Anthropic trained Fable 5 by analyzing its reasoning traces, showing complex search even for simple math problems. This explains why the model is efficient in cybersecurity tasks but exhibits inefficient search patterns in mathematical reasoning.
Anthropic revealed how it trained Fable 5 by analyzing its reasoning traces. When solving a simple Summle puzzle (reach 397 from 1,1,6,12,50,100), Fable 5 performed ~60k tokens of greedy depth-first search before switching to systematic root-split enumeration and solving it.
Fable 5 excels at cybersecurity tasks (recon→CVE→exploit→privesc→lateral→exfil) but shows inefficient search in mathematical reasoning (prime-check→partition→recurse→memoize). This is because the model memorizes canonical kill chains with branching factor ~1 from pretraining data, but cannot systematically explore diverse branches in math problems.
This analysis suggests Anthropic is improving training data by analyzing model reasoning traces. It highlights that while models can be efficient in specific domains, general reasoning ability remains limited. This may drive future training towards more diverse and efficient reasoning processes.