I didn't set out to become someone who writes papers. I came back to academic study to learn — to catch up on a field I'd stepped away from, to read, to understand. What I didn't expect was that reading would turn into wanting.
I went through a lot of papers, some careful and precise, some brilliant in ways their authors probably never got enough credit for. Somewhere in that stack, a thought that had been sitting quietly in my mind — about how confidently multimodal LLMs classify things they've never explicitly been trained on, and how little we could trust the reasoning they gave for it — stopped feeling idle and started feeling like something I had to write down. I wanted to see it exist as a real paper. I wanted, honestly, to one day see someone cite it.
That wanting turned out to be the easy part. What made this one, "Grounding Reduces Citation Hallucination in Multimodal LLM Zero-Shot Classification of IoT-Botnet," stick was a problem I kept running into across the literature: ask a multimodal LLM to zero-shot classify something like IoT-botnet traffic and explain itself, and it will often justify the answer with citations that sound plausible and don't exist. The model was frequently right about the classification and confidently wrong about why. That gap — good at pattern-matching malicious traffic, unreliable about its own reasoning — became the paper: testing whether grounding the model in real reference material could shrink that gap.
What I wasn't prepared for was how much of "writing a paper" has nothing to do with writing. Coming back into research without a current institutional home, the first real wall wasn't methodological — it was administrative. Endorsement requirements assume someone is already vouching for you, a strange loop when you're trying to re-establish yourself. I wrote to authors whose work I'd been reading for months, cold, asking if they'd endorse someone they'd never met. Some didn't reply; some did. I wrote to universities about funding to cover publication costs, learning how much of academic publishing quietly assumes an institutional budget behind you. Each rejection or silence was small, but they added up into a test I had to pass before the research itself even began. I kept writing — more emails, more explanations, more patience than I thought I had. Eventually enough doors opened.
Then came the work I actually thought writing a paper would be. It wasn't linear. Running multimodal LLM experiments across enough conditions to say something meaningful about grounded versus ungrounded classification meant testing across multiple providers, each with its own pricing and quirks. I started with one provider, watched token costs climb faster than I'd budgeted as a self-funded researcher, and reworked the experiment around a cheaper one without changing what it actually measured. Every switch meant re-validating that results still held — that a drop in citation hallucination was really about grounding, not just which model happened to be answering. That paranoia paid off: it's what surfaced the paper's most interesting finding.
Grounding each model's citations in a small, verified reference corpus took citation correctness for Claude Sonnet from 43.0% to 72.6% (n = 1,020), closely replicating an earlier 141-example pilot (44.1% to 69.9%). For GPT-4o-mini, it went from effectively unmeasurable to 57.1%. In both cases, grounding also converted an enormous, unauditable pile of fabricated citations into a small, checkable one — a claim nobody can verify versus one someone actually can.
The twist I didn't expect: the same technique that left Claude's classification accuracy nearly untouched (100.0% to 99.6%) cost GPT-4o-mini 28.1 points, concentrated almost entirely in one category, Mirai, where the model appeared to borrow vocabulary straight from the retrieved reference text instead of its required taxonomy. That divergence held at matched scale, with confidence intervals between models that didn't overlap in any condition — not a small-pilot artifact. The lesson: a grounding strategy validated on one model can't be assumed safe for another without re-testing classification accuracy, not just citation quality, at comparable scale. (A direct ablation test of that vocabulary-leakage hypothesis is the natural next step.)
For a field increasingly leaning on LLMs to triage security threats, that distinction matters. Trusting a model's classification is one risk. Trusting its fabricated justification is another, worse one. And a model quietly getting worse at the task while getting better at citing itself is a third risk nobody was watching for.
If you're coming back to research, or starting late, without the institutional scaffolding that makes this easier: the paperwork and gatekeeping are real, and cost more time than the actual thinking does. Ask anyway. Write to the authors. Write to the universities. Expect silence and don't take it personally. The research itself, once you're finally allowed to sit down and do it, is the reward you were chasing all along.