Reflexive Multiple Access (RMA) Protocol

• RMA Protocol is a family of MAC protocols that uses an agent-based ORDE loop and RL-driven tree splitting to optimize data freshness, bandwidth efficiency, and collision avoidance.
• It dynamically adjusts transmission strategies based on real-time feedback, supporting heterogeneous nodes and prioritizing low age-of-information and minimal delay.
• Experimental results show up to 14.9% AoI reduction, 35% throughput improvement, and 44% lower packet delay compared to conventional protocols.

Reflexive Multiple Access (RMA) Protocol is a modern family of medium access control (MAC) protocols that optimize information freshness, bandwidth efficiency, and collision avoidance in heterogeneous wireless and IoT network environments. RMA incorporates either an LLM-agent–based "Observe–Reflect–Decide–Execute" (ORDE) closed-loop architecture for AoI minimization in environments with complex node heterogeneity (Liu et al., 26 Jan 2026), or a reinforcement learning–driven, belief-MDP–guided tree-splitting reservation procedure with optimal coding for contention resolution and reservation bandwidth minimization (Chen et al., 3 Apr 2025). Both approaches realize significant advances in scalability, responsiveness, and overhead reduction compared to conventional random access (ALOHA, CSMA/CA) and centralized scheduling solutions.

1. System Model and Assumptions

RMA is applicable to time-slotted wireless networks (single channel), where a central access point (AP) manages accesses from $m$ nodes (potentially heterogeneous, e.g., TDMA, ALOHA, RMA-enabled) (Liu et al., 26 Jan 2026). The standard scenario assumes:

• Time-slotted channel: All access operations are synchronized to slot boundaries.
• “Generate-at-will”: Every node always has a fresh packet available at the beginning of each slot.
• Collision model: Any simultaneous transmissions result in failure. The AP broadcasts per-slot feedback, enabling nodes to update their local transmission strategies.
• Node heterogeneity: Environments comprise TDMA nodes (fixed slot assignment), ALOHA nodes (random access with probability $q$), and RMA (“heteronodes”) leveraging agent-based adaptive decision-making.
• Feedback: The channel provides explicit 0/1/e feedback per slot (idle, success, or collision, respectively) (Chen et al., 3 Apr 2025), or richer feedback including current Age-of-Information (AoI) per node (Liu et al., 26 Jan 2026).

In the tree-splitting RMA reservation regime (Chen et al., 3 Apr 2025), up to $N_\text{max}$ terminals contend per contention cycle, with no inter-terminal signaling. Each makes local transmit decisions based on its cluster state and global feedback ($c_t\in\{0,1,e\}$).

2. Key Protocol Mechanisms

2.1 LLM-Agent–Based RMA (AoI-Optimization) (Liu et al., 26 Jan 2026)

The protocol is structured around the ORDE loop:

• Observe: Aggregates per–slot statistics (AoI gradients, collision/idle rates) over an observation period of $N$ slots, producing perturbations $\Delta p_i(o) = f_\mathrm{obs}(F_o)$ to node transmit probabilities.
• Reflect: At coarser timescales (every $O$ observation periods), an LLM agent self-diagnoses strategy effectiveness, generating semantic reflections $R_r$ and storing in reflection memory.
• Decide: Decides new global transmit probability for each node type, via updates of the form $$p_i(t+1) = p_i(t) + \beta f_\mathrm{refl}(\mathrm{Reflection}(t))$$.
• Execute: On the slot level, nodes sample actions $$a_i(n) \sim \mathrm{Bernoulli}(p_i^\text{final}(n))$$, with $p_i^\text{final}(n) = p_i(t) + \Delta p_i(o)$, and log experience to short-term memory.

2.2 RL-Driven Tree-Splitting RMA (Reservation Optimization) (Chen et al., 3 Apr 2025)

The protocol proceeds in reservation “trees,” with the following rules:

• Initial clustering: All $N$ active terminals are assigned to cluster 1.
• At each reservation slot $t$: Each cluster $i$ is assigned probability $p_{t,i}$; all members act independently. Feedback $c_t$ is observed.
  • $c_t=1$: One terminal wins, leaves contention, clusters unchanged.
  • $c_t=0$: Idle, clusters unchanged.
  • $c_t=e$: Collision, transmitting nodes (and only those) are reflexively split into a new cluster.
• Cluster memory: Each terminal maintains its cluster index $j_t$ and only requires minimal labeling ($\lceil\log_2 M_t\rceil$ bits per packet).
• Expected resolution time: Optimized by tuning $p_{t,i}$ using RL over the POMDP belief state.

3. Optimization Formulations and Learning Components

3.1 MDP/POMDP Model (AoI- and Reservation-Optimal RMA)

• State space:
  • LLM-RMA: $s(n)$ includes instantaneous AoI $\delta_i(n)$, transmission outcome, etc.
  • RL-RMA: $s = (\eta_1, ..., \eta_M)$ with $\sum \eta_i = n$ (active node counts per cluster).
• Actions:
  • LLM-RMA: Slot-level transmit/sample, reflection-level global strategy update.
  • RL-RMA: Cluster-wise attempt probabilities $\mathbf{p}^s = (p_1, ..., p_M)$.
• Observation space: Broadcast feedback string, either per-slot AoI and collision statistics (Liu et al., 26 Jan 2026) or cluster-based summary feedback (Chen et al., 3 Apr 2025).
• Transition: Determined by the outcome of the slot (success, failure, or collision) and subsequent re-clustering.
• Reward/cost:
  • LLM-RMA: Negative change in average AoI, $$r_t = -(\Delta_\mathrm{sys}^{\mathrm{after}} - \Delta_\mathrm{sys}^{\mathrm{before}})$$.
  • RL-RMA: Unit negative cost per unresolved state, $C(s)=1$, until all packets scheduled (Chen et al., 3 Apr 2025).

3.2 Learning and Strategy Optimization

• LLM-RMA: Supervised Fine-Tuning (SFT) to encode high-quality reflection responses, followed by PPO to maximize AoI-reduction reward. SFT loss: $$\mathcal{L}_\mathrm{SFT} = -\sum_{i=1}^N \log P_\theta(y_i | x_i)$$; PPO uses clipped policy optimization on the action space.
• Tree-Splitting RMA: Real-Time Dynamic Programming on the belief MDP (RTDP-Bel), iteratively minimizing expected resolution time by Bellman backups in the quantized belief and action space.

4. Priority and Heterogeneity Support

The LLM-agent RMA supports differentiated service via explicit priority encoding (Liu et al., 26 Jan 2026):

• Priority encoding: Each node is assigned Priority(i) $\in$ {High, Low}.
• Transmit probability initialization: $p_i^\mathrm{initial}$ set higher for High-priority nodes.
• Perturbation and semantic reasoning: Policy perturbations and LLM-suggested updates weighted more aggressively for High-priority nodes, whose strategy snapshots are also favored in long-term memory.
• Convergence: Priority-based RMA achieves a 15–20% improvement in AoI convergence rate versus non-priority LLM-agent baselines.

In the tree-splitting RMA context, all terminals are treated equivalently during reservation, but can be assigned order-dependent transmission slots post-resolution, guaranteeing FIFO service (Chen et al., 3 Apr 2025).

5. Protocol Overhead, Coding, and Practical Performance

RMA protocols achieve significant efficiency advantages over classical reservation or random access mechanisms:

• Reservation message size: Only a cluster-ID index ($\lceil\log_2 M_t\rceil$ bits) and a 2-bit feedback per slot; versus $O(n)$-byte RTS/CTS overheads and per-terminal backoff in DCF (Chen et al., 3 Apr 2025).
• Bandwidth/throughput: RMA reduces reserved bandwidth (slots) by 30–50% compared with DCF for $n \leq 5$. Under $\lambda=0.2$ packets/slot and $\rho=3$, sustainable throughput is increased by 35% ($\gamma \approx 0.92$ for RMA, $0.68$ for DCF).
• Delay: Tree-splitting RMA reduces average packet delay from 75 slots (DCF) to 42 slots, 44% lower under mid-to-high load (Chen et al., 3 Apr 2025). LLM-agent RMA reduces system AoI by up to 14.9% versus LLMA baselines.
• Response to dynamic topology: In early (low-load) stages, some deep learning MAC protocols perform better, but in later heterogeneous regimes, RMA outperforms with 5–11% AoI improvements.
• Ablation studies: Gains of 18–23% in AoI attainable with full ORDE + PPO pipeline; incremental improvements measured for each module (Liu et al., 26 Jan 2026).

6. Implementation and Extensibility

RMA frameworks are suited to modern edge computing environments:

• Deployment: Agent engines can be hosted on edge servers or on-device LLMs with moderate parameter sizes (e.g., 7B, with 8-bit quantization) (Liu et al., 26 Jan 2026).
• Control timescales: Multi-timescale architecture decouples real-time slot execution from slower strategy reflection/optimization, mitigating slot-level latency.
• Offloading: Integration with SDR/IoT gateways is feasible via standard APIs.
• Potential extensions: Hierarchical multi-objective reward design (combining AoI, energy, latency, reliability), adversarial robustness via reflection validation, federated RMA across cellular domains, and real-time reasoning constraint enforcement. External knowledge integration (such as 5G scheduling or network slicing information) is a supported avenue.

7. Theoretical Underpinnings and Analytical Results

Both main RMA paradigms leverage rigorous analytical frameworks:

• Tree-splitting analysis: $E[T(n)]$ for binary splitting is empirically minimized near $p=0.5$; expected slot count grows linearly with $n$.
• POMDP solution: Belief-state optimization via RTDP-Bel converges to optimal transmit probabilities $\pi^*$ over the space of cluster partitions.
• Throughput-delay tradeoff: Steady-state throughput $\gamma$ scales as $\lambda \rho / (1 + \lambda \rho E[T(N)])$ under Poisson arrivals; RMA protocols saturate close to $\gamma=1$ with minimal collisions as system load increases (Chen et al., 3 Apr 2025).
• AoI dynamics: LLM-RMA systematically reduces time-averaged AoI, with practical network scenarios exhibiting 10–15% reductions across a range of baseline configurations.

8. Summary of Comparative Features

Feature	LLM-Agent RMA (Liu et al., 26 Jan 2026)	Tree-Splitting RMA (Chen et al., 3 Apr 2025)
Network Scope	Heterogeneous (TDMA/ALOHA/heteronode)	Homogeneous (reservation contention)
Optimization Target	Age of Information	Reservation bandwidth, delay
Control Approach	Observe–Reflect–Decide–Execute (LLM/SFT/PPO)	RL–driven belief-MDP optimization via RTDP-Bel
Priority Support	Explicit, via semantic and policy tiers	FIFO ordering only
Achievable Gain	Up to 14.9% AoI reduction, 20% faster convergence	35% throughput, 44% lower packet delay vs. DCF

RMA protocols, as instantiated by both LLM-agent and reinforcement learning reservation frameworks, redefine adaptability and efficiency in next-generation wireless multiple access, successfully addressing challenges in freshness, delay, bandwidth, and heterogeneity in IoT and related emerging network paradigms (Liu et al., 26 Jan 2026, Chen et al., 3 Apr 2025).