@inproceedings{v-etal-2026-much,
title = "How Much Should the Agent See? A Disclosure Axis for {LLM} Data-Analyst Agents",
author = "V, Athira M and
Challa, Aditya and
Tripathi, Jaya and
Mohanty, Saroj and
Choudhary, Santosh Kumar and
Danda, Sravan and
Jain, Ruchika Mehra and
Saha, Snehanshu",
editor = "Atanasova, Pepa and
Haffari, Gholamreza and
Wang, Minghan",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-industry.9/",
pages = "124--144",
ISBN = "979-8-89176-460-6",
abstract = "Existing LLM data-analyst benchmarks fix data access at full disclosure and evaluate only on answer accuracy. But real deployments rarely grant full access. No prior work, to our knowledge, asks *{''}what{'}s the least amount of information needed for the agent to work?''*. In this article, we treat dataset disclosure as a controlled treatment along a *graduated information-disclosure* (GID) axis: $C_0$ (column names) and $C_\infty$ (full table) bracket a nested chain $C_1 \subseteq \dots \subseteq C_5$ that adds per-column metadata, dataset purpose, summary statistics, sampled cells, and sampled rows. We maintain agent architecture, tool registry, and decoding fixed, and score the final reports on a 14-attribute rubric across a $6 \times 4 \times 12$ panel of frontier models, planted-signal datasets, and disclosure levels. At a high level, from $C_0$ to $C_\infty$, seven attributes rise and five fall. However, the rise/fall pattern has nuance, the optimal intermediate level depends on several factors including models, datasets and attributes. Level $C_1$ - one sentence of types and units beyond column names - matches (statistically) $C_\infty$ on signal recovery and decoy rejection at ${\sim}49\times$ lower prompt-token cost. The key takeaway from this work is - **current LLM data-science agents can act on partial information alone**, so row-level records need never leave the institution. Hence, the current framework also offers a way to enforce data boundary constraints (required by GDPR) without compromising on quality."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="v-etal-2026-much">
<titleInfo>
<title>How Much Should the Agent See? A Disclosure Axis for LLM Data-Analyst Agents</title>
</titleInfo>
<name type="personal">
<namePart type="given">Athira</namePart>
<namePart type="given">M</namePart>
<namePart type="family">V</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Aditya</namePart>
<namePart type="family">Challa</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jaya</namePart>
<namePart type="family">Tripathi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Saroj</namePart>
<namePart type="family">Mohanty</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Santosh</namePart>
<namePart type="given">Kumar</namePart>
<namePart type="family">Choudhary</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sravan</namePart>
<namePart type="family">Danda</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ruchika</namePart>
<namePart type="given">Mehra</namePart>
<namePart type="family">Jain</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Snehanshu</namePart>
<namePart type="family">Saha</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track</title>
</titleInfo>
<name type="personal">
<namePart type="given">Pepa</namePart>
<namePart type="family">Atanasova</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Gholamreza</namePart>
<namePart type="family">Haffari</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Minghan</namePart>
<namePart type="family">Wang</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-460-6</identifier>
</relatedItem>
<abstract>Existing LLM data-analyst benchmarks fix data access at full disclosure and evaluate only on answer accuracy. But real deployments rarely grant full access. No prior work, to our knowledge, asks *”what’s the least amount of information needed for the agent to work?”*. In this article, we treat dataset disclosure as a controlled treatment along a *graduated information-disclosure* (GID) axis: C₀ (column names) and C_ınfty (full table) bracket a nested chain C₁ \subseteq \dots \subseteq C₅ that adds per-column metadata, dataset purpose, summary statistics, sampled cells, and sampled rows. We maintain agent architecture, tool registry, and decoding fixed, and score the final reports on a 14-attribute rubric across a 6 \times 4 \times 12 panel of frontier models, planted-signal datasets, and disclosure levels. At a high level, from C₀ to C_ınfty, seven attributes rise and five fall. However, the rise/fall pattern has nuance, the optimal intermediate level depends on several factors including models, datasets and attributes. Level C₁ - one sentence of types and units beyond column names - matches (statistically) C_ınfty on signal recovery and decoy rejection at \sim49\times lower prompt-token cost. The key takeaway from this work is - **current LLM data-science agents can act on partial information alone**, so row-level records need never leave the institution. Hence, the current framework also offers a way to enforce data boundary constraints (required by GDPR) without compromising on quality.</abstract>
<identifier type="citekey">v-etal-2026-much</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-industry.9/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>124</start>
<end>144</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T How Much Should the Agent See? A Disclosure Axis for LLM Data-Analyst Agents
%A V, Athira M.
%A Challa, Aditya
%A Tripathi, Jaya
%A Mohanty, Saroj
%A Choudhary, Santosh Kumar
%A Danda, Sravan
%A Jain, Ruchika Mehra
%A Saha, Snehanshu
%Y Atanasova, Pepa
%Y Haffari, Gholamreza
%Y Wang, Minghan
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-460-6
%F v-etal-2026-much
%X Existing LLM data-analyst benchmarks fix data access at full disclosure and evaluate only on answer accuracy. But real deployments rarely grant full access. No prior work, to our knowledge, asks *”what’s the least amount of information needed for the agent to work?”*. In this article, we treat dataset disclosure as a controlled treatment along a *graduated information-disclosure* (GID) axis: C₀ (column names) and C_ınfty (full table) bracket a nested chain C₁ \subseteq \dots \subseteq C₅ that adds per-column metadata, dataset purpose, summary statistics, sampled cells, and sampled rows. We maintain agent architecture, tool registry, and decoding fixed, and score the final reports on a 14-attribute rubric across a 6 \times 4 \times 12 panel of frontier models, planted-signal datasets, and disclosure levels. At a high level, from C₀ to C_ınfty, seven attributes rise and five fall. However, the rise/fall pattern has nuance, the optimal intermediate level depends on several factors including models, datasets and attributes. Level C₁ - one sentence of types and units beyond column names - matches (statistically) C_ınfty on signal recovery and decoy rejection at \sim49\times lower prompt-token cost. The key takeaway from this work is - **current LLM data-science agents can act on partial information alone**, so row-level records need never leave the institution. Hence, the current framework also offers a way to enforce data boundary constraints (required by GDPR) without compromising on quality.
%U https://aclanthology.org/2026.emnlp-industry.9/
%P 124-144
Markdown (Informal)
[How Much Should the Agent See? A Disclosure Axis for LLM Data-Analyst Agents](https://aclanthology.org/2026.emnlp-industry.9/) (V et al., EMNLP 2026)
ACL
- Athira M V, Aditya Challa, Jaya Tripathi, Saroj Mohanty, Santosh Kumar Choudhary, Sravan Danda, Ruchika Mehra Jain, and Snehanshu Saha. 2026. How Much Should the Agent See? A Disclosure Axis for LLM Data-Analyst Agents. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: Industry Track, pages 124–144, Budapest, Hungary. Association for Computational Linguistics.