Hybrid Reasoning Agents: Which Monitor Still Works
At 14B parameters, HybridAgent's estimated generation cost drops from $470 to $176 per million questions (2.7x cheaper), as steps fall from about 2,200 to 900 per question.
Source
Adham Khaled
Published
Author
Adham Khaled
Read
1 min read
Takeaways
At 14B parameters, HybridAgent's estimated generation cost drops from $470 to $176 per million questions (2.7x cheaper), as steps fall from about 2,200 to 900 per question.
About 35% of HybridAgent's steps at 14B are hidden, with visible text shrinking 43% (from 497 to 284 tokens) while forward passes fall only 6% ( 497 to 467 ).
In the paper's monitoring test on 560 math problems, a strong checker ( GPT-5.6-Luna ) scored 0.989AUROC on HybridAgent's traces versus 0.990 on a text-only agent's, while a weaker checker ( GPT-4o-mini ) scored 0.860 versus 0.818 .
HybridAgent outperforms the matched text-only baseline ARPO by 4% to 6% on math/knowledge tests and 27% to 38% on agent benchmarks like GAIA and WebWalker , while scoring only 16.0 versus Kimi-K3 's 31.4 on Humanity's Last Exam.
Goodfire's activation probes caught 3.1% more reward hacks than a chain-of-thought monitor on Kimi K3 , but 7.9% fewer on GLM 5.2 , and its cyber-misuse probes hit 93% recall at a 5.5% false-positive rate for under $200 per million exchanges.
Reporting is indexed from AlphaSignal. Rights remain with the original publisher and cited sources.