A deep research agent does not return an answer. It returns a report — pages of structured, confident, well-cited prose on a question that has no single correct response. That is exactly what makes it useful, and exactly what makes it hard to grade. You cannot diff it against a key, because there isn't one. You cannot run a unit test, because "a good market analysis of solid-state batteries" is not a value that equals true or false. So most teams reach for the only tool that scales — hand the report to an LLM judge, ask "how good is this, 1 to 10," and ship the average.
That number is worse than useless, because it answers the wrong question. It tells you how the report reads. It tells you nothing about whether a single sentence in it is true.
The output is a report, so the evaluation has two axes#
The 2026 deep-research benchmarks all converged on the same structural insight: a research report has to be scored on two axes that have almost nothing to do with each other. One is the quality of the writing-and-reasoning. The other is whether the claims are grounded in sources that actually say what the report says they say. A report can ace the first and fail the second, and when it does, it is the most dangerous artifact a research agent can produce: plausible, comprehensive, fluent, and wrong.
Axis one — report quality. DeepResearch Bench (Du et al., 2025) grades this with a framework called RACE: Reference-based Adaptive Criteria-driven Evaluation. The "adaptive criteria" part is the load-bearing idea. Instead of one fixed checklist applied to every report, RACE generates a task-specific rubric for each question and scores the report against a strong reference along four dimensions — Comprehensiveness, Insight/Depth, Instruction-Following, and Readability. A fixed rubric can't fairly compare a report on tax policy to one on protein folding; a regenerated rubric can. The benchmark itself is 100 PhD-level tasks, split 50 English and 50 Chinese across 22 fields, which is small but deliberately hard — and, like BrowseComp and the broader deep-research benchmark field, built specifically for agents that browse, gather, and synthesize rather than answer from memory.
Axis two — grounding. This is the one teams skip, and it splits again into two numbers that get conflated constantly:
Citation accuracy is precision — of the sources you cited, how many actually back the claim. Citation coverage is recall — how many supported facts you surfaced at all. They are not the same axis, and the best agents prove it.
On the DeepResearch Bench leaderboard, Gemini-2.5-Pro Deep Research leads the overall RACE score (48.88) and surfaces the most effective citations — roughly 111 verifiably-supported facts per task. But the highest citation accuracy belongs to Perplexity Deep Research, at 90.24%. One agent finds the most. A different agent is the most trustworthy about what it found. If you collapse those into a single "research quality" score, you erase precisely the distinction a person relying on the report needs most.
Build the rubric from how the tool is actually used#
Perplexity's DRACO benchmark makes the same two-axis split but sources its tasks differently — from millions of real Deep Research production requests, sampled and then run through a five-stage pipeline that strips personal information, adds scope, filters for objectivity and difficulty, and ends in domain-expert review. Each of its 100 tasks carries a bespoke rubric averaging ~40 criteria across four axes: factual accuracy, breadth and depth, presentation quality, and citation quality. The detail worth stealing isn't the leaderboard (Perplexity, unsurprisingly, reports leading three of four dimensions on its own benchmark — read that with the appropriate eyebrow). It's the construction: rubrics grounded in real queries, written by experts, regenerated per task.
That construction also solves the quiet problem with every fixed benchmark: it rots. A frozen answer key is wrong about a fast-moving topic within months (temporal drift), and the moment a benchmark is public, its tasks leak into the next training run and scores inflate (contamination). Regenerating rubrics per task slows the first; keeping the hardest evaluation human — arena-style pairwise voting, as in DR-Arena and the Deep Research Comparator — sidesteps the second. A deep-research eval is not a thing you build once.
What to actually do on Monday#
You will not stand up a 100-task expert benchmark this week. You don't need to. Steal the structure:
- Score quality and grounding separately. Never report one number. A report's readability says nothing about its truth, and one figure lets the first impersonate the second.
- Generate the rubric per task, not once. Have a judge model produce the criteria for this question before it grades the answer. A fixed checklist over-rewards reports that are merely thorough.
- Verify citations by hand, on a sample. Pull ten cited sentences, open the links, and check each source actually supports the claim. Report precision (how many held up) and coverage (how many supported facts appeared) as two columns. This is the cheapest high-signal eval you can run, and almost nobody runs it.
- Run each task more than once. A single flattering trace hides an inconsistent agent. Look at the spread, not the best draft — the same pass@k-versus-pass^k gap that bites tool-use agents bites research agents too.
The temptation with deep research is to trust the artifact because it looks like work. It looks like a report a competent analyst would hand you, so the instinct is to grade it like prose. Grade it like evidence instead. The question is never "is this well-written." The question is "if I act on the third paragraph, will the source hold."
