ARC-Forcing in Multidisciplinary Systems

• ARC-Forcing is a term describing various technical mechanisms that constrain the evolution of dynamical systems in fields such as Arctic wave forecasting, graph theory, and machine learning.
• In Arctic forecasting, it specifically refers to the sea-ice forcing used by the ARC MFC wave-ice model to modulate wave attenuation by ice characteristics such as thickness.
• Across disciplines, ARC-Forcing involves auxiliary states or forcing objects that shape model behavior, from zero forcing in graphs to autoregressive training in diffusion-based navigation.

ARC-Forcing is a context-dependent technical expression rather than a single standardized construct. In Arctic wave forecasting, it denotes the sea-ice forcing used by the ARC MFC wave-ice model, where the ocean/ice analysis supplies the ice state that governs modeled wave attenuation in ice-covered waters (Nose et al., 2023). In graph theory, a closely related “arc-forcing” viewpoint represents zero forcing through directed forcing arc sets rather than through the step-by-step coloring process (Cameron et al., 2024). In machine learning, the distinct term “AR Forcing” refers to an autoregressive training strategy for diffusion-based robot navigation world models (Yang et al., 29 May 2026). The commonality is structural rather than terminological: in each case, an auxiliary state, forcing object, or rollout regime constrains the evolution of a downstream dynamical system.

1. Scope and principal usages

The expression appears in multiple, non-equivalent technical contexts. The Arctic wave-ice usage is the most literal instance of “ARC-Forcing” in the cited literature, whereas graph-theoretic and machine-learning usages are adjacent formulations built around “forcing arc sets” or “AR Forcing.”

Domain	Meaning	Representative source
Arctic wave-ice forecasting	Sea-ice forcing used by the ARC MFC wave-ice model	(Nose et al., 2023)
Zero forcing on graphs	Encoding forcing moves as directed forcing arc sets	(Cameron et al., 2024)
Twisted hypercubes	Constructing zero forcing sets through forcing arc sets	(Collier et al., 3 May 2025)
Diffusion world models	“AR Forcing” as autoregressive training under rollout-generated contexts	(Yang et al., 29 May 2026)

This terminological spread matters because the word “forcing” denotes different mathematical and physical roles. In the Arctic setting it is an input field supplied to an operational forecast system. In zero forcing it is a combinatorial encoding of propagation rules on a graph. In diffusion world models it is a training schedule that exposes the model to its own inference-time state distribution.

2. ARC-Forcing in the ARC MFC wave-ice model

In the Arctic forecasting literature, ARC-Forcing is the sea-ice forcing used by the ARC MFC wave-ice model. The operational product is distributed through CMEMS and uses WAM as the wave model core, ECMWF HRES atmospheric winds as forcing, lateral wave boundary conditions from ECMWF HRES wave forecasts, and sea ice forcing from the ARC MFC ocean analysis. Its wave-ice interaction follows the parameterization of Sutherland et al. (2019), which is designed for thin ice and marginal ice zone conditions (Nose et al., 2023).

Within that framework, wave attenuation due to sea ice is parameterized through a dissipation rate that depends on ice thickness and wave number,

$\alpha = \frac{1}{2}\Delta_0 \epsilon h_i k^2,$

and the dissipated wave spectrum is modeled as

$S(f) = S_0(f)e^{-\alpha x}.$

Here, $\alpha$ is the wave dissipation rate, $\Delta_0 \approx 1$, $\epsilon$ relates to ice permeability and microscopic properties, $h_i$ is sea ice thickness, $k$ is wave number, $S_0(f)$ is the incoming spectrum, and $x$ is propagation distance. Because dissipation increases with sea ice thickness, any forcing field that makes $h_i$ too large will over-damp wave energy and underpredict wave heights.

The observational basis for the analysis was unusually strong for the central Arctic Ocean. Two drifting wave buoys were deployed north of the Laptev Sea, where there had historically been no wave observations available. An experimental buoy was deployed alongside a commercial buoy, and the inter-buoy comparison showed that the experimental buoy measured wave heights and periods accurately. The buoy data were then used to assess the predictability of the ARC MFC wave-ice model product in a region where validation had previously been limited.

3. Fetch geometry, thin-ice resolution, and model failure modes

The study isolates two distinct failure modes of ARC-Forcing. The first involved a sudden decrease in significant wave height $S(f) = S_0(f)e^{-\alpha x}.$0 observed by both buoys when wind direction changed from along the ice edge to off-ice wind. Before the drop, the ARC MFC product underestimated $S(f) = S_0(f)e^{-\alpha x}.$1, and the analysis attributed this to an inaccurate model representation of an ice tongue upwind of the buoys, which constrained the available fetch for wave growth. After the wind shifted, the model agreed better with observations. The ECMWF wave forecast matched the buoy observations better before the drop, and its sea-ice representation did not show the same upwind ice-tongue sheltering, supporting the conclusion that the ARC MFC sea-ice edge geometry was inaccurate (Nose et al., 2023).

The second case was an on-ice wave event on 29 September, when the buoys were in newly forming ice. Zeni-v2021, the downwind buoy, measured a peak of $S(f) = S_0(f)e^{-\alpha x}.$2, while SPOT-1386, about 30 km away, detected no waves. The ARC MFC wave-ice model strongly underestimated the waves and simulated essentially none at the buoy locations. This was notable because the model’s dissipation scheme was intended for thin ice cover, yet the forcing field did not resolve the thin-ice thickness distribution that controlled attenuation in this regime.

To diagnose the failure, the study compared ARC MFC forcing with the neXtSIM sea-ice product. neXtSIM has a 3-category thermodynamic representation, explicitly includes newly formed ice, assigns newly formed ice a thickness range of 0.05 m to 0.275 m, and assimilates SIC daily via nudging. By contrast, ARC MFC sea-ice forcing uses a 1-thickness-category CICE formulation, with newly formed ice having a minimum thickness of 0.5 m. The comparison showed that ARC MFC had poor resolution below 0.5 m, whereas neXtSIM resolved a thin thickness distribution between 0 and 0.3 m; neXtSIM’s ice edge and 0.80 SIC contour were also closer to what was needed to explain the buoy-observed waves.

This thickness mismatch is physically consequential because the Sutherland dissipation law scales linearly with $S(f) = S_0(f)e^{-\alpha x}.$3. If the forcing constrains newly formed ice to thicknesses of order 0.5 m rather than resolving young and grey ice, typically less than 30 cm thick, then $S(f) = S_0(f)e^{-\alpha x}.$4 is systematically too large and the model overestimates dissipation. The paper states that for a 7 s wave period, ARC MFC dissipation could exceed neXtSIM’s by a factor of more than 3 along parts of the wave path. The central conclusion is therefore specific: better sea-ice forcing that resolves thin-ice thickness distributions is needed for improved wave predictability in marginal ice zones, especially when new ice forms.

4. Arc-forcing in zero forcing theory

In graph theory, an arc-forcing viewpoint recasts zero forcing as a directed arc-set problem. Starting from the standard zero forcing rule on a graph $S(f) = S_0(f)e^{-\alpha x}.$5, where a blue vertex $S(f) = S_0(f)e^{-\alpha x}.$6 can force a white vertex $S(f) = S_0(f)e^{-\alpha x}.$7 if $S(f) = S_0(f)e^{-\alpha x}.$8 is the unique white neighbour of $S(f) = S_0(f)e^{-\alpha x}.$9, the paper "An approximation algorithm for zero forcing" defines an arc set $\alpha$0 whose arc $\alpha$1 records the move “$\alpha$2 forces $\alpha$3.” A forcing arc set is an arc set such that $\alpha$4 is a disjoint union of directed paths and there exists a zero forcing process on $\alpha$5 whose forcing moves are exactly the arcs in $\alpha$6. The sources of $\alpha$7 are exactly the initial blue vertices, and the correspondence satisfies

$\alpha$8

so larger forcing arc sets correspond to smaller zero forcing sets (Cameron et al., 2024).

The central structural obstruction is the chain twist. The main theorem states that an arc set $\alpha$9 satisfying the disjoint-dipath condition is a forcing arc set if and only if it contains no chain twist. This characterization makes it possible to manipulate zero forcing processes algebraically through arc sets rather than through explicit forcing schedules. The framework also yields closure and symmetry properties: any subset of a forcing arc set is again a forcing arc set, and $\Delta_0 \approx 1$0 is a forcing arc set if and only if its reverse arc set $\Delta_0 \approx 1$1 is a forcing arc set.

The same viewpoint powers a pathwidth-based approximation algorithm. Using a nice path decomposition, the algorithm incrementally builds a zero forcing set $\Delta_0 \approx 1$2, a forcing arc set $\Delta_0 \approx 1$3, and a fort packing $\Delta_0 \approx 1$4. Theorem $\Delta_0 \approx 1$5 shows that if the decomposition has width $\Delta_0 \approx 1$6, then the returned zero forcing set satisfies

$\Delta_0 \approx 1$7

which yields a $\Delta_0 \approx 1$8-approximation algorithm, and the runtime is $\Delta_0 \approx 1$9 for a connected graph.

The forcing-arc-set method has also been used to construct explicit low zero forcing sets in nontrivial graph families. "The Zero Forcing Number of Twisted Hypercubes" uses forcing arc sets to build a family of twisted hypercubes of dimension $\epsilon$0 with zero forcing sets of size

$\epsilon$1

which is below the minimum zero forcing number of the hypercube (Collier et al., 3 May 2025). In that construction, correctness again reduces to showing that the relevant arc set contains no chain twist.

5. Forcing as quasirandom extremality

A different graph-theoretic meaning of forcing arises in extremal combinatorics. In the Sidorenko and forcing conjectures, a bipartite graph $\epsilon$2 is Sidorenko if

$\epsilon$3

for every graph $\epsilon$4, and $\epsilon$5 is forcing if asymptotic equality at fixed edge density implies quasirandomness. "Forcing Graphs to be Forcing" extends the family of bipartite graphs for which the forcing conjecture is known to hold to include balanced blow-ups of Sidorenko graphs and subdivisions of Sidorenko graphs by a forcing graph; it also shows that the box product of a Sidorenko graph with an edge is forcing, and in particular that cubes are forcing (Kiem et al., 2024).

This usage is conceptually separate from zero forcing. There is no color-change process and no forcing arc set. Instead, “forcing” means that a graph acts as a quasirandomness certificate: any asymptotically extremal sequence must in fact be $\epsilon$6-quasirandom. The proof strategy in that paper is algebraic and builds on Razborov’s flag algebra framework. The resulting closure properties establish that forcing can be preserved under balanced blow-up, suitable subdivision, and specific graph products.

A common misconception is to conflate these two graph-theoretic meanings because both use the word “forcing.” Zero forcing concerns propagation under the unique-white-neighbour rule, whereas the forcing conjecture concerns uniqueness of quasirandom minimizers in homomorphism density inequalities. The overlap is lexical, not definitional.

6. Distinct machine-learning and systems usages

In robot navigation world models, the relevant term is “AR Forcing,” not ARC-Forcing. It addresses a train-test mismatch in diffusion-based world models: training is usually parallel and conditioned on ground-truth contexts, while inference is autoregressive and conditioned on the model’s own predictions. AR Forcing changes only the training loop. Starting from a real context window, the model predicts the next latent or image step, feeds its own prediction back into the context, repeats autoregressively, and applies the same standard diffusion noise-prediction loss at every step. The resulting objective keeps the ordinary diffusion loss but changes the conditioning distribution from $\epsilon$7 to rollout contexts drawn from $\epsilon$8 (Yang et al., 29 May 2026).

The reported empirical results span RECON, SCAND, HuRoN, and TartanDrive. The paper states that AR Forcing improved the consistency of generated images during long-horizon navigation and the accuracy of predicted trajectories, and it reports improvements in LPIPS, DreamSim, FID, ATE, RPE, PosErr, and YawErr. It also reports a 7.67× training-only overhead, longer autoregressive training trajectories improving long-horizon robustness, and more stable zero-shot generalization on Go Stanford.

A neighboring development is JaxARC, a high-performance JAX-based environment for the Abstraction and Reasoning Corpus. That paper does not explicitly use the term “ARC-Forcing,” but it presents an infrastructure layer that makes ARC experimentation far more scalable: the environment is functional, stateless, JAX-native, compatible with jit, vmap, and pmap, and achieves matched-batch speedups of 38× on CPU, 903× on RTX 3090, and 5,439× on H100, with peak throughput of 790M steps/second on H100 at 2 million environments (Aadam et al., 24 Jan 2026). A plausible implication is that such an environment is an enabling substrate for ARC-Forcing-style experimentation in ARC reasoning, because it supports large batches, many tasks, and many algorithmic variants without the Python/Gymnasium bottlenecks emphasized in earlier environments.

A further non-equivalent use of “ARC” and “forcing” appears in cache replacement theory. "Analyzing Adaptive Cache Replacement Strategies" studies the Adaptive Replacement Cache algorithm ARC and asks how badly it can be forced by adversarial request sequences. The paper proves that the competitiveness ratio of ARC has a lower bound of $\epsilon$9 and an upper bound of $h_i$0, while CAR has lower and upper bounds of $h_i$1 and $h_i$2, respectively (Consuegra et al., 2015). Here “forcing” means adversarially inducing misses in an online algorithm, not supplying a physical forcing field or constructing forcing arcs.

7. Terminological boundaries and non-equivalent “ARC” usages

Several additional papers contain “ARC,” “arc,” or “forcing” but do not define ARC-Forcing. "Formation of Self-Organized Anode Patterns in Arc Discharge Simulations" studies a free-burning arc sustained by a constant DC current between a conical cathode and a flat anode, with no auxiliary gas flow, no imposed magnetic field, and no external forcing. Its central result is that a time-dependent, 3D thermodynamic nonequilibrium simulation can spontaneously form self-organized anode attachment spot patterns, and that heavy-species–electron energy equilibration has a dominant role in the formation of anode spots (Trelles, 2012). This is an arc-discharge self-organization problem, not an ARC-Forcing formulation.

Likewise, "ARC: A compact, high-field, fusion nuclear science facility and demonstration power plant with demountable magnets" uses ARC as the name of a tokamak concept. The reactor is designed around a high magnetic field, demountable REBCO toroidal field coils, non-inductive current drive, and an all-liquid FLiBe blanket; the abstract reports a plasma fusion gain of $h_i$3, fully non-inductive operation, and a tritium breeding ratio $h_i$4 (Sorbom et al., 2014). Although one may speak interpretively of design choices that force the plasma toward a steady-state operating regime, the paper does not present ARC-Forcing as a named concept.

The principal misconception to avoid is therefore terminological collapse. In the cited literature, ARC-Forcing most directly denotes the sea-ice forcing used by the ARC MFC wave-ice model. Nearby usages include forcing arc sets in zero forcing, forcing graphs in quasirandom extremal theory, AR Forcing in diffusion world models, and adversarial forcing of the ARC cache algorithm. These are connected only by vocabulary; their mathematical objects, governing equations, and operational goals are distinct.