An AI Research Agent Scores 18.6% at the One Job Research Agents Exist to Do. It's Also the Best Model on the Leaderboard.

18.59%. That's how often the top-ranked AI research agent on a new benchmark correctly combines evidence from multiple scientific sources into one structured answer — the single task a research agent exists to perform. The same system, OpenAI's Deep Research, posts the highest overall score of any model tested: 49.39%. Isolate the step that actually requires synthesis instead of retrieval, and the leaderboard's best performer barely beats one answer in five. Human experts hit a related wall on the same benchmark: accuracy on multi-step database navigation falls from 66.7% at two reasoning hops to 25% at six. Finding information was never the hard part. Combining it is.

A benchmark built to separate finding from combining

Most AI-for-science benchmarks measure a single skill at a time: can a model find the right paper, answer a question about a dataset, or write correct code. SciExplore (arXiv 2607.20926, Tang, Fang, Sun, Liu, Zhang, Liu, Liu, Zhang, and Chen, submitted July 23) instead chains four task types together on the same 103 expert-curated tasks across more than ten scientific disciplines, deliberately increasing in difficulty: T1 tests multi-hop navigation through a structured scientific database; T2 tests retrieval of a specific paper from an ambiguous, underspecified description; T3 tests completing a citation or reference an earlier step left missing; and T4, the final and hardest stage, tests whether an agent can integrate what it found across multiple sources into one coherent, structured answer. The paper evaluated more than ten frontier models and agents, including OpenAI Deep Research, on all four.

The headline number is an average across four very different jobs

OpenAI Deep Research's 49.39% overall average is the highest score any evaluated system posts on SciExplore. Reported alongside it, in the same table, is a blunt summary from the authors: performance degrades sharply as task complexity increases, with "extremely low accuracy on the most challenging structured synthesis tasks" across every system tested, not just the leader. An average that blends an easier retrieval-style task with a much harder synthesis task will always look better than the synthesis task alone. The number worth reading is not the average. It's what happens when you isolate the step where the actual value of a research agent is supposed to show up.

Isolate synthesis, and the best model in the world clears one cell in three

T4 asks a model to pull facts scattered across multiple source documents into a single structured table or answer — the exact operation a working scientist does by hand when reconciling numbers from three papers into one comparison. Graded at the level of individual cells, OpenAI Deep Research gets 30.14% right. Graded at the level of a full, usable row — the unit that actually has to be correct for the answer to be trustworthy — that same model scores 18.59%. The paper's own framing of the gap is direct: partial, cell-level correctness "does not scale to coherent structured outputs." Getting three of ten facts right in a row is not partial credit toward a usable answer. It's a wrong answer with some right numbers buried in it.

Even human experts hit a version of the same wall

One honest objection to a low AI score is that the task itself might just be unreasonably hard, for anyone. SciExplore checks this directly with a human baseline on T1, its multi-hop navigation task, varying the number of reasoning hops required. Human expert accuracy starts at 66.7% for a 2-hop lookup and falls steadily as the hop count rises: 60.0% at 3 hops, 50.0% at 4, 37.5% at 5, and 25.0% at 6 hops — a 41.7-point collapse as the chain of reasoning steps lengthens.

That's a genuine data point in the benchmark's favor: some of this difficulty is structural, not an AI-specific failure. But it also sharpens the comparison rather than excusing it. The hardest human condition tested, 6 hops of pure navigation, still beats the best model's row-level synthesis score by more than 6 points — and synthesis is a different, arguably less mechanical task than navigation. The task getting harder for humans too doesn't explain why the leading AI system does worse than a human at their worst measured condition.

What "research agent" pitches leave out

Nearly every autonomous-research-agent product demo emphasizes the parts of the pipeline SciExplore's T1 through T3 measure: searching a huge corpus fast, tracking down an obscure citation, filling in a reference a human missed. Those are real, useful capabilities, and the benchmark's own numbers suggest models are meaningfully better at them than at T4. But a working scientist doesn't need an assistant that finds ten papers. They need one that can tell them, correctly, what those ten papers say when read together — a table that reconciles conflicting units, a synthesis of three methods sections into one comparable summary, a claim checked against multiple sources at once. That is exactly the operation SciExplore isolates, and exactly the operation where even the best-scoring model on the leaderboard drops to 18.59%.

What this means for reproducible, local-first science

SciExplore doesn't test MegaBrain BioScience, and its numbers don't transfer to any specific product — it's a benchmark of general-purpose research agents as a category. But the generalizable lesson lands on a distinction this blog keeps coming back to: retrieval and synthesis are not the same skill, and a benchmark (or a vendor) that reports only the blended average is hiding the number that actually matters. A synthesis step that has to reconcile numbers from multiple sources is exactly the kind of operation that benefits from being handed to a verifiable, re-runnable computation rather than free-text reasoning alone — and from a result that ships with a reproducibility record showing which sources fed which output, so a wrong row is traceable instead of invisible. The fix for a synthesis cliff isn't a bigger model averaging over more tasks. It's making the synthesis step itself checkable.