SecondSourceJudgment rebuilt from primary sources
Research · Jul 22, 2026

A claim worth knowing: a model's "generalization" may live not in its weights but in the orchestration layer around it — and the same mechanism is a new source of benchmark inflation.

Research Notes · This week Trend (published around 07-20)

From the Jul 22, 2026 daily brief

Alex Zhang, author of the RLM framework paper, argues that carefully designed task orchestration (the loops and tool-dispatch logic around the model) can reduce superficially different tasks to similar execution traces, letting training on short tasks generalize to tasks 8 to 32 times longer (X / @a1zhang, 07-20). Commentator swyx supplied the dark side: you don't need to train on the test itself — train on data that merely looks like the test and you can farm impressive scores. The same mechanism is at once a capability source and a benchmark-contamination source (X / @swyx, 07-21). Caveat: the 8–32× figure is from a single paper, not independently reproduced. The takeaway question: when an agent benchmark score jumps, first ask whether the jump is in the model or in the orchestration.

Subscribe free — first issue lands tomorrow morning

Just an email address, unsubscribe anytime. This is the only thing we ask of you.

More in this section