Skip to main content

Twelve posters at GCASR 2026

Talks & Insights · · 5 min read
Gnosis Research Center
Gnosis Research Center

GRC researchers and collaborators presented twelve posters at the 13th Greater Chicago Area Systems Research Workshop (GCASR 2026) on May 11 at Northwestern University in Evanston. Taken together, the posters give a snapshot of the lab's work in spring 2026: how AI agents read and find scientific data, how multi-agent workflows run and recover, how workflows and logs move data, and how inference uses memory.

How agents read and find scientific data​

Correct is not I/O efficient. Izzet Yildirim, with Jie Ye, Jaime Cernuda, Xian-He Sun and Anthony Kougkas, traced the file operations of scientific workflows written by LLM coding agents. Of 105 agent-generated workflows across five tasks, 46 passed validation. Even the correct ones issued a median of 3.8 times more POSIX operations and read 6.5 times more data than expert implementations. The conclusion: storage behavior has to be measured directly before agent-written code runs on shared HPC or cloud storage.

Agentic search. Shazzadul Islam, with Cernuda, Sun and Kougkas, presented CLIO Search, a retrieval engine that profiles each data collection and then runs lexical, vector, graph and scientific-unit searches in parallel. The same LLM answered ten discovery queries over 341 National Data Platform datasets three ways. All three answered every query correctly, but the CLIO-backed agent used 52% fewer tokens and 91% less wall time than an agent relying on general web tools.

Semantic search for agent discovery. Rajni Pawar, with Luke Logan, Sun and Kougkas, embedded search inside the storage layer: file metadata is extracted, summarized by an LLM and indexed when data is written. On 273 astrophysics simulation files and 20 natural-language queries, a vector backend found the right file first 95% of the time, while BM25 keyword search inside the storage system answered in 0.047 milliseconds, 350 times faster than the same algorithm running as an external service.

Running and recovering multi-agent workflows​

Speculative dispatching. Ye, with Islam, Cernuda, Sun and Kougkas, applied speculative execution to the decision of which agent runs next. Agents start ahead of time while the full plan is computed; a reconciliation step keeps correct work, salvages partial matches and discards mispredictions. In preliminary results, the speculative planner produced dispatch plans with 0.50 to 0.78 similarity to the optimal plan at 4 to 10 times lower latency.

AgentRecall. Keith Bateman, with Sun and Kougkas, presented a provenance and rewind system for multi-agent workflows. It records tool calls, messages and prompts, stores them in IOWarp's hierarchical storage, and lets a person or an agent roll back to an earlier context when a mistake surfaces. Preliminary results showed rollback improving accuracy by up to 84% over baselines, with tool-call errors easier to fix than logical or formatting errors.

Trigger, Render, Reason. Hua Xu presented work on multimodal LLM agents that monitor scientific simulations with awareness of the science being simulated.

Workflows, logs and streams​

Storage mismatches. Meng Tang, with Luanzheng Guo and Nathan Tallent of PNNL, Kougkas and Sun, studied genomics and climate-tracking workflows on a production HPC cluster. I/O took 42% to 98% of total execution time, and choosing the fastest storage for each task on its own could make the whole workflow up to 6.8 times slower than a storage-aware choice, because of the data movement it forced between dependent tasks.

ChronoLog. Eneko Gonzalez, Kun Feng, Inna Brodkin, Kyle Chard of the University of Chicago, Kougkas and Sun presented ChronoLog, the lab's distributed log store that orders events by physical time, appends without locks, tiers data automatically and serves range queries. The poster showed two uses: live HPC telemetry in Grafana, and an MCP server that gives LLM agents a durable, replayable memory of past sessions.

StreamGuard. Hai Duc Nguyen, Bogdan Nicolae, Tekin Bicer, Amal Gueroudji, Matthieu Dorier, Chard and Ian Foster, of Argonne National Laboratory and the University of Chicago, presented low-overhead resilience for real-time HPC data streams.

Where files go. Neeraj Rajesh, with Sun and Kougkas, presented "Applications do I/O. What if files actually went... somewhere else?"

Inference and memory​

Pre-RoPE versus post-RoPE. Zia Uddin Chowdhury, with Cernuda, Ye, Kougkas and Sun, measured the trade-off between storing KV cache keys before or after rotary position embeddings. Storing them before costs compute at every decoding step but lets the cache survive truncation without re-running the prompt.

HyPIM. Jinlin Wu, Ferit Ozdaban, Selvanathan Sadhasivam, Xiaoyang Lu and Minxuan Zhou of Illinois Tech, with Yunhui Guo of the University of Texas at Dallas, moved the memory-bound operations of hyperbolic machine learning models into processing-in-memory units, leaving compute-dense work on the GPU. On HypFormer datasets the design ran 3 to 7 times faster than an A100 GPU and cut total energy by an average of 30% for a single inference task.

Update (November 2026): The twelve posters, with their PDFs, boards and sessions, are gathered on the GCASR 2026 event page. GCASR posters are not peer reviewed, so GRC now lists them in Studio with its other posters and talks rather than among its publications.

Learn more​