LEAD-RL: credit each subagent by how useful its report was
TL;DR. Ant Group, UESTC, Zhejiang and Southeast University jointly train a main research agent and its subagents. The usual joint training gives every agent the same team advantage, so a misleading subagent report gets rewarded when the team happens to win. LEAD-RL scores each subagent report with a training-time rubric for research utility and uses that score to gate how much team advantage the subagent inherits. Final-answer correctness stays the only team reward. There is no token, step or speed reward. On BrowseComp it reaches 73.5%, 2.49 points above shared-credit joint training, and cuts critical-path cost from 323.3 to 244.2 steps per question (24.5%).
Delegation only pays after the team learns to coordinate
A rubric decides how much of the team's reward each subagent report deserves.
flowchart LR
Q["Question<br/><small>BrowseComp task</small>"] --> M["Main agent<br/><small>issues subtasks</small>"]
M --> S["Subagents<br/><small>search in parallel</small>"]
S --> RB["Rubric<br/><small>scores each report</small>"]
S --> M
M --> A["Final answer<br/><small>only team reward</small>"]
RB -->|gates advantage| S
classDef input fill:#d0ebff,stroke:#1971c2,color:#1b1b1b,stroke-width:2px
classDef core fill:#e5dbff,stroke:#6741d9,color:#1b1b1b,stroke-width:2px
classDef loop fill:#fff3bf,stroke:#f08c00,color:#1b1b1b,stroke-width:2px
classDef exit fill:#d3f9d8,stroke:#2f9e44,color:#1b1b1b,stroke-width:2px
class Q input
class M,S core
class RB loop
class A exit
linkStyle 5 stroke:#f08c00,stroke-width:2px
Blue is input, purple is the agents, amber is the training-time credit rubric, green is the result.
Key findings
- 73.5% on BrowseComp, +2.49 over shared credit, with 24.5% fewer critical-path steps.
- Swarm effect: turning on subagents moved the untrained base model from 59.0% to 58.0%, while the trained system went from 66.5% to 73.5%. Delegation helps only after coordinated training.
- Efficiency fell out of better credit, not an explicit cost reward.
How it relates to prior wiki pages
- Confirms the 11-12 "single agent is the right default" reading in multi-agent-systems (WikiTrace digest): untrained delegation is neutral or harmful.
- Credit assignment is the shared fix. CoReMem (today) credits steps that used memory anchors. Harness-S (11-15) routes failed RL attempts to sibling harnesses. All three are about who gets credit inside an agent system.
Gaps
- The rubric is a training-time judge. Its own biases transfer into the subagent policy, and that is not audited.
- Only "critical-path" steps are reported. Total tokens across parallel subagents, the real bill, are not in the abstract.
Links
- Paper: https://arxiv.org/abs/2611.17041
- Raw:
raw/dair/2026-11-16-123502-dair-academy.md - Concepts: multi-agent-systems