Interaction-Centric Knowledge Transfer

• Interaction-centric knowledge transfer is a paradigm that emphasizes structured, context-aware interactions to determine what, when, and between whom knowledge is transferred.
• It employs formal models such as knowledge transfer graphs with dynamic gating and attention mechanisms to optimize learning and communication between agents.
• Empirical evidence shows that this approach improves performance in tasks like multi-agent reinforcement learning, image classification, and human-AI collaboration.

Interaction-centric knowledge transfer refers to a paradigm in which the mechanisms of transferring knowledge—within or between artificial agents, models, or human-AI teams—are explicitly constructed around structured, context-aware interactions, rather than through one-way or globally uniform signal propagation. This approach encompasses representation, learning, and evaluation of knowledge transfer as an emergent property of agent-agent, module-module, or human-machine interaction dynamics, leveraging explicit modeling of relationships, dependencies, communication policies, or attention mechanisms. The interaction-centric view sharpens both the scope and effectiveness of transfer by focusing not only on "what" is transferred (content) but also on "how," "when," and "between whom" transfer should occur.

1. Formal Models and Architectures for Interaction-Centric Knowledge Transfer

Interaction-centric knowledge transfer departs from classic teacher-student formulations by enabling flexible, compositional, and context-sensitive transfer graphs or protocols. In deep collaborative learning, the knowledge transfer graph (KTG) formalism (Minami et al., 2019) describes a directed graph $G=(V,E)$ in which each node $m_i$ is a model (with a pseudo-node $m_0$ for one-hot labels), and each directed edge $(m_s \rightarrow m_t)$ encodes a transfer relation. For a batch $B$, the loss per edge is

$$L_{s,t} = \sum_{x_n \in B} G_{s,t}(\mathrm{KL}(p_s(x_n)\,\|\,p_t(x_n)))$$

where $G_{s,t}$ is a gate controlling gradient flow. Each node accumulates losses over its incoming edges, and gradients are only back-propped into the destination node.

The framework generalizes past knowledge transfer methods:

• In classic KD, the graph is a one-way $m_{teacher} \rightarrow m_{student}$ with fixed uniform weighting.
• In deep mutual learning (DML), edges are bidirectional between identical students with homogeneous KL objectives.
• KTG allows arbitrary edge selection, directionality, loss weighting, selective gating, and non-tree-structured interactions.

Interaction-centricity also underpins frameworks in other domains:

• KIX constructs a multi-level policy architecture in RL where a meta-policy operates over transferable interaction types and object categories, selecting type-level interaction goals which are instantiated via low-level policies, all composed through explicit graphs and mappings (Kumar et al., 8 Feb 2024).
• Bridged-GNN/KBL in transfer learning replaces domain-level assumptions with per-sample "knowledge bridges," learning for each target instance a personalized set of sources and constructing a message-passing graph to propagate information accordingly (Bi et al., 2023).

2. Interaction Mechanisms: Gating, Attention, and Selective Routing

Effective interaction-centric transfer relies on mechanisms that modulate when and how information flows between entities. In KTG (Minami et al., 2019), "gates" on each edge can:

• Allow all transfer (Through), block transfer (Cutoff), warm-up transfer over time (Linear), or pass transfer only for samples where the source is correct (Correct).
• Search algorithms (e.g. ASHA) are employed to discover optimal graph structures and gating schedules.

In multi-agent RL, the Parallel Attentional Transfer framework (Liang et al., 2020) employs shared attention networks ("Attention Teacher Selector") to dynamically select relevant advice from peers based on the current agent state and peer history, weighting and combining teacher policy parameters. Each agent independently learns when to engage in self-learning vs. to solicit (and aggregate) advice via attention.

These interaction modulations enable:

• Selective transfer based on state, sample correctness, or phase of training.
• Robustness to noise and scalability by avoiding rigid or indiscriminate knowledge propagation.

3. Empirical Outcomes and Comparison to Non-Interaction-Centric Approaches

Interaction-centric transfer regularly exceeds traditional paradigms across domains:

• For deep collaborative image classification, KTG achieves gains of up to +4% (CIFAR-100), outperforming DML, teacher-student KD, and recent ensemble-based methods. These improvements arise from the learned flexibility: the system discovers, for example, cascading multi-phase pipelines (teacher $\to$ assistant $\to$ student), partial ensembles, or selective routing regimes unreachable by manual pattern design (Minami et al., 2019).
• In multi-agent RL, PAT yields 53% higher rewards than prior advising methods (e.g., LeCTR, AdHocTD) and scales gracefully to large teams, unlike centralized or vote-based transfer strategies (Liang et al., 2020).
• In transfer learning for data-scarce domains, Bridged-GNN increases macro F1-scores by 4.5–11.9 points by targeting sample-wise message passing, a result robust to noise and unrelational data (Bi et al., 2023).

Table 1: Representative Gains from Interaction-Centric Transfer

Domain	Method	Baseline	Interaction-Centric	Gain
CIFAR-100	ResNet32 (Van)	70.71%	KTG (Graph Search)	+4.00%
Multi-agent RL	LeCTR (Reward)	29.9	PAT (Reward)	+53%
Twitter_UD	S³D (F1-macro)	78.4%	Bridged-GNN_KTGNN	+4.5 pt

These results consistently demonstrate that leveraging structured, adaptive, or attention-based interaction protocols can actualize stronger and more targeted transfer, especially in complex or noisy learning environments.

4. Organizational and Societal Perspectives: Local Rules and Topology in Human Systems

Interaction-centric transfer also underlies organizational learning models. Cellular automata-based studies (Kowalska-Styczeń et al., 2017, Paradowski et al., 2017) reveal that knowledge diffusion is governed by local rules (knowledge can only be transferred from a neighbor with exactly one more "chunk"), and transfer effectiveness is strongly influenced by the topology and density of the interaction network:

• Increasing the size of each agent’s neighborhood (from 4 to 12) eliminates non-monotonic diffusion failures and halves the time to full coverage.
• The initial distribution of partial knowledge (number of seeds, parameter $p$) and the "social distance" threshold determine whether transfer propagates broadly or deadlocks occur.
• Organizational interventions (shortening social distances, broad seeding, flexible incentives) map directly onto manipulation of local rules and topology in the CA, translating to more rapid and complete knowledge transfer in real-world deployments.

This suggests that both biological and artificial systems benefit from granular, topology-aware interaction design for efficient and robust knowledge dissemination.

5. Specialized Protocols: Human-AI and Human-Human Knowledge Transfer

Interaction-centricity is critical in designing AI systems that transfer knowledge to humans. The KITE framework (Shi et al., 5 Jun 2025) defines the transfer not just as model accuracy but as the projection operator $\Pi_{M \rightarrow H}$: how model-internal knowledge is communicated and understood by humans.

• A two-phase experimental protocol—collaborative ideation (allowing open discussion, no copying) followed by independent task execution—isolates genuine transfer from rote copying.
• Behavioral and structural characteristics (e.g., stepwise explanations, adaptiveness to user skill, analogy use) predict transfer effectiveness more than static performance benchmarking.
• The empirical slope of collaborative improvement vs. model solo accuracy is consistently less than one, indicating that transfer capabilities must be explicitly optimized and cannot be assumed to scale with model competence.

Within organizational settings, Socially Interactive Agents (SIA) leverage multimodal, trust-calibrated dialogue and chain-of-thought (CoT) prompting to surface and encode tacit knowledge via human-like, incremental interaction (Benderoth et al., 27 Aug 2025). The RAG pipeline grounds these interactions in existing organizational memory, and the agent adapts style parameters based on warmth-competence models and user feedback, fostering privacy and disclosure.

6. Open Problems and Future Directions

Several persistent challenges and open directions characterize research on interaction-centric knowledge transfer:

• Automated Search: As the number of entities and transfer edges increases, the space of possible interaction graphs grows super-polynomially, necessitating scalable, possibly RL-based search or meta-optimization strategies (Minami et al., 2019).
• Evaluation: Distinguishing knowledge transfer per se from performance gains requires task design that isolates projection and internalization (e.g., the two-phase KITE protocol) (Shi et al., 5 Jun 2025).
• Cross-Domain and Cross-Modal Transfer: Frameworks such as Learning from Interactions (LFI) transfer cross-modal attention patterns from VLMs to VFMs, yielding improved alignment to human prediction patterns and better out-of-domain generalization (Gao et al., 23 Sep 2025).
• Robustness to Label/Schema Mismatches: Interaction-centric frameworks (e.g., TIN for HOI detection (Li et al., 2021), ACC for scene graph generation (Li et al., 8 Nov 2025)) enable transfer across varying taxonomies or backgrounds by decoupling universal interaction cues from task-specific label vocabularies.
• Multi-level and Hierarchical Interaction: KIX and Bridged-GNN illustrate the utility of decomposing transfer via hierarchical or type-instance policies and per-sample knowledge bridges, supporting generalization and adaptation without catastrophic interference (Kumar et al., 8 Feb 2024, Bi et al., 2023).

A plausible implication is that future AI and organizational systems will increasingly instantiate explicit, search-optimized interaction protocols—at both micro (sample, agent) and macro (topology, interface) scales—in order to maximize the fidelity, interpretability, and adaptability of knowledge transfer.

7. Summary Table: Core Mechanisms in Interaction-Centric Transfer

Mechanism	Example Framework	Function
Transfer graph/gates	KTG (Minami et al., 2019)	Control edge-wise transfer pathways
Attention selector	PAT (Liang et al., 2020), LFI	Select peers/interactions adaptively
Per-sample bridges	Bridged-GNN (Bi et al., 2023)	Connect each target to relevant knowledge nodes
Local transfer rule	CA models (Kowalska-Styczeń et al., 2017)	Restrict transfer by neighborhood and gap
Projection operator	KITE (Shi et al., 5 Jun 2025)	Mediate human-AI knowledge transformations
Hierarchical modules	KIX (Kumar et al., 8 Feb 2024)	Separate meta-policy and interaction policy

In conclusion, interaction-centric knowledge transfer systems operationalize transfer as a function of structured, context-sensitive interactions—through graphs, attention, gating, or communication protocols—yielding measurable gains in accuracy, efficiency, interpretability, and generalizability compared to one-shot or uniform approaches. The empirical and theoretical results across modalities, tasks, and organizational scales collectively validate the centrality of explicit interaction structuring for effective knowledge propagation in both artificial and hybrid human-AI systems.