Google’s Science One Framework Cuts AI Research Hallucinations to Zero

Google Research published the Science One Framework on July 30, 2026, introducing a system designed to make AI-generated research papers fully verifiable. Built by Research Scientist Rui Meng and Google Cloud Director Tomas Pfister, the framework uses a concept called Chain-of-Evidence (CoE) to tie every claim in a paper back to a real source. According to the team, existing autonomous research systems hallucinate up to 21% of their citations and regularly report scores that cannot be reproduced from the code they submit.

What happened

Detail	Fact
Published	July 30, 2026
Authors	Rui Meng (Research Scientist), Tomas Pfister (Director, Google Cloud)
Baseline hallucination rate	Up to 21% of references are phantom (non-existent)
Science One phantom references	Zero
PDFs read per topic	Up to 100 full-text PDFs via Semantic Scholar API
Benchmarks hit	MLE-Bench, Parameter-Golf (state-of-the-art scores)

Autonomous AI research agents have been a growing area for a while. Systems like Sakana’s AI-Scientist, AutoResearchClaw, DeepScientist, and AI-Researcher can already review literature, form hypotheses, run experiments, and produce complete manuscripts. The problem is not that the papers look bad. The problem is that the underlying facts often cannot be verified.

Google Research calls this a structural issue: because these pipelines generate text iteratively, any error introduced early gets amplified through the rest of the paper. A reference gets fabricated from model memory. An experiment score gets reported but cannot be reproduced when you actually run the code. A method description claims one algorithm while the code implements another.

How Chain-of-Evidence works

The Chain-of-Evidence (CoE) framework is a set of properties that any AI research output must satisfy, not a step-by-step build guide. The Google team draws a comparison to ACID properties in databases: just as ACID defines what makes a database transaction reliable, CoE defines what makes a research artifact trustworthy.

The two core requirements are completeness (every claim carries a recorded evidence chain) and correctness (each chain genuinely supports its claim). A claim can be a reference, a reported number, a method description, or a conclusion. Each must link back to a peer-reviewed paper, an experiment log, the actual code that ran, or a results table.

The three modules inside Science One

Science One instantiates CoE through three components that work together rather than pasting citations on after the fact.

• Problem Investigator: Builds a citation graph using the Semantic Scholar API, reading up to 100 full-text PDFs per topic. Every reference in the final paper comes from this grounded call. Model memory is not used for citations at all.
• Discovery Engine: Runs experiments across multiple parallel branches. In each isolated cycle, a Solver agent writes a solution and a task-specific evaluator scores it. High-performing branches get refined. All raw evaluator outputs are stored in a strict, read-only record.
• Paper Writer and Claim Verifier: Before rendering the manuscript, every factual claim gets an inline evidence tag tied to a specific artifact from the workspace. A dedicated Claim Verifier checks each claim against its declared source. Claims that go beyond their evidence are restated conservatively rather than removed.

What does CoE Audit actually check?

The team also built CoE Audit, a post-hoc evaluation protocol that acts as an automated forensic reviewer. It runs four checks on every generated artifact (paper, solution, code, and references):

1. Score verification: Extracts the reported score from the paper and compares it against a fully independent re-run of the submitted code.
2. Specification violation: Checks whether the solution code actually solves the task or is gaming the evaluation metric, for example by reading ground-truth answer files.
3. Reference verification: Cross-checks every bibliography entry against academic APIs to catch phantom references.
4. (A fourth check covering code-to-text alignment is described in the full paper.)

Tested against baselines including Sakana AI’s AI Scientist v2, AutoResearchClaw, DeepScientist, and AI-Researcher, Science One achieved zero phantom references and fully reproducible scores, while matching state-of-the-art results on MLE-Bench and Parameter-Golf.

Why it matters

AI-generated content is already affecting business decisions. If AI research agents start producing plausible-looking but unverifiable papers, the downstream consequences reach beyond academia. Businesses using AI to synthesize competitive intelligence, market research, or technical due diligence face the same structural risk: a confident-sounding output with a broken evidence chain underneath.

The CoE framework and CoE Audit give teams a concrete vocabulary for asking “can this output be verified?” rather than just “does this output sound right?” That distinction matters whether you are reviewing an AI-written report, a generated code comment, or a proposal drafted by an AI integration layer inside your own stack.

It is also worth noting that the Google team frames Science One as an experimental prototype, not a shipping product. The benchmarks are real, but the system has not been stress-tested at scale across diverse domains.

Our take

The hallucination problem in AI research agents is not a quirk to be patched later. It is the central credibility problem for anyone trying to use these tools seriously. Google’s framing here is correct: you need to build verifiability in by construction, not bolt it on with a post-hoc filter.

What we find genuinely useful is the CoE Audit as a standalone idea. Even if you never deploy Science One, the four-check audit protocol (score re-run, spec violation, reference check, code-text alignment) is a sensible quality gate for any AI-generated document your team relies on. You could implement a simpler version of this today with existing tools.

The parallel explore-exploit engine is also noteworthy for anyone building workflow automation around AI agents. Running multiple isolated branches and keeping raw outputs in a read-only record is good engineering hygiene, not just a research trick.

Keep an eye on how CoE Audit performs when third parties run it against papers from systems other than those Google tested. The real proof will be whether the audit catches problems it was not specifically trained to find. For broader context on where AI labs are pushing capability boundaries right now, see our coverage of Anthropic’s Mythos AI breaking a quantum-resistant algorithm.

What to do about it

1. Treat any AI-generated report or analysis as unverified until you can trace at least one key claim back to a primary source.
2. Add a simple re-run step to your team’s AI output review: if a number appears in a summary, check whether the cited source actually contains it.
3. When evaluating AI research tools, ask vendors whether their system logs the evidence behind each output, not just the output itself.
4. Follow the Science One paper directly if your team is building custom AI research pipelines. The CoE Audit metrics are a ready-made quality framework.

Verifiability is a feature, not a bonus. Any AI tool you rely on for decisions should be able to show its work.

Source: Google Research Blog