Can AI judges tell a strong research report from a polished one?

AI research assistants search the web, gather information and assemble lengthy reports. But who checks their work? Expert assessment is costly, especially when a report combines specialist knowledge, numerous claims and extensive references. Delegating that assessment to another AI creates a second problem: the evaluator may mistake fluent presentation for strong research.

In an arXiv preprint, Shudong Liu and colleagues introduce DeepResearch RewardBench to test these evaluators. Their results favor several large language models over conventional reward models, while exposing important weaknesses in both. The study measures agreement with benchmark preferences—not a guarantee that a selected report is factually correct.

Building a test for the judges

A reward model evaluates an AI response, often by assigning a numerical score. Such scores can help select outputs or guide model training. Another approach uses a generative language model as a judge: it reads competing responses and produces an assessment or ranking.

The researchers collected reports answering questions from two existing research benchmarks, DeepResearch Bench and ResearcherBench. Candidates came from dedicated research agents and language models equipped with web search, covering areas including AI, finance and technology.

To establish reference rankings, the team adopted each source benchmark’s evaluation procedure. One assesses comprehensiveness, depth, instruction-following and readability using task-specific criteria and a reference report. The other measures coverage of expert-designed, weighted rubric items—specific points a successful answer should address. These procedures use AI judges to score the reports.

The researchers then filtered candidate groups for factors including similar lengths and relatively close scores, aiming to avoid easy comparisons or a simple preference for longer answers. Human annotators checked the rankings, discussed disagreements and discarded unresolved cases. The appendix reports initial agreement of 63%, rising to approximately 85% after discussion.

The resulting benchmark contains 160 two-report comparisons, 138 three-report comparisons and 33 four-report comparisons: 331 instances comprising 866 report entries. Most reports run between 1,000 and 5,000 tokens, the text units processed by language models.

Choosing a winner is not the same as ranking everyone

The team evaluated 19 models using two questions: can the evaluator identify the highest-ranked report, and can it reproduce the entire reference ordering?

Several large generative models performed better than conventional score-producing reward models. But this was not a universal advantage for every generative evaluator: the smaller Prometheus model performed poorly.

Results also depended on the task. GPT-4.1 matched the preferred answer in about 91.9% of two-report comparisons, yet selected the preferred report in only about 60.1% of three-report comparisons. Qwen3-Instruct reproduced the complete four-report ordering in about 60.6% of cases.

These subsets contain different comparisons, so the figures do not isolate the effect of adding another report to an otherwise identical task. Nevertheless, they show why success at picking between two answers should not be treated as proof of reliable multi-report ranking.

The researchers also compared judging reports together with scoring each report separately. In the models and settings tested, comparative judgment performed better. Independent scores often clustered around values such as 80 or 85, producing ties or incorrect orderings. Without a detailed rubric or anchoring reference, an absolute score could conceal differences that became clearer in direct comparison.

When the judge uses the wrong yardstick

Changing report order shifted performance by as much as seven percentage points in the tested models. Removing URLs had comparatively little effect, although this experiment did not establish whether citations actually supported the reports’ claims.

An analysis of 52 wrong judgments from selected difficult cases found that roughly 63.5% involved a mismatch between the judge’s criteria and the benchmark rubric. For example, a judge might favor recent references and concrete experimental results while the rubric rewards explanation of a particular technical argument. Other failures involved knowledge gaps, overlooked sourcing problems and preferences for surface organization.

Those percentages describe a curated error sample, not all research reports. More fundamentally, disagreement with a rubric is not automatically evidence that every aspect of the losing report is worse.

A useful test, not a certificate of truth

The study remains a preprint. Its dataset draws on only two source benchmarks, with particularly few four-report instances. Reference preferences depend partly on AI scoring, although human verification provides an additional check. GPT-4.1 also had a limited role in early scoring for part of the source data.

The practical finding is therefore narrower than “AI can judge research.” Several models can reproduce these reference preferences reasonably well, especially through comparison. Reliable evaluation still requires explicit criteria and attention to evidence—not confidence in polished prose or a precise-looking score.