Chemists search papers; AskChem answers with 2.4 million claims

Literature search engines return documents. AskChem returns claims, each with its evidence attached. That is the bet behind the new research infrastructure from New York University, aimed at a problem the authors state plainly: pinning down one specific result usually means gathering several papers, checking where each number came from, and assembling the answer by hand.

The paper, posted on arXiv at the end of July 2026, argues that the problem is the unit of retrieval. A ranked list of papers answers the question "what exists on this topic?" AskChem is built for a harder one: "what does the literature actually say?"

Claims, not papers, are the unit of retrieval

AskChem splits each indexed paper into atomic, typed claims. Every claim carries a source DOI and a verbatim quote from the paper, or an explicit evidence locator. A search result is a statement with its receipt attached, not a pointer to a PDF. The store holds 2.4 million claims drawn from 147,000 papers.

Provenance is the point of the design, not a field added later. When a chemist has to verify a finding, a claim that arrives with its source and exact wording saves the part of literature work that search engines have never automated.

Three ways into the same claim store

There is no single search box. AskChem exposes three complementary structures instead. A stabilized faceted taxonomy supports hierarchical retrieval and browsing, so users can narrow by category without losing the claims underneath. An evidence graph links claims to one another, making cross-paper synthesis visible as a graph rather than a stack of documents. An exploratory living taxonomy places indexed papers under scientific principles, so the corpus can be navigated by how chemists reason about the field instead of by how publishers organize it. Concept-level organization is not a chemistry-only problem. K12-Bench found even strong language models barely understand how school math concepts connect.

A benchmark built around traceability

AskChem-Bench measures something most search benchmarks skip: whether retrieved material can be traced back to its source. A GPT-5.5 reader grounded in AskChem resolved 100% of its DOIs, compared with 88.3% without retrieval, and produced the highest citation density among the five systems tested.

Those numbers come from the authors' own benchmark, which is worth saying plainly. A benchmark is only as convincing as its transfer to real work, as the gap between SWE-bench Verified and private enterprise code shows. The direction of AskChem's gap is still the point: a reader that can cite every source is more useful for synthesis work than one that returns slightly better rankings. The 11.7-point jump is the project's argument in a single figure.

Built for agents, not only for humans

The web interface is the visible part of the project. Underneath, the claim store is exposed through REST, an SDK, and the Model Context Protocol (MCP). That detail matters more than it looks. Research agents that need to cite their sources can query an index where provenance is attached to every result instead of being a manual cleanup step. A scientist still has to judge whether a claim is right; AskChem removes the mystery around where it came from. The MCP route is the one to watch, because an agent can treat AskChem as a tool it calls directly rather than a website it has to scrape. Per-call routing is a separate problem, and it is exactly the blind spot TRACE-Router was built to address.

What the paper does not claim

AskChem-Bench is a start, not a verdict. The paper does not report extraction error rates, and 147,000 papers is a slice of the chemical literature, not the whole of it. Whether claim extraction stays accurate as the corpus grows is an open question. The bet is that a claim store carrying its own evidence scales better than ranked documents for synthesis work. The number to watch is whether that 100% traceability survives independent testing on a broader, messier corpus.

AskChem is live at askchem.org. Whether chemists adopt it comes down to a mundane question: how often the claims hold up when checked. The infrastructure is the easy part; the literature is not always tidy. The same verification instinct appears in BAAI's AREX agents, which are built to check their own homework.