saved
The Mismanaged Geniuses Hypothesis
gist
Zhang, Li, and Khattab argue that frontier language models already have the competence for the next capability jump, but human-engineered agent scaffolds waste it. Their mismanaged-geniuses hypothesis says the bottleneck is the space of decompositions a model can express, not more pretraining. Coding agents and recursive language models hint that models can manage other models if the scaffold splits work into in-distribution calls. They recommend training the composer, citing a 4B RLM that reached 100% on a 1M-context needle task after RL on a much smaller setting.
ideas
- The bottleneck is management, not scale. Frontier LMs already beat humans on IMO and IOI, yet fail long-horizon work because human-written scaffolds waste individual calls.
- Define the space of decompositions first. How a model is allowed to split work has exponential impact on what composition can solve.
- Keep every call in-distribution. A capable system can solve OOD tasks if each LM call, including the act of decomposing, stays in-distribution.
- Train the composer, not a bigger genius. A 4B RLM went from near-zero to 100% on a 1M-context 8-needle task after RL on a 32k, 1-needle setting.
- Orchestrator-subagent is only a start. Claude Code-style plans work for software, but recursive code-with-subcalls may be needed for length and long-horizon work.
quotes
“existing frontier language models are severely underutilized due to sub-optimal use of individual language model calls”
“Nearly all modern agent scaffolds are human-engineered, task-specific decomposition strategies that use language models.”
“directly learning the operator to compose LMs is a significantly more efficient strategy”
“the correct decompositions are likely already within the distribution of what LMs can generate”