CoVeR Cuts 62-68% of Agentic Retrieval Verifier Calls Without Losing Accuracy
_reachsumit · x · 2026-09-23
arXiv paper "CoVeR: Coverage-Based Routing of Verifier Calls in Agentic Retrieval" (Daeyoung Roh, Donghee Han).
- Problem: in agentic retrieval, LLM verifiers decide at each step whether evidence suffices to stop, but they reprocess the ever-growing evidence after every retrieval step — a substantial repeated cost.
- Method: CoVeR uses a single threshold on a frozen sentence-embedding coverage margin as a gate: when evidence is plainly incomplete, the verifier call is skipped; only ambiguous states get verified. The router distills into a 921k-parameter head, leaving no LLM in the routing loop.
- Results: across three multi-hop QA benchmarks, accuracy matches the always-verify baseline within a fraction of an EM point while cutting 62-68% of verifier calls (93% in a saturated regime); count/lexical/BM25-based routers trade off worse, and the gate transfers without re-tuning across deciders and agent scales.
- Caveat: the same signal can't replace verification — matching a claim is far easier than deciding it's supported.
More from coding & agent
- Open-source GUI agents top out at 8% task success on composite cross-device tasks — maier_ak · 2026-09-23
- JarvisGUI benchmark tests GUI agents across Android, Windows and Ubuntu in one workflow — maier_ak · 2026-09-23
- Cross-DURIAN: A New Benchmark for Multi-Device GUI Agents — maier_ak · 2026-09-23
- "Proactive" personal agents: reminder, proposal, or acting without asking? — sujingshen · 2026-09-23
- AWS Open-Sources Strands Harness, Claims 28% Fewer Agent Tokens — shashib · 2026-09-23
- Before your AI agent pays for you: four questions about authorization and billing — sujingshen · 2026-09-23