Helping Students Become Independent Researchers
Updated: 2 days ago
What should happen when a student asks for an experiment to be fixed? A practical account of clarifying the problem, choosing an informative test and giving the student's reasoning a greater role in supervision.
When an experiment refuses to make sense
"Could you give me an idea that will make this experiment work?"
Imagine a student asking me this after several unsuccessful attempts. They are upset. They have changed the settings, repeated the experiment and tried another method, yet the results still refuse to make sense. What began as a comparison between two approaches now seems to produce a different answer each time. They have come to me hoping I can see a way forward.
I would take that frustration seriously. Repeated disappointment can make a student question their own ability, especially when they cannot explain what is going wrong. I would want us to leave the meeting with a manageable next step. We would begin by putting the results beside the original plan and looking carefully at what happened.
In my earlier essay on supervising PhD researchers, I described the gradual shift from following a project to helping direct it. Here I want to stay with one ordinary meeting and ask what the supervisor can actually do to support that shift.
First, agree what “working” would mean
The word "work" would be worth pausing over. Does it mean that the apparatus behaves as intended? That the measurements are repeatable? Or that the proposed method performs better than the baseline? These are different problems. An experiment can be carefully conducted and still fail to support the result we hoped for.
In this example, we might notice that the two methods were tested in separate sessions, with one always tested after the other. That would not explain the inconsistent results by itself. But it would raise a question about whether the comparison was measuring the difference between the methods or a change in the experimental conditions.
I could help the student examine that possibility. Then I would want to hear how they would investigate it. They might suggest reversing the order of the trials, or repeating a reference measurement at the beginning and end of a session. We could discuss what each test would establish and choose one that was practical. The next experiment would have a clear purpose, even though its outcome remained uncertain.
What I would value in that conversation is the student beginning to formulate an idea of their own. They might still need help with the test, and the original problem might remain unresolved. But they would have a reason for the next step and a clearer sense of what its result could tell them. That is something they can build on when another experiment becomes difficult.
AI makes this especially important. It can offer a long list of possible fixes within seconds. Those suggestions may be useful, but the student still needs to decide which possibility deserves attention and how to examine it. I want my guidance to help them develop that judgement, so that asking for help becomes part of learning to think for themselves.
Guidance whose effects are easy to miss
An old Chinese story helps me think about why this kind of guidance matters. In the Heguanzi, the physician Bian Que is asked which of his brothers is the best doctor. He describes his eldest brother as treating illness before it becomes apparent, and his other brother as treating it while it is slight. His own interventions are more visible, which helps explain his greater reputation. [1]
What interests me is the relationship between a contribution and its visibility. Correcting a paper or helping solve an experimental problem produces something we can point to. Helping a student recognise a weak assumption before they design the experiment is harder to count. Yet that earlier guidance may influence every subsequent project they undertake.
I would not turn the story into a ranking of teachers, or treat every difficulty as something education should prevent. Research depends on encountering uncertainty. The useful analogy concerns when guidance helps: at the moment of confusion, during the development of a method, or earlier still, as a student learns what to notice and question.
In the imagined meeting, an early question about trial order could save the student from treating every inconsistent result as a reason to redesign the algorithm. The contribution is small and specific: noticing a possible confound before the project becomes committed to an explanation.
Explain enough for the student to make a choice
This is an important part of direct teaching. Students entering a new field need explanations and examples. Giving a useful explanation can help someone ask a better question the next time. There is no educational virtue in withholding the knowledge they need to begin.
For example, in a force-estimation exercise using tactile images, I could ask a student to compare model errors across different contact conditions. They would first say where they expect the model to struggle and why, then inspect whether the results support that expectation. A concentration of errors at particular forces might lead them to examine the training data or the behaviour of the sensor.
AI could help produce the analysis code or suggest a plot. The student's contribution would include deciding what comparison is informative, checking that it has been implemented correctly and interpreting the result. That is a teaching activity I would choose because it joins a prediction to an observation and gives us something substantive to discuss.
Our TacEva work gives this discussion a basis in a real research problem. Vision-based tactile sensors differ in their physical designs and sensing mechanisms. TacEva defines performance metrics and experimental procedures that support quantitative comparison across different sensors, including their operating performance and robustness. [2]
For me, the interesting educational question is how to decide what should be compared. A sensor can be useful for one task and less suitable for another. The choice of measurement and test conditions therefore deserves as much thought as the method used to analyse the data. A student designing an evaluation should be able to explain which capability it measures and what a different laboratory would need to repeat it.
This is also an example of work that helps others ask more precise questions. Its value extends beyond a result for a particular device. For a student, recognising that an evaluation method can be a research contribution is part of learning where useful work is needed.
Leave the meeting with a testable next step
Returning to the two methods tested in separate sessions, I would want a short record of the decision we reached. It need not be a polished report. A few sentences could make clear what the student is trying to find out:
Observation: the relative performance of the two methods changes between sessions.
Possible explanation: trial order or changing experimental conditions may be affecting the comparison.
Proposed check: reverse the order of the methods, or include a reference measurement at the beginning and end of each session.
Interpretation: specify beforehand which result would support that explanation, which would weaken it, and which would remain inconclusive.
Practical agreement: decide what can be completed before the next discussion and where help is needed.
Reversing the order would not by itself establish the cause of every difference. The point is to make one uncertainty easier to examine. If the check remains inconclusive, that is a result to discuss rather than a reason to pretend the problem has been solved.
For routine meetings, I would find a short record more useful than a complete transcript of every interaction with an AI tool. The student could note a consequential suggestion, what they checked and what they eventually decided. That preserves the reasoning we need to examine without turning research into a reporting exercise.
Make room for a well-founded disagreement
Han Yu's essay On Teachers describes teaching in terms of transmitting understanding, imparting learning and resolving doubts. It also recognises that a pupil need not remain less accomplished than the teacher and that expertise differs between people. I read these ideas together: the relationship can support the student's development without requiring the teacher to possess the final answer to every question. [3]
In supervision, I would welcome a student's well-founded challenge to an assumption I have made. I would ask them to explain the concern, examine the evidence with them and discuss how we could investigate it. My responsibility includes giving them a fair hearing and helping turn a disagreement into useful work.
Suppose the student thinks that repeating a reference measurement is more informative than my suggestion to change the trial order. I would ask them to explain what each option could reveal. If their argument is convincing and the test is feasible, the decision should reflect it. That gives intellectual ownership a practical consequence: their reasoning changes what we do next.
A relationship that can change after the degree
One of the contributors to TacEva is a former Master's student of mine who is now pursuing a PhD with a collaborator at another university. We continue to work together. For me, this is a useful example of how an educational relationship can develop beyond the original degree. A student can move to a new institution, acquire expertise elsewhere and bring a different perspective to a shared problem. Continued collaboration can give that independence room to grow.
That is the kind of independence I want doctoral education to support. A graduate should be able to learn from many people and use powerful tools while maintaining a direction of their own. They should also recognise when another person's expertise would help and be able to build a productive collaboration around it.
I hope that, by the end of a PhD, a student can bring me a question I had not considered and explain why they want to pursue it. We may still work together. I may still offer advice. But the investigation can now proceed through decisions they are able to make and defend for themselves.
Sources and further reading
Continue reading
AI assistance was used in developing and editing this article.
Illustrations were generated with AI. People, settings and research sketches are illustrative; they do not depict actual individuals or experimental results.
Join the discussion
What kind of supervision conversation has helped you decide what to investigate next? Leave a comment below, or use the Like button if you found the essay useful.
Conceptual illustrations
The illustrations below were generated with AI. They are conceptual and do not document actual people, events or experimental results.
Figure 1. Separating questions about the apparatus, the measurements and the comparison before choosing the next test. Conceptual illustration generated with AI; the sketches are not experimental data.
Comments