Independent verifiers improve agent output - but frontier verifiers are expensive. So we trained our own. Results on @GoogleDeepMind's FACTS…

Independent verifiers improve agent output - but frontier verifiers are expensive. So we trained our own. Results on @GoogleDeepMind's FACTS-Search: > No verifier: 89.4 (previously published SOTA) > With a frontier verifier: 93.4 > With our trained 8B verifier: 92.9 (at 3.2× lower cost than the frontier verifier) Another example of when frontier-only isn't worth it. Full write up: https://www.ai21.com/blog/you-need-a-verifier/

Related

• Adding a verifier improves every agentic AI system. But frontier verifiers are expensive - so we trained our own. Matching quality, at a fra…
• 1/5 We’re seeing 4 common agent optimization methods for hitting the right accuracy-cost or accuracy-latency tradeoff. We tried them all out…
• Notes from evaluating a customer support chat agent system: heuristic evaluators give false signal, retrieval bugs masquerade as LLM failures, and the cost/quality Pareto frontier is rarely where you think [D]

Source: AI21 Labs (X) | 2026-09-02