LOCAA: An Agentic System for Automated Lossy Compressor Tuning

What it claims

LOCAA replaces a fixed error-bound (EB) search algorithm, such as binary search or FRaZ's derivative-free optimizer, with a single tool-calling LLM agent. The agent picks each next EB by reasoning over the measured compression ratio (CR), PSNR and SSIM from earlier trials. Every candidate is actually compressed and decompressed, so no surrogate model is involved and no prediction error is introduced.

Beyond the summary's headline figures, the abstract gives one more result. On CR maximization under joint PSNR and SSIM constraints across six CESM fields, average trials fall from 61.5 with binary search to 17, a 72.4% reduction. The fixed-ratio comparison (1.98× fewer trials than binary search, 5.03× fewer than FRaZ) is an average over twelve fields from two applications.

How it works

• Stack. GPT-5.5 is the underlying LLM. The workflow is built on the Pocket Agent framework, and compressor and evaluation tools are exposed over MCP.
• Single agent. One agent runs the whole select, execute, observe, decide loop. The authors chose this deliberately over separate planner and analyzer agents so the search state stays in one context.
• Structured prompt. The prompt encodes a search policy:
• how to pick the first EB
• how to keep a trial history
• how to read compressor-specific response curves
• how to project the next candidate
• feasibility checks, budget-aware stopping rules and final selection
• an explicit observe, interpret, project, act chain-of-thought before each evaluation
• PSNR-driven initialization. When a PSNR target exists, the first EB is ε₀ = V_r / 10^(PSNR_target/20), where V_r is the data value range. This treats RMSE = ε as the boundary case. The authors call it an order-of-magnitude anchor rather than a feasible bound, because RMSE is usually smaller than ε. Without a PSNR target, the agent samples a random relative EB from [10⁻⁹, 10⁻¹] and converts it with ε_ABS = ε_REL·V_r for absolute-mode compressors.
• Staircase-aware navigation for ZFP. ZFP's CR and quality only change at bit-plane boundaries. An agent that interpolates between points keeps landing on plateaus it has already seen. The prompt therefore tells it to:
• step the EB multiplicatively
• skip regions that would repeat already-observed metric values
• stop once observed levels bracket the target
• if the target falls in a gap between levels, return the closest feasible level and report that the gap is unattainable
• Multiple compressors. These are tuned one after another. History from earlier compressors seeds later ones, and the best feasible configuration across all of them is selected at the end.
• Persistent memory. This is the mechanism behind the 27.9% reduction across timesteps.

flowchart TD
  A[User objective + constraints] --> B{PSNR target given?}
  B -- yes --> C[EB from PSNR formula]
  B -- no --> D[Random rel. EB]
  C --> E[Compress/decompress via MCP tools]
  D --> E
  E --> F[Observe CR, PSNR, SSIM]
  F --> G{Feasible & budget/stop rule met?}
  G -- no --> H[Project next EB<br/>staircase-aware for ZFP]
  H --> E
  G -- yes --> I[Best feasible config]

Limits

• The fetched text is cut off mid-methodology. The detailed results, the ablation of which components matter, the quality-plus-time-constraint use case, and the authors' Limitations section (Section V) are not visible here.
• The three applications are not fully named in the fetched text; only CESM is. The five compressors are not enumerated either. Related work covers SZ2, SZ3, SZx, ZFP and SPERR, but the text does not confirm that these are the evaluated set.
• Efficiency is reported only as a count of evaluation trials. The fetched text gives no LLM latency, token cost or wall-clock figures, so check these before assuming fewer trials means less total time.
• Results depend on GPT-5.5 and the hand-written prompt policy. Other models are not reported in the visible text.
• The visible text does not state whether code is released, or under what licence.