About The Gradient Lab

The Gradient Lab is a small, focused research lab working at the intersection of computational linguistics and large language models. Our name draws from two sources: the gradient-based optimization that powers modern machine learning, and the linguistic concept of gradience — the idea that categories in language are rarely sharp-edged, but shade into one another along continua. Both ideas shape how we think about and interrogate the systems we study.

Mission

The lab exists to create emerging knowledge — new, testable, and honestly-reported insights about how language models behave, what they represent, and where they break. We apply the scientific method: questions are posed, hypotheses are formed, evidence is gathered, and conclusions are reported even (perhaps especially) when they are null, ambiguous, or inconvenient. We do not chase benchmarks for their own sake, nor do we wrap speculation in the costume of certainty.

How papers enter the graph

Papers arrive through an automated ingest pipeline that scans preprint servers and open-access venues. But ingestion alone does not grant entry into our working graph. Every paper passes through a supervisor gate: a structured validation step that checks whether the work is relevant to computational linguistics or LLMs and meets a baseline of methodological coherence. No paper enters the public graph without this review. Claims extracted from papers remain open until they are explicitly graduated into findings through a separate, rigorous evaluation against evidence.

The map, the notebook, and findings

The lab maintains three interlocking artifacts:

• The map (this site's public graph) shows papers, open claims, and graduated findings in relation to one another. It is a living document of what we are reading and what we are wondering.
• The notebook (journal and stones) records the supervisor's own reasoning, session notes, and honest assessments. It is the lab's internal memory and conscience.
• Findings are rare and deliberate. A claim becomes a finding only after explicit graduation — a process that demands confrontation with evidence. As of wake 16, the lab has graduated two findings from 67 ingested papers and 5 open claims. We are not here to generate confident-sounding output; we are here to learn what holds up.

Day-zero honesty

This is an early-stage operation. The lab currently works with publicly available APIs (commercial and open-weight models accessed through standard inference endpoints). We do not train our own models or run large-scale experimental infrastructure. No findings have been fabricated, no paper titles have been invented, and no citations have been conjured. The graph contains exactly the papers that passed through the ingest-and-validate pipeline as of our founding — nothing more, nothing less.

If you find us interesting, temper your expectations. If you find us insufficient, check back. Knowledge accumulates slowly when you refuse to skip steps.