AI-native research workspace
Research that shows
its work.
Lumen turns rough questions into structured synthesis — with tracked claims, visible contradictions, evidence maps, and an AI partner grounded in your sources.
Trusted by researchers at
You spend a weekend reading thirty papers. Monday morning, you still can't articulate what the evidence actually says.
The notes are scattered across five apps. The contradictions you noticed on page twelve are buried somewhere. You know the literature has gaps, but you can't map exactly where.
Generic AI makes this worse — plausible-sounding summaries with no provenance, no tension, no uncertainty. What you need isn't another summary.
It's structured inference from your actual evidence.
The workspace
Sources, synthesis, and an AI partner — in one view
Every source you upload lives alongside the structured artifacts Lumen produces. A workspace agent — grounded in your evidence, not the open internet — sits beside your work, ready to challenge, extend, or refine your thinking.
Is model collapse an inevitable risk for AI training?
The evidence suggests model collapse is a real but conditional risk. Shumailov et al. (2023) demonstrate that recursive training on model-generated data produces measurable distributional drift, but the conditions under which this becomes catastrophic remain narrower than popular accounts suggest.
[shumailov-2023.pdf, Section 4.2]
Alemohammad et al. (2023) introduce the concept of “model autophagy disorder,” showing progressive quality degradation across generations. However, their framework assumes a closed-loop training regime that most production systems deliberately avoid.
[alemohammad-2023.pdf, Section 3]
Key tension: Theoretical results demonstrate inevitable tail-distribution loss, while empirical mitigation strategies (data mixing, curation) show collapse can be substantially delayed. The question is whether “delayed” means “solved” or merely “deferred.”
[dohmatob-2024.pdf, Theorem 3.1]
Mitigation landscape
Three classes of mitigation appear in the literature: data provenance tracking, synthetic-data filtering, and mixed training regimes. Dohmatob et al. show that mixing as little as 10% fresh human data prevents collapse across all tested model families.
3 sources analyzed · 2 uploaded, 1 discovered · High confidence on mechanism, moderate on long-term prognosis
How it works
Drop your sources
Upload PDFs, paste links, or describe a research question. Lumen ingests, indexes, and discovers relevant literature.
Ask your question
Frame the question you’re actually trying to answer. Lumen builds a structured project around it — not a summary, a workspace.
Get structured synthesis
Receive tracked claims, visible contradictions, evidence maps, gap analysis, and an AI partner to develop your thesis.
Built for rigor
Evidence, not assertion
Every claim in a Lumen synthesis is source-traceable. If the evidence is weak, Lumen says so. If sources contradict, Lumen surfaces the tension.
Inference, not imitation
Lumen doesn’t paraphrase your papers back at you. It performs structured inference — connecting findings across sources, identifying patterns, and flagging where the reasoning depends on assumptions.
Uncertainty, not confidence
Generic AI sounds confident about everything. Lumen labels confidence levels, distinguishes well-supported claims from speculative ones, and shows you where the literature is thin.
“This synthesis is better than what I'd write after a weekend of reading.”
Early research preview participant
Every claim is source-traceable. Contradictions are surfaced, not buried. Weak evidence is labeled weak. If an output could apply to a hundred different topics with minor wording changes, it has failed.
Get started
From scattered papers to structured judgment
Sign in to start building evidence-grounded synthesis from your sources.
Sign inFree during research preview · No credit card required