ACM Conference on AI and Agentic Systems · CAIS’26

Trace-Level Analysis of Information Contamination in Multi-Agent Systems

Anna Mazhar¹, Huzaifa Suri², Sainyam Galhotra¹

¹Cornell University · ²University of Illinois Urbana-Champaign

Abstract

Reasoning over heterogeneous artifacts (PDFs, spreadsheets, slide decks) increasingly occurs within structured agent workflows that iteratively extract, transform, and reference external information. In these workflows, uncertainty is not merely an input-quality issue: it can redirect decomposition and routing decisions, reshape intermediate state, and produce qualitatively different execution trajectories. We study this by treating uncertainty as a controlled variable — injecting structured perturbations into artifact-derived representations, executing fixed workflows under comprehensive logging, and quantifying contamination via trace divergence in plans, tool invocations, and intermediate state. Across 614 paired runs on 32 GAIA tasks with three language models, we find a decoupling: workflows may diverge substantially yet recover correct answers, or remain structurally similar while producing incorrect outputs.

At a Glance

614

paired runs under controlled perturbation

GAIA tasks spanning heterogeneous artifacts

language models evaluated across the workflows

Three Manifestations of Contamination

Silent semantic corruption

The workflow looks structurally normal — same plan, similar tool calls — yet produces an incorrect answer. The damage never surfaces in the control flow.

Behavioral detours with recovery

The trace diverges meaningfully (rerouting, extended execution) but the system still arrives at the correct result, decoupling divergence from failure.

Combined structural disruption

Both the execution path and the output break down, with control-flow signatures like rerouting, extended execution, and early termination.

Contributions

A formal taxonomy of contamination manifestations in structured agent workflows.
A trace-based measurement framework for detecting and localizing contamination across agent interactions — over plans, tool invocations, and intermediate state.
Empirical evidence with implications for targeted verification, defensive design, and cost control, including an analysis of why common verification guardrails fail to intercept contamination.

BibTeX

@inproceedings{mazhar2026trace,
  title     = {Trace-Level Analysis of Information Contamination
               in Multi-Agent Systems},
  author    = {Mazhar, Anna and Suri, Huzaifa and Galhotra, Sainyam},
  booktitle = {ACM Conference on AI and Agentic Systems (CAIS)},
  year      = {2026},
  eprint    = {2604.27586},
  archivePrefix = {arXiv},
  primaryClass  = {cs.AI}
}