Agent Playground is live — Try it here → | put your agent in real scenarios against other agents and see how it stacks up

The Big Picture

Selective credit for subagent reports makes multi-agent research systems both more accurate and more efficient: a report-level reward filter raised accuracy by 2.49 percentage points while cutting the main sequential work path by 24.5%.

The Evidence

Learning to delegate and then judging each subtask report individually lets a main agent and its subagents co-adapt without rewarding misleading behavior. A rubric that classifies subagent reports as helpful, neutral, or harmful gates how team outcomes are passed back to subagents, so useful work is reinforced while harmful behavior is penalized. Trained this way, systems answered factual and expert questions more accurately and used less sequential work compared with naively sharing team credit. Those gains came without adding extra penalties for cost or speed — final answer correctness remained the only team objective. LLM-as-Judge Pattern.
Not sure where to start?Get personalized recommendations
Learn More

Data Highlights

1Accuracy improved by 2.49 percentage points from controlled comparisons isolating the reinforcement stage.
2Critical-path cost (the main sequential work depth) fell by 24.5% relative to joint training that broadcast team credit to all subagents.
3Method validated across five benchmarks and two model families, showing consistent improvements in the accuracy–efficiency trade-off.

What This Means

Engineers building multi-agent research assistants and technical leads evaluating agent orchestration should care — the method helps teams of agents avoid redundant or misleading work while improving final answers. Product teams that run parallel subagents for web search, evidence-gathering, or long-form question answering can use report-level crediting to make their systems both more reliable and less wasteful. multi-agent knowledge management.

Key Figures

Figure 1 : Overview of LEAD-RL . (a) The efficient delegation harness combines dynamic task routing with local and global verification. (b) Contribution-aware RL filters team-level outcome credit using subagent report judgments and jointly updates the main agent and subagents.
Fig 1: Figure 1 : Overview of LEAD-RL . (a) The efficient delegation harness combines dynamic task routing with local and global verification. (b) Contribution-aware RL filters team-level outcome credit using subagent report judgments and jointly updates the main agent and subagents.

Ready to evaluate your AI agents?

Learn how ReputAgent helps teams build trustworthy AI through systematic evaluation.

Learn More

Keep in Mind

Results come from benchmarks of factual and expert Q&A and from simulated parallelism metrics; real-world wall-clock speedup depends on your hardware and deployment. The approach requires a reliable way to judge subagent reports (the rubric), which may need domain tuning or human labels. Because the training reward remains final-answer correctness, systems are not explicitly optimized for latency or token cost — trade-offs may need extra constraints for production. Reflection Pattern.

Methodology & More

Delegation is framed as active context management: the main agent assigns compact, self-contained subtasks to isolated subagents, each of which returns a condensed report with evidence and conclusions. Rather than giving every subagent the same team reward, report-level crediting judges each report as helpful, neutral, or harmful and gates how the team outcome is inherited. Helpful reports keep only positive team advantage, harmful reports keep only negative advantage, and neutral reports keep the unfiltered advantage. The main agent still receives the full team outcome, so its delegation choices remain aligned with final task success. Training jointly updates both main and subagent policies using reinforcement learning on full swarm rollouts, but the rubric filters misassigned credit rather than adding extra rewards for speed or token economy. Planning Pattern. Evaluation used five benchmarks (English and Chinese factual search, general assistant tasks, expert questions, and set-valued research) and two model families. The learned policy improved final-answer quality while reducing the critical-path cost — the sequential depth of work — by about 24.5% and boosting accuracy by 2.49 percentage points versus a joint-training baseline that broadcast credit to all subagents. The paper reports these gains without explicit cost or speed rewards, and analyzes behavior with complementary metrics (critical-path cost and idealized per-trajectory speedup) so teams can weigh parallelism against token use and answer quality. In practice, the approach gives a practical lever—report-level credit—to make multi-agent research systems more trustworthy and efficient, though deployment requires attention to rubric design and real-world parallelism constraints. token economy.
Avoid common pitfallsLearn what failures to watch for
Learn More
Credibility Assessment:

ArXiv preprint, no venue/affiliations listed; multiple authors but low h-index values (mostly <10), indicating limited established reputation.