ProfBench: Multi-Domain Rubrics requiring Professional Knowledge to Answer and Judge

AI summary

Overview

• Research area: Large language model evaluation, specifically rubric-based benchmarking of long-form professional report generation in specialized domains.
• Technical level: Intermediate. The paper assumes familiarity with LLM benchmarking conventions (Macro-F1 agreement, LLM-as-judge setups, contamination mitigation) but the core idea is explainable without deep technical background.
• Scope: Introduces ProfBench, a 7,347 response-criterion-pair benchmark across Chemistry, Physics, Finance, and Consulting, together with a cheap, low-bias LLM judge to score model-generated professional reports.

What This Paper Is About

Most LLM benchmarks can only test tasks with easily verifiable answers, like math problems, code that passes unit tests, or multiple-choice questions. This narrows evaluation to "exam-style" problems and leaves out the messy, open-ended work that real professionals actually do, such as writing an investment memo, a research report, or a consulting recommendation. ProfBench addresses that gap by building a benchmark where expert humans (PhD holders and MBAs) wrote the grading rubrics themselves, so that model responses to realistic professional tasks can be scored on many independent criteria rather than a single correct answer.

Key Contributions

1. ProfBench, a multi-domain expert-rubric benchmark. Over 7,000 response-criterion pairs across 80 tasks split evenly across four professional domains: Physics PhD, Chemistry PhD, Finance MBA, and Consulting MBA. All rubrics were written by domain professionals, not generated by a model.

2. A large-scale evaluation of 40+ models in two roles. The paper measures models both as report generators (producing answers to the professional prompts) and as LLM judges (deciding whether a response satisfies each rubric criterion), enabling analysis by model family, size, open vs. closed weights, and reasoning effort.

3. Bias-aware judge evaluation. The authors define a bias-index that measures how unevenly a judge scores responses from different model families (o3, Grok4, DeepSeek R1-0528) relative to human labels. This explicitly penalizes self-enhancement bias in the overall score.

4. A cheap, competitive judge. The final recommended judge (GPT-OSS-120B with adaptive reasoning effort) matches the best proprietary judge at 78.2% overall while costing roughly 1.68% as much, reducing full benchmark evaluation from around $1,320 (PaperBench JudgeEval) to about $0.70, or roughly $12 per model run using o3.

Main Findings

• Top models still struggle. GPT-5 with high reasoning achieves only 65.9% overall, the best of any report generator. That is comparable to HealthBench difficulty and far below GPT-5's performance on AIME 2025 (94.6%), GPQA-Diamond (87.0%), and SWE-Bench Verified (72.4%).

• Domain difficulty varies sharply. Physics is hardest (49.3%), then Finance (63.8%), Chemistry (70.6%), and Consulting (80.0%). The authors suggest financial reasoning and physical reasoning are relatively under-served by current training data and benchmarks compared to code and math.

• Closed-weight leads, but the gap is uneven. The largest open–closed gap is in Finance (15.0%), moderate in Chemistry and Consulting (roughly 9%), and nearly absent in Physics (under 1%). The authors speculate that open-weight models have over-focused on code and math-adjacent skills, which overlap with physics-style problem solving.

• Bigger models help, but with diminishing returns. Within each family (GPT-4.1, GPT-5, Gemini-2.5, Llama, Qwen), larger models score better, but gains flatten. The jump from Llama-3.1-70B to Llama-3.3-70B (+3.4%) beats the jump to Llama-3.1-405B (+0.9%), pointing to post-training recipe quality over raw scale.

• Thinking helps judges more reliably than generators. Turning on reasoning improves judge agreement consistently and improves generator scores by only 0.3% to 2.3%. For the judge role, more thinking raised alignment with human labels but also raised bias toward specific providers.

• Reasoning rubrics dominate the benchmark. 62.9% of criteria test logical validity and correctness, 34.1% test factual extraction or recall, and only 3.0% test style and formatting. This shows the benchmark's emphasis is on substance rather than presentation.

• Response length has limited effect. Longer answers correlate with covering more criteria up to a point, but beyond a threshold the benefit vanishes. GPT-5 with low verbosity lost only 3.0% overall while producing 30.7% shorter responses, and high verbosity gained just 1.1% for 32.4% more text.

• Human annotators agree strongly. A validation subset of 1,127 response-criterion pairs re-annotated by additional domain experts produced a Fleiss' kappa of 0.912, indicating the rubrics and labels are reliable.

• Judge-predicted scores track human scores closely. Using the recommended weighted scoring scheme, judge-predicted model performance differed from human-annotated ground truth by only 0.7% to 1.3% across the three reference models.

Methodology in Plain English

The authors started by recruiting professionals with genuine domain credentials rather than general annotators. Thirty-eight annotators from eight countries contributed, 44.7% holding PhDs and 18.4% holding MBAs, with an average of 5.24 years of post-degree experience. Each annotator was limited to five tasks, spent 10 to 20 hours per task, and worked through three stages: coming up with a hard realistic prompt (something they might hand to a junior colleague), writing 15 to 60 grading criteria for that prompt, and scoring sample model responses criterion by criterion. Reviewers checked every stage, and 41.4% of criteria were sent back for improvement at some point. No LLMs were allowed at any stage of annotation.

To make grading affordable and fair, the evaluate-by-rubric step is framed as a binary entailment decision: given a response and one criterion, does the response satisfy it? Following prior work, the judge is not shown the original prompt, because criteria are meant to be independently interpretable. Judges are scored on Macro-F1 agreement with human labels, then penalized by the bias-index, which is the spread between the highest and lowest mean scoring deviation a judge shows for responses from o3, Grok4, and R1-0528. This discourages judges that boost their own model family.

For report-generation evaluation, the authors provide grounding documents alongside each prompt, mirroring how professionals actually work, and score each generated report on the weighted fraction of criteria satisfied. Importance weights range from 1 for additional criteria to 4 for critical ones. Each generation is repeated 16 times to estimate variance.

Two steps protect the benchmark's future use: only half the data is released publicly, with the other half held back as a private set to limit test contamination, and the final judge was selected by testing many candidates for both accuracy, bias, and monetary cost.

Why This Matters

Impact on research. ProfBench extends rubric-guided evaluation beyond health and machine learning into physics, chemistry, finance, and consulting, which are domains with real economic value but weak measurement infrastructure. By releasing a bias-controlled, low-cost judge recipe, it lowers the barrier for smaller labs to evaluate rubric-following behavior, and by holding back half the data, it gives future evaluations a contamination-resistant anchor. The strong human agreement (kappa 0.912) also validates rubrics as a trustworthy reward signal, which matters for reinforcement learning from verifiable rewards where such signals are scarce.

Real-world applications:

• Financial analysis and investment research. Models can be stress-tested on tasks like assessing whether a financing mechanism such as IFFIm is a viable blueprint for funding global health, including a review of past bond issuances and investor participation.
• Scientific research support. Physics and chemistry PhD-level tasks measure whether models can synthesize literature and produce defensible technical reports rather than just recall facts.
• Consulting and strategy deliverables. The consulting track measures whether models can produce structured, well-reasoned recommendations in the style clients expect, moving beyond bullet-point summaries.
• Automated grading of professional training and certification. The rubric-plus-judge setup could be adapted to evaluate junior analysts, graduate students, or trainees at scale, where expert human grading is too expensive.

Industry relevance. The Finance gap between open-weight and closed-source models is the largest in the benchmark, which gives enterprise adopters concrete evidence about where proprietary APIs still buy a meaningful advantage. Meanwhile the judge cost reduction from roughly $1,320 to $0.70, and the finding that a 120B open-weight model matches the best proprietary judge, has direct implications for companies that want to run rigorous internal evaluations without large API bills.

Future Directions

• Closing the Finance and Chemistry gap in open-weight models. The 15.0% Finance gap and 9.2% Chemistry gap suggest under-representation of these domains in current training and post-training data, raising the question of how to curate such data deliberately.
• Understanding why post-training recipes beat scale. The Llama-3.1-70B to 3.3-70B result (+3.4%) outrunning the jump to 405B (+0.9%) implies there is unexplored room in post-training and data quality rather than parameter count.
• Reducing judge bias further. Reasoning effort improved agreement but also increased provider-specific bias, so the trade-off between alignment and fairness is unresolved. Better debiasing methods or ensembling across providers is an open direction.
• Determining how many inference samples are needed per task. The paper shows task-level variance differs substantially, so an open question is whether adaptive sampling can cut evaluation cost without losing robustness.
• Extending rubrics to more professional domains. The framework is domain-agnostic in principle, so law, medicine beyond physician-patient conversation, and engineering design are natural next additions.

Target Audience

The paper is most useful to LLM evaluation researchers and benchmark designers, particularly those working on rubric-based reward signals or reinforcement learning from verified rewards. Applied AI teams at companies in finance, consulting, and the physical sciences will find the domain breakdowns and the cost figures directly relevant to model selection and internal evaluation. Data annotation and dataset construction practitioners will benefit from the detailed recruitment and quality-control protocol. The paper is less suited to readers seeking a general introduction to LLMs, since it assumes comfort with benchmark terminology such as Macro-F1, entailment-based judging, and contamination controls.

Authors’ abstract

Evaluating progress in large language models (LLMs) is often constrained by the challenge of verifying responses, limiting assessments to tasks like mathematics, programming, and short-form question-answering. However, many real-world applications require evaluating LLMs in processing professional documents, synthesizing information, and generating comprehensive reports in response to user queries. We introduce ProfBench: a set of over 7000 response-criterion pairs as evaluated by human-experts with professional knowledge across Physics PhD, Chemistry PhD, Finance MBA and Consulting MBA. We build robust and affordable LLM-Judges to evaluate ProfBench rubrics, by mitigating self-enhancement bias and reducing the cost of evaluation by 2-3 orders of magnitude, to make it fair and accessible to the broader community. Our findings reveal that ProfBench poses significant challenges even for state-of-the-art LLMs, with top-performing models like GPT-5-high achieving only 65.9% overall performance. Furthermore, we identify notable performance disparities between proprietary and open-weight models and provide insights into the role that extended thinking plays in addressing complex, professional-domain tasks. Data: https://huggingface.co/datasets/nvidia/ProfBench and Code: https://github.com/NVlabs/ProfBench and Leaderboard: https://huggingface.co/spaces/nvidia/ProfBench