[research] KV-Cache Grafting: smaller model beats larger one, 93.3% vs 89.2%

[github-actions[bot]]

🔬 The Finding

Researchers introduce KV-Cache Grafting — pre-computing verified knowledge as a byte-exact KV-state artifact and "grafting" it into fresh inference contexts with zero weight changes. On AIME 2025, a frozen Gemma-4-12B jumps from 80.0% → 93.3% accuracy after grafting a verified solution library, surpassing its 31B sibling's 89.2% — all with dramatically lower per-query compute. The restore is cryptographically verified (SHA-256 bit-exact, zero KL divergence).

⚙️ What It Means for Agentic Workflows

1. Pre-cache expensive context once, reuse forever. System prompts, tool schemas, RAG snippets, and domain knowledge can be computed once and grafted into every agent invocation — cutting TTFT and token costs for any workflow that shares a fixed prefix.
2. Downsize your backbone model without sacrificing quality. Grafting domain-specific KV artifacts lets a smaller deployed model outperform a larger one, reducing inference costs for high-volume agentic pipelines.

🔗 Source

Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting — 15 Jul 2026

Generated by Daily Agentic AI Research Digest · 66.9 AIC · ⌖ 11.5 AIC · ⊞ 24.2K · ◷

• expires on Jul 27, 2026, 10:12 AM UTC

[github-actions[bot]]

This discussion was automatically closed because it expired on 2026-07-27T10:12:12.042Z.

Closed by Workflow