Agent·

LEAD-RL multi-agent deep research — Contribution-aware subagent credit — What does it mean?

The news. On November 12, 2026, Yuyang Wu, Yang Liu and co-authors posted From Delegation to Coordination: Learning Efficient Multi-Agent Deep Research (arXiv 2611.17041). It introduces LEAD-RL, which trains a research lead and its subagents as one shared policy and gates each subagent's credit with a training-time rubric. On BrowseComp, it reports +2.49 percentage points of accuracy and a 24.5% lower critical-path cost than the same joint training with shared team credit, with final-answer correctness as the only reward. Read the paper →

Picture a newsroom that pays bonuses by one rule: if the front-page story is right, every reporter gets the bonus; if it is wrong, every reporter is docked. It sounds fair, and it is badly broken. A reporter who filed a solid tip that the editor ignored gets docked. A reporter who filed a made-up quote that the editor happened to catch and throw out gets paid. Over many stories, the reporters learn that accuracy does not change their pay.

That is shared team credit in a multi-agent system. A main agent splits a research question into briefs and hands them to parallel subagents — the orchestrator-workers pattern. Each subagent searches in its own context and returns a report. When the main agent and the subagents are trained together, the only reward is whether the final answer was correct. LEAD-RL points out the gap: a winning rollout can contain a misleading report, and a losing rollout can contain the one report that eliminated the wrong candidate. Broadcasting the team score to both treats them the same.

The fix is to hire a fact-checker. During training only, a separate contribution judge reads each subagent report together with its brief, the subagent's own search evidence, and the reference answer, and labels it helpful (+1), neutral (0) or harmful (−1). Importantly, the judge never sees the main agent's final answer or whether the team won, so it grades the report on its own evidence, not on the team's luck. The judge is also cautious: a helpful or harmful label with confidence below 0.8 is turned into neutral, and "harmful" requires a materially misleading claim.

Report labelTeam did well (advantage +)Team did badly (advantage −)Effect
Helpful (+1)keeps the +gets 0 instead of −never docked for the editor's mistake
Neutral (0)keeps the +keeps the −plain shared credit where usefulness is unclear
Harmful (−1)gets 0 instead of +keeps the −never paid for a story the editor rescued

The gate only blocks credit that points the wrong way; it never invents a new reward. In the paper's notation, a helpful report gets max(A, 0), a harmful one gets min(A, 0), and a neutral one gets A unchanged, where A is the team's group-relative advantage. The main agent itself is not gated at all: it always receives the full team advantage, so its choices about what to delegate and how to combine the reports stay supervised by the final answer — the job of the supervisor in a supervisor/worker team. Each subagent's gated advantage is then divided by the number of subagents in that rollout that still carry a non-zero signal, so a rollout that spawns more subagents does not get a bigger total update just for spawning them.

Here is the arithmetic on one question (illustrative numbers). Training samples K = 4 rollouts; two answer correctly and two do not, so after group normalization the winners sit at roughly +1 and the losers at roughly −1. In losing rollout 3, one subagent's report was helpful — it ruled out a plausible wrong candidate, but the main agent ignored it. Shared credit gives that subagent −1; the gate gives max(−1, 0) = 0. In winning rollout 1, three subagents reported: one harmful report the main agent overrode, two neutral ones. Shared credit pays all three +1. The gate gives the harmful one min(+1, 0) = 0, which leaves two subagents with a non-zero signal, so each gets +1 ÷ 2 = 0.5 before a training-schedule weight w(t), which the authors reduce over training, is applied. Net effect: the two mis-credited updates (−1 and +1) become 0, and nothing else changes sign.

Why would a credit rule change cost when cost is not in the reward? The paper measures efficiency as critical-path cost: the main agent's steps plus, for each delegation, only the slowest subagent branch, because parallel branches wait for the longest one. With illustrative numbers, a main agent that takes 8 steps and dispatches three subagents that take 10, 12 and 30 steps has a sequential workload of 60 steps but a critical path of 8 + 30 = 38; one long branch dominates — the trade-off covered in coordination costs. The authors' hypothesis is that more reliable reports mean fewer compensating searches, so the path shrinks on its own. They say directly that they did not measure that mechanism, only the outcome below.

Training variant (BrowseComp)AccuracyHelpful reportsHarmful reportsCritical-path steps per questionSource
Starting point (supervised fine-tuning, SFT)72.50%27.11%12.89%403.37Table 3
Main-only RL (subagents frozen)70.39%27.42%11.21%421.76Table 3
Joint RL, shared team credit71.00%21.15%13.84%323.30Table 3
LEAD-RL (gated credit)73.49%31.52%9.87%244.22Table 3

The table shows the failure the gate targets. Under shared team credit, the share of helpful reports fell from 27.11% at the starting point to 21.15%, and harmful reports rose to 13.84% — the reporters learned that accuracy did not change their pay. Both plain RL variants ended below the starting checkpoint's accuracy. With the gate, helpful reports rose to 31.52%, harmful ones fell to 9.87%, accuracy reached 73.49%, and critical-path cost dropped from 323.30 to 244.22 steps per question (−24.5%). The report rates are graded with the same LLM-judged rubric used in training, so they are the paper's own measurement, not an independent audit. The lesson carries beyond this paper: when several agents share one reward, decide per agent which direction of credit it has earned before you train on it.

Goes deeper in: Agent Engineering → Agent Teams → Supervisor / Worker

Related explainers

Frequently Asked Questions

Check what you knowMap your AI & GPU knowledge across every track — free, role-based