General Collaborative Intelligence: Architecting Cognition for Resilient Multi-Agent Ecosystems

Lei Zhang; Chun Ye; Le Yang; Zhaozhong Wang; Deng-Ping Fan; Hang Dai; Binglu Wang

cs.CVSep 19, 2026

General Collaborative Intelligence: Architecting Cognition for Resilient Multi-Agent Ecosystems

Authors: Lei Zhang, Chun Ye, Le Yang, Zhaozhong Wang, Deng-Ping Fan, Hang Dai, Binglu Wang

Organizations: School of Astronautics, Northwestern Polytechnical University, Xi’an, China. · School of Computer Science, Wuhan University, Wuhan, China. · School of Electronics and Control Engineering, Chang’an University, Xi’an, China. · School of Automation, Northwestern Polytechnical University, Xi’an, China. · College of Computer Science, Nankai University, Tianjin, China.

Abstract

Multi-agent unmanned systems are moving from isolated, ego-centric sensing toward collaborative intelligence, in which distributed agents exchange compact features to overcome a local observation trap that no single agent can escape: occlusions, finite sensor range, and environmental degradation. The field has matured across architectural, communication, embodied, resilience, and trust dimensions, yet existing surveys examine these dimensions in isolation and rarely expose their dependencies. This review offers a unified synthesis through two complementary lenses. The first is a five-dimensional taxonomy spanning collaboration stage, communication paradigm, fusion architecture, learning strategy, and application domain. The second is three cognitive synergy conditions, Semantic Disambiguation, Pragmatic Information Exchange, and Proactive Informational Foraging, that turn cognitive synergy into operational criteria. Across these lenses we survey collaboration architectures and topologies, neural-communication co-design that treats the channel as a differentiable pipeline component, embodied action-perception loops via multi-agent reinforcement learning, and resilience mechanisms for synchronization, uncertainty quantification, and label-efficient learning. We then map these advances onto four operational domains, V2X, unmanned aerial, industrial logistics, and smart cities, and onto the safety-privacy-utility triad. To counter benchmark saturation and evaluation fragmentation, we propose GCI-Bench, a five-pillar scoring protocol with a maturity model that makes the trade-offs of collaborative methods comparable across studies. A critical reflection on reproducibility, the sim-to-real gulf, and conditions under which collaboration degrades performance identifies open challenges and charts directions toward general collaborative intelligence under real-world uncertainty.

Figures & tables

Figure 1: Vision of general collaborative intelligence for multi-agent unmanned systems. The transition from isolated autonomy to GCI: heterogeneous agents exchange compact feature tokens through communication-aware fusion, evolving from deterministic data fusion toward cognitive synergy in a distributed ecosystem supporting open-world adaptation, pragmatic communication, embodied information foraging, and trustworthy synergy.

		Application Domain		Technical Core		Tasks & Robustness		System & Reflection
Ref.	Year	V2X	UxS/ISC	Co-Des.	Formal.	Tasks	Chall.	Embod.	Trust	Refl.
Han et al. [ 14 ]	2023	$\bullet$	–	–	–	$\bullet$	$\circ$	–	–	–
Yazgan et al. [ 11 ]	2024	$\bullet$	–	$\circ$	–	$\bullet$	$\circ$	–	$\circ$	$\circ$
Bai et al. [ 15 ]	2024	$\bullet$	$\circ$	–	–	$\circ$	–	–	–	–
Song et al. [ 10 ]	2025	$\bullet$	–	$\bullet$	–	$\circ$	$\bullet$	–	–	–
Huang et al. [ 3 ]	2025	$\bullet$	–	$\circ$	–	$\circ$	$\bullet$	–	$\circ$	–

Table 1: Systematic comparison of this review with representative prior surveys. A filled circle (

\bullet

) denotes comprehensive coverage; an open circle (

\circ

) denotes partial coverage; a dash (–) indicates the dimension is not addressed.

Condition	Observable Criterion	Typical Technical Realization	Operational Failure Test (metric; benchmark)
C1 Semantic Disambiguation	Resolve entity identity, spatial location, and trajectory intent beyond any single agent’s line-of-sight	Cross-attention Transformers [ 23 , 24 ] ; generative map priors (CoGMP) [ 25 ]	On a dynamic-occlusion split of OPV2V [ 26 ] : identity-switch or class-misattribution rate above a set threshold (e.g., $>$ 5%) as targets cross multi-agent occlusions
C2 Pragmatic Exchange	Bandwidth scales inversely with spatial confidence; high-entropy, high-surprise regions transmitted first	Extended information bottleneck (InfoCom [ 27 ] ); spatial confidence masking (Where2comm [ 21 ] )	On OPV2V [ 26 ] or V2X-Sim [ 28 ] : communication volume scaling near-linearly with agent density (log–log slope $\gtrsim 0.8$ ), or a non-negligible share of redundant background bytes
C3 Proactive Foraging	Trajectory adjustments or sensor re-aiming triggered by localized state-uncertainty peaks	Active perception loops; planning-integrated perception (Plan2comm [ 29 ] , CooperDrive [ 30 ] )	On an active-perception testbed (e.g., CSAOT [ 31 ] ): no measurable trajectory or sensor reconfiguration, and hence no entropy reduction, in regions of high epistemic uncertainty

Table 2: Operational criteria for cognitive synergy. The three conditions translate the conceptual claim of GCI into observable requirements that can be evaluated experimentally.

Figure 2: A multi-dimensional taxonomy of collaborative perception. The taxonomy organizes the field along five orthogonal dimensions. Dimension 1 (Collaboration Stage) determines the abstraction level at which information is exchanged, directly governing the rate–fidelity trade-off formalized in Eq. 1 . Dimension 2 (Communication Paradigm) spans the spectrum from periodic broadcasting to neural-communication co-design (Section 4 ). Dimension 3 (Fusion Architecture) specifies the topological structure of information aggregation. Dimension 4 (Learning Strategy) determines how the collaborative model acquires its parameters. Dimension 5 (Application Domain) captures the operational context that drives domain-specific design choices (Section 7 ).

Panel of Table 3 — Table 3: Taxonomy of collaboration architectures and cognitive synergy conditions. Methods are mapped onto the proposed five-dimensional framework.

Figure 3: Illustration of three primary collaboration strategies in multi-agent perception systems. The collaboration stage dictates the abstraction level at which information is exchanged. Early Collaboration (solid line) shares high-bandwidth raw data before the feature backbone. Intermediate Collaboration (dashed line) shares compressed features after the backbone, offering a balance between performance and communication cost. Late Collaboration (dotted line) shares low-bandwidth local outputs (e.g., object detections) after the perception head. Lower panels illustrate the fusion process within each strategy, where data from multiple agents (differentiated by color) is aggregated into a unified representation before being re-injected into the ego-agent’s pipeline.

Figure 4: Canonical collaborative perception architecture topologies. (a) Cloud-based centralized: remote data center aggregation, global coverage at high latency. (b) Edge/infrastructure-based: RSU local fusion, low latency with infrastructure dependency. (c) Decentralized P2P: on-board V2V fusion, high resilience at the cost of compute and synchronization overhead. The SDCN paradigm (Section 3.3 ) targets adaptive transitions among these topologies.

Panel of Table 4 — Table 4: Taxonomy of neural-communication co-design methods and cognitive synergy conditions.

Figure 5: Neural-communication co-design pipeline for collaborative perception. The pipeline transforms raw multi-agent observations into a collaborative perception output through five stages. (1) Multi-modal sensor observations from heterogeneous agents (LiDAR, camera, radar). (2) Neural feature encoding that compresses raw streams into compact latent representations. (3) An information bottleneck (IB) gate that enforces pragmatic communication by jointly minimizing transmitted redundancy while maximizing task-relevant mutual information given neighbors’ knowledge (Eq. 2 ). (4) A differentiable channel model (Deep JSCC) that directly maps latent codes to channel-modulated symbols, enabling end-to-end gradient flow from the task loss back through the physical layer and thereby avoiding the cliff effect. (5) Collaborative fusion that produces the multi-agent BEV detection, tracking, segmentation, and prediction output. The top feedback loop encodes the unified joint perception–communication Lagrangian (Eq. 3 ), enabling the entire stack to be optimized as a single differentiable system.

Panel of Table 5 — Table 5: Taxonomy of embodied synergy methods and cognitive synergy conditions.

Figure 6: The “three shadows” of distributed collaborative perception and their corresponding structural resilience mechanisms. The taxonomy addresses three foundational systemic vulnerabilities in collective intelligence: (1) Spatiotemporal Asynchrony (The Temporal Shadow): Communication latency $\Delta t$ and coordinate misalignments inherently induce catastrophic ghosting artifacts and localization degradation. This is mitigated through proactive neural manifold warping or predictive motion compensation networks that dynamically project stale historical features onto the current temporal baseline. (2) Epistemic Uncertainty (The Spatial Shadow): Localized line-of-sight occlusions (e.g., dynamic obstacles or blind spots) result in high-variance, uncertified feature volumes. The framework counteracts this by quantifying evidential uncertainty distributions and performing inverse-variance-weighted belief consensus to neutralize perceptual blind spots. (3) Semantic Scarcity (The Annotation Shadow): The high cost of multi-view cooperative annotation creates a severe training bottleneck. This is structurally resolved via the Dual Teacher-Student (CoDTS) paradigm, which distills mainstream and supplementary pseudo-labels from massive unlabeled data streams to achieve label-efficient collaborative optimization.

Panel of Table 6 — Table 6: Taxonomy of resilience methods and cognitive synergy conditions.

Systemic Vulnerability	Root Cause	Key Mechanisms	Representative Methods	Open Challenges
Spatiotemporal Asynchrony	Communication latency, heterogeneous clocks, coordinate misalignment	Neural manifold warping (PNAT), latent motion compensation (LMC), motion-aware robust communication	V2X-ViT [ 23 ] , CoBEVT [ 24 ] , SyncNet [ 69 ] , MRCNet [ 97 ] , V2X-INCOP [ 98 ]	Real-time inference under ms latency budgets; formal alignment guarantees for safety-critical settings
Epistemic Uncertainty	Sensor degradation, adversarial agents, occlusion	Evidential deep learning (EDL), conformal prediction, inverse-variance Bayesian aggregation, Byzantine-robust fusion	MOT-CUP [ 99 ] , GevBEV [ 47 ] , ROBOSAC [ 96 ] , Krum/Median [ 106 ]	Provable Byzantine resilience with formal convergence guarantees; runtime uncertainty calibration on embedded hardware
Semantic Scarcity	High multi-view annotation cost, domain shift across deployments	Consensual contrastive learning (CCL), dual teacher-student (CoDTS), domain adversarial adaptation (DANN)	MA-SUE [ 100 ] , CoDTS [ 102 ] , domain generalization [ 110 ]	Closing the sim-to-real generalization gap; scaling to diverse real-world edge cases with minimal supervision

Table 7: The “Three Shadows” of collaborative resilience: systemic vulnerabilities, mechanisms, and open challenges. Each shadow exposes a distinct failure surface in distributed perception; the corresponding mechanisms represent the current state of the art in addressing them.

Figure 7: Domain-specific synergy across collaborative-intelligence deployments. The illustration links four application contexts (industrial logistics, connected mobility, smart-city public services, and environmental monitoring) to the shared requirement for distributed sensing, communication-aware fusion, and proactive information acquisition.

Panel of Table 8 — Table 8: Taxonomy of domain-specific methods and cognitive synergy conditions.

Panel of Table 9 — Table 9: Taxonomy of trustworthy CI methods and cognitive synergy conditions. Methods are mapped onto the proposed five-dimensional framework.

Technique	Protection Granularity	Comm. Overhead	Accuracy Penalty	Real-Time	Representative Work
Differential Privacy (DP)	Feature / trajectory	Low	Moderate	High	Feature-map DP [ 166 ] , trajectory DP [ 168 ] , DP-aware fusion [ 172 ]
Federated Learning (FL)	Model	Moderate	Low	Moderate	FedDrive [ 156 ] , FedCP [ 157 ] , Hierarchical FL [ 170 ]
Homomorphic Encryption (HE)	Raw data / feature	Very high	None	Low	Secure MKHE [ 158 ]
Secure Multi-Party Computation (SMPC)	Raw data / feature	High	None	Low–Mod.	Oblivious Fusion [ 155 ] , SMPC-based fusion [ 171 ]

Table 10: Comparative overview of privacy-preserving techniques for collaborative perception. Each paradigm occupies a distinct position in the trade space between protection strength, communication overhead, and real-time feasibility.

Panel of Table 11 — Table 11: Taxonomy of open-world CI methods and cognitive synergy conditions.

Figure 8: Roadmap toward general collaborative intelligence. Four converging research trajectories, i.e., collaborative foundation models, pragmatic game-theoretic communication, hardware-agnostic universal semantics, and proactive swarm embodiment, connect the current state of collaborative perception with the long-term vision of GCI.

Challenge	Severity	Scope	Near-Term Solvability	Representative Evidence	Potential Mitigation Pathways
Benchmark Saturation	High	V2X domain	Moderate	AP clustering within few percentage points [ 180 ] ; unreliable cross-paper comparisons [ 181 ]	Multi-dimensional metrics (accuracy–bandwidth–latency–energy); benchmarks with 10+ agents
Reproducibility Gap	High	All domains	Low	Missing code and incomplete training configurations; multi-agent training instability [ 183 ]	Standardized protocols [ 182 ] ; containerized implementations; locked test-set leaderboards
Sim-to-Real Gulf	Critical	All domains	Low	AWGN channel models vs. real mmWave [ 185 ] ; sensor fidelity gaps [ 184 ] ; MDrive planning-perception gap [ 179 ]	Domain randomization; hardware-in-the-loop; large-scale real-world multi-agent benchmarks
Under-Reported Failure Modes	Moderate	All domains	Moderate	LE/CE error taxonomy [ 181 ] ; Byzantine-robust fusion rarely benchmarked [ 106 ] ; latency-induced degradation [ 186 ]	Failure-analysis tracks; negative-result venues; standardized stress-test suites
Scalability & Agent Churn	High	Large-scale deploy.	Low	$O(N^{2})$ communication growth [ 187 ] ; cold-start problem unaddressed [ 190 ]	Hierarchical fusion [ 189 ] ; sparse attention [ 191 ] ; incremental onboarding protocols

Table 12: Open challenges in collaborative perception: severity, scope, and mitigation pathways. Challenges are ordered approximately by their impact on the field’s transition from laboratory benchmarks to robust real-world deployment.

Figure 9: Overview of the GCI-Bench evaluation framework. (a) Scoring pipeline: raw operating points reported by a method on standard benchmarks are normalized into the five pillars of Table 13 , each scored on a 0–100 scale against the disclosed single-agent and oracle anchors, and aggregated into a composite under the explicit missing-data rule and maturity caps. (b) Case-study scorecard: pillar profiles of the five representative methods of Table 15 together with the aspirational GCI Target profile; unreported pillars are plotted at 0 only to close the polygons and denote missing rather than measured values.

Pillar	What It Measures	Primary Metrics	Normalization	Benchmark Sources
P1. Perception Fidelity (PF)	Collaborative detection and tracking accuracy relative to single-agent baseline	[email protected], [email protected], NDS, AMOTA	$\mathrm{PF}=100\times(m_{\mathrm{collab}}-m_{\mathrm{single}})/(m_{\mathrm{oracle}}-m_{\mathrm{single}})$ $m$ = primary detection metric; oracle = perfect-communication upper bound	OPV2V [ 26 ] , V2XSet [ 23 ] , DAIR-V2X [ 32 ] , V2V4Real [ 117 ]
P2. Communication Efficiency (CE)	Perceptual gain per unit of communication cost	Bandwidth (KB/frame), latency (ms), bits per AP gain	$\mathrm{CE}=100\times\max\bigl(0,\,1-C_{\mathrm{method}}/C_{\mathrm{ref}}\bigr)$ , clamped to $[0,100]$ $C$ = communication cost; $C_{\mathrm{ref}}$ = fixed reference cost, so $\mathrm{CE}=0$ at $C_{\mathrm{ref}}$	All benchmarks with communication measurements
P3. Resilience (RS)	Performance degradation under adverse conditions	AP drop under packet loss (10–50%), pose noise ( $\sigma=0.1$ – $0.3$ m), adversarial perturbation	$\mathrm{RS}=100\times\Bigl(1-\frac{1}{K}\sum\nolimits_{k=1}^{K}\delta_{k}\Bigr)$ $\delta_{k}$ = relative AP drop under perturbation type $k$ ; averaged over $K$ types	OPV2V [ 26 ] , V2X-Sim [ 28 ] , custom stress-test suites [ 181 ]
P4. Scalability (SC)	Performance retention as the number of collaborating agents grows	AP at $N=2,5,10$ ; $O(N^{2})$ communication growth factor	$\mathrm{SC}=100\times\min\bigl(1,\,m_{N=10}/m_{N=2}\bigr)$ , clamped to $[0,100]$ accuracy retention as the agent count grows from $N{=}2$ to $N{=}10$	Benchmarks supporting 2–10 agents
P5. Trustworthiness (TW)	Privacy preservation, uncertainty calibration, and safety compliance	Privacy budget $\varepsilon$ , expected calibration error (ECE), ISO 26262/21448 [ 159 , 160 ] compliance level	$\mathrm{TW}=w_{\varepsilon}\,\mathrm{TW}_{\varepsilon}+w_{\mathrm{ECE}}\,\mathrm{TW}_{\mathrm{ECE}}+w_{\mathrm{safety}}\,\mathrm{TW}_{\mathrm{safety}}$ $w$ = domain-adjustable weights (summing to 1)	DAIR-V2X [ 32 ] , V2V4Real [ 117 ] , protocol-level tests

Table 13: GCI-Bench five-pillar evaluation system. Each pillar is scored on a 0–100 scale and normalized against a single-agent baseline. A weighted aggregate yields the GCI Composite Score.

Level	Designation	Score Range	Characteristics	Representative Archetype
L1	Isolated	$<20$	No effective collaboration; performance indistinguishable from or worse than single-agent baseline	Single-agent perception; naive late fusion that disregards neighbors’ raw features
L2	Reactive	$20$ – $40$	Basic information sharing without adaptivity; fixed communication schedule	Broadcast-based intermediate fusion (V2VNet [ 33 ] , F-Cooper [ 39 ] )
L3	Adaptive	$40$ – $60$	Communication-aware collaboration with selective information transmission	Saliency-driven fusion (Where2comm, Which2comm)
L4	Proactive	$60$ – $80$	Embodied collaboration closing the perception-action loop; proactive information foraging	MARL-driven embodied systems (CSAOT, Plan2comm, CoPnP)
L5	Cognitive	$80$ – $100$	Trustworthy, self-optimizing, cross-domain distributed cognitive system; plug-and-play interoperability across heterogeneous fleets	Aspirational target for GCI; no current system fully achieves this level

Table 14: GCI maturity model. Five capability levels defined by composite score ranges, with representative method archetypes.

Method	Stage	Comm.	PF	CE	RS	SC	TW	Comp.	[email protected]	KB/f	Lat.
Late Fusion [ 32 ]	Late	Metadata	25	99	20	75	–	55 (L2) †	75	$<1$	Low
V2VNet [ 33 ]	Intermed.	Broadcast	62	59	30	25	–	44 (L3)	82.4	26	Med
V2X-ViT [ 23 ]	Intermed.	Broadcast	66	6	35	30	–	34 (L2)	88.2 ‡	60	Med
Where2comm [ 21 ]	Intermed.	Saliency	63	92	40	45	–	60 (L3)	82.5	5	Med
InfoCom [ 27 ]	Intermed.	IB-pruned	–	98	35	50	–	61 (prov.) §	–	$\sim 1$	Med
GCI Target	–	–	95	90	90	85	85	89 (L5)	$>95$	$<5$	Low

Table 15: GCI-Bench case study: pillar scores and reported operating points for five representative methods. Normalised pillar scores (PF/CE/RS/SC/TW, 0–100) are derived from the disclosed illustrative operating points under the protocol of Table 13 ; the three right-hand columns list the assumed raw operating points of the demonstration. Scores illustrate the framework rather than provide definitive rankings. “–” indicates that the original publication did not report data for that pillar.

Explore similar work

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

Multi-Agent Collaboration LLM Agent Self-Improvement Multi-Agent LLMs Self-Evolving Agents

Ask PDF

Sep 17, 2026cs.AI

Rethinking Multi-Agent Collaboration: When More Is Less

The rapid advancement of large language models and single-agent harnesses has reshaped the landscape of autonomous systems, raising a critical question of when multi-agent collaboration offers genuine value. As individual agent capabilities continue to scale, multi-agent collaboration faces diminishing returns while incurring growing context overhead. Through systematic analysis, we delineate the capability boundaries of multi-agent collaboration relative to single-agent alternatives, showing that it confers systematic benefits specifically in long-horizon tasks with sparse dependencies, while single-agent harnesses remain superior in tightly coupled, sequential workflows. Building on these insights, we propose SAIGE, a lightweight multi-agent collaboration mechanism based on Semantic-Aware Incremental Graph Evolution. SAIGE models collaboration as a dynamically evolving graph, where nodes are agent instances spawned on demand and edges encode semantic dependencies established through content-based information retrieval. Experiments on long-horizon, complex task benchmarks show that SAIGE achieves a favorable trade-off between context efficiency and task performance, and that scaling the agent pool or deepening the recursion level does not consistently improve outcomes. Our findings suggest that multi-agent superiority is bounded by task structure rather than universal, and that more agents do not necessarily make a system more intelligent.

Yishuo Yuan, Yibo Wu, Yihan Zhang +5

Nanjing University · Shanghai Jiao Tong University

Multi-Agent Orchestration Multi-Agent Collaboration Multi-Agent Systems

Ask PDF

May 19, 2026cs.MA

What Do Agents Communicate? Characterizing Information Exchange in Multi-Agent Systems

Large Language Models (LLMs) have enabled collaborative Multi-Agent (MA) systems, where interacting agents improve performance through diverse reasoning and iterative refinement. However, these systems remain vulnerable to error propagation, where early-stage information degrades downstream reasoning. To address this, we conduct a systematic analysis of inter-agent communication to identify which information drives MA performance. We find that the absence of reasoning and verification in inter-agent communication significantly degrades performance. Based on these insights, we propose Category-Aware Recovery Augmentation (technique), which enforces the presence of critical information during communication. recovers up to 86.2% of failed cases. Our results highlight the key role of information quality in effective MA collaboration. Our code is available at https://anonymous.4open.science/r/cara_mas

Yong Jin Chun, Iftekhar Ahmed

Department of Informatics University of California, Irvine Irvine, CA, 92618

Multi-Agent Collaboration Multi-Agent Systems Multi-Agent Reasoning Multi-Agent Communication LLM Agent Reliability

Ask PDF