Source

arXiv:2603.19733 — PoC: Performance-oriented Context Compression via Performance Prediction (March 20, 2026)

Key Contribution

Inverts context compression: developer specifies a performance floor, a lightweight predictor finds the most aggressive compression ratio that stays within it, then drives an off-the-shelf compressor. Avoids over-compression that degrades reasoning quality.

Relevance to Zeph

Applies to zeph-memory context compaction pipeline:

• Current Zeph compaction: fires at threshold, applies structured summarization without quality floor
• This approach maps onto the existing CompactionProbe — probe score could serve as the performance floor
• The predictor could be trained on Zeph's compression_failure_pairs data (ACON guidelines)

Implementation Sketch

1. Treat probe.hard_fail_threshold as the performance floor
2. Train a lightweight predictor on (compression_ratio, probe_score) pairs from audit log
3. At compaction time: iterate candidate ratios, predict probe score, select most aggressive ratio ≥ floor
4. Store ratio→score pairs in compression_failure_pairs for predictor training