CollabFlow: Recursive Self-Improvement of Agent Collaboration

Xiao Huang; Mingda Zhang; Junming Zhang; Qiang Huang; Hanwen Zhang; Yue Dai; Zijia Wang; Xiaoying Tang

cs.MASep 29, 2026

CollabFlow: Recursive Self-Improvement of Agent Collaboration

Authors: Xiao Huang, Mingda Zhang, Junming Zhang, Qiang Huang, Hanwen Zhang, Yue Dai, Zijia Wang, Xiaoying Tang

Organizations: The Chinese University of Hong Kong, Shenzhen · Fudan University · University of Oxford

Abstract

Recursive self-improvement (RSI) lets a system improve from its own outcomes; in LLM-based multi-agent systems, Agents refine one another within a task, and outcomes improve how they collaborate across tasks. However, existing multi-agent collaboration leaves this loop open: collaboration is pre-defined at the operator level, topology-only learning keeps verbatim exchange that propagates errors, and reward maximization on a system's own outcomes concentrates on a few teams. To address these challenges, we propose CollabFlow, an RSI system of Learned Agent Collaboration: a trainable Collab-Director constructs teams of complete Agents, a frozen executor runs them, and each round's outcomes retrain the director. Within each round, the edges of a collaboration graph carry protocols of Evidence-Conditioned Communication: a receiver adopts a differing answer only when the sender's evidence is stronger by a margin, so the director learns who communicates and how. Across rounds, we further propose Collaborative Trajectory Balance (CTB), a flow-based objective that credits each team once across its construction orders and targets a reward-proportional distribution over teams, so several good teams stay in play. We also bound how far this self-generated target moves between rounds, which shrinks as records accumulate. On twelve datasets, CollabFlow outperforms all baselines and keeps improving across rounds. Code is available at https://anonymous.4open.science/r/CollabFlow-631E.

Figures & tables

Figure 1: CollabFlow overview. (a) The director

\pi_{\theta}

builds a team for executor

K_{\chi}

. (b) An evidence rule gates each message. (c) CTB retrains

\pi_{\theta}

Figure 2: Multi-agent orchestration paradigms as self-improvement loops. Each paradigm runs the same loop of candidate designs, execution, outcome, feedback assessment, update, and re-execution, but updates a different object: (a) workflow orchestration updates operators and their order for one executor; (b) adaptive organization updates which Agents are used and linked; (c) communication design adjusts messages and channels; (d) CollabFlow jointly updates members, links, and edge protocols, trains the Collab-Director with CTB, and admits each message through an evidence gate.

Figure 3: The self-improvement loop of CollabFlow. (a) Within a round, the Collab-Director builds a team from fixed Agent profiles through checked actions, the frozen executor runs it, and the evaluated outcome becomes a replay record (Section 4.1 ). (b) On each edge, the protocol sets how many evidence gates run, and a fixed evidence rule keeps, adopts, or revises the receiver’s candidate (Section 4.2 ). (c) Across rounds, CTB merges construction histories only at the complete canonical team and updates the director with the length-normalized residual and a KL term (Section 4.3 ).

		Baseline	SFT	GRPO	AFlow	Agent+RL		Skill evolution		Ours
Dataset	Metric	Qwen3.5	Qwen3.5	Qwen3.5	Qwen3.5	FlowSteer	Evolving	SkillFlow	SkillRL	CollabFlow ( $\Delta\uparrow$ )
(a) In-Distribution (IID) benchmarks
HotpotQA	EM	46.25 ${}_{\pm\text{0.86}}$	47.34 ${}_{\pm\text{0.89}}$	49.53 ${}_{\pm\text{1.62}}$	53.75 ${}_{\pm\text{1.78}}$	53.28 ${}_{\pm\text{2.02}}$	60.16 ${}_{\pm\text{2.14}}$	60.78 ${}_{\pm\text{0.86}}$	58.28 ${}_{\pm\text{2.04}}$	63.91 ${}_{\pm\text{1.69}}$ (+17.7)
	F1	65.59 ${}_{\pm\text{0.89}}$	67.32 ${}_{\pm\text{1.46}}$	67.80 ${}_{\pm\text{2.02}}$	69.97 ${}_{\pm\text{1.50}}$	71.38 ${}_{\pm\text{2.18}}$	72.95 ${}_{\pm\text{1.05}}$	78.37 ${}_{\pm\text{1.95}}$	75.71 ${}_{\pm\text{1.96}}$	79.72 ${}_{\pm\text{1.96}}$ (+14.1)
TriviaQA	EM	69.69 ${}_{\pm\text{1.28}}$	69.53 ${}_{\pm\text{1.66}}$	69.84 ${}_{\pm\text{1.96}}$	87.34 ${}_{\pm\text{1.60}}$	87.34 ${}_{\pm\text{1.02}}$	86.88 ${}_{\pm\text{1.16}}$	86.56 ${}_{\pm\text{1.50}}$	82.97 ${}_{\pm\text{0.35}}$	87.97 ${}_{\pm\text{1.88}}$ (+18.3)
	F1	74.67 ${}_{\pm\text{2.11}}$	77.03 ${}_{\pm\text{2.10}}$	76.62 ${}_{\pm\text{1.52}}$	89.04 ${}_{\pm\text{0.43}}$	88.55 ${}_{\pm\text{1.87}}$	89.71 ${}_{\pm\text{1.10}}$	89.41 ${}_{\pm\text{1.18}}$	86.28 ${}_{\pm\text{1.62}}$	91.23 ${}_{\pm\text{2.14}}$ (+16.6)

Table 1: Main results on twelve IID and OOD benchmarks. SFT and GRPO are applied on Qwen3.5-9B; all trainable orchestration methods use Qwen3.5-9B as the frozen executor under the same per-task budget.

\Delta\uparrow

is the gap to direct Qwen3.5-9B. Every cell is the mean over five runs; the subscript is the standard deviation across those five, propagated across benchmarks for the Avg. rows. QA averages are taken separately for EM and F1, and Acc. averages the primary metrics (accuracy, success rate, or Pass@1) of the other four benchmarks in each part.

Panel of Table 2 — Table 2: Paradigm comparison and ablation of CollabFlow on the six IID and six OOD benchmarks (Qwen3.5-9B executor; EM for QA; one run). Paradigm rows fix how Agents collaborate before execution: one tool-using Agent, parallel voting, a fixed chain, fully connected debate, and the evolving orchestrator. Team-level rows remove director training or recursion across rounds, or replace the learned team with a fixed team or a random legal graph; message-level rows remove communication or the evidence gate, force every edge to OneWay , or revise on every update.

Figure 5: Solution diversity, team-level training, skill evolution and cost. (a) Successful solutions on 1,438 items, with SkillOpt ( Yang et al., 2026b ) ; SW: success-weighted. (b) First 64 prompt groups of training (8 steps, 256 rollouts per method); TV between consecutive team distributions. (c) Success over 8 rollouts before and after the twenty evolution steps with the largest one-step gain; dashed: means. (d) Gain at the next step against five steps later. (e) The 55 skill edits by type (inner) and by whether their step still gains five steps later (outer). (f) Tokens per item.

Appendix figures & tables8 assets

Supplementary material from the paper’s appendix.

Appendix

Object / field	Meaning
Agent identity and role	Stable profile identifier, role description, and local objective.
Compatibility	Supported task families and capability labels.
Tool permissions	Permitted tools and whether an operation can modify protected environment state.
Visibility	Task evidence, messages, and observations visible to the Agent.
Execution mode	Stateless Agent, mutable Executor , or read-only Advisor .

Appendix

Table A.1: Conceptual schemas used by the method. Exact registry identifiers and serialization are configuration-dependent.

Action	Structural effect	Required conditions
add_agent $(a)$	Add $a$ to $V$ .	$a\in\mathcal{A}_{x}\setminus V$ ; profile, permission, capacity, and budget checks pass.
add_edge $(a_{i},a_{j},p)$	Add $(a_{i},a_{j},p)$ to $E$ .	Both endpoints have been selected; $p\in\mathcal{P}$ is supported for the task and endpoint modes; the edge is not already present; relation and budget checks pass.
bind_skill $(a,k)$	Attach library skill $k$ to Agent $a$ .	$a\in V$ ; the skill grants no tool permission and never overrides write permission.
set_output $(o^{\prime})$	Set output configuration to $o^{\prime}$ .	The output is not yet set; the configuration is supported, all referenced Agents exist, and its execution is compatible with task permissions and budget.
stop	Commit the terminal organization.	The team is nonempty, output is defined, and all task-specific terminal constraints are satisfied.

Appendix

Table A.2: Structural-action semantics and local validity conditions. Terminal constraints are checked separately from partial-state invariants.

Protocol	Execution meaning
FinalOnly	Make a completed result available without rerunning the receiver.
OneWay	Send evidence or a proposal to the receiver, which may revise once under the configured resource bound.
Interactive	Allow a bounded sequence of mutual correction rounds where the task and runtime support it.
Stateful restriction	Internal Advisors provide bounded one-way guidance to the sole mutable Executor, and write permission stays with the Executor.

Appendix

Table A.3: Communication semantics. Bounds and task-specific availability are set by the frozen runtime configuration.

Benchmark	Suite	Task	Metric	Eval. items
HotpotQA	IID	Multi-hop question answering	EM / F1	128
TriviaQA	IID	Factual question answering	EM / F1	128
AIME 2026	IID	Mathematical reasoning	Accuracy	30
MedXpertQA-Text	IID	Medical reasoning	Accuracy	128
ALFWorld	IID	Embodied task completion	Success rate	128
MBPP+	IID	Function-level code generation	Pass@1	128

Appendix

Table D.1: Benchmarks, metrics, and evaluation items.

Method	Role in comparison	Configuration
Direct Qwen3.5-9B	Shared-backbone reference	Checkpoint, tools, prompt, decoding and submission limits.
SFT	Task-level supervised baseline	Demonstration source, training split, trainable components, optimization budget.
GRPO	Task-level RL baseline	Policy target, reward, sample group, optimizer and rollout budget.
AFlow	Workflow-search baseline	Search budget, backend, candidate components and task adapter.
FlowSteer	Workflow-orchestration baseline	Policy/backend versions, action interface, reward and adapter changes.
Evolving orchestrator	Adaptive-organization baseline	Orchestrator policy, Agent pool, reward and adapter changes.

Appendix

Table E.1: Baselines and their configurations.

Row	Intervention
Single agent with tools	One tool-using Agent answers alone.
Parallel vote	Independent Agents answer; a majority vote selects the output.
Fixed chain	A fixed researcher–solver–verifier chain with verbatim messages.
Fully connected debate	Every Agent receives every other Agent’s full output each round.
Evolving orchestrator	The orchestrator of Dang et al. (2026) activates Agents in sequence.
$-$ Director training	The untrained Collab-Director builds the teams.

Appendix

Table E.2: Paradigm and ablation rows of Table 4 .

(a) In-Distribution (IID) benchmarks
	HotpotQA	TriviaQA	AIME 2026	MedXpertQA	ALFWorld	MBPP+	Avg. (IID)
Executor	EM	EM	Acc.	Acc.	SR	Pass@1	Direct	CollabFlow
GPT-5.6-Luna	55.47 (+33.59)	88.28 (+35.16)	93.33 (+16.67)	100.00 (+62.50)	89.06 (+16.41)	92.97 (+5.47)	58.22	86.52 (+28.30)
DeepSeek V4	58.59 (+28.91)	90.63 (+14.84)	83.33 (+33.33)	100.00 (+21.09)	81.25 (+65.63)	89.84 (+3.91)	55.99	83.94 (+27.95)
GLM	65.63 (+38.28)	87.50 (+53.13)	56.67 (+13.33)	100.00 (+10.16)	96.09 (+7.03)	84.38 (+19.53)	58.13	81.71 (+23.58)
DeepSeek V4.1	64.06 (+24.22)	89.06 (+23.44)	83.33 (+13.33)	100.00 (+18.75)	94.53 (+36.72)	88.28 (+7.03)	65.96	86.55 (+20.58)

Appendix

Table H.1: CollabFlow with other frozen executors. Each cell gives the score with CollabFlow and, in pink, its gain over direct inference by the same executor; averages cover each part’s six benchmarks.

	Mean per item			Spread of the total
Method	Total $\downarrow$	Input $\downarrow$	Output $\downarrow$	P90 $\downarrow$	Max $\downarrow$	Std $\downarrow$	CV $\downarrow$
Single agent	12,010.0	10,889.8	1,120.2	27,507.4	39,279	13,892.5	1.157
Fixed team	10,404.4	9,119.6	1,284.8	13,091.0	14,437	2,373.8	0.228
Ensemble	24,081.8	20,894.2	3,187.6	37,974.8	47,942	12,557.5	0.521
SkillFlow	10,451.6	9,396.0	1,055.6	16,639.4	18,039	5,063.7	0.484
CollabFlow	9,676.6	8,645.8	1,030.8	11,490.0	11,986	1,903.2	0.197

Appendix

Table H.2: Tokens per item. Single agent: one tool-using Agent; Fixed team: the original hand-built team; Ensemble: independent Agents whose answers are combined. CV is the standard deviation divided by the mean total.

Explore similar work

May 13, 2026cs.LG

Reinforced Collaboration in Multi-Agent Flow Networks

Multi-agent systems provide a powerful way to extend large language models (LLMs) by decomposing a complex task into specialized subtasks handled by different agents. However, their performance is often hindered by error propagation, arising from suboptimal workflow design or inaccurate agent outputs, which can propagate through the agent collaboration process and degrade final results. To address the challenges, we present MANGO (Multi-Agent Network Gradient Optimization), a data-driven framework that organizes and refines agent collaboration via a flow network constructed from past successful workflows. MANGO integrates reinforcement learning and textual gradients to jointly optimize workflow paths and agent behaviors, while a skipping mechanism prevents redundant updates to well-optimized agents for improving efficiency. Extensive experiments on seven benchmarks show that MANGO achieves up to 12.8% performance improvement over state-of-the-art baselines, enhances efficiency by 47.4%, and generalizes effectively to unseen domains. Our code and datasets are publicly available at https://github.com/openJiuwen-ai/agent-store/tree/main/community/mango.

Zheng Wang, Yuang Liu, Yangkai Ding

Huawei Technologies Co., Ltd.

Collaboration Flow Map

Ask PDF

May 14, 2026cs.AI

Beyond Individual Intelligence: Surveying Collaboration, Failure Attribution, and Self-Evolution in LLM-based Multi-Agent Systems

LLM-based autonomous agents have demonstrated strong capabilities in reasoning, planning, and tool use, yet remain limited when tasks require sustained coordination across roles, tools, and environments. Multi-agent systems address this through structured collaboration among specialized agents, but tighter coordination also amplifies a less explored risk: errors can propagate across agents and interaction rounds, producing failures that are difficult to diagnose and rarely translate into structural self-improvement. Existing surveys cover individual agent capabilities, multi-agent collaboration, or agent self-evolution separately, leaving the causal dependencies among them unexamined. This survey provides a unified review organized around four causally linked stages, which we term the LIFE progression: Lay the capability foundation, Integrate agents through collaboration, Find faults through attribution, and Evolve through autonomous self-improvement. For each stage, we provide systematic taxonomies and formally characterize the dependencies between adjacent stages, revealing how each stage both depends on and constrains the next. Beyond synthesizing existing work, we identify open challenges at stage boundaries and propose a cross-stage research agenda for closed-loop multi-agent systems capable of continuously diagnosing failures, reorganizing structures, and refining agent behaviors, extending current coordination frameworks toward more self-organizing forms of collective intelligence. By bridging these previously fragmented research threads, this survey aims to offer both a systematic reference and a conceptual roadmap toward autonomous, self-improving multi-agent intelligence.

Shihao Qi, Jie Ma, Rui Xing +15

MOE KLINNS Lab · School of Cyber Science and Engineering, Xi’an Jiaotong University · School of Software Engineering, Xi’an Jiaotong University +6

Model-Based Multi-Agent Systems Autonomous Agents Self-Evolving Agents Collaboration Self-Evolution General Intelligence

Ask PDF

May 26, 2026cs.MA

AgensFlow: A Coordination-Policy Substrate for Multi-Agent Systems

Multi-agent systems built on large language models (LLMs) require many coordination choices that are difficult to fix a priori: which skill protocol to invoke, which agent role should perform a subtask, which model to bind to each role, how roles should interact, when to use retrieval or verification, and when to omit a step entirely. These choices interact with task regime and operational constraints, so static pipelines and one-off model comparisons provide only a limited view of the design space. This paper introduces AgensFlow, an open-source framework that treats multi-agent coordination as an online policy-learning problem under partial observability. The framework makes coordination decisions observable and learnable from repeated trajectories, rather than treating skill, role, model, topology, and evaluation choices as fixed pipeline design. AgensFlow is evaluated on two corpora: distributed-systems incident tasks and security-advisory tasks. The evaluation shows three main results: learned routing reaches a higher-quality operating point than a fixed pipeline baseline on coordination-heavy classes; skip:X isolates topology compression as a meaningful part of the substrate; and warm-started policy graphs can reduce exploration cost while preserving plateau quality. Overall, the results support that learned, auditable routing can improve coordination-heavy multi-agent workflows over static wiring.

Nicole Koenigstein

Independent researcher.

Multi-Agent Workflows Incident Response

Ask PDF