Personalized Local-Global Collaboration (PLGC)

Updated 9 July 2026

PLGC is a design paradigm that integrates client-specific local signals with overarching global objectives to resolve optimization conflicts.
It employs adaptive credit assignment, dual representations, and structured fusion to balance fine-grained local details with global system coherence.
Its applications span personalized memory systems, recommendations, federated learning, and human-AI collaboration to enhance accuracy and efficiency.

Personalized Local-Global Collaboration (PLGC) denotes a class of learning and inference designs in which user-, client-, agent-, or task-specific local signals are coupled with broader global signals, rather than optimized in isolation. In the most explicit formulation, PLGC is motivated by personalized LLM memory systems in which independent optimization of specialized agents can create local conflicts and global objective ambiguity, so local improvements must be aligned with end-to-end query-answer accuracy (Mao et al., 13 Mar 2026). Closely related formulations appear in recommendation, federated and decentralized learning, speech enhancement, edge-agent interoperability, and human-AI collaboration, where “local” may refer to user history, client data, short-range context, or module-specific objectives, while “global” may refer to shared structure across users, system-level accuracy, network-wide efficiency, or long-range semantic context (Yang et al., 2023, Yu et al., 2024, Huang et al., 20 Jan 2025, Kelley et al., 31 Oct 2025, Wang et al., 18 May 2026).

1. Definition and scope

In the cited literature, PLGC is used either explicitly or as a direct interpretive framing for methods that must reconcile two competing demands: preserving local specificity and exploiting global regularities. The local side is specialized and heterogeneous. It may be fine-grained conversational evidence and user profiles in personalized memory systems, local textual and entity semantics in news recommendation, class-wise prompts or personalized heads in federated learning, short-time time-frequency structure in personalized speech enhancement, or individual work style in human-AI collaboration. The global side is correspondingly broader: end-to-end answer accuracy, cross-user behavioral graphs, server-aggregated prompts or prototypes, utterance-level speaker traits, collective cognition, or network-wide load balancing (Mao et al., 13 Mar 2026, Yang et al., 2023, Yu et al., 2024, Huang et al., 2023, Huang et al., 20 Jan 2025, Wang et al., 18 May 2026).

The concept is therefore not a single architecture. It is a recurring systems principle: local modules remain specialized, but their outputs, rewards, or parameters are coordinated through a mechanism that makes local progress useful at the system level. In some works the coordinating object is a sequential MDP with adaptive credit assignment; in others it is a graph, a hypergraph, a prompt pool, a low-rank global representation, a collaboration matrix, a Bayesian routing game, or a hypernetwork-conditioned client descriptor (Mao et al., 13 Mar 2026, Yang et al., 2023, Mukherjee et al., 2024, Liang et al., 19 Aug 2025).

A useful antecedent appears in work on competitive learning, which showed that global topology can be uncovered through strictly local interactions when local map quality is equalized across agents. That result is not personalized in the later federated or user-adaptive sense, but it established an important precursor: local interaction rules can induce globally coherent structure without explicit global interaction (Zantedeschi et al., 2019).

Setting	Local component	Global component
Personalized memory systems	Extraction, profile, and retrieval agents with task-specific rewards	End-to-end query-answer accuracy and inter-agent credit assignment
News recommendation	Local text and entity semantics from user history	Global news and entity graphs from population behavior
Federated continual learning	Fine-grained local prompts or client-specific heads	Coarse global prompts or shared backbone knowledge
Personalized speech enhancement	Short-time time-frequency cues	Utterance-level speaker and spectral context
Edge-agent interoperability	Local personalized agent and user demand	Network-wide specialization and load balancing

2. Recurring mathematical structure

A first recurring pattern is the decomposition of optimization into local terms, global terms, and an explicit coupling term. In CoMAM, the personalized memory pipeline is regularized as a 3-step sequential MDP whose agents receive local rewards for extraction, profile abstraction, and retrieval, plus a global reward for exact-match answer accuracy. Each agent is optimized with an integrated reward

$r_{\mathrm{final},n}^{(i)} = r^{(i)}_{(n)} + w_n \cdot r^{(i)}_{(3)},$

where the adaptive weight $w_n$ is derived from ranking consistency between local and global rewards via NDCG (Mao et al., 13 Mar 2026). In FedSLR, the same local-global logic is expressed structurally rather than sequentially: the shared model is a low-rank Global Knowledge Representation, while each client learns a sparse additive personalization $p_i$ , so inference is performed with $f(x; \mathrm{GKR} + p_i)$ (Huang et al., 2023).

A second pattern is the use of dual representations followed by fusion. GLORY learns local representations from titles and entities, then enriches them with a global news graph and a global entity graph before scoring candidate news with

$s(u,i)=\mathrm{emb}_{\mathrm{user}}^\top \mathrm{emb}_{\mathrm{cand}}.$

The user and candidate embeddings are each formed by attentively aggregating local and global views (Yang et al., 2023). GPFL applies the same principle in federated learning at the feature level: a shared feature extractor is routed through a Conditional Valve into a global-view feature $f_i^G$ and a personalized-view feature $f_i^P$ , then trained with a joint loss combining personalized cross-entropy, angle-level global guidance, and magnitude-level global guidance (Zhang et al., 2023).

A third pattern is adaptive collaboration coefficients. Learn2pFed defines a client-specific diagonal participation matrix $\Lambda_i$ , from which a derived participation degree

$s_{i,j}=\frac{(\Lambda_i)_{jj}}{(\Lambda_i)_{jj}+\rho_i}$

controls how strongly each parameter participates in collaboration (Lv et al., 2024). PGFedSplit blends a client head with the periodically aggregated head through

$\phi_i^{(t)}=\alpha_i^{(t)}\phi_i^{(t-1)} + \bigl(1-\alpha_i^{(t)}\bigr)\bar{\phi}^{(t)},$

where $w_n$ 0 is learned from a distillation-regularized objective and then fed back into adaptive aggregation scheduling (Kang et al., 26 May 2026). FedPAC similarly computes a client-specific convex combination of classifier heads

$w_n$ 1

with $w_n$ 2 obtained from a bias-variance quadratic program (Xu et al., 2023). Across these formulations, PLGC is not mere interpolation; it is usually a learned or data-dependent allocation rule.

3. Representative system designs beyond federated optimization

In personalized memory systems, PLGC is instantiated as collaboration among specialized agents. CoMAM defines an Extraction Agent, a Profile Agent, and a Retrieval Agent that operate sequentially during training, even though construction and retrieval may be asynchronous in deployment. The states $w_n$ 3, $w_n$ 4, $w_n$ 5, and $w_n$ 6 embed inter-agent dependencies directly into the transition structure, so the system can align coverage, abstraction quality, retrieval precision, and final answer accuracy (Mao et al., 13 Mar 2026). The key point is that memory quality is not assessed only locally; it is judged by whether downstream retrieval and answer generation improve.

In recommendation, PLGC typically means combining user-specific local signals with cross-user global structure. GLORY is exemplary: the Global-aware Historical News Encoder augments a user’s clicked history with a directed global news graph built from reading transitions, while the Global-aware Candidate News Encoder augments candidate items with an undirected global entity graph built from inter-news entity transitions. Both encoders fuse local textual or entity semantics with graph-enhanced global representations, yielding improved ranking and diversity (Yang et al., 2023). LGMRec applies a comparable decomposition in multimodal recommendation: collaborative-related local interests and modality-related local interests are explicitly decoupled, then fused with modality-aware global hypergraph embeddings that model higher-order dependencies beyond local pairwise edges (Guo et al., 2023).

In personalized speech enhancement, SEF-PNet treats PLGC as collaboration between local short-time structure and global utterance-level context. Interactive Speaker Adaptation aligns enrollment and mixture signals, while Local-Global Context Aggregation uses a Global Attention branch and a Local Attention branch to build a channel-attention mask

$w_n$ 7

so denoising remains locally precise but globally speaker-consistent (Huang et al., 20 Jan 2025). The same structural idea appears in SpecSteer for personalized generation, where the device-side Specialist drafts with private local context, the cloud-side Generalist verifies through a ratio

$w_n$ 8

and recovery re-injects local intent through contrastive steering logits

$w_n$ 9

without exposing raw local context to the cloud (Lv et al., 17 Mar 2026).

Human-AI collaboration work pushes the term in a cognitive direction. In the creative-work study, personalization is built from psychometrics and an AI-guided interview, but the relevant global effect is not a shared model parameter. It is a multi-turn collaborative process: collective memory, attention, and reasoning improve because the assistant begins with a richer partner model. The reported mediation results treat PLGC as a causal pathway from local profile information to global collaborative cognition (Kelley et al., 31 Oct 2025). PPAI extends the same logic to edge-agent interoperability: the local personalized agent delegates when another agent is semantically better suited, and the global coordination problem becomes one of specialization-aware routing and load balance in a churn-prone P2P network (Wang et al., 18 May 2026).

4. Federated and decentralized realizations

Federated learning provides the densest body of explicit PLGC mechanisms. FedMGP separates knowledge by granularity: a coarse-grained global prompt $p_i$ 0 is learned and selectively fused at the server, while fine-grained class-wise local prompts $p_i$ 1 remain private and are injected into ViT attention layers. This is designed to mitigate Spatial-Temporal Catastrophic Forgetting by keeping shared knowledge coarse and local adaptation fine-grained (Yu et al., 2024). FedSLR instead separates knowledge structurally, training a low-rank global component with nuclear-norm regularization and a sparse local add-on learned via $p_i$ 2-proximal updates, thereby improving personalization while reducing downlink communication (Huang et al., 2023). PGFedSplit adopts a split architecture, aggregates representation layers every round, aggregates personalization heads periodically, and further stabilizes local heads with Gaussian-guided synthetic representations under label imbalance and missing-class conditions (Kang et al., 26 May 2026).

A second federated family treats PLGC primarily as feature-space alignment. GPFL simultaneously learns global and personalized feature information on each client through a Conditional Valve and Global Category Embedding module; only shared modules are aggregated, while personalized heads remain local (Zhang et al., 2023). FedPAC likewise uses a shared global encoder but adds explicit local-global feature alignment through class-wise global centroids and a client-specific combination of classifier heads, computed from a quadratic bias-variance criterion (Xu et al., 2023). Learn2pFed moves the adaptation granularity down to individual parameters, learning which local parameters should participate in collaboration and to what extent through algorithm unrolling over an ADMM-based inner loop (Lv et al., 2024).

Decentralized work shows that PLGC does not require a server. MAPL separates Personalized Model Learning from Collaborative Graph Learning: local models are trained with intra-client and inter-client contrastive objectives, while a decentralized collaboration graph is updated from neighbor similarities and then used to aggregate prototypes (Mukherjee et al., 2024). Earlier fully decentralized work jointly learned personalized boosting models and collaboration graphs through peer-to-peer exchanges, with local model updates restricted to neighbors and graph updates performed against a few random users, thereby making the global collaboration structure itself an object of decentralized optimization (Zantedeschi et al., 2019). FedSheafHN generalizes the server-side descriptor idea to graph federated learning: clients send graph-level embeddings, the server builds a collaboration graph over clients, applies sheaf diffusion to enrich those descriptors, and uses a hypernetwork to generate personalized backbones that are returned to each client (Liang et al., 19 Aug 2025).

These systems differ sharply in implementation, but they share a common principle: the global part is not simply “the average model.” It is a structured collaborative object—prompt pool, centroid set, low-rank subspace, adaptive schedule, collaboration graph, or sheaf-enriched descriptor—that exists to improve personalized local behavior rather than replace it.

5. Empirical patterns and common misconceptions

Reported results are consistently favorable when local and global components are jointly optimized rather than separately tuned. In personalized memory systems, CoMAM reports main query-answer accuracies of 0.64, 0.70, and 0.66 for Qwen at 32K, 128K, and 1M context lengths, and 0.62, 0.68, and 0.69 for Llama, with improvements over Memory-R1 and Mem1 ranging from 8.5–16.7% for Qwen and 8.8–13.1% for Llama; removing or leaving any agent untrained reduces performance by 2–8% (Mao et al., 13 Mar 2026). In news recommendation, GLORY reaches AUC 68.15 and nDCG@10 42.78 on MIND-small, AUC 69.54 and nDCG@10 44.19 on MIND-large, and AUC 74.31 and nDCG@10 49.51 on Adressa, while also improving diversity metrics such as ILAD@5 and ILMD@5 over NRMS (Yang et al., 2023). In personalized speech enhancement, SEF-PNet improves a 2-speaker Libri2Mix baseline from 11.62 dB SI-SDR to 13.00 dB SI-SDR and from 2.76 to 3.05 PESQ, with fewer parameters than the baseline separator (Huang et al., 20 Jan 2025).

Federated results show the same trend. FedMGP reports 90.56% average global accuracy in asynchronous PFCL and 83.46% in synchronous PFCL on CIFAR-100 with ViT-B/16, outperforming the listed baselines and showing near-zero temporal forgetting when local prompts are kept class-wise (Yu et al., 2024). PGFedSplit reports 97.62% on Fashion-MNIST, 90.57% on CIFAR-10, 64.41% on CIFAR-100, and 38.45% on Tiny-ImageNet under Dir $p_i$ 3, with stable convergence and strong partial-participation performance (Kang et al., 26 May 2026). GPFL reports gains of up to 8.99% in accuracy over baselines, while also improving fairness and stability under heterogeneous client participation (Zhang et al., 2023). FedSheafHN achieves 82.35 on Cora, 92.99 on Amazon-Photo, and 71.75 on ogbn-arxiv in the reported non-overlapping setting, and its ablation isolates gains from the collaboration graph, sheaf diffusion, attention, and dynamic embedding (Liang et al., 19 Aug 2025).

Two additional patterns are notable. First, PLGC often improves system properties other than raw accuracy. PPAI reports up to 7.96% accuracy improvement and 16.34% latency reduction over its baseline while balancing load in a P2P edge-agent network (Wang et al., 18 May 2026). SpecSteer reports a 2.36× speedup over standard baselines while improving personalized generation quality (Lv et al., 17 Mar 2026). Second, the literature repeatedly shows that simple equal mixing is usually suboptimal. CoMAM finds “Ours” superior to local-only, global-only, and equal local-global mixing (Mao et al., 13 Mar 2026); Learn2pFed learns per-parameter participation rather than fixing a collaboration mask (Lv et al., 2024); PGFedSplit adjusts its aggregation interval based on the evolution of $p_i$ 4 rather than using a static schedule (Kang et al., 26 May 2026).

These results also clarify three common misconceptions. PLGC is not equivalent to naive interpolation between a local model and a global model; most successful variants learn collaboration weights, routing policies, or structured decompositions. It is not restricted to federated averaging; graph-based, hypergraph-based, sheaf-based, and sequential-credit formulations are prominent. Nor does “global” necessarily mean raw-data centralization: many systems retain private prompts, local heads, private context, or raw histories on-device and exchange only compressed statistics, prompts, prototypes, capability vectors, or sparse corrections (Yu et al., 2024, Huang et al., 2023, Lv et al., 17 Mar 2026).

6. Limitations, controversies, and open directions

Recurring limitations in the cited work concern the fidelity of the local-global interface itself. In CoMAM, mis-specified evidence $p_i$ 5 can distort local rewards, and the method assumes availability of ground-truth evidence and answers for reward construction (Mao et al., 13 Mar 2026). GLORY uses static global graphs built from training logs, so temporal drift, popularity bias, and noisy transitions remain concerns (Yang et al., 2023). PGFedSplit depends on Gaussian class-conditional representations and a simple $p_i$ 6 adaptive period update, which may be too coarse in strongly multimodal or rapidly changing settings (Kang et al., 26 May 2026). FedSLR inherits the computational burden of layer-wise SVD and may over-compress global structure under aggressive nuclear regularization (Huang et al., 2023). SpecSteer reduces privacy exposure but still transmits drafted token IDs and accepts a calibration trade-off through $p_i$ 7 and $p_i$ 8; PPAI explicitly leaves trust, adversarial agents, DoS, and rate limiting to future work (Lv et al., 17 Mar 2026, Wang et al., 18 May 2026).

A second limitation is domain dependence. Some PLGC mechanisms assume compact client sets, common label spaces, or stable shared backbones. FedMGP requires a shared frozen ViT and a proxy dataset for selective prompt fusion (Yu et al., 2024). GPFL assumes trainable category embeddings over a shared label space (Zhang et al., 2023). FedPAC relies on comparable classifier heads and class-wise feature statistics (Xu et al., 2023). The competitive-learning precursor shows another boundary condition: strictly local interaction can uncover global topology on continuous manifolds, but disconnected clusters remain difficult because units in empty gaps do not receive samples and cannot relay corrective forces across components (Zantedeschi et al., 2019).

A broader implication is that future PLGC systems will likely move toward more explicit uncertainty handling, richer global structures, and tighter privacy controls. The surveyed literature already points in that direction: Bayesian routing under partial observability in PPAI, Bayesian knowledge fusion in SpecSteer, adaptive credit assignment in CoMAM, sheaf diffusion over client descriptors in FedSheafHN, and learned parameter-participation degrees in Learn2pFed all replace fixed collaboration heuristics with data-dependent control (Wang et al., 18 May 2026, Lv et al., 17 Mar 2026, Liang et al., 19 Aug 2025, Lv et al., 2024). This suggests that PLGC is evolving from a broad intuition—combine local specificity with global structure—into a more precise design discipline centered on calibrated interfaces, adaptive coordination, and modular privacy-preserving collaboration.