Baseline Comparison

Baseline Comparison

Results (8-prompt heretic set)

Method	exact_keep_pct	keep_rate	avg ms
kompress-v8 (ours)	0.993	0.936	97.0
Random eviction	0.910	0.835	0.0
LLMLingua-2	0.867	1.550†	238.9
TextRank	0.599	0.543	23.1

† LLMLingua-2's keep_rate >1.0 reflects token expansion from special boundary markers — expected behavior.

Analysis

• The gap between kompress-v8 (0.993) and random eviction (0.910) is the learned component's contribution: +0.083 over chance
• TextRank's 0.599 confirms that extractive summarization is unsuitable for must-keep preservation
• LLMLingua-2's 0.867 at 1.55 keep_rate shows that token-budget compression without must-keep awareness is insufficient

How to reproduce

python baselines/run_baselines.py

Results save to baselines/baseline_results.json.

Interactive comparison

https://peterlodri-sec.github.io/longrun-eval-kompress/baselines.html

Missing baselines

AutoCompressors and Gisting require separate training on large corpora and are not runnable on a single M1 Pro. They are deferred to a revision with GPU resources.