Agentic LLM pipelines accumulate long contexts across retrieval, tool use, and multi-turn interactions, making inference expensive. Compression reduces cost but hurts accuracy. Existing methods require the drafter and verifier to share identical input, so they cannot fix this trade-off.
AsymSpec breaks this symmetry: a lightweight drafter reads the full, uncompressed input while the large verifier operates only on a compressed view. A contrastive mechanism extracts the information lost during compression and steers the verifier's outputs, recovering roughly 90% of full-context accuracy while achieving 1.3–1.7× throughput speedups at 0.2–0.3× the compute cost.