inferwire
/
AI·5 min read

Naive Graders Missed Most Correct Answers From a Data Agent

Columbia statisticians tested how well automated graders score a data-analysis agent. A last-number parser found 26% of its correct answers; with nudges, an LLM grader found 97%.

TL;DR

  • Columbia statisticians graded a data-analysis agent on 153 tasks; a last-number parser found only 26% of its correct answers, an LLM grader 97% [^1].
  • Benchmark scores mix agent errors with grading errors, so the authors urge checking automated graders against human labels and reporting that audit [^1].

Background

Data-analysis agents take a question and a data file, write and run code, and report an answer wrapped in code, numbers and commentary [^1]. Scoring them means finding that answer in the output and comparing it with a known result. Many evaluations hand this job to a language model. The study behind MT-Bench found GPT-4 as a judge agreed with human preferences over 80% of the time, but also documented biases, such as favoring longer, more verbose responses [^2].

What happened

Tian Zheng and Kai-Tai Hsu of Columbia University's Department of Statistics posted the study on arXiv on June 23 [^1]. They ran LAMBDA, an open-source multi-agent data-analysis system, on the 153 QRData tasks in the DSGym benchmark that have a single numerical answer. In LAMBDA, one agent writes and runs code, and an inspector agent checks for execution errors and suggests fixes; here it ran on gpt-5-mini and gpt-4.1-mini [^1]. The starting problem is mundane: DSGym's default scorer treats "0.2" and "0.20" as different answers, which produces false negatives [^1].

To separate the agent's mistakes from the grader's, the authors built a three-layer cascade. A strict grader pulls candidate numbers from the output and accepts one within 2% of the ground truth. Every strict failure goes to a lenient grader, GPT-4o-mini, which extracts the intended answer from the full output and allows 3%. A human then reviewed all 153 transcripts [^1]. By that human label, the agent was right on 72 tasks, or 47.1% [^1]. Neither automated grader accepted a wrong answer. They differed in how many right ones they found: a parser taking the last number in the output found 26%, a parser looking for the block that matches the question's keywords found 86%, and the lenient grader found 97% [^1].

The biggest fix was on the agent's side. LAMBDA's code-writing agent is built to end each turn with "suggestions for next steps", which reads naturally in a chat but scatters extra numbers through an answer that a benchmark expects as a single value [^1]. The authors added a nudge: up to two follow-up prompts asking the agent to state a single numerical answer. Of the 153 tasks, 98 needed at least one, and 94 of those recovered [^1]. The share of runs that produced a gradable answer rose from 36% to 97%, and the lenient pass rate from 16% to 46% [^1]. Repeating the original question in the nudge did not help, because the agent had not forgotten what to compute, only how to report it [^1].

Why it matters

The study's central point is that a score mixes two things: whether the agent got the answer, and whether the grader could find it [^1]. With the last-number parser, the agent would have been credited with about a quarter of its correct answers. The authors warn that developers who rely on uncalibrated automated metrics may end up optimizing the wrong factors [^1]. Their remedy is to calibrate each automated metric against human labels on a representative sample and report that audit next to the metric. A quick review of 100-character snippets matched the final human label in 136 of 153 cases, or 89%, which keeps such audits cheap [^1].

The layers also fail in different ways. The keyword parser is deterministic and can only extract numbers that actually appear in the output, so it cannot hallucinate; the LLM grader handles cases the parser can't, but inherits risks such as hallucination [^1]. Some errors no grader explains. On one hypothesis test, the agent reported a one-sided p-value of 0.006 where the ground truth was the two-sided 0.012. Both graders correctly marked it wrong, but only a human could see why [^1]. The authors keep their claims narrow: one agent, 153 tasks, and an agent and grader from the same model family, so shared biases are possible [^1].

Practical example

Say Tomás leads analytics at a regional insurer and is testing an open-source data-analysis agent before the analysts get access. Tomás writes a few dozen questions with known numerical answers, runs the agent and scores it with a script that takes the last number in each reply. The score comes back dismal.

The paper suggests the agent may not be the main problem. Like LAMBDA, it may close each answer with suggestions for next steps full of other numbers [^1]. In one of the study's cases, the agent's code printed the correct probability, 0.61, but also displayed the first rows of a 55-column table, and the strict grader picked the 2 from the rounding command instead. The lenient LLM grader found 0.61 [^1].

So Tomás changes the setup: a follow-up prompt that asks for a single number, a parser that looks for the part of the reply matching the question's keywords, and an LLM grader as a second pass on strict failures [^1]. The score rises, but Tomás still has a statistician read the misses. That is where the study's p-value case lives: an agent can confidently report 0.006 from a one-sided test when the question expected the two-sided 0.012, and no automated layer explains the gap [^1]. Tomás then shares the score internally together with the human audit, as the authors recommend [^1].

Related gear

We recommend this book to help users understand the foundational coding and statistical logic that agentic systems are attempting to automate and that graders must evaluate.

AdvertisementAmazon

Data Science from Scratch: First Principles with Python

Sources

  1. [1]arXiv — Grading the Grader: Lessons from Evaluating an Agentic Data Analysis System
  2. [2]arXiv — Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena