RCORE: Dual Frameworks in Video and Cloud

• RCORE is an acronym used in two distinct domains, representing frameworks for zero-shot compositional video recognition and renewable-driven cloud scheduling.
• In video understanding, RCORE employs Composition-Aware Augmentation and Temporal Order Regularization to mitigate object-driven shortcut biases and enhance unseen verb-object generalization.
• In cloud systems, RCORE dynamically adjusts CPU core power profiles based solely on renewable energy availability, reducing VM eviction incidents and overall server power consumption.

Searching arXiv for "RCORE" to ground the article in the relevant papers. RCORE is an overloaded acronym in recent arXiv literature rather than a single, unified framework. In the available papers, it denotes at least two technically distinct systems: RCORE as Robust COmpositional REpresentations for zero-shot compositional action recognition in video understanding, and RCORE as Renewables-driven CORES for carbon-aware real-time workload management in cloud systems (Ahn et al., 22 Jan 2026, Hewage et al., 2024). The former addresses object-driven shortcut learning in compositional video models; the latter treats renewable-energy availability as a per-core scheduling and placement constraint in cloud infrastructure. The shared acronym masks substantially different objectives, assumptions, and evaluation protocols.

1. Nomenclature and scope

In zero-shot compositional action recognition, RCORE is introduced in the context of Compositional Video Understanding (CVU), where a model must recognize verbs and objects and compose them to generalize to unseen verb-object combinations (Ahn et al., 22 Jan 2026). The paper identifies a specific failure mode—object-driven verb shortcuts—and proposes RCORE as a training framework that enforces temporally grounded verb learning.

In cloud systems, RCORE appears as a framework for carbon-aware real-time workload management using renewables-driven cores (Hewage et al., 2024). Here the term refers to a systems mechanism that dynamically switches CPU cores between real-time and low-power profiles according to renewable availability, while preserving static compute guarantees for real-time virtual machines.

A common misconception is to treat “RCORE” as if it referred to one research line. The literature does not support that reading. The acronym is reused across at least these two domains, and the underlying methods are unrelated except for the name.

2. RCORE in zero-shot compositional action recognition

The video-understanding RCORE begins from an analysis of Zero-Shot Compositional Action Recognition (ZS-CAR). The task assumes that only a subset of possible verb-object compositions is observed during training, and the model must generalize to unseen combinations. The paper attributes current model failures to two intertwined factors: severe sparsity and skewness of compositional supervision, and asymmetric learning difficulty between verbs and objects (Ahn et al., 22 Jan 2026).

The quantitative characterization in the paper is explicit. Only a small fraction of all possible verb-object pairs are seen during training: ~14% for Sth-com and ~7.5% for EK100-com. Objects are easier to recognize from static spatial cues, whereas verbs require temporal, multi-frame reasoning. As training progresses, the model increasingly ignores visual evidence and overfits to co-occurrence statistics. The result is shortcut behavior: predicting the most frequent verb for a given object regardless of the actual motion.

This diagnosis is tied to the paper’s use of Compositional Gap,

$\Delta_\text{CG} = Acc^C - (Acc^V \times Acc^O),$

which measures whether joint composition accuracy exceeds what would be expected from independently predicting verbs and objects. The paper reports that the baseline C2C model yields a negative compositional gap on unseen compositions, which it interprets as evidence that the model does not gain the benefit of compositional recognition in unseen verb-object compositions (Ahn et al., 22 Jan 2026).

This suggests that the central problem is not merely insufficient model capacity. The paper’s account is that the supervision structure itself pushes the learner toward co-occurrence bias unless verb learning is made explicitly temporal.

3. Mechanisms of the video-understanding RCORE

RCORE introduces two principal components: Composition-Aware Augmentation (VOCAMix) and Temporal Order Regularization (TORC) (Ahn et al., 22 Jan 2026). Their roles are complementary: VOCAMix addresses compositional sparsity and skewness, whereas TORC penalizes shortcut behaviors by requiring the verb encoder to be sensitive to temporal order.

VOCAMix synthesizes plausible unseen verb-object combinations without corrupting motion cues. For a video $X_i$ with verb $y_i^V$ and object $y_i^O$, and another video $X_j$ with object $y_j^O$, the augmentation is described as

$$h(\hat{X}_i(k,:)) = (1-\lambda)h(X_i(k,:)) + \lambda h(X_j(\lfloor T/2\rfloor,:)),$$

$\tilde{y}^O_i = (1-\lambda)y^O_i + \lambda y^O_j.$

The verb label is left unchanged, so temporal dynamics are preserved while the object region is augmented. The stated purpose is to diversify verb-object combinations while breaking strong training-time associations between specific verbs and specific objects.

TORC operates by computing verb features for the original sequence, a time-reversed sequence, and a temporally shuffled sequence. Its two terms are

$L_{\cos} = \cos(f^V, f^V_\text{rev}),$

which is minimized so that temporally opposite actions are distinguishable, and

$$L_\text{ent} = \sum_{i=1}^{|\mathbb{Y}^V|} p_i \log p_i,$$

which discourages confident verb predictions when temporal order has been destroyed by shuffling. The combined regularizer is

$X_i$0

The paper also includes a margin loss that penalizes high confidence for frequent but incorrect verb-object compositions.

A notable empirical claim is that TORC alone gives the largest boost in verb and composition accuracy, while VOCAMix alone gives modest gains; the combination yields the best compositional generalization. By contrast, MixUp/CutMix are reported as detrimental because they do not preserve temporal action cues (Ahn et al., 22 Jan 2026).

4. Empirical behavior of the video-understanding RCORE

The video RCORE is evaluated on Sth-com and EK100-com. Sth-com is based on Something-Something V2 and contains 161 verbs, 248 objects, ~79K videos; EK100-com is derived from EPIC-Kitchens-100 and contains 82 verbs, 228 objects, ~71K videos (Ahn et al., 22 Jan 2026). Metrics include Top-1 accuracy for verbs, objects, and compositions, together with Compositional Gap.

On Sth-com, the baseline C2C achieves unseen Verb 54.36, Object 56.10, Comp 30.08, Comp H.M. 36.47, and Comp Gap -0.42. RCORE achieves unseen Verb 57.29, Object 55.56, Comp 32.23, Comp H.M. 38.75, and Comp Gap +0.40 (Ahn et al., 22 Jan 2026). The positive shift in compositional gap is presented as evidence that RCORE yields genuine compositional benefit rather than merely improving one primitive classifier.

On EK100-com, the baseline C2C reports unseen Verb 49.71, Object 47.48, Comp 22.38, Comp H.M. 29.38, and Comp Gap -1.22. RCORE reports unseen Verb 52.07, Object 47.11, Comp 24.24, Comp H.M. 31.23, and Comp Gap -0.29 (Ahn et al., 22 Jan 2026). The gap remains negative, but the paper presents the reduced negativity as a narrowing of shortcut-driven failure.

The paper also states that RCORE reduces the False Co-occurrence Prediction (FCP) rate and makes fewer “opposite verb” confusions such as opening versus closing. A plausible implication is that the framework is not only improving classification accuracy but changing the failure mode itself, from static object bias toward temporally grounded verb discrimination.

5. RCORE in carbon-aware cloud workload management

The cloud-systems RCORE is a framework for harvesting green renewable energy for delay-intolerant real-time workloads in cloud systems (Hewage et al., 2024). Its central mechanism is the use of renewables-driven cores in servers to switch CPU cores dynamically between a real-time power profile and a low-power profile, matching renewable energy availability.

The paper defines the high-performance real-time profile using examples such as C-state = POLL at 2.7 GHz, while unused or unallocated cores may be placed in a deep sleep state such as C-state = C6. At 100% renewable capacity, all cores are in real-time mode; at 0%, eligible cores are put to sleep. Importantly, these power-profile transitions depend solely on renewable supply, not on workload demands (Hewage et al., 2024).

The framework couples this mechanism to a VM Execution Model. Each VM core is pinned to a real-time-profile core for the VM’s lifetime, using OpenStack’s dedicated-cores feature. If insufficient real-time cores are available, the system does not degrade a running real-time VM by shrinking its physical core allocation. Instead, it performs criticality-aware VM evictions, evicting best-effort VMs before critical ones. The paper justifies this design by arguing that dynamic shrinking harms real-time latency bounds.

The same work introduces the Green Cores abstraction, which converts renewable energy usage into a server inventory attribute. This inventory is then used by a VM Packing Algorithm to jointly optimize renewable energy utilization and reduction of VM eviction incidents.

6. Architecture, objectives, and results of the cloud RCORE

The cloud RCORE formalizes a joint optimization problem over a time window $X_i$1: $X_i$2 where $X_i$3 is harvested renewable energy and $X_i$4 is the number of VM eviction incidents (Hewage et al., 2024). The renewable energy term is

$X_i$5

subject to real-time latency constraints

$X_i$6

The server power model is given as

$X_i$7

with $X_i$8 the pinned real-time cores and $X_i$9 the sleeping cores. The paper also states the Green Cores relation

$y_i^V$0

where $y_i^V$1 increases renewable use but also eviction risk.

Placement uses a two-dimensional feature vector. For each server,

$y_i^V$2

capturing headroom for additional green cores and additional safe grid-backed cores. The placement rule assigns a VM to the server closest, by Euclidean distance, to an ideal point chosen according to VM criticality (Hewage et al., 2024).

Implementation is provided as openstack-gc, an open-source prototype that modifies the OpenStack control plane and adds an on-node daemon using the Intel Power Optimization Library. The paper reports both hardware and simulation results. In the prototype, the system achieves a 22% reduction in peak server power and, relative to a temporal-flexibility baseline, a 6.52× spike in coefficient of variation is avoided when shrinking physical core allocations is replaced by eviction-based management. In large-scale simulation with 8,000+ servers and a 14-day Azure VM arrival trace, the proposed packing algorithm achieves a 34.83% increase in renewable energy harvest over crit-aware packing and a 79.64% reduction in VM eviction incidents compared to best-fit packing (Hewage et al., 2024).

7. Comparative significance

The two RCORE frameworks illuminate different uses of structured inductive bias. In video understanding, RCORE constrains learning with temporally faithful augmentation and temporal order regularization to suppress object-driven shortcuts (Ahn et al., 22 Jan 2026). In cloud systems, RCORE constrains execution with renewables-driven cores, criticality-aware evictions, and Green Cores-aware packing to reconcile renewable intermittency with real-time guarantees (Hewage et al., 2024).

The resemblance is therefore nominal rather than methodological. One RCORE is a representation-learning and regularization framework for ZS-CAR; the other is a systems framework for carbon-aware cloud scheduling. Treating either as the canonical meaning of the acronym would erase this domain-specific distinction. A more accurate reading is that “RCORE” currently functions as a reused acronym for different research programs, each defined by its local technical context rather than by a shared underlying formalism.