The industry's default way to make agents better is to keep stuffing skills into the context: retrieval tips, tool usage notes, checklists. A team led by the Chinese University of Hong Kong, Shenzhen (with collaborators at Fudan, UIUC, Tsinghua and NUS) took the opposite path in R2Orch, a paper posted in mid-October and revised on October 22: skills are scaffolding. Once the orchestrator has internalized them into its parameters, the scaffolding is removed piece by piece, and the final orchestrator deploys with zero skills. On November 2 one of the authors submitted it to Hugging Face Daily Papers, where it sat at #2 with 86 upvotes.
Two nested loops: one trains, one removes
The protagonist is a Director. For every task it starts from an empty canvas and issues one atomic edit at a time — add an agent, set a prompt, assign a model route, connect a hand-off — until an agent graph is built for frozen Worker models to execute. Training nests two loops.
The inner loop is self-play. A Proposer picks the tasks, driven by an objective called Recursive Learnability: it rewards tasks whose rollouts still reach different outcomes after the candidate skill is withheld, with learnability estimated as l(p)=4p(1−p), where p is the hidden part's success rate. This targets the known weak point of group-based RL like GRPO — when every rollout in a group succeeds or fails together, there is no advantage left to learn from, what the paper calls advantage collapse.
The outer loop is Skill Withdrawal. Each task runs G=5 graphs: 3 in a hidden part that cannot see the candidate skill, 2 in a visible part that can. Because hiding is random, the visible-minus-hidden mean difference is an unbiased estimate of the skill's net benefit; these contrasts accumulate into a Bayesian-bootstrap posterior, and once the posterior expected benefit falls below threshold, the skill is withdrawn. The paper reports all 12 withdrawals passed the criterion, with posterior E[Δ+] at most 0.019. The loops close into one cycle: failures are distilled into new skills, skills into parameters, and each withdrawal points the Proposer at new failures.
Twelve benchmarks, zero skills at test time
The setup: a frozen Qwen3.5-9B backbone, Director and Proposer both rank-64 LoRA adapters, one node with 4 A100 64GB GPUs, 200 cycles, only 18 seed skill cards, and a training pool of 6,154 tasks. In the paper's own tables, tested with zero skills, R2Orch tops all 12 benchmarks against six families of baselines: the in-distribution average rises from 86.0 (baseline PATS) to 89.8, out-of-distribution from 76.1 to 78.8; AIME 2026 goes from 74.0 to 81.3, MedQA 94.06, ALFWorld 90.63, WebShop 91.09, SWE-Bench Verified 42.81. Ablations show both loops matter — a fixed scaffold, where skills are never withdrawn, loses the most at zero-skill testing — and halfway through training, zero-skill accuracy already overtakes the with-skills state.
Transfer is real; so are the limits
Without retraining, the same Director improves all 6 other Worker backbones on every benchmark, and weaker backbones gain the most: the in-distribution spread across backbones narrows from 25.7 to 6.4 points, suggesting the graph supplies planning and checking rather than one model's private ability. The paper's own stated limitation is blunt: the initial skills are hand-customized per dataset, and deriving them automatically for a new domain remains open. The repository is freshly open-sourced (10 stars) under a research-use license, with the R2Orch-9b model and dataset released.
The endgame of skill libraries may not be an ever-thicker prompt, but something like a textbook: used, absorbed, discarded. When "still works after withdrawal" becomes the training objective itself, an agent has truly internalized the knowledge.
References: paper arXiv:2610.22740; code github.com/beita6969/R2Orch