Skip to main content

Autumn papers at eScience and PACT 2026

AI & Data · · 6 min read
Gnosis Research Center
Gnosis Research Center

AI agents are starting to do scientific work: finding data, running tools and producing results. That raises systems questions. Does a workflow of agents recover when one of them makes a mistake? How does an agent find the right file? Can storage use the time an agent spends thinking? GRC students answered all three in papers at IEEE eScience 2026, along with a fourth on in situ analysis, and a fifth paper at PACT 2026 addressed transformers on GPUs.

The 22nd IEEE International Conference on eScience took place in Naples, Italy, from September 28 to October 2, and all four GRC papers appeared in its main track. PACT 2026, the International Conference on Parallel Architectures and Compilation Techniques, ran from October 19 to 22 at the Discovery Partners Institute in Chicago.

Testing how agent workflows recover from errors​

Workflows built from several large language model (LLM) agents now run multi-step computational tasks, but little is known about how they cope with bad inputs, failed tools or small numerical errors. In Agent Error Simulator, Keith Bateman presents a benchmark framework that injects faults on purpose. It sits as a transparent proxy between the agents and the model and injects format, logic and tool-call faults at chosen points in a planner, worker and aggregator workflow.

The team ran 194 jobs in 10 test groups on four open models from 2 billion to 32 billion parameters. Format and tool-call faults were recovered by every model. Logic faults were not: the 27 and 32 billion parameter models recovered from a wrong value injected into an earlier step, while the 7 and 2 billion parameter models carried it forward. A logic error injected into the aggregator, the agent that combines the others' results, caused every model to fail. The benchmark suite, injection tools and result data are released for others to reuse. The work is part of the DTIO project, with Jaime Cernuda, Luke Logan, Bogdan Nicolae and Franck Cappello of Argonne, Xian-He Sun and Anthony Kougkas.

Search that agents do not have to guess​

LLM agents usually look for files with guesses: glob patterns, grep and file reads, repeated until something matches. This costs tool calls and tokens and ignores the metadata inside scientific file formats. Acropolis, by Rajni Pawar, adds content indexing to a tiered file system, summarizing each file as it is written, and offers search to agents through an MCP server, with lexical, vector and graph search engines behind one interface.

On a 70-query benchmark over CMIP6 climate data, an agent using Acropolis found files as accurately as one using shell tools, with about one third fewer tool calls per query and about 3.1 times fewer LLM tokens. The work builds on IOWarp and LABIOS.

Staging data while the agent thinks​

Agents working on scientific data spend most of their time inside tool calls that read from storage. Before each call, a reasoning model thinks, and storage sits idle while the files the next tool needs start to appear in that reasoning. AgentStage, by Izzet Yildirim, reads the agent's streaming reasoning and its file system probes to infer which files come next. It copies them to fast local storage, decompressing on the way, so the tool's reads are served from fast storage.

Across four reasoning models, AgentStage made curated scientific tasks 1.74 times faster on average and up to 2.34 times faster in a single run, and sped up tasks from the MLE-bench, KramaBench and DSBench benchmarks by up to 1.36 times. The paper also shows how to estimate in advance whether a task will benefit, since compute-bound tasks or tasks that touch many files gain less.

Derived quantities inside the I/O path​

Simulations produce more data than can be written as often as scientists would like, so analysis has to happen in situ, while the data is produced. Hua Xu's paper, Rich I/O-centric Derived Quantity Operators, adds pluggable operators to the ADIOS2 I/O library that compute derived quantities as data is written. Some operators need only the current timestep, such as one that finds vortex cores. Others need earlier timesteps or earlier runs, such as a temporal derivative or a checksum that confirms a rerun matches the original. A metadata layer runs the operators at each step and records every result, with CLIO Core from IOWarp as the I/O manager that keeps history on tiered storage. Each operator declares what it will need, and that drives which data is kept on the fastest tier.

The team tested the framework on three platforms, including the Frontier supercomputer at Oak Ridge, and released the code as part of Coeus. Xu wrote the paper with Cernuda, Kougkas and Sun and with collaborators from Oak Ridge, Argonne, Los Alamos and Sandia national laboratories, the University of Tennessee and Rochester Institute of Technology.

Block-sparse transformers on GPUs​

At PACT 2026, Lihan Hu presented "Improving Data Reuse across Blocks for Efficient Block-Sparse Transformers on GPUs" on October 22, in the session on performance optimization, scheduling and communication. The paper, with Xiaoyang Lu, Sun and Peng Jiang, reuses data across the blocks of block-sparse transformer models to make them run more efficiently on GPUs.

Learn more​