Context-CF Co-Evolution: Multidisciplinary Insights

• Context-CF co-evolution is a framework where evolving contexts and CF objects continuously influence each other, forming a feedback loop across various disciplines.
• It spans domains such as cooling flows in astrophysics, AGN covering factors, collaborative filtering in recommendations, and contextual focus in cultural evolution, each with distinct state variables.
• Researchers employ diverse methods—from bi-level optimization and neural collaborative filtering to temporal point processes and cognitive models—to quantify and manage these co-evolutionary dynamics.

Context-CF co-evolution denotes a family of coupled-dynamics problems in which a changing context and a second evolving object abbreviated CF are treated as mutually conditioning rather than as fixed background and response. In the cited literature, CF refers to several different technical objects: cooling flows in cluster cores, covering factor in AGN studies, collaborative filtering in recommendation, and contextual focus in cultural-evolution models. Closely related work uses the same bidirectional logic without the acronym, through co-evolving context functions, context artifacts, or context-dependent fitness landscapes. Across these literatures, the common structure is that context shapes what can be selected, transmitted, or inferred, while the evolving CF object reshapes future context (Soker, 2015, Rałowski et al., 2023, Du et al., 2016, Gabora et al., 2013).

1. Terminological scope and shared structure

The expression spans heterogeneous research programs, and its first technical feature is therefore polysemy rather than doctrinal unity. The same two letters designate different state variables in different fields, but each usage embeds them in a feedback loop rather than a one-way pipeline.

CF usage	Domain	Representative formulation
Cooling flow	Cluster astrophysics	“Maintain ICM temperature”
Covering factor	AGN/quasar studies	$\mathrm{CF}=L_{\rm IR}/L_{\rm agn}$
Collaborative filtering	Recommendation	co-autoregression over UUC and IIC
Contextual focus	Cultural evolution	shift between divergent and convergent thought

In Soker’s comparative jet-feedback framework, CF clusters are the benchmark case for accretion, jets, bubble inflation, and negative feedback (Soker, 2015). In quasar torus studies, CF is the luminosity-ratio covering factor, and the central question is whether apparent redshift evolution is physical or selection-driven (Rałowski et al., 2023). In recommendation, CF appears in the ordinary sense of collaborative filtering, but the most relevant models are those that make user-side and item-side dependencies jointly context-sensitive rather than independent (Du et al., 2016). In EVOC, CF means contextual focus, the capacity to shift between divergent and convergent thought, and its role is explicitly tested in a changing fitness landscape (Gabora et al., 2013).

Other papers instantiate the same logic without using CF as the acronym. NCCE introduces an explicit Context-CF Co-Evolution stage in which a context catalog and a neural preference model improve each other through instance-wise routing and context generation (Zhu et al., 15 May 2026). MCE formalizes a related bi-level system in which CE skills and context functions co-evolve, with the outer loop optimizing skill design and the inner loop optimizing the induced context artifact (Ye et al., 29 Jan 2026). A plausible implication is that the phrase is best understood as a structural motif—bidirectional dependence between contextual state and an adaptive object—rather than as a single field-specific formalism.

2. Cooling flows, SMBHs, and astrophysical feedback loops

In cluster astrophysics, cooling-flow systems are presented as one of the clearest empirical realizations of co-evolution between a compact accretor and an extended gaseous reservoir. Soker characterizes cluster/CF systems by jet-inflated X-ray bubbles, an energy per jets-launching episode of order $10^{60}\,\mathrm{erg}$, system mass $\sim 10^{12} M_\odot$, size $\sim 100\,\mathrm{kpc}$, and activity duration $10^{7}-10^{8}\,\mathrm{yr}$. The accretor is a SMBH with $M_a \sim 10^{8}-10^{10} M_\odot$, the bubbles have $T_{\rm bubble}\sim 10^{9}-10^{10}\,\mathrm{K}$, the ambient reservoir has $T_{\rm res}\sim 10^{7}-10^{8}\,\mathrm{K}$, the jets’ main effect is “Heating ICM,” and the JFM role is “Maintain ICM temperature” (Soker, 2015). The potential contrast is correspondingly large: $R_a \simeq 10^{13}-10^{16}\,\mathrm{cm}$, $\Phi_a \simeq c^2$, $10^{60}\,\mathrm{erg}$0, $10^{60}\,\mathrm{erg}$1, and $10^{60}\,\mathrm{erg}$2. Within that framework, cooling feeds SMBH accretion, accretion powers jets, jets inflate bubbles and heat the ICM, and heating suppresses further cooling; the explicit fizzle outcome is “Cooling catastrophe” (Soker, 2015).

A second astrophysical line complicates any simple reading of co-evolution as necessarily feedback-driven. In NIHAO zoom-in simulations, viscous-disc and gravitational torque-driven accretion models—both with weak or nearly absent explicit $10^{60}\,\mathrm{erg}$3 dependence—naturally generate a common $10^{60}\,\mathrm{erg}$4–$10^{60}\,\mathrm{erg}$5 growth track even without AGN feedback, whereas Bondi-Hoyle accretion with $10^{60}\,\mathrm{erg}$6 yields a two-step history with early stellar growth followed by rapid SMBH catch-up (Soliman et al., 2023). The authors therefore distinguish co-existence, in which BH and stellar growth remain close because both are supplied by the same central gas reservoir, from stronger causal co-evolution claims based on feedback. Their conclusion is not that feedback is irrelevant: feedback still regulates the final BH and stellar masses. It is that scaling relations by themselves do not prove that feedback is the leading driver of the cosmic relation between those quantities (Soliman et al., 2023).

This debate sits naturally beside the broader black-hole/galaxy review literature. Schawinski’s review emphasizes that co-evolution is real in a broad statistical sense, but probably proceeds through several distinct pathways rather than one universal merger-driven channel. The paper highlights the roughly $10^{60}\,\mathrm{erg}$7 parallelism between the cosmic star-formation history and the cosmic black-hole accretion history, the $10^{60}\,\mathrm{erg}$8-$10^{60}\,\mathrm{erg}$9 and black-hole–bulge relations, the downsizing of both galaxy assembly and BH growth, and the fact that the impact of BH feedback on host-galaxy evolutionary trajectories remains less secure observationally than the link between BH accretion and specific host phases (Schawinski, 2012). Taken together, these astrophysical works make “context-CF co-evolution” a problem of which reservoir, accretion prescription, and coupling mechanism define the relevant context, not a single universally closed feedback law.

3. AGN covering factor and the problem of apparent evolution

In AGN unification studies, CF denotes the covering factor of circumnuclear dust, operationally defined as

$\sim 10^{12} M_\odot$0

The central question is whether CF genuinely evolves with redshift or whether apparent evolution is induced by luminosity evolution, flux-limited truncation, and infrared photometric systematics. The study constructs large Low-$\sim 10^{12} M_\odot$1 and High-$\sim 10^{12} M_\odot$2 quasar samples from SDSS, WISE, UKIDSS, GALEX, and SPITZER, computes $\sim 10^{12} M_\odot$3 and $\sim 10^{12} M_\odot$4 by model-independent all-points SED integration, and then applies the Efron & Petrosian test to the joint samples (Rałowski et al., 2023).

The empirical result is that both $\sim 10^{12} M_\odot$5 and $\sim 10^{12} M_\odot$6 evolve strongly with redshift, whereas robust evidence for CF evolution is absent. The EP statistics are $\sim 10^{12} M_\odot$7 and $\sim 10^{12} M_\odot$8 without SNR cuts, $\sim 10^{12} M_\odot$9 and $\sim 100\,\mathrm{kpc}$0 for the SNR$\sim 100\,\mathrm{kpc}$1 sample, and $\sim 100\,\mathrm{kpc}$2 and $\sim 100\,\mathrm{kpc}$3 for the SPITZER sample, rejecting independence in every case (Rałowski et al., 2023). By contrast, the preferred high-quality SPITZER medians are

$\sim 100\,\mathrm{kpc}$4

and are therefore consistent within the errors with no significant evolution. In the matched high-mass subset with $\sim 100\,\mathrm{kpc}$5, the one-dimensional KS test gives $\sim 100\,\mathrm{kpc}$6, again failing to support redshift evolution of CF (Rałowski et al., 2023).

The paper therefore relocates co-evolution claims from torus geometry to sample construction and measurement context. It identifies WISE W4 as the weakest photometric element, recommends SPITZER MIPS $\sim 100\,\mathrm{kpc}$7 whenever feasible, and interprets much of the apparent low-luminosity CF excess as instrumental contamination or polar-dust contamination rather than torus evolution (Rałowski et al., 2023). A plausible implication is that in this literature “context-CF co-evolution” is often a selection problem masquerading as a physical one: the context that co-varies with CF is not only cosmic epoch, but also luminosity, survey depth, and photometric reliability.

4. Collaborative filtering, context routing, and context-engineering loops

In recommendation, CF resumes its conventional meaning of collaborative filtering, but the technically relevant question is whether context is modeled on one axis or jointly across both user and item structure. CF-UIcA answers this with a co-autoregressive factorization over the whole user-item matrix,

$\sim 100\,\mathrm{kpc}$8

and conditions each target entry on both User-User Correlations (UUCs) and Item-Item Correlations (IICs) (Du et al., 2016). The resulting conditional distribution adds user-side and item-side activations before the softmax, and chronology can be enforced by restricting the ordering in top-$\sim 100\,\mathrm{kpc}$9 recommendation. The model therefore captures co-dependent collaborative context, but the paper is explicit that it is not a full latent state-space co-evolution model (Du et al., 2016).

NCCE makes the context loop explicit and renames the problem. It formulates context engineering as recommendation, where instances are “users,” context strategies are “items,” and observed task accuracy is the interaction signal. The target is instance-wise routing,

$10^{7}-10^{8}\,\mathrm{yr}$0

approximated by a neural collaborative filtering model

$10^{7}-10^{8}\,\mathrm{yr}$1

Its Context-CF Co-Evolution stage then alternates between learning instance-context preferences and generating new contexts for failure cases. The failure set is

$10^{7}-10^{8}\,\mathrm{yr}$2

and latent context embeddings are optimized by

$10^{7}-10^{8}\,\mathrm{yr}$3

The final reported test accuracies are 74.7 on HoVer, 89.7 on SCONE, and 60.1 on HotpotQA, for an average of 74.8; the ablation from 74.8 to 72.4 under “Cluster-only routing” isolates the benefit of the co-evolutionary stage itself (Zhu et al., 15 May 2026).

MCE generalizes the same architecture from routing to context production. It defines a context function

$10^{7}-10^{8}\,\mathrm{yr}$4

and a bi-level optimization

$10^{7}-10^{8}\,\mathrm{yr}$5

where CE skills and context artifacts/context functions co-evolve (Ye et al., 29 Jan 2026). The paper reports 5.6–53.8% relative improvement over state-of-the-art agentic CE methods, with a mean of 16.9%, and explicitly treats context as flexible files and code rather than a fixed prompt schema (Ye et al., 29 Jan 2026). In this branch of the literature, context-CF co-evolution is no longer about selecting among static contexts; it becomes a joint search over how contexts are represented, updated, and routed.

5. Point processes, graph learning, event-driven adaptation, and contextual fitness

A closely related family of models studies co-evolution as interleaved event streams. COEVOLVE couples information diffusion and network growth with temporal point processes: $10^{7}-10^{8}\,\mathrm{yr}$6 Retweet intensity depends on the current follow graph, while link-creation intensity depends on prior diffusion exposure, so that network topology shapes exposure and exposure reshapes topology (Farajtabar et al., 2015). The learning problem is convex, the simulator exploits local updates and exponential kernels for efficiency, and on Twitter data the model yields more accurate prediction than alternatives for both link and activity prediction (Farajtabar et al., 2015). This is one of the cleanest continuous-time realizations of joint contextual and structural evolution.

In heterophilous graph learning, CO-EVOLVE translates the same logic into dual-view representation learning. A GNN sends structural context to an LLM as soft prompts, the LLM returns semantic embeddings that induce a dynamic semantic graph, and the two views are updated with Gauss-Seidel alternating optimization under a total loss

$10^{7}-10^{8}\,\mathrm{yr}$7

The stabilizers are a hard-structure conflict-aware contrastive loss, an adaptive node gating mechanism, an uncertainty-gated consistency term, and entropy-aware adaptive fusion at inference. On public text-attributed graph benchmarks, the paper reports average improvements of 9.07% in Accuracy and 7.19% in F1-score over state-of-the-art baselines (Xing et al., 20 Mar 2026). The point here is not only that structure helps semantics or vice versa, but that each view is treated as a mutable latent variable conditioned on the current state of the other.

Event-driven adaptation in ML_CoDa makes the same issue concrete at the programming-language level. Context is a Datalog knowledge base, computation can query it or update it, and asynchronous events change the context independently of the current control flow. The semantics therefore records the goal $10^{7}-10^{8}\,\mathrm{yr}$8 that justified a context-dependent branch and restarts the relevant expression if, after event handling, $10^{7}-10^{8}\,\mathrm{yr}$9 (Degano et al., 2016). In optimization, CEPS adopts a competitive form of contextual fitness: a configuration population and an instance population are co-evolved so that configurations are evaluated in the context of hard synthetic instances, while instances are evaluated in the context of the current portfolio through $M_a \sim 10^{8}-10^{10} M_\odot$0. Its generalization objective is

$M_a \sim 10^{8}-10^{10} M_\odot$1

and the paper derives the upper bound

$M_a \sim 10^{8}-10^{10} M_\odot$2

to motivate alternating evolution of portfolios and instances (Tang et al., 2020). These works differ in domain, but each makes context itself part of the evolving state space.

6. Contextual focus, co-evolutionary niches, and broader theoretical generalizations

In cultural-evolution modeling, CF means contextual focus, not cooling flow, covering factor, or collaborative filtering. EVOC tests the hypothesis that open-ended cultural evolution required two cognitive transitions: chaining and contextual focus. CF is implemented through the rate of creative change (RCC), with

$M_a \sim 10^{8}-10^{10} M_\odot$3

and

$M_a \sim 10^{8}-10^{10} M_\odot$4

High RCC corresponds to more divergent search and low RCC to more convergent search. Across 500 runs, both mean fitness and diversity of actions increase with chaining, and even more so with CF, but the paper is explicit that CF is only effective when the fitness function changes (Gabora et al., 2013). In this usage, context-CF co-evolution means that the search mode itself is context-sensitive: when inherited solutions fail, the system broadens search; when they succeed, it narrows it.

A more abstract generalization appears in the morphogenesis literature. Raimbault argues that morphogenetic systems can be understood through co-evolutionary niches and explicitly postulates that “the presence of morphogenetic processes in a system is equivalent to a decomposition in niches in which its components are co-evolving” (Raimbault, 2018). Context here is supplied by signals, boundaries, niches, and relative subsystem autonomy rather than by a single scalar variable. In multi-relational social networks, a related mesoscale view appears in the study of trust and trade communities: the paper reports that co-evolution rates in trust-based communities are approximately 60% higher than in trade-based communities, that trust and trade connectivity reduce as community size increases, and that in tightly knit communities unusual trade dynamics are followed by unusual trust dynamics (Singhal et al., 2014). These papers move the topic away from any single acronym and toward a broader systems view in which context is boundary structure and co-evolution is circular causation within niches.

Taken together, these theoretical extensions suggest that context-CF co-evolution is best treated as a general problem of mutually conditioned adaptation under structured environments. What changes from field to field is the ontology of the evolving variables—ICM thermodynamics, torus luminosity ratio, user-item behavior, semantic graphs, program control flow, cognitive search breadth, or territorial niches. What persists is the rejection of a static backdrop: context is itself dynamic, and the CF object is meaningful only through its recursive coupling to that dynamics.