Keeping Up With LLM Systems Research Without Losing Your Mind
If you've ever tried to follow the LLM systems literature, you know the feeling: a dozen new papers drop every week, and the ones that actually matter get buried under the ones that don't. You bookmark a few, forget the rest, and six months later you're rediscovering the same ideas. The llmsys-paperlist repository is an attempt to fix that—a curated list of LLM systems papers organized so you can actually see where the field is heading.
What It Does
At its core, this is a curated reading list. It collects academic papers, articles, tutorials, slides, and projects related to Large Language Model systems, organized into a fairly detailed taxonomy. The README covers training (pre-training, post-training and RLHF, fault tolerance), serving (including agent systems, edge serving, and system-model co-design), multi-modal training and serving, LLM-for-systems work, industrial technical reports, frameworks, benchmarks, surveys, MLSys courses, and more.
What sets it apart from a plain link dump is the trends section. The maintainers track the list over time and present a three-era framing from 2024 to 2026, organized around a single question: what unit of work is the system optimizing? The answer shifts from a request in 2024, to a session or reasoning trace in 2025, to a whole agent trajectory in 2026. That framing is backed by charts and per-area paper counts, and the repo includes reproduction scripts under trends/ if you want to dig into the numbers yourself.
Why It's Cool
The trend analysis is the real draw. Most awesome-lists are static—you scroll, you star, you never return. This one treats the list itself as data. Serving is still the biggest area in absolute terms, but its share dropped from 49% to 33% between 2024 and 2026. That growth had to go somewhere, and the README shows where: kernel and model co-design went from 3 to 29 papers, agentic systems from 4 to 22, AI-for-systems from 5 to 16, and edge from 2 to 14.
It tracks techniques, not just topics. The fastest-rising techniques by share of the year's papers are listed explicitly: agentic and multi-agent work (1.0% to 13.1%), compiler/kernel/megakernel (0.0% to 10.2%), speculative decoding (1.0% to 5.7%), sparse attention (1.0% to 4.5%), and energy/power (1.0% to 3.3%). If you're deciding where to spend your reading time, that's a useful signal—and it's the kind of thing you'd otherwise have to assemble yourself from scattered sources.
The taxonomy is practical, not academic. Sections like "LLM for Systems" and "System Efficiency Optimization - Model Co-design" reflect how people actually work on these problems. The training section splits pre-training from post-training/RLHF and fault tolerance, which mirrors how teams are actually organized. There's also a dedicated industrial technical report section, which is often where the most concrete engineering details show up.
It's opinionated about scope. The list deliberately spans papers, slides, tutorials, and projects rather than just arXiv links. That makes it useful whether you're writing a survey, prepping a talk, or just trying to understand what a megakernel is and why everyone started talking about them.
How to Try It
There's nothing to install—it's a reading list, and that's the point.
- Go to the repository: github.com/amberljc/llmsys-paperlist
- Star it. The README explicitly suggests this as the way to keep up with updates.
- Start with the "Trends at a Glance" section at the top to get oriented, then use the Table of Contents to jump to whatever area you care about.
- If you want to verify or extend the trend analysis, look in the
trends/directory for the full writeup, per-technique numbers, and reproduction scripts. - Use the collapsible
<details>sections (like "Before 2024" and "2024" under pre-training) to avoid getting overwhelmed—the list is long, and the folding helps.
Since it's a GitHub repo, the usual workflow applies: fork it if you want your own version, or open a PR if you spot something missing.
Final Thoughts
This isn't a tool you run—it's a resource you return to. It's best suited for graduate students, systems researchers, and engineers who need to stay current with LLM infrastructure work without reading everything. The trend framing is genuinely useful because it gives you a mental model for a field that's moving fast and in several directions at once. The honest caveat: curated lists live or die by their maintenance, and the value here depends on the maintainers keeping the trends updated as 2026 plays out. But the structure they've built—tracking units of work rather than just topics—is a smart lens, and it's worth borrowing even if you never open the repo again.
Follow @githubprojects for more developer tools and open source projects.