Rong He’s study proposes a model that shows what share of results from lower-level agents reaches the final agent. The model describes task decomposition as a tree: an agent given b results keeps each one with probability r(b).
When r(b)=1/b, every tree delivers exactly one result to the final agent regardless of task size or tree shape. In the r(b)=Cb^−δ model, a flat scheme produces the largest number of results.
Depth reduces the amount of data in the final agent’s context from N to N^(1/k). At equal cost, a two-tier scheme begins to outperform a flat one at 403 results.
Claim check:
- Rong He proposed a model in a study that shows what share of results from lower-level agents reaches the final agent. (confirmed by the publication itself: evidence; «Authors: Rong He»)
- The model represents task decomposition as a tree in which an agent given b results keeps each one with probability r(b). (confirmed by the publication itself: evidence; «Model a decomposition as a tree in which an agent handed $b$ items keeps any one with probability $r(b)$.»)
- When r(b)=1/b, every tree delivers exactly one result to the final agent regardless of task size or tree shape. (confirmed by the publication itself: evidence; «If $r(b)=1/b$, every tree delivers exactly one finding, for every task size and every shape»)
- In the r(b)=Cb^−δ model, a flat scheme produces the largest number of results. (confirmed by the publication itself: evidence; «so flat is optimal for yield»)
- Depth reduces the amount of data in the final agent’s context from N to N^(1/k). (confirmed by the publication itself: evidence; «depth cuts its exposure from $N$ items to $N^{1/k}$»)
- At equal cost, a two-tier scheme begins to outperform a flat one at 403 results. (confirmed by the publication itself: evidence; «at equal spend two tiers overtake flat at 403 findings»)
Primary sources:
score 78.1 out of 100 · kind: research