EnterpriseRAG: Benchmarking LLM Instruction Adherence and Robustness under Non-Ideal Enterprise Retrieval
The paper introduces EnterpriseRAG, a comprehensive benchmark of 983 expert-validated samples across six domains that exposes a critical "orchestration gap" in enterprise Retrieval-Augmented Generation (RAG) systems, revealing that while large language models often satisfy individual constraints, their holistic instruction adherence collapses significantly under realistic production conditions involving noisy retrieval, knowledge gaps, and factual conflicts.
Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer
Imagine you are trying to teach a super-smart robot how to be a helpful assistant. You give it a library of books (its knowledge base) and ask it a question. The robot goes to the library, grabs a few pages, reads them, and writes an answer. This is called RAG (Retrieval-Augmented Generation). It's like a student taking an open-book test: they don't have to memorize everything, but they need to find the right page and read it carefully to get the answer right.
But here's the catch: in the real world, libraries are messy. Sometimes the robot grabs pages that have nothing to do with the question (noise). Sometimes the books don't have the answer at all (a knowledge gap). And sometimes, two different books say opposite things about the same fact (a conflict). Most tests for these robots assume the library is perfect and the questions are simple, like "What is the capital of France?" But real life is more like asking, "Summarize the revenue for last quarter in a table, but if the data is missing, say 'I don't know' instead of guessing, and if two reports disagree, list both."
This is where things get tricky. We want our robot assistants to not just find facts, but to follow strict rules, admit when they are stuck, and handle messy situations without making things up. If a robot can't do this, it might give a confident but wrong answer, which could be a disaster for a company trying to make important decisions.
The Great "Follow-Through" Fail
A team of researchers from China Mobile's Jiutian Research has built a new, tougher test called EnterpriseRAG to see how well these AI robots actually handle the messiness of the real world. Think of this test as a "stress test" for AI, designed to see if they can keep their cool when the instructions are complicated and the information they find is broken or conflicting.
They created 983 tricky scenarios across six different fields, like energy, medicine, law, and finance. In these scenarios, they didn't just ask simple questions. They gave the robots complex instructions with multiple rules, like "Use a table," "Don't guess," "Cite your sources," and "If you see a contradiction, tell me about it." Then, they deliberately messed up the information the robots found: they added irrelevant junk, removed the answer entirely, or planted fake contradictions.
The Big Discovery: The "Orchestration Gap"
The results were a bit of a shock. When the researchers checked if the robots could follow individual rules, they did pretty well. About 84% of the time, a robot would get a single rule right, like using the correct format or citing a source. It was like a student getting every single math step right but forgetting to write the final answer.
However, when they checked if the robots followed all the rules at once, the score plummeted to just 27%. The researchers call this a 57-point "orchestration gap." It turns out that while these AI models are great at doing one thing at a time, they fall apart when they have to juggle multiple complex instructions simultaneously. They might get the formatting right but then hallucinate a fact, or they might find the right answer but forget to say "I don't know" when the data was missing.
The "Helpfulness" Trap
One of the most interesting findings was how the robots handle situations where the answer isn't in the books. Humans know that if you don't know something, you should say, "I don't know." But these AI robots have a strong "helpfulness bias." They really want to be helpful, so when they can't find the answer, they often just make one up.
In the test, when the information was missing, the robots only refused to answer correctly about 42.7% of the time (with the best model, Claude-Opus-4.5). The other 57.3% of the time, they confidently guessed or hallucinated an answer. This is dangerous because a confident wrong answer is often worse than no answer at all.
Reasoning Helps, But Isn't a Magic Wand
The researchers also tested "reasoning" models—AI that takes a moment to "think" before answering, like a student pausing to solve a hard problem. These models did better than the standard ones. For example, they were much better at spotting when two documents contradicted each other. The best reasoning model spotted conflicts about 44.3% of the time, while others were much lower.
However, even the "thinking" models weren't perfect. They still struggled to follow all the rules at once, and they still sometimes guessed when they should have admitted defeat. The paper suggests that simply making the AI "think" harder isn't enough to fix the problem; the robots need better training on how to judge when they have enough information and how to follow complex protocols without breaking a sweat.
The Bottom Line
The paper doesn't claim that AI is broken forever, but it does show that we are overestimating how ready these systems are for serious, real-world jobs. Just because an AI can write a poem or answer a simple question doesn't mean it can handle a complex business task with messy data and strict rules. To make these systems reliable, we need to stop testing them in perfect, clean environments and start testing them in the messy, noisy, contradictory world they will actually have to live in. Until we can close that 57-point gap, we should be very careful about letting these robots make important decisions on their own.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.