John Garger’s Post

Verbalized sampling, a fancy name for asking LLMs to generative creative outputs and unlock model diversity instead of typical boring responses! Asking the right question is as important as writing a good prompt. write “generate five answers with their probability” if you want more creative answers! Read this paper:

#Mitigate Mode Collapse and Unlock LLM Diversity Saeed Al Dhaheri Christopher Manning Mode collapse is a commonly seen phenomena in LLMs that stems from the human (annotators) preferences for a certain type of output. This in turn, constricts diversity thereby restricting the creativity of the model. Hence the model may have issues dealing with prompts that call for creative writing or suggesting innovations. This cognitive preference is termed as typicality bias. One of the way out could be to have a LLM post-trained by another ML model that was trained on synthetic data and without human feedback. We can also use a hybrid model that uses both the previously discussed ML model and RLHF (reinforcement learning from human feedback) together. This may to a large extent mitigate the typicality bias already inherent in the system. #The paper uses “verbalized sampling” as a prompting methodology to output multiple responses instead of a singular one. These multiple responses are mapped against the probabilities thus providing model with diversity derived from the data on which it is trained. This removes the bias to a large extent. We can also leverage verbalized sampling by #infusing in synthetic data which does not carry any bias. These augmented data sets could be augmented (artificial) human responses too that can be used to train models. #Another idea could be to use ensembles of models that use a group of (slave) models to tackle tasks based on specialized domains. This will also promote diversity in its replies. #Alternative could be to use, conditions inside the output mechanism to enforce diversity and choke singular responses. We can also use reinforcement learning to reward creativity and innovation by focusing rewards accordingly. #Yet another idea could be to use new heuristic sampling techniques (search) that promote diverse exploration in the reply.

Why editors miss typos (and what’s actually happening in our brain) You’d think editors catch everything. But here’s the truth: we sometimes miss the most obvious typos — right under our nose. It’s not laziness. It’s not a distraction. It’s your brain doing its job too well. When you read a familiar text, your brain doesn’t process every letter. It predicts what comes next. It autocorrects the world before your eyes can check it. So when you see “teh” instead of “the”, your brain just fixes it silently and moves on. That’s the paradox of expertise: the better you know the language, the more your brain fills in the gaps. So no, editors don’t miss typos because we’re bad at our job. We miss them because our brains are too good at it. And thank god I was trained to always (ALWAYS!) run an automatic spellcheck — even on a 300-word project 😉

Start with “Why” – The Foundation of Every Successful Journal Before you start thinking about submission systems or indexing, take a step back and ask: Why does your journal need to exist? A strong sense of purpose shapes everything that follows: from your editorial policy to how you attract authors and reviewers. Define your journal’s scope clearly, choose your niche wisely, and make sure your mission adds value to the field rather than replicating what’s already out there. The most successful independent journals are built on clarity and consistency, not size or speed. 👉 Begin with your purpose. The rest will follow. Learn more about how JAMS - Journal & Article Management System supports small publishers: https://lnkd.in/eVd53qhH #AcademicPublishing #OpenAccess #ScholarlyCommunication #JAMSforJournals

🎓 How to End Strong: The Power of Conclusion Paragraph Starters ✍️ A powerful conclusion can make or break your academic paper. Professors often jump straight to the conclusion — so what you write there truly matters! Here’s what makes a strong conclusion stand out 👇 ✅ Use clear conclusion starters like “To sum up,” “In essence,” or “Finally” ✅ Restate your thesis and main arguments ✅ Add a thought-provoking final line ✅ Avoid introducing new information ✅ Maintain flow and clarity till the last word Remember — a well-structured conclusion isn’t just a summary, it’s the final impression you leave on your reader. 🧠 📘 Want to master the art of writing impactful conclusions? Read our latest blog: https://lnkd.in/dcbYQi2z #writingskills #conclusionparagraph #assignmenthelp #essaywriting #thesishelp #assignmentdotworld

Stop treating citations like sprinkles on your research sundae. They're the foundation, not the finishing touch. I used to think citing references was just a final checkbox in a manuscript. List. Format. Move on. But over time, I’ve realized—how you cite reveals how deeply you think. Messy, irrelevant, or bloated references don’t just look bad. They undermine your argument and your credibility. Here’s how to do it right 👇 ↓ ↓ ↓ How are you building good references for your paper? --- P.S. Use research boost to get relevant high quality pubmed references. Sign up here to get 5000 words FREE: https://researchboost.com/

How do journal publishing services help streamline the journey from submission to publication? Reliable publishing processes mean faster dissemination of research. At Siliconchips Services, we offer complete journal publishing services that cover submission management, editing, typesetting, and metadata delivery. Key Highlights: - XML-first production and LaTeX support - Integrated peer review and submission tracking - DOI and indexing setup for visibility Efficient workflows turn research into accessible knowledge. Which stage of journal publishing do you think has the most room for innovation? #JournalPublishingServices #AcademicPublishing #PublishingWorkflow 📺 Explore more insights on our YouTube channel: 👉 https://lnkd.in/dFNeC43Q Don't forget to subscribe!

Stop saying "the author"...please ... This post will probably be like Marmite, but seriously, research needs to start sounding human. I've been doing a lot of research for a book I am working on and it never stops amazing me how stale academic language can be. You read paper after paper filled with phrases like "the authors argue" or "the authors suggest"; as if the people behind the words don't actually exist. How did we get to the point where research sounds detached and robotic? Removing "I" or "we" doesn't make our ideas more credible. It makes them harder to read. Research and its dissemination should be about clarity and connection not hiding behind outdated habits. If you wrote it, own it. Be proud of it! Say "I found" "we discovered" or 'our analysis shows". It will make your research more honest and relatable. Readers need to be reminded that knowledge is created by real people. And if the editors or reviewers of the journals where you submit disagree with this, then to hell with that journal. Maybe they don't deserve your amazing work and effort. Just blame 'the author' of this post 🤭

Today I reviewed a paper, and honestly, if I had written it myself, I wouldn’t have been as tired as I got from reviewing it. I pointed out their mistakes and gave suggestions on how to fix them, but I had to reject it because of methodological errors. Please: Make your papers clear and easy for everyone to understand. Always download and read one or two sample papers from reputable journals that match your methodology, so you can become more familiar with the proper reporting and writing style.

Evidence-based. Transfer-focused. Actually doable. That’s what makes authentic assessment worth talking about. Whether you’re just starting to rethink your quizzes or fully redesigning for relevance and rigor, this blog breaks it down into 10 practical steps, backed by research. Because sometimes, the best assessment doesn’t come with a bubble sheet. https://lnkd.in/eAtzQqNR