Hierarchical Cyclic Merging Regulation

• HCMR is a family of regulated, hierarchical merge-and-regulate procedures characterized by staged merging, cyclic refinement, and explicit admissibility constraints derived from topological, dynamic, or computational principles.
• It is applied in multi-robot double coverage planning to optimize collision-free trajectories using Morse-theoretic constraints and balanced partitioning, achieving significant improvements in path length and task time.
• The paradigm also extends to sequencing-enabled CAV on-ramp merging and long-context LLM inference, demonstrating its versatility in addressing challenges in robotics, autonomous vehicles, and efficient large-scale language modeling.

Searching arXiv for the cited papers and closely related work to ground the article. Hierarchical Cyclic Merging Regulation (HCMR) denotes a class of hierarchical merge-and-regulate procedures in which local structures are aggregated under explicit rules, revisited cyclically or recursively, and then converted into executable plans or compressed representations. In the literature supplied here, the term is used most explicitly for optimal multi-robot double coverage planning, where HCMR is a named algorithm grounded in Morse-bounded graph search and balanced partitioning (Zheng et al., 7 Aug 2025). The same label also serves as a broader interpretive paradigm for two other settings: sequencing-enabled connected automated vehicle (CAV) on-ramp merging control, which uses event-triggered upper-layer sequencing and time-stepped lower-layer MPC regulation (Li et al., 2023), and a proposed cyclic extension of HOMER for long-context inference in pre-trained LLMs, where hierarchical chunk merging and calibrated token reduction provide the underlying primitives (Song et al., 2024). This usage pattern suggests that HCMR is not a single canonical construction, but a family of hierarchical regulation schemes whose common elements are staged merging, constrained refinement, and explicit control of feasibility, memory, or collision structure.

1. Conceptual scope and defining characteristics

In its most concrete formulation, HCMR is an optimal planning algorithm for the double coverage problem: multiple robots must simultaneously cover a workspace area and a set of embedded linear features while remaining collision-free (Zheng et al., 7 Aug 2025). The algorithm reduces the search space of Eulerian tours by imposing topological admissibility conditions derived from Morse theory, then reconstructs edge-level traversals and partitions them across robots.

Two additional usages broaden the term’s scope. In CAV on-ramp control, a sequencing-enabled hierarchical framework is described as aligning with an HCMR paradigm because an upper-level regulator computes a merging order and lower-level distributed MPC controllers repeatedly regulate longitudinal and lateral states as vehicles arrive and resequencing events occur (Li et al., 2023). In long-context LLM inference, HOMER itself is a training-free hierarchical context merging method rather than an HCMR method, but its chunking, pre-merge token reduction, propagative refinement, and depth-first scheduling have been presented as the natural substrate for a cyclic merge-and-refine extension (Song et al., 2024).

Across these settings, several defining motifs recur. First, the system is organized hierarchically: cycle clusters, vehicle sequences, or chunk trees are processed at multiple levels. Second, merging is regulated rather than arbitrary: admissible cycle intersections, density-aware vehicle ordering, or token-pruning rules constrain the evolution of state. Third, regulation is cyclic or recursive: search progresses component by component, sequencing is refreshed on arrivals while MPC runs continuously, or compressed representations can be periodically re-expanded. A common misconception is that HCMR refers only to cyclic schedule design; in the supplied literature, the more fundamental issue is regulated merging under hierarchy, with cyclic behavior arising either from recursive search or from explicit alternation between compression and refinement.

2. Formal HCMR in multi-robot double coverage

The 2025 HCMR formulation addresses a known rectangular workspace $W \subset \mathbb{R}^2$ of size $l \times w$, a set of line-like structures $C \subset W$, and $N$ robots $R=\{R_1,\dots,R_N\}$. Each robot has two roles: exploration with circular footprint $H=B_r(0)$ and service with circular footprint $F=B_a(0)$, with $r \gg a$. For robot trajectories $\pi_i$, simultaneous double coverage is defined by area coverage

$$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$

and service coverage

$l \times w$0

where $l \times w$1 indicates whether the robot is exploring or servicing at location $l \times w$2 (Zheng et al., 7 Aug 2025).

Collision freedom is role-dependent. Exploration paths of different robots are strictly disjoint, whereas service paths may intersect only at set-measure zero, such as crack junctions. The global objective combines total path length and makespan: $l \times w$3 The planning model simplifies switching costs by assuming service motion can be executed quickly relative to exploration, while still accounting for both path length and task completion times through edge weights and role-specific speeds.

HCMR operates on an augmented and simplified graph $l \times w$4 with edges $l \times w$5. Here, $l \times w$6 are Reeb-graph edges produced by Morse cellular decomposition of the exploration subspace, and $l \times w$7 are service edges covering the linear features. Cracks are decomposed into branches, endpoints closer than $l \times w$8 are merged, and inserted points are added so that footprint-constrained service traversals cover each branch. Critical points of the Morse decomposition and crack endpoints are connected by augmentation, after which matched pairs are contracted to obtain $l \times w$9 (Zheng et al., 7 Aug 2025).

The planning variables are an Eulerian tour $C \subset W$0 of $C \subset W$1 and a partition $C \subset W$2 that cuts $C \subset W$3 into contiguous subpaths $C \subset W$4. Each subpath is mapped to a continuous trajectory that performs zig-zag coverage within Reeb cells and service coverage along crack edges. The fixed sweep direction assumption is critical: it fixes the family of cells, the Reeb graph, and the leftmost/rightmost edges used in hierarchy construction, and the stated optimality guarantees hold under this assumption.

3. Morse-bounded search, cyclic merging, and edge-sequence reconstruction

The distinctive theoretical ingredient of HCMR is its Morse-theoretic restriction of the admissible tour set. The paper lifts the Euclidean decomposition to manifold objects $C \subset W$5 and $C \subset W$6 using quotient maps $C \subset W$7 and $C \subset W$8, with stitching at task-transition points along shared boundaries. Topological complexity is quantified by

$C \subset W$9

where $N$0 is genus and $N$1 is the number of boundary components. For a subpath $N$2, the Morse bound is

$N$3

and continuity is enforced through

$N$4

The candidate set is therefore

$N$5

Any collision-free minimum-length, or minimum-time under fixed sweep, solution is claimed to belong to this Morse-bounded collection (Zheng et al., 7 Aug 2025).

Algorithmically, HCMR has four principal stages. The first is cycle hierarchy construction. A disjoint cycle basis $N$6 is computed, and cycles are partially ordered by containment of their sweep-coordinate intervals $N$7. Local cycles become children at deeper hierarchy levels, and each level induces connected subgraphs $N$8 augmented at leftmost and rightmost Reeb edges.

The second stage is Cyclic Merging Search (CMS), a depth-first search with pruning that enumerates segment sequences $N$9, where a segment is either a cycle or a bridge. If the current segment is a bridge, the next segment may be any unvisited vertex-adjacent bridge or cycle, and the merged-cycle accumulator is reset. If the current segment is a cycle, the search prefers unvisited cycles that are vertex-adjacent to the current cycle and edge-adjacent to the accumulator, updating the merged cluster by symmetric difference $R=\{R_1,\dots,R_N\}$0. Pruning occurs whenever the remaining graph becomes unspannable or when an unvisited cycle is enclosed by visited segments, which would create holes. These rules enforce the paper’s structural lemmas: cycles preserve Morse boundedness iff they belong to a disjoint cycle basis, two cycles may merge iff their intersection is exactly one simple path, and a bridge may merge with a cycle only through a Reeb edge.

The third stage is Edge Sequence Back Propagation (ESBP). Given a legal segment sequence, ESBP greedily produces an Eulerian edge sequence. On bridges it performs a direct walk from the current vertex. On cycles it first covers internal edges shared by the current merged cluster and the current cycle, then continues to the next segment through remaining unvisited edges. This preserves the cycle-level regulations while ensuring edge-level completeness.

The fourth stage is Balanced Partitioning (BP). For each tour $R=\{R_1,\dots,R_N\}$1, BP partitions the tour into $R=\{R_1,\dots,R_N\}$2 contiguous subpaths to minimize

$R=\{R_1,\dots,R_N\}$3

The paper uses an FHK-style balanced partitioning method over weighted cycles with $R=\{R_1,\dots,R_N\}$4 time and $R=\{R_1,\dots,R_N\}$5 space per tour, independent of the number of robots (Zheng et al., 7 Aug 2025).

4. Guarantees, empirical behavior, and limitations of the planning algorithm

Under a fixed sweep direction, known environment, holonomic robot motion with negligible turn cost, and role-dependent traversal speeds encoded in edge weights, HCMR is stated to attain the global optimum for the double coverage objective. The argument has three parts: any optimal collision-free solution must be Morse bounded; CMS, ESBP, and hierarchy concatenation enumerate the entire Morse-bounded collection $R=\{R_1,\dots,R_N\}$6; and BP selects the minimum-cost partition over all tours in $R=\{R_1,\dots,R_N\}$7 (Zheng et al., 7 Aug 2025). The same framework yields a collision-free guarantee except possibly at odd-node pairings introduced during augmentation, which are non-manifold connectors and do not induce interior overlaps.

The empirical evaluation uses a $R=\{R_1,\dots,R_N\}$8 environment, omni-wheel robots with $R=\{R_1,\dots,R_N\}$9, $H=B_r(0)$0, $H=B_r(0)$1, and $H=B_r(0)$2, and synthesized crack maps with densities from $H=B_r(0)$3 to $H=B_r(0)$4. Against G-mSCC, S-mSCC, and mGreedy baselines, HCMR improves planned path length by at least $H=B_r(0)$5, reduces average task time by at least $H=B_r(0)$6, achieves $H=B_r(0)$7 filling coverage and full area coverage, and yields essentially zero sensing conflicts across densities. From 1 to 40 robots, total path length decreases and task time saturates depending on feature distribution; uniform maps permit deeper balancing, while Gaussian maps saturate earlier. Enumeration size $H=B_r(0)$8 and CMS time grow exponentially with density, but ESBP remains constant-scale millisecond per sequence, and end-to-end trajectory generation is reported as $H=B_r(0)$9–$F=B_a(0)$0 orders of magnitude faster than SCC/Greedy trajectory generators per path because it uses single-shot Boustrophedon on merged polygons with manifold-aware stitching (Zheng et al., 7 Aug 2025).

The limitations are sharply stated. HCMR assumes a known environment and fixed global sweep direction. Its discretization depends on crack branch insertion and footprint dilation. The size of the candidate set grows exponentially with density, so the method is not polynomial-time in the worst case. These caveats matter because the optimality claim is conditional rather than universal. It does not imply optimality for arbitrary sweep choices, dynamic environments, nonholonomic robots without reformulated edge weights, or settings in which the Morse-bounded collection is itself computationally infeasible to enumerate.

5. HCMR as a paradigm for sequencing-enabled CAV on-ramp merging

A distinct but related use of the HCMR idea appears in sequencing-enabled hierarchical cooperative CAV on-ramp merging control. The framework is not formally titled HCMR, yet it has been explicitly described as aligning with an HCMR paradigm because it combines hierarchical regulation, merging-order assignment, and cyclic lower-level correction (Li et al., 2023). Its architecture has two layers.

The upper layer is a roadside coordinator that collects vehicle positions and velocities in the control area and solves a mixed-integer linear program over a binary assignment matrix $F=B_a(0)$1. The MILP enforces permutation constraints, roadwise non-overtaking constraints, linearized absolute spacing deviations $F=B_a(0)$2, binary sign indicators for spacing and speed differences, and integer indicators $F=B_a(0)$3 that penalize pairings in which the speed difference would increase the spacing deviation. Its objective combines microscopic and macroscopic regulation: $F=B_a(0)$4 with the accompanying text stating that the cost is intended to use $F=B_a(0)$5 rather than $F=B_a(0)$6. The macroscopic term harmonizes densities across mainline and ramp segments by penalizing early merging from the lower-density road.

The lower layer is a distributed longitudinal MPC supplemented by an auxiliary lateral MPC. Longitudinal dynamics use the state $F=B_a(0)$7 and Euler-discretized model

$F=B_a(0)$8

with terminal equalities $F=B_a(0)$9 and $r \gg a$0. The stage cost includes control effort, quadratic state penalties, and an exponential safety term that activates when spacing becomes dangerously small and relative speed is negative. The auxiliary lateral MPC uses a linearized kinematic bicycle model and predicted longitudinal velocities to regulate lane keeping and merging smoothness along the ramp centerline.

The cyclic structure is explicit. Sequencing is event-triggered only when a new CAV enters the control area, whereas longitudinal and lateral MPCs run every sampling interval $r \gg a$1. Predecessor predicted trajectories are transmitted via V2V, so the control topology is re-instantiated whenever the upper-level sequencing changes. In the paper’s terminology, this realizes merging regulation through density harmonization, asymptotic local stability, $r \gg a$2-norm string stability, and safety enforcement (Li et al., 2023).

The theoretical guarantees are stronger than mere feasibility. Under a constant-speed leader and inactive safety cost, Proposition 2 states that CAV2 is asymptotically locally stable if $r \gg a$3, $r \gg a$4, and $r \gg a$5, with analogous extension to downstream vehicles. For string stability, the sufficient condition $r \gg a$6 reduces to

$r \gg a$7

with the stated $r \gg a$8-$r \gg a$9 conditions defining the string-stable region. The controller also expands the initial feasible set relative to zero-terminal and invariant-set-terminal MPC designs, especially along the spacing-deviation axis.

Numerically, upper-level solve times are $\pi_i$0 and $\pi_i$1 in the two reported scenarios, while lower-level longitudinal MPC averages $\pi_i$2 and $\pi_i$3 with maxima $\pi_i$4 and $\pi_i$5, all within $\pi_i$6. The lateral controller averages $\pi_i$7 with maximum $\pi_i$8. Relative to a FIFO distance-based sequencing baseline, the MILP produces lower acceleration peaks, shorter time for spacing deviations to enter the safe range $\pi_i$9, and lower accumulated longitudinal stage cost before convergence (Li et al., 2023). A necessary clarification is that these results establish a hierarchical cyclic regulation scheme for traffic merging, but they do not define a graph-theoretic HCMR algorithm in the sense used for double coverage planning.

6. HOMER primitives and a proposed HCMR extension for long-context LLM inference

HOMER, “Hierarchical cOntext MERging,” is a training-free method for long-context inference in pre-trained decoder-only LLMs. It addresses context-length limitations by dividing an input of length $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$0 into $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$1 chunks of length at most $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$2, processing chunks independently in lower transformer layers, and then progressively merging adjacent chunks at higher layers. Before each merge, low-importance tokens are pruned so that chunk length remains bounded, and a depth-first computation order with propagative refinement yields logarithmic peak memory growth in input length (Song et al., 2024).

The core mechanics are precise. Each chunk receives a shared prefix and suffix, position IDs are reused across chunks, early layers attend only within chunks, and token reduction keeps the top $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$3 tokens using calibrated attention from the chunk’s last token while exempting affix positions. Adjacent chunks are then concatenated, with duplicate affixes averaged rather than repeated. Pruning choices from upper layers are propagated down to lower-layer embeddings, producing fixed-length per-layer embeddings that can replace a standard kv-cache during generation. The appendix-level memory bound is

$$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$4

and the peak bound is

$$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$5

so with $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$6, peak memory scales as $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$7. The forward attention cost is near-linear in $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$8, approximately $$\bigcup_{i=1}^N \bigcup_{x \in \pi_i} (x \oplus H) \supseteq W$$9, rather than quadratic in the full-sequence length (Song et al., 2024).

Empirically, HOMER materially improves long-context operation. On passkey retrieval with Llama-2-7B-chat, HOMER reaches $l \times w$00 at $l \times w$01, and YaRN+HOMER reaches $l \times w$02 at $l \times w$03 and $l \times w$04 at $l \times w$05; for 13B-chat, HOMER reaches $l \times w$06 at $l \times w$07, while YaRN+HOMER reaches $l \times w$08 at $l \times w$09 and $l \times w$10 at $l \times w$11. On QuALITY validation, plain Llama-2-7B-chat scores $l \times w$12, HOMER $l \times w$13, NTK $l \times w$14, and NTK+HOMER $l \times w$15. On PG-19, HOMER yields perplexities $l \times w$16 at $l \times w$17 and $l \times w$18 at $l \times w$19 for Llama-2-7B, and $l \times w$20 at $l \times w$21 and $l \times w$22 at $l \times w$23 for Llama-2-13B; YaRN+HOMER improves these further. At $l \times w$24, baselines exceed $l \times w$25 on A100 with FlashAttention-2, whereas HOMER uses $l \times w$26, and at $l \times w$27 it uses $l \times w$28. Generating 100 tokens at $l \times w$29 takes $l \times w$30 for baselines versus $l \times w$31 for HOMER, reported as a $l \times w$32 speedup (Song et al., 2024).

HOMER does not itself define a cyclic regulation scheme. The supplied technical narrative instead proposes HCMR as a principled extension that alternates merge-and-prune phases with controlled refinement or expansion phases. Using token set notation $l \times w$33, the compressive step is written

$l \times w$34

and the refinement step

$l \times w$35

with small expansion ratio $l \times w$36 and period $l \times w$37. The intended benefit is selective recovery of fine-grained details lost during pruning, potentially helping tasks such as code comprehension. The stated risk is that unconstrained expansion can erode the logarithmic-memory property and complicate kv-cache compatibility. With strict caps, the narrative claims peak memory remains $l \times w$38 plus bounded overhead $l \times w$39. Here, a key clarification is essential: these cyclic HCMR rules are a proposed extension built from HOMER’s primitives, not a contribution claimed by HOMER itself (Song et al., 2024).

7. Synthesis and interpretive boundaries

Taken together, these works define HCMR less as a single domain-specific algorithm than as a regulated hierarchy of merge operations with explicit admissibility constraints. In multi-robot coverage, the constraints are topological: only Morse-bounded cycle mergers are admissible, and balanced partitioning chooses the optimal deployment over the resulting tour set (Zheng et al., 7 Aug 2025). In cooperative traffic merging, the constraints are dynamical and safety-critical: the upper layer selects a virtual sequence, and lower layers repeatedly regulate toward stable, feasible, and collision-safe trajectories (Li et al., 2023). In long-context inference, the constraints are computational: token reduction, hierarchical merging, and depth-first scheduling compress long contexts while bounding memory, and a cyclic extension has been proposed to selectively reintroduce detail when pruning is too aggressive (Song et al., 2024).

This comparison also clarifies the term’s limits. HCMR is not synonymous with hierarchical merging in general. The graph-theoretic HCMR of double coverage has a specific Morse-theoretic definition and optimality claim under fixed sweep direction. The CAV framework is better understood as an HCMR-aligned control architecture rather than a formal HCMR algorithm. The HOMER-based formulation is a conceptual extension, not an established nomenclature in the original long-context paper. What unifies them is the regulated alternation between aggregation and correction: cycles and bridges are merged only when manifold complexity remains bounded; vehicle orders are assigned so that downstream MPC can stabilize the platoon; and chunks are merged only after calibrated token pruning keeps representation size under control.