Dual-Route Models in AI & Neuroscience

• Dual-route Models are frameworks with two parallel pathways—one for fast pattern-matching and one for slower, structured reasoning.
• They support specialized operations in language modeling, object detection, memory retrieval, and transportation systems.
• Empirical studies show that isolating token-level and concept-level processes enhances performance, flexibility, and interpretability.

Dual-route models refer to systems, both in cognitive neuroscience and artificial intelligence, that partition processing into two or more parallel pathways––often with one route tuned for fast, surface-level or pattern-driven tasks, and another for slower, more deliberative, or structured operations. Across computational modeling, LLMs, memory retrieval, object detection, and transportation systems, dual-route mechanisms enable the selective allocation of representational and computational resources, yielding gains in performance, flexibility, and interpretability.

1. Fundamental Definitions and Taxonomy

Dual-route systems arise when models incorporate two segregated but coactive pathways for information processing. In the cognitive tradition, dual-process theories (e.g., Kahneman's System 1/System 2 distinction) hypothesize separate fast/intuitive and slow/analytical reasoning subsystems. In neural architectures and algorithmic frameworks, dual-route models instantiate two explicit classes of operations:

• Sublexical/token-level vs. lexical/concept-level: For sequence models, e.g., LLMs with distinct mechanisms for verbatim token copying versus semantic-level or concept-based operations (Feucht et al., 3 Apr 2025).
• Fast (“non-thinking”/“System 1”/pattern-matching) vs. slow (“thinking”/“System 2”/reasoning): For cognitive routing in LLMs and memory, with dedicated infrastructure for each route (Du et al., 17 Aug 2025, Zhang et al., 3 Jul 2025, Tang et al., 17 Feb 2026).
• Parallel query-processing branches: In object detection, different attention bias branches for query suppression and delegation (Zhang et al., 15 Dec 2025).
• Dual dynamical regimes in transport networks: Physically distinct state spaces and update rules in, e.g., bus route models (Ngoc et al., 2024).

Dual-route models are thus characterized by a bifurcation in representational scope and functional specialization, often permitting causal experimental isolation of each pathway.

2. Dual-Route Models in LLMs

2.1 Token- and Concept-level Induction in LLMs

The dual-route model of induction in LLMs posits that two functionally independent attention subsystems arise during training:

• Token-level induction heads copy individual tokens verbatim, using positionally constrained attention to promote the immediate next element of a sequence ("prefix-matching" or "induction heads"). Their causal contribution is measured by

$$\text{TokenCopying}(l, h) = \frac{1}{|\mathcal{R}|} \sum_{r \in \mathcal{R}} \big[ P(r_1 \mid a_{p_{\text{clean}}}^{(l,h)} \rightarrow p_{\text{corrupt}}) - P(r_1 \mid p_{\text{corrupt}}) \big]$$

where $\mathcal{R}$ is a set of random subtoken sequences and $a_{p_{\text{clean}}}^{(l,h)}$ is the activation patch inserted into a “corrupt” prompt (Feucht et al., 3 Apr 2025).

• Concept-level induction heads copy multi-token lexical units (words or entities) in a single operation, attending not to the next token but to the last subtoken of the coming word. The corresponding causal metric:

$$\text{ConceptCopying}(l, h) = \frac{1}{|\mathcal{C}|} \sum_{c \in \mathcal{C}} \big[ P(c_2 \mid a_{p_{\text{clean}}}^{(l,h)} \rightarrow p_{\text{corrupt}}) - P(c_2 \mid p_{\text{corrupt}}) \big]$$

where $\mathcal{C}$ is a set of multi-token concepts.

Ablation experiments reveal their causal independence: lesioning token induction heads abolishes nonce copying but leaves translation and semantic tasks largely intact; ablating concept heads cripples translation and synonym/antonym induction but not verbatim copying. Patching concept head outputs across language-pair translation prompts shows that concept heads carry language-independent semantic representations; patched activations enable cross-lingual word transfer at near the same level as the model's own translation performance.

2.2 Dynamic Reasoning Routing

Dual-route frameworks in LLM decision processes (e.g., Cognitive Decision Routing [CDR]) explicitly partition inference between:

• Fast (“System 1”) reasoning: Single forward pass, direct answer generation; triggered for tasks with high input–output correlation, low uncertainty, and homogeneity.
• Slow (“System 2”) reasoning: Multi-step, chain-of-thought, decomposition with deliberative synthesis; invoked for cross-domain, ambiguous, or stakeholder-diverse queries.

Routing is performed by a meta-cognitive decision module operating over features like correlation strength ($C_s$), domain boundaries ($D_c$), stakeholder multiplicity ($S_m$), and output uncertainty ($U_l$), with route assignment:

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

where $\mathcal{R}$0 may be linear, neural, or tree-based (Du et al., 17 Aug 2025).

End-to-end experiments show CDR increases both accuracy and consistency versus uniform deep reasoning, with a 34% reduction in computational cost. In medical QA, analogous dual-mode routing using “thinking” (chain-of-thought) and “non-thinking” (one-step answer) yields both higher accuracy and 37–40% lower inference latency and token cost compared to always invoking high-cost reasoning (Zhang et al., 3 Jul 2025).

3. Dual-Route Memory and Retrieval

In long-term LLM memory systems, as exemplified by the Mnemis framework, dual-route retrieval leverages two structurally orthogonal access strategies (Tang et al., 17 Feb 2026):

• System-1 similarity search: Retrieval of semantically relevant nodes (episodes, entities, edges) by vector or BM25 similarity, supporting fast, high-recall access to items most similar to the current query.
• System-2 global selection: Hierarchical, top-down graph traversal controlled by LLM queries to select category nodes at each semantic level, eventually yielding all structurally necessary memory elements—even those not adjacent in embedding space.

Formally, System-1 operates via cosine or BM25 ranking and reciprocal-rank fusion, while System-2 uses a multi-level, many-to-many hierarchical graph with category nodes and explicit LLM-driven exploration:

$\mathcal{R}$1

Complementary strengths yield state-of-the-art performance on long-horizon and structurally complex memory tasks, with ablation showing System-2 adds +4.2 points to combined retrieval metrics not reachable by similarity search alone.

4. Dual-Route Mechanisms in Object Detection

Route-DETR implements a dual-route scheme at the attention bias level for object query routing in transformer-based detection networks (Zhang et al., 15 Dec 2025). The two routes are:

• Suppressor route: Decouples attention for queries that are deemed to compete for the same object instance, thereby reducing redundant predictions.
• Delegator route: Increases attention flow for queries targeting different objects or unexplored regions, promoting broad spatial coverage.

These routes are parameterized via low-rank learnable attention bias terms $\mathcal{R}$2 added to the self-attention logits:

$\mathcal{R}$3

with gating by query similarity, output confidence, and geometric statistics. During training, only the auxiliary branch uses routed attention, keeping inference unimpaired. Empirical results show consistent +1–2 mAP improvements across major detection benchmarks, demonstrating that dual routing resolves the query competition/duplication inefficiencies in vanilla DETR architectures.

5. Dual-Route Models in Statistical Physics and Transportation

The exactly solvable dual bus route model (Ngoc et al., 2024) generalizes classical bus and exclusion process models by introducing dual states and neighboring effects in a one-dimensional lattice:

• States $\mathcal{R}$4: “Bus present,” “passenger waiting,” and “empty.”
• Dual routes in dynamics: Particles in “empty” or “waiting” states have different movement and conversion rules, parameterized by rates $\mathcal{R}$5, $\mathcal{R}$6, $\mathcal{R}$7 with neighbor-dependent modifications ($\mathcal{R}$8, $\mathcal{R}$9).

The stationary distribution admits an exact Gibbs measure solution under parameter constraints, and the steady-state current and velocity can be computed analytically:

$a_{p_{\text{clean}}}^{(l,h)}$0

where $a_{p_{\text{clean}}}^{(l,h)}$1 encode rate combinations and headway probabilities. Strong neighbor effects result in non-monotonic and reentrant transport behaviors, revealing how dual-route coupling governs collective dynamics. Limiting cases recover the TASEP, cooperative exclusion, and RNA polymerase models.

6. Experimental and Practical Implications

Dual-route models permit granular ablation and patching experiments that precisely localize representational or functional specialization:

• Causal lesioning in LLMs can destroy token or concept-based copying and translation independently (Feucht et al., 3 Apr 2025).
• Adaptive routing in LLM inference and memory directly reduces cost without sacrificing accuracy, with robust empirical gains (Zhang et al., 3 Jul 2025, Du et al., 17 Aug 2025, Tang et al., 17 Feb 2026).
• In object recognition, suppressor/delegator routes address known architectural pathologies in large transformer decoders (Zhang et al., 15 Dec 2025).

In each application domain, the layered, independent operation of the two routes facilitates system robustness, specialization, and scalability.

---

Dual-route models represent a unifying architectural and theoretical paradigm in contemporary cognitive and machine intelligence, supporting both structure-sensitive and form-sensitive processing. Their formal separation, proven causal independence, and robust empirical advantages point to their foundational role in next-generation inference, memory, retrieval, and dynamical systems across fields.