Cognitive Decision Routing (CDR) Architecture

• Cognitive Decision Routing (CDR) is an architectural pattern that explicitly infers decision context to route tasks and queries among specialized cognitive operations.
• CDR applies across diverse domains such as network routing, LLM control, clinical support, and embodied navigation, using adaptive, context-based policies.
• Empirical evidence shows CDR improves accuracy, reduces processing costs, and localizes failures by dynamically calibrating routing decisions based on real-time signals.

Cognitive Decision Routing (CDR) is an architectural pattern in which a system explicitly infers decision context and routes a task, query, flow, or intermediate state to an appropriate next action—such as a path, reasoning mode, retrieval step, tool, clarification request, or executable rule—instead of leaving control implicit inside a monolithic inference or routing procedure. The term is stated directly for large-language-model reasoning in "Cognitive Decision Routing in LLMs: When to Think Fast, When to Think Slow" (Du et al., 17 Aug 2025), but closely related mechanisms appear under neighboring labels such as cognitive networks, cognitive routing, decision-centric design, cognitive topology control, and proactive decision support across networking, search, navigation, and clinical systems (Nair et al., 2010, Sun, 1 Apr 2026, Zhai et al., 12 Mar 2026, Huang et al., 15 Jul 2025, 2505.23055).

1. Conceptual origins and scope

Earlier networking literature framed cognition as a routing-relevant systems property before the explicit CDR label emerged. "Transformation of Networks through Cognitive Approaches" defines a cognitive network as one whose elements, through learning and reasoning, dynamically adapt to varying network conditions in order to optimize end-to-end performance, and it organizes this behavior around perception, acting and planning, learning, reasoning, and decision-making, with the OODA loop as the principal feedback model (Nair et al., 2010). In that formulation, routing is not a static shortest-path lookup but a decision function embedded in a broader adaptive control process.

Subsequent work translated that view into more concrete router and packet architectures. "Cognitive Routing with Stretched Network Awareness through Hidden Markov Model Learning at Router Level" equips each router with a Forward Channel Performance Index and an HMM-based learner to predict the likely condition of forward routers and channels, thereby extending route choice beyond immediate outgoing links (Nair et al., 2010). "Routing Diverse Evacuees with Cognitive Packets" adapts Cognitive Packet Network ideas to building evacuation by treating evacuees as traffic whose routes should be tailored to time, safety, or energy objectives under dynamic congestion and hazard spread (Bi et al., 2013). These papers do not use the phrase Cognitive Decision Routing, but they already instantiate its central thesis: route selection should depend on learned or estimated future context, not only current local state.

The scope of CDR later widened beyond communication networks. In contemporary LLM and agent systems, routing refers to choosing among cognitive operations rather than physical paths: fast versus slow reasoning, answer versus clarify, stop versus expand retrieval, accept versus backtrack, or select-and-execute a domain-specific rule (Du et al., 17 Aug 2025, Sun, 1 Apr 2026). In search and embodied systems, routing likewise extends to evidence channels, tools, and user-guidance moves, as in multi-agent e-commerce decision support and demand-driven robotic navigation (Zhai et al., 12 Mar 2026, Huang et al., 15 Jul 2025). This suggests that CDR is best understood as a control-layer abstraction spanning both network routing and cognitive workflow orchestration.

2. Core architectural elements

A recurrent architectural pattern across CDR systems is the separation of signals, policy, and execution. "Decision-Centric Design for LLM Systems" formalizes a decision point as a tuple $(A,c,\delta)$, where $A$ is a finite action set, $c$ is the decision context, and $\delta$ is the decision function. It further gives the generic policy form

$$\delta(c)=a^*=\arg\max_{a\in\mathcal{F}(c)} U(a,c),$$

with feasibility set $\mathcal{F}(c)$ and utility $U(a,c)$, optionally decomposed into reward minus cost terms (Sun, 1 Apr 2026). This decomposition makes routing an explicit, inspectable layer rather than an implicit by-product of generation.

The explicit CDR formulation for LLMs follows the same structure. The query analyzer estimates a four-dimensional complexity profile—correlation strength, domain crossing, stakeholder multiplicity, and uncertainty level—and the routing module chooses between fast and slow reasoning according to

$$R(q)= \begin{cases} \text{Fast} & \text{if } f(C_s,D_c,S_m,U_l)<\tau \ \text{Slow} & \text{otherwise.} \end{cases}$$

The resulting action is then executed by an adaptive reasoning engine that either performs direct generation or structured multi-step analysis (Du et al., 17 Aug 2025). Here, routing is neither token-level decoding nor simple confidence thresholding; it is a meta-cognitive policy over explicitly represented task characteristics.

In domain-specific systems, the same architecture appears in more specialized form. CogSearch decomposes e-commerce search into Planner, Executor, Guider, Decider, and Memory. The Planner maps a dynamic context

$$\mathcal{C}_t = \{q_t,\mathcal{H}^{search}_{<t},\mathcal{H}^{click}_{<t}\}$$

to a task graph

$\mathcal{G}=(\mathcal{V},\mathcal{E}),$

with task types ProductSearch, WebSearch, and ToolInvocation generated according to $A$0; downstream agents acquire evidence, steer the user’s next decision step, and synthesize a final recommendation with rationale (Zhai et al., 12 Mar 2026). Architecturally, this is an explicit routing topology over cognitive operators and evidence channels.

3. Routing criteria and control policies

CDR systems differ chiefly in the signals they expose and the policies they use to map those signals to actions. In the LLM-specific CDR framework, the core signals are $A$1, $A$2, $A$3, and $A$4, with additional threshold adaptation

$A$5

so routing itself is calibrated over time (Du et al., 17 Aug 2025). In decision-centric LLM control, the simplest sequential signal is a sufficiency score $A$6, optionally paired with a correctness score $A$7, enabling policies such as clarify-versus-execute or expand-retrieval-versus-stop (Sun, 1 Apr 2026).

Embodied CDR often uses hard symbolic routing rules rather than learned gating. CogDDN performs demand matching to produce semantically relevant object candidates $A$8 and then switches between Explore and Exploit according to

$A$9

If hindrance is detected, control is routed to an Analytic Process, and corrected experience is added to the knowledge base via

$c$0

This is explicit arbitration among semantic target selection, fast heuristic control, and slow reflective correction (Huang et al., 15 Jul 2025).

Clinical note routing uses a different policy family. CDR-Agent embeds the note $c$1 and each rule description $c$2, computes

$c$3

fits a Gaussian to the note-specific similarity scores, and selects candidate rules by anomaly detection at $c$4, followed by exclusion-rule filtering and deterministic Python execution

$c$5

Routing is therefore retrieval-plus-screening rather than free-form generation over the full rule set (2505.23055).

Network-oriented CDR often optimizes over forwarding probabilities, access probabilities, link weights, or cluster roles. "Statistical Routing for Multihop Wireless Cognitive Networks" defines link success probabilities using collision structure and outage statistics, then jointly optimizes routing probabilities $c$6, access probabilities $c$7, and transmit powers under PU interference constraints (Dall'Anese et al., 2012). "Towards a Cognitive Routing Engine for Software Defined Networks" instead uses Random Neural Networks and Reinforcement Learning to update path preferences while selective monitoring limits sensing cost (Francois et al., 2016). "Intelligent Routing as a Service (iRaaS)" abstracts routing logic as a pair $c$8, allowing telemetry-derived and RNN-predicted reliability signals to alter path computation at the application plane (Ghosh et al., 2024). These variants differ algorithmically, but all treat routing as a policy over explicit context rather than a fixed protocol rule.

4. Domain-specific instantiations

The CDR label spans several non-identical but structurally related problem classes.

Domain	Routed object	Representative systems
Communication networks	Next hop, link weight, forwarding probability, protocol mode, cooperative group	(Nair et al., 2010, Francois et al., 2016, Dall'Anese et al., 2012, Guirguis et al., 2016, Guan et al., 2011)
LLM control	Fast/slow reasoning, clarify/answer, stop/expand retrieval, accept/backtrack	(Du et al., 17 Aug 2025, Sun, 1 Apr 2026)
Search and decision support	Evidence channels, follow-up guidance, tool use, final recommendation	(Zhai et al., 12 Mar 2026)
Clinical support	Selection and execution of applicable clinical decision rules	(2505.23055)
Embodied navigation	Goal-object binding, explore/exploit switching, reflective recovery	(Huang et al., 15 Jul 2025)
Cyber-physical evacuation	QoS-tailored path choice under hazard and congestion	(Bi et al., 2013)

In communication networks, CDR retains the literal meaning of routing but expands the decision space. Undercover chooses not only a next hop but possibly a cooperative transmit group that nulls interference at primary receivers; its link metric

$c$9

explicitly trades effective capacity against the number of harmed secondary flows (Guirguis et al., 2016). Prediction-based Cognitive Topology Control in CR-MANETs instead filters the topology using predicted link availability under mobility and PU interference, so that conventional routing protocols operate on a cognitively pre-shaped graph (Guan et al., 2011).

In search and advisory systems, the routed object is the user’s evolving decision process. CogSearch treats search as a collaborative progression from need articulation to decision formation, not merely retrieval and ranking. Its Planner decomposes latent intent into a DAG of ProductSearch, WebSearch, and ToolInvocation tasks; Guider generates dynamic facets, convergent or stimulative query suggestions, and a Personalized Purchase Strategy; Decider performs LLM-based multi-criteria decision making over unified context and returns an optimal item plus rationale (Zhai et al., 12 Mar 2026). This is a particularly clear instance of routing among information channels, cognitive stages, and interaction moves.

In embodied and clinical settings, CDR becomes selection among specialized subprocedures. CogDDN routes vague demands to candidate objects, then routes control between heuristic search, object-directed exploitation, and reflective repair (Huang et al., 15 Jul 2025). CDR-Agent routes emergency notes to executable clinical decision rules, then routes extracted structured variables into deterministic code rather than relying on unconstrained generation (2505.23055). These systems show that the same CDR principle applies when the action space consists of domain rules, object hypotheses, or recovery modes rather than paths through a network.

5. Empirical evidence and operational trade-offs

Reported results across domains show that CDR tends to matter most when tasks are complex, underdetermined, or cost-sensitive.

Paper	Setting	Reported outcome
"CogSearch" (Zhai et al., 12 Mar 2026)	JD.com offline and online decision support	Simple Needs: $\delta$0 vs $\delta$1; Complex Needs: $\delta$2 vs $\delta$3; Consultative Queries: $\delta$4 vs $\delta$5; online Decision Cost reduced by 5%, overall UCVR increased by 0.41%, and decision-heavy intents improved by over 30%
"Cognitive Decision Routing in LLMs" (Du et al., 17 Aug 2025)	Adaptive fast/slow reasoning	CDR (Neural): 81.4 ± 1.7 accuracy, 0.81 ± 0.03 consistency, 226 ± 37 tokens, 2.5 ± 0.5s; Uniform Slow: 78.9 ± 1.8, 0.71 ± 0.05, 342 ± 45 tokens, 3.8 ± 0.7s; routing accuracy 87.3%; 67% of queries sent to fast processing
"CDR-Agent" (2505.23055)	Clinical decision-rule selection and execution	Synthetic exact-match accuracy 0.983 vs 0.420; CDR-Bench 0.513 vs 0.426; total time 1.22s vs 4.12s on synthetic and 1.16s vs 8.70s on CDR-Bench
"CogDDN" (Huang et al., 15 Jul 2025)	Demand-driven embodied navigation	Unseen scene/unseen instruction: NSR 34.5, SPL 17.1, SSR 27.5; DDN baseline: 16.1, 8.4, 6.0
"Towards a Cognitive Routing Engine for Software Defined Networks" (Francois et al., 2016)	SDN traffic engineering	Near-optimal paths with 1.65% average optimality gap while using 9.5 times less monitoring; average additional convergence time 30.8s
"Cooperation-based Routing in Cognitive Radio Networks" (Guirguis et al., 2016)	Cognitive radio multihop routing	Network goodput improvement up to 250% compared to other popular cognitive routing protocols, with minimal added overhead

Two regularities stand out. First, gains concentrate where fixed policies are weakest: complex search, consultative queries, cross-domain reasoning, professional judgment, sparse or noisy clinical notes, and spectrum-uncertain routing. In contrast, simple factual retrieval or low-complexity requests often show only marginal improvement, as in CogSearch’s flat ACC@5 on Simple Needs and the negligible factual-query gain in LLM CDR (Zhai et al., 12 Mar 2026, Du et al., 17 Aug 2025). Second, CDR frequently trades one resource for another. CogSearch increases average end-to-end latency from 800ms to 1.3s but uses streaming output to preserve responsiveness (Zhai et al., 12 Mar 2026). CRE reduces monitoring overhead dramatically but converges more slowly (Francois et al., 2016). LLM CDR reduces token cost relative to uniform slow reasoning but still costs more than uniform fast inference (Du et al., 17 Aug 2025). These trade-offs are intrinsic to routing among cognitive modes with different quality and cost profiles.

The controlled experiments in decision-centric LLM design make a related point with unusually clear attribution. In the calendar task, explicit clarify-versus-execute routing maintained 100% success under missingness levels where prompt-only baselines degraded sharply, and the architecture exposed whether failures originated in signal estimation, policy, or question generation (Sun, 1 Apr 2026). This suggests that one empirical advantage of CDR is not only better aggregate performance but also better failure localization.

6. Limitations, misconceptions, and open problems

A common misconception is that CDR denotes a single algorithm. The literature instead presents a family of architectures: HMM-based forward-state prediction, RNN+RL path selection, utility-based action policies, hard symbolic switches, anomaly-based rule selection, and staged multi-agent orchestration (Nair et al., 2010, Francois et al., 2016, Sun, 1 Apr 2026, Huang et al., 15 Jul 2025, 2505.23055). Another misconception is that CDR is simply adaptive routing or cross-layer design. Earlier cognitive-network work explicitly distinguishes cognition from memoryless adaptation by emphasizing learning from prior decisions, reasoning over multiple goals, and end-to-end optimization under imperfect information (Nair et al., 2010).

A second limitation is formal incompleteness. Many influential systems demonstrate the value of the architecture while omitting exact internals. CogSearch does not provide a route-selection loss, a planner training objective, an explicit definition of $\delta$6, prompt text, or a memory update rule (Zhai et al., 12 Mar 2026). The LLM CDR paper leaves classifier architecture details, oracle construction procedure, and full calibration methodology only partially specified, and it notes that about 8% of queries lie near routing boundaries while static front-loaded analysis misses emergent complexity (Du et al., 17 Aug 2025). CDR-Agent’s selection criterion is described as anomaly detection, but the printed Gaussian rule is ambiguous if read literally, and the evaluation is limited to 15 trauma-related rules with proxy execution labels on CDR-Bench (2505.23055). In embodied navigation, CogDDN lacks long-term exploration memory, uses expensive GPT-4-based reflection, and has not been validated on real robots (Huang et al., 15 Jul 2025).

A third limitation is environmental assumption load. Network-oriented CDR often presumes statistical channel knowledge, common control channels, stationary or slow-fading conditions, localization, or reliable estimation of PU-related variables (Dall'Anese et al., 2012, Guirguis et al., 2016, Guan et al., 2011). Router-centric stretched-awareness designs rely on coarse exchanged summaries such as 8-bit FCPI values and first-order HMM assumptions (Nair et al., 2010). SDN-based cognitive routing studies often validate only on small or stylized topologies, leaving scalability and robustness under non-stationary traffic unresolved (Wu et al., 2020, Francois et al., 2016).

Open problems are correspondingly consistent across domains. The literature points toward dynamic routing during inference rather than only before it begins, long-term cognitive memory, multimodal perception, continual learning of routing policies, stronger uncertainty estimation, better separation of “insufficient information” from “wrong trajectory,” and more explicit treatment of cost-quality-risk trade-offs (Du et al., 17 Aug 2025, Zhai et al., 12 Mar 2026, Huang et al., 15 Jul 2025, Sun, 1 Apr 2026). A plausible implication is that the most mature form of CDR will not be a single router but a layered control stack: signal estimation, policy selection, execution, and post-action diagnosis, each independently inspectable and improvable.