Toolkit SenSys 2027 submission · under double-blind review

Scaled learning runs on data that is easy to store, share and read back. Peacodec lets agents design expert-level codecs for every modality you care about, beyond images and video.

Paper Data Compression through Agentic Recursive Experimentation

Code Paper BibTeX

Data

Any data you can load as arrays. The nine datasets below are the ones in our experiments, across sensing, science and media. Click a card to see why its data matters.

Sensor FMCW radar

Radar range over time: bright V-shaped paths on a pale background

Human MotionRange over time with static clutter removed. A moving person traces the bright paths.NMSE at 4–32×

Sensor Why it matters

Human Motion

Radar senses people in the dark, through smoke and without a camera's privacy cost. Activity and gesture models learn from hours of raw chirps, which few groups store or share.

Peacodec vs. best baseline

pending

Sensor 60 GHz radar

Radar range over time: one bright band that pulses with each breath

Vital SignalRange over time with static clutter removed: a seated child, the band pulsing with each breath over 40 seconds.NMSE at 4–32×

Sensor Why it matters

Vital Signal

Contactless breathing and heart-rate monitoring, for sleep studies and patient care. The signal is a sub-millimetre chest motion carried in phase, so a codec must keep phase, not just power.

Peacodec vs. best baseline

pending

Sensor Wi-Fi CSI

Heatmap of Wi-Fi subcarrier amplitude over time

OPERAnetAmplitude of 30 subcarriers over time, relative to each subcarrier's median.NMSE at 4–32×

Sensor Why it matters

OPERAnet

Channel state turns ordinary Wi-Fi links into motion and presence sensors. Every packet carries a full subcarrier matrix, so recordings grow quickly and are rarely shared raw.

Peacodec vs. best baseline

pending

Sensor 12-lead ECG

PTB-XLLeads II, V1 and V5 over 6 seconds.NMSE at 16–128×

Sensor Why it matters

PTB-XL

The ECG is among the most common clinical tests, and large public archives train diagnostic models. Diagnosis rests on waveform shape, beat by beat, so the error must stay small everywhere.

Peacodec vs. best baseline

pending

Science Climate simulation

A global atmospheric field showing continents and an equatorial band

CESM-ATMOne atmospheric field on a global 1800 × 3600 grid.bits/value under a max-error bound

Science Why it matters

CESM-ATM

Climate models produce far more output than archives can keep, so runs and variables get discarded. A bounded error per value lets modelling centres keep more of what they simulate.

Peacodec vs. best baseline

pending

Science X-ray laser detector

Mosaic of 32 detector panels with intensity gradients

EXAFELOne event across the detector's 32 panels, log intensity.bits/value under a max-error bound

Science Why it matters

EXAFEL

X-ray free-electron lasers image samples one pulse at a time, faster than storage and networks can absorb. Detector data has to shrink near the instrument, within a strict error bound.

Peacodec vs. best baseline

pending

Science Molecular dynamics

Horizontal bands: lattice planes in atom positions over time

EXAALTDensity of atom positions along one axis over 117 time steps. Lattice planes appear as bands.bits/value under a max-error bound

Media Radiographs

ChestX-ray148-bit 1024 × 1024 images, coded one at a time.bits/pixel · PSNR at 0.25–1 bpp

Media Why it matters

ChestX-ray14

Chest radiographs are among the most common medical images and a standard set for medical vision models. Archives keep them lossless; training can often use smaller lossy copies.

Peacodec vs. best baseline

Losslesspending

PSNRpending

Media Read speech

LibriSpeechSpectrogram, 0–4 kHz, 4 seconds of 16 kHz mono speech.ViSQOL MOS at 6–24 kbps

Media Why it matters

LibriSpeech

Thousands of hours of read speech train recognition and synthesis models. Audio has mature expert codecs, which makes it the reference point: how close do agents get to decades of design?

Peacodec vs. best baseline

pending

Yours Any modality

dataset.toml · convert/ · data_io.py

Your dataWrite a loader, a converter and a neutral description. The goals, evaluators, agents and search are reused as they are.How to add a dataset →

How it works

The agent experiments. The harness judges. Everything the agent may touch runs in a sandbox; the evaluator and the test split stay outside it.

01 Data Recordings and a neutral description: sample format only, never the dataset's name or subjects.
02 Agent, sandboxed Inspects the data, writes probes, and builds an encoder and a decoder.
03 Fixed evaluator Checks the reconstruction, counts every byte the decoder needs, and enforces the time budget.
04 Score Dev score during the run; the held-out test split is scored afterwards.

Search · E-RSI

Each session starts fresh and builds on the evidence. Experiment-Driven Recursive Self-Improvement keeps every version and a report from every session, instead of one ever-growing conversation.

Baseline Linear

One long conversationA single chain of versions; every turn re-reads the whole history.

E-RSI Tree

Fresh sessions, shared evidenceOne idea per session, started from any earlier version and guided by the reports.

Agents

Bring your own agent. Peacodec works with most coding agents, proprietary or open-source. An agent is a scaffold plus a model; the data, prompt and evaluator stay the same for every one.

Proprietary Tested

Claude Code

ModelClaude Opus 5.5

Accessvendor CLI

Proprietary Tested

Codex CLI

ModelGPT-6.1 Sol

Accessvendor CLI

Open weights Tested

Qwen Code

ModelQwen3.8-2.4T-A95B

AccessOpenRouter

Open weights Tested

OpenCode

ModelDeepSeek V4 Pro

AccessOpenRouter

Yours Any CLI agent

Your agent

Any coding agent that runs headless in a terminal, with the model of your choice. Add one adapter file, agents/scaffolds/<name>.sh, that says how to start it, where its session log lives and how it counts tokens.

The four tested in the paper run in the same container, with the same tools, network rules and evaluator. Vendor models run at high effort.

Results

Leaderboards arrive once the test splits are scored. Per-task rankings of agent-designed codecs and baselines, with each codec's search tree.

coming soon

Get started

Prepare a run with one command. Set your data and image paths in config.toml, then render a task without starting the agent; drop --dry-run to start it.

bash

$ python3 harness/launch.py --dataset human-motion --goal lossy-nmse \
    --scaffold claude-code --model opus --effort high --prompt default \
    --network tools --hours 3 --max-evals 0 --dry-run

Cite

@inproceedings{peacodec2027,
  title     = {Data Compression through Agentic Recursive Experimentation},
  author    = {Anonymous Author(s)},
  booktitle = {Under review},
  year      = {2027}
}