Discussion Solved

Tools and tips for collaborative writing.Overleaf, Google Docs, or Git? (98)

Back to Forum
Starting 0 pts 0 followers
R.V. College of Engineering · Posted
Political Science & Governance Finance & Banking Accounting & Auditing

My Setup

I run a machine learning lab with 8 PhD students. Over the past 18 months, I have noticed a significant shift in how my students (and I) work, largely driven by large language models.

Things that have genuinely improved:

  • Literature search: I use LLMs to get a quick conceptual map before diving into papers
  • Code debugging: GitHub Copilot has reduced boilerplate time by ~30%
  • Writing: Grammar and clarity checks are now instant

Things I am cautious about:

  • Students are sometimes getting confidently wrong answers about recent results
  • Risk of AI-generated text without proper attribution
  • I worry that relying on AI for literature summarization may weaken reading comprehension skills

The question: How are others balancing the productivity benefits against these risks? And what explicit policies has your department adopted for AI use in research?

Sign in to interact

Sign in to join the discussion.

39 Replies

Oldest Top
0
Anita Rao · Accepted answer

The IEEE TPAMI turnaround in my case was 7 months for the first review. Plan accordingly.

0
Replying to Anita Rao
Sowmya Narayanan Active 310 pts · Accepted answer

This is very reassuring. I was worried the null result would kill the paper but your framing around transparency and effect sizes makes a lot of sense.

0
Fatima Sheikh Growing 95 pts · Accepted answer

Flower (flwr) is the most actively maintained and research-friendly federated learning framework right now. My group has been using it for 18 months on a healthcare project with 4 hospital sites.

Strengths:

  • Clean simulation API that lets you test on a single machine before deploying
  • Good support for custom aggregation strategies (FedProx, FedNova)
  • Active Discord community with quick responses from the maintainers

For non-IID data specifically, look at the FedProx strategy.it handles heterogeneous data distributions much better than FedAvg.

0
Replying to Fatima Sheikh
Admin NITK · Accepted answer

This is a known issue with XLM-R on code-mixed data. The MuRIL preprint has a section specifically comparing these.

0
Pooja Nambiar Growing 115 pts · Accepted answer

This is a known issue with XLM-R on code-mixed data. The MuRIL preprint has a section specifically comparing these.

0
Replying to Pooja Nambiar
Hema Suresh Starting 20 pts · Accepted answer

CSIR-UGC NET preparation: past papers are the most important resource. Everything else is secondary.

0
Replying to Pooja Nambiar
Meghana Rao Growing 170 pts · Accepted answer

Strongly agree with the point about choosing venues carefully. One paper in NeurIPS > five papers in obscure workshops.

0
Amit Joshi Starting 15 pts · Accepted answer

Open access is the right direction but the APCs are prohibitively expensive for many Indian researchers without institutional funding.

0
Farhan Mirza · Accepted answer

Check whether your institution has a Springer/Elsevier waiver agreement.many NITs and IITs do.

0
Wajid Khan Growing 110 pts · Accepted answer

Respectfully, I think this framing misses the systemic dimension. Individual researchers cannot solve a collective action problem. If you submit to open-access journals and your colleague submits to Nature, your colleague gets the promotion. Until evaluation criteria change at the institutional level.which requires policy intervention.individual choices have minimal impact.

0
Zara Hussain Growing 55 pts · Accepted answer

Highly recommend checking out the PMRF portal for the actual numbers.they update the stipend structure annually.

0
Prakash Iyer Distinguished 640 pts · Accepted answer

The IEEE TPAMI turnaround in my case was 7 months for the first review. Plan accordingly.

0
Quamar Ahmed Growing 65 pts · Accepted answer

Great question.I went through something very similar in my second year.

The key insight for me was that LoRA is actually quite well-suited for NER tasks, especially in low-resource settings. I would recommend:

  1. Use LoRA with r=8 or r=16.don't go higher for 8K samples
  2. Apply LoRA to attention layers only, not the feed-forward layers
  3. Use a cosine learning rate schedule with warm-up (10% of steps)

For Telugu-English code-mixed NER specifically, you might also look at MuRIL.it is pretrained on Indian language data and often outperforms XLM-R on Indic tasks even with less fine-tuning data.

0
Chirag Mehta Starting 15 pts · Accepted answer

Open access is the right direction but the APCs are prohibitively expensive for many Indian researchers without institutional funding.

0
Replying to Chirag Mehta
Anita Rao · Accepted answer

Thank you for this detailed answer! A quick follow-up: when you mention cosine learning rate with warm-up, are you warming up the LoRA parameters specifically or the entire model including the frozen backbone?

0
Replying to Chirag Mehta
Ravi Patel Starting 25 pts · Accepted answer

Overleaf + GitHub integration is underrated. You get version control and collaboration in one place.

0
Replying to Chirag Mehta
Quamar Ahmed Growing 65 pts · Accepted answer

If you are not already using a reading group format with your students, I would strongly recommend it. Forces everyone to articulate what they read.

0
Bhavana Reddy Starting 30 pts · Accepted answer

Highly recommend checking out the PMRF portal for the actual numbers.they update the stipend structure annually.

0
Gopal Naidu · Accepted answer

Highly recommend checking out the PMRF portal for the actual numbers.they update the stipend structure annually.

0
Varsha Pillai Growing 60 pts · Accepted answer

This question comes up a lot. The answer really depends on your specific field and what your committee values.

0
Replying to Varsha Pillai
Yogesh Pandey Growing 70 pts · Accepted answer

This is exactly what I needed to hear. One clarifying question: you mentioned FedProx handles non-IID data better.does that mean I should switch to FedProx by default, or only if I observe poor convergence with FedAvg first?

0
Replying to Varsha Pillai
Xena D'Souza Starting 25 pts · Accepted answer

If you are not already using a reading group format with your students, I would strongly recommend it. Forces everyone to articulate what they read.

0
Replying to Varsha Pillai
Gopal Naidu · Accepted answer

This is exactly what I needed to hear. One clarifying question: you mentioned FedProx handles non-IID data better.does that mean I should switch to FedProx by default, or only if I observe poor convergence with FedAvg first?

0
Priya Nair · Accepted answer

I have a slightly different take from my experience in industry research. The reproducibility crisis is real but unevenly distributed. Fields with strong engineering culture (computational biology, ML with benchmarks) have actually improved significantly in the last 5 years. The bigger problem is in fields where data sharing is structurally difficult.clinical medicine, behavioral economics.

The ML community's move toward open code and reproducibility checklists has been genuinely effective.

0
Replying to Priya Nair
Vikram Bhatia Active 420 pts · Accepted answer

If you are not already using a reading group format with your students, I would strongly recommend it. Forces everyone to articulate what they read.

0
Replying to Priya Nair
Aditya Sharma Starting 45 pts · Accepted answer

This is very reassuring. I was worried the null result would kill the paper but your framing around transparency and effect sizes makes a lot of sense.

0
Hema Suresh Starting 20 pts · Accepted answer

I can answer your question about IEEE Transactions review timelines from my experience submitting to IEEE TNNLS.

For Q1 IEEE journals, expect:

  • Initial desk review: 2–4 weeks
  • Full peer review: 3–6 months
  • Major revision decision: another 2–3 months for re-review

As for code and data, most IEEE journals now request code upon acceptance, not submission. Regarding your 2.3% improvement.that is borderline. You will need very strong experimental rigor: multiple runs, confidence intervals, and significance tests.

0
Replying to Hema Suresh
Kiran Yadav Growing 130 pts · Accepted answer

Thank you for this detailed answer! A quick follow-up: when you mention cosine learning rate with warm-up, are you warming up the LoRA parameters specifically or the entire model including the frozen backbone?

0
Replying to Hema Suresh
test tester Growing 165 pts · Accepted answer

Seconding the recommendation for Zotero. Game changer for managing references across multiple projects.

0
Replying to Hema Suresh
Prakash Iyer Distinguished 640 pts · Accepted answer

+1 to everything said above. My experience was identical.

0
Vikram Bhatia Active 420 pts · Accepted answer

I am in a very similar situation. Would you be willing to share the outline of your PMRF proposal? Not the content.just the section headings and approximate word allocation.

0
Replying to Vikram Bhatia
Shiva Prasad Active 480 pts · Accepted answer

The IEEE TPAMI turnaround in my case was 7 months for the first review. Plan accordingly.

0
Replying to Vikram Bhatia
Bhavana Reddy Starting 30 pts · Accepted answer

Highly recommend checking out the PMRF portal for the actual numbers.they update the stipend structure annually.

0
Replying to Vikram Bhatia
Asha Pillai Distinguished 880 pts · Accepted answer

This is exactly what I needed to hear. One clarifying question: you mentioned FedProx handles non-IID data better.does that mean I should switch to FedProx by default, or only if I observe poor convergence with FedAvg first?

0
Farhan Mirza · Accepted answer

Great question.I went through something very similar in my second year.

The key insight for me was that LoRA is actually quite well-suited for NER tasks, especially in low-resource settings. I would recommend:

  1. Use LoRA with r=8 or r=16.don't go higher for 8K samples
  2. Apply LoRA to attention layers only, not the feed-forward layers
  3. Use a cosine learning rate schedule with warm-up (10% of steps)

For Telugu-English code-mixed NER specifically, you might also look at MuRIL.it is pretrained on Indian language data and often outperforms XLM-R on Indic tasks even with less fine-tuning data.

0
Prakash Iyer Distinguished 640 pts · Accepted answer

Thank you for this detailed answer! A quick follow-up: when you mention cosine learning rate with warm-up, are you warming up the LoRA parameters specifically or the entire model including the frozen backbone?

0
Rajesh Kumar · Accepted answer

This is exactly what I needed to hear. One clarifying question: you mentioned FedProx handles non-IID data better.does that mean I should switch to FedProx by default, or only if I observe poor convergence with FedAvg first?

0
Replying to Rajesh Kumar
Disha Malhotra Starting 8 pts · Accepted answer

This is very reassuring. I was worried the null result would kill the paper but your framing around transparency and effect sizes makes a lot of sense.

0
Farhan Mirza · Accepted answer

I have a slightly different take from my experience in industry research. The reproducibility crisis is real but unevenly distributed. Fields with strong engineering culture (computational biology, ML with benchmarks) have actually improved significantly in the last 5 years. The bigger problem is in fields where data sharing is structurally difficult.clinical medicine, behavioral economics.

The ML community's move toward open code and reproducibility checklists has been genuinely effective.