Research update: Aiming for RLC

2 views

Skip to first unread message

Ram Rachum

unread,

Jan 3, 2026, 9:52:42 AM (5 days ago) Jan 3

to ram-rachum-res...@googlegroups.com

Hi everyone, and happy new year!

Experience Breakdown paper

I'm working on my explainability project which I renamed to Experience Breakdown. I've decided to target RLC 2026 which will take place in August in Montreal. Deadline is March 5th, which is plenty of time assuming nothing goes wrong. I'm excited about RLC because I know that they have an innovative review process.

I haven't been working on the actual experiments for the last month, because I've been distracted with other things :( I think that in two weeks I'll start working on LLM experiments again.

I started iterating on writing the paper. I've written the abstract, method section and some of the background section. I have the general structure with summaries of the sections I haven't written yet.

Building tools with Claude Code

I've been building a lot of tools with Claude Code lately. The productivity boost is hard to overstate. It has completely changed how I work.

Some examples of what I've built recently: A unified search tool that queries my emails, files, GitHub repos (including issues and PRs), and past Claude Code conversations. It uses embeddings for semantic search, so I can ask "that email about the conference deadline" and it finds relevant results even if I don't remember the exact words. The underlying database has about 200,000 indexed items now, and queries take a few seconds. I also built a service manager for all the background processes I run, a process inspector for debugging stuck Python programs, and many more tools. I also like to have Claude Code build TUIs for these programs.

I should mention that building tools is somewhat addictive. There's satisfaction in having exactly the tool you need, built exactly the way you want it. But I have to be careful not to spend all my time building tools instead of doing the actual research.

Literature review infrastructure

The other big project I've been working on is infrastructure for doing literature reviews. The goal is to make it easy to explore a research area starting from a few seed papers.

Here's the core workflow: I start with a paper I care about. The tool downloads the PDF and LaTeX source from arXiv, parses the bibliography to find all the papers it cites, and adds those to a queue. Then it does the same for each of those papers, building out a citation graph. I can also go backwards: for any paper, find all the papers that cite it (using OpenAlex's API). This gives me a comprehensive view of what a paper builds on and what it influenced.

Here is a screenshot of its TUI:

Parsing bibliographies from LaTeX is difficult. There are multiple formats (natbib, biblatex, plain bibtex), multiple ways to specify references, and lots of edge cases.

The grand vision for this project goes beyond just collecting papers. I want to build concept graphs: for each paper, extract the key concepts it introduces or uses, and link them across papers. "This paper introduces advantage estimation, that paper builds on it, this other paper critiques it." The raw materials are there (I have the full text of every paper), but extracting structured concepts from unstructured text is hard. I'm thinking about using LLMs to help with this, but I haven't gotten there yet.

For now, the tool is useful for quickly building up a reading list around a topic. I point it at a few seed papers and it gives me a comprehensive bibliography. That's already a significant improvement over manually chasing citations.