanism.

167. Viva question: “Why not a qualitative study?”

A qualitative study would be valuable for experience and mechanism hypotheses, but it would not estimate the between-condition performance contrast that is central to this question.

168. Viva question: “Why not a larger sample?”

A larger sample would improve precision. The relevant question is whether current precision is adequate for the claim. For our primary estimate, the interval remains wide, so magnitude remains uncertain and our conclusion reflects that.

169. Viva question: “What assumption worries you most?”

Name the assumption with the greatest inferential consequence.

170. Viva question: “What would you redesign?”

Do not pretend the study is perfect.

171. Strong redesign answer

I would add a longer unaided follow-up and a direct measure of the proposed mediator. That would separate durability from mechanism rather than asking the current design to answer both.

172. Viva question: “What is original here?”

State contribution narrowly and verifiably.

173. Viva question: “What did you learn that surprised you?”

Do not rewrite a surprise as a prespecified hypothesis.

174. Strong answer

The factual-reliability trade-off was more pronounced than we expected. Because that analysis was secondary, I treat it as a finding that changes interpretation rather than as the original confirmatory hypothesis.

175. Viva question: “What is the strongest alternative explanation?”

A mature answer identifies a real competitor.

176. Viva question: “How would a critic interpret this differently?”

Represent the critic fairly before responding.

177. Viva question: “What evidence would falsify your interpretation?”

State what result would make you revise the claim.

178. This is a C2 research-English skill

Advanced academic language includes controlled revision of one’s own position.

179. Drill 1 | 30-second research job

Explain question, design and primary conclusion in 30 seconds without jargon.

180. Drill 2 | 60-second Methods

Preserve only details needed to judge the inference.

181. Drill 3 | One-figure narration

Orient → pattern → estimate → uncertainty → meaning.

182. Drill 4 | Null result

Explain a non-significant result without saying “no effect” unless the design supports that conclusion.

183. Drill 5 | Significant but small result

Separate statistical evidence from practical importance.

184. Drill 6 | Hostile causality question

Answer without sarcasm and without expanding the design.

185. Drill 7 | “Why not method X?”

Defend your estimand/design logic, then acknowledge valid alternatives.

186. Drill 8 | “I don’t know”

Practise five different forms: forgotten value, unmeasured variable, design boundary, mixed evidence, unfamiliar literature.

187. Drill 9 | Generalisation

Answer questions about a population your study did not sample.

188. Drill 10 | Mechanism

Turn an unmeasured mechanism from “because” into “one possible explanation.”

189. Drill 11 | Three-part question

Segment it aloud and answer in order.

190. Drill 12 | Interrupted answer

Recover without restarting the whole explanation.

191. Drill 13 | Slide failure

Explain the primary figure verbally.

192. Drill 14 | Time cut

Deliver the conclusion when the chair says “one minute left.”

193. Drill 15 | Valid criticism

Say “you’re right” and explain the consequence accurately.

194. Drill 16 | Invalid premise

Correct the premise respectfully before answering.

195. Drill 17 | Contradictory study

Compare population, design, outcome and timing.

196. Drill 18 | Adoption question

Separate efficacy from implementation and policy.

197. Drill 19 | Biggest weakness

Name the real inferential boundary.

198. Drill 20 | What would change your mind?

State a concrete evidential condition.

199. Chinese-to-English pressure drill | 中文先想，英文再守边界

Chinese thought: 这个问题我们的研究其实回答不了。

English: That question goes beyond what this design can establish. We can answer X from these data; Y would require a different design.

200. Chinese thought: “这个结果不显著，所以没效果。”

Repair: The estimate is uncertain; the interval still contains effects that could matter, so we should not equate non-significance with no effect.

201. Chinese thought: “这个 reviewer 说得不对。”

Live defence: I agree that X is a limitation. I would separate it from Y, however, because the randomised comparison still supports the within-sample causal contrast.

202. Chinese thought: “我忘记数字了。”

I don’t want to invent the exact number. The direction was X; I would need to check the table for the precise estimate.

203. Chinese thought: “我没看过这篇论文。”

I’m not familiar enough with that paper to compare it responsibly. I’d want to read it before making a claim about the difference.

204. Seven-day presentation cycle

Day 1: 30-second research job.
Day 2: five-slide scientific spine.
Day 3: figure narration and uncertainty.
Day 4: 20 difficult questions.
Day 5: hostile-tone simulation.
Day 6: timed full talk with interruptions.
Day 7: unseen Q&A benchmark.

205. 100-point presentation rubric

Dimension	Points
research question and route	10
design explanation	15
evidence hierarchy	15
uncertainty and scope	15
visual narration	10
spoken clarity and timing	10
Q&A directness	10
method defence	5
intellectual honesty	5
recovery under pressure	5

206. 90–100 | C2-ready research presenter

The speaker makes the research logic audible, preserves outcome hierarchy, states uncertainty without loss of authority, answers questions directly, accepts valid criticism and never expands a claim merely because of live pressure.

207. 75–89 | Strong C1 presenter

The science is mostly clear, but the speaker may over-explain Methods, hedge excessively or need more concise Q&A boundaries.

208. 60–74 | Knowledgeable but manuscript-bound

The speaker knows the study but reads slides, uses written syntax aloud, loses the main claim in detail or becomes defensive under questioning.

209. Below 60 | Performance without evidence control

The talk may sound fluent but confuses causality, significance, mechanism, generalisation or outcome hierarchy.

210. Final presentation gate

Can I state the research job in one sentence?
Does the title match the evidence?
Is the primary outcome visibly primary?
Can I explain the design without reading?
Can I narrate each figure without the slide?
Do I give magnitude and uncertainty?
Are secondary/exploratory findings labelled?
Do limitations change the claim?
Does the conclusion answer the opening question?
Can I answer “I don’t know” precisely?
Can I distinguish unmeasured from unknown?
Can I defend a method without claiming it is the only method?
Can I accept a valid criticism?
Can I correct a false premise respectfully?
Can I keep causality within design?
Can I keep population within sample?
Can I keep time within follow-up?
Can I keep mechanism within measured evidence?
Can I separate efficacy from adoption?
Can I finish on time?

211. Research and public guidance floor

212. Canonical eduKate research-writing route

213. SEO language map

This lesson naturally covers research presentation, academic presentation, research presentation example, how to present research, thesis defence, viva questions, dissertation defence questions, conference presentation, academic Q&A, difficult questions after presentation, how to answer research questions, research presentation slides, present research findings, explain statistics orally, defend methodology, avoid overclaiming, C1 academic English, C2 academic English, 研究汇报, 学术汇报, 答辩英语, viva 英语, research presentation 怎么讲 and 中文母语学术英语.

214. Final principle | The microphone does not increase the evidence

A research presenter’s authority comes from keeping the claim stable when time is short, the room is watching and the questions are difficult.

真正高级的答辩，不是每个问题都“赢”。而是在压力下仍然知道：什么是 data，什么是 interpretation，什么是 limitation，什么必须留给下一项研究。

215. Exit standard

You are ready to move on when you can give a timed research talk, lose a slide, receive an unexpected challenge, admit an unknown, defend a justified method, accept a valid criticism and finish with exactly the same evidence boundaries you would defend in writing.

Next lesson reserved | 下一课

EDKS-ADV-ZH-0026 · Lesson No.026 · Write and Deliver an Academic Poster Pitch That Makes the Research Legible in Two Minutes | 做 Academic Poster Pitch：两分钟讲清研究问题、证据与边界

Back to Advanced English Chinese Edition Hub · 返回高级英语中文版主页

Small Group Tutorials

How to Learn Advanced English (Chinese Edition) | Lesson No.025 | Give a Research Presentation and Defend It Under Questions Without Overclaiming | 第025课：做研究汇报并在问答中守住证据边界