Research map / Multi-agent systems and games / LLM agents

EnterpriseBench: Evaluating LLM Agents on End-to-End Spreadsheet Tasks in Finance

NeurIPS 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent collaboration and safety

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces MBABench, a benchmark of end-to-end financial spreadsheet construction tasks graded by a rubric-based LLM judge (Accuracy, Formula, Format), evaluating 18+ agents across GUI, CODE, and API harnesses.

Problem

Existing spreadsheet benchmarks evaluate atomic tasks (QA, single formula edits) via exact match, not end-to-end construction of complete multi-sheet financial models whose quality includes readability and modifiability.

Approach

A three-dimension quality taxonomy (Accuracy, Formula, Format) with fine-grained sub-criteria, graded by an LLM judge (Gemini-2.5-pro) validated against expert annotations, applied to end-to-end financial workbook tasks from FMWC, ModelOff, and WSP, with agents accessed via GUI, CODE, and API harnesses.

Key results

• Judge agrees with 408 expert annotations at 0.92 accuracy, 0.88 balanced accuracy, and 0.85 F1 on agent attempts.
• Top agents score 72.6/100 at level-4 tasks (Codex) versus 92.4 at level 2 (Claude Code), showing performance degradation with difficulty.
• GUI and CODE agents on proprietary harnesses substantially outperform API agents built on the same underlying models, suggesting harness quality drives much of the gap.

Surprising findings

• Excel-native agents (Claude for Excel, ChatGPT for Excel) trail Claude Web in spreadsheet quality and often hardcode values instead of writing formulas.
• Leading API models perform similarly under the in-house harness, while differing more on proprietary harnesses, suggesting harness matters more than model.

Limitations

• Benchmark lacks many of the hardest level 5 and 6 tasks.
• Cannot determine the source of performance differences because proprietary harnesses are opaque.
• Criteria weighting is subjective and some criteria (e.g., number notation) are finance-specific.

Experimental setup

Datasets: FMWC tasks ModelOff tasks Wall Street Prep training tasks SpreadsheetBench

Benchmarks: MBABench

Metrics: composite score accuracy balanced accuracy F1 score completion time

Baselines: SpreadsheetBench GDPval ChatGPT Agent Claude (Excel) ChatGPT (Excel) Codex Claude Code

Builds on: SpreadsheetBench GDPval ModelOff

Closest papers by meaning

• BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows (NeurIPS 2026 · similarity 0.91)
• FinSearchComp: Towards a Realistic, Expert-Level Evaluation of Financial Search and Reasoning (ICLR 2026 · similarity 0.87)
• FinBen: A Holistic Financial Benchmark for Large Language Models (NeurIPS 2024 · similarity 0.87)
• InfiAgent-DABench: Evaluating Agents on Data Analysis Tasks (ICML 2024 · similarity 0.86)
• FinRpt: Dataset, Evaluation System and LLM-based Multi-agent Framework for Equity Research Report Generation (AAAI 2026 · similarity 0.86)
• InsightBench: Evaluating Business Analytics Agents Through Multi-Step Insight Generation (ICLR 2025 · similarity 0.86)