GCASR is a one-day workshop that brings together systems researchers from universities and national laboratories in the Chicago area. Its posters are not peer-reviewed publications, so GRC keeps them in Studio with the rest of its posters and talks. The recap article tells the longer story of the day.
What LLM agents cost the storage system when they search for and read scientific files, and how storage can answer them faster.
Board L1Lake, session 2
Of 105 scientific workflows written by LLM coding agents, 46 passed validation, and even the correct ones issued a median 3.8 times more POSIX operations than expert code.
Board R9Rock, session 2
CLIO Search profiles each data collection and runs lexical, vector, graph and unit-aware searches in parallel; over 341 National Data Platform datasets, the CLIO-backed agent used 52% fewer tokens than one using web tools.
Board D10Dillo, session 1
Indexing file metadata inside the storage layer as data is written let a vector backend rank the right astrophysics file first for 95% of 20 queries.
Starting agents before the plan is final, rewinding them when it goes wrong, and watching simulations for science that has stopped making sense.
Board L10Lake, session 1
Speculative dispatch starts agents while the full plan is still being computed; preliminary plans reached 0.50 to 0.78 similarity to the optimal plan at 4 to 10 times lower latency.
Board L18Lake, session 2
AgentRecall records tool calls, messages and prompts in IOWarp's hierarchical storage so a person or an agent can roll a workflow back to an earlier context.
Board N26Northwestern, session 1
Multimodal LLM agents watch a running simulation for science-aware triggers, render the flagged data as images and judge whether the results are still valid.
Where data lands between workflow stages, how a log store keeps time, and how real-time streams survive failures.
Board N15Northwestern, session 2
In genomics and climate-tracking workflows on a production cluster, choosing the fastest storage per task made whole workflows up to 6.8 times slower than a storage-aware choice.
Board R5Rock, session 1
ChronoLog orders events by physical time, appends without locks and tiers data automatically; the poster showed live HPC telemetry and a durable memory for LLM agents.
Board R4Rock, session 1
StreamGuard, from Argonne and the University of Chicago, adds low-overhead resilience to real-time HPC data streams that must keep producing timely results through faults.
Board R5Rock, session 2
I/O Router intercepts file operations through LD_PRELOAD, FUSE or eBPF and sends the data to other backends, such as HDF5 or ADIOS2, without changing the application.
Memory in machine learning: a KV cache that survives truncation, and hyperbolic models run inside memory.
Board N14Northwestern, session 1
Storing KV cache keys before rotary position embeddings costs compute at every decoding step, but lets the cache be truncated without re-running the prompt.
Board N8Northwestern, session 2
HyPIM moves the memory-bound operations of hyperbolic models into processing-in-memory units; on HypFormer datasets it ran 3 to 7 times faster than an A100 GPU.
Boards and sessions are from the GCASR 2026 poster program (opens in a new tab), checked November 13, 2026. Titles and authors follow the posters. Three posters have no public PDF. The news item from the day has the short version.
