ReLOAD: Multidomain Mechanisms & Models

• ReLOAD is a multifaceted research term covering techniques in computer vision, optics, reinforcement learning, and other domains.
• Research on ReLOAD demonstrates specialized implementations like boundary-guided semantic segmentation, optical angle amplification, and reward annotation in offline RL with notable performance gains.
• Applications span from enhancing image segmentation and laser scanning to optimizing database queries and mitigating cache attacks through tailored methodological innovations.

Searching arXiv for papers on “ReLOAD” and closely related usages to ground the article. ReLOAD is a research term used for several unrelated methods, mechanisms, and models across computer vision, optics, reinforcement learning, software testing, database systems, privacy-preserving machine learning, and graph optimization. In recent arXiv usage it denotes the Reload mechanism in semantic segmentation, “Relayed-Loop Optically Amplified Deflection” in laser scanning, “Reinforcement Learning with Optimistic Ascent-Descent” in constrained MDPs, “Reinforcement Learning with Offline Reward Annotation via Distillation” in offline RL, a self-adaptive reinforcement-learning–driven load testing agent, a robust learned query optimizer, and a partially-blind unlearning framework (Chan et al., 16 Jun 2026, Jayakumar et al., 22 Sep 2025, Moskovitz et al., 2023, Chaudhary et al., 17 Jul 2025, Moghadam et al., 2021, Lee et al., 16 Apr 2026, Newatia et al., 12 Apr 2026).

1. Terminological scope

The term appears in both acronymic and non-acronymic forms. In some papers it expands to a specific phrase; in others, “reload” denotes a domain-specific operation such as a traversal cost, a reactor refueling pattern, or a reload-style cache attack. The table summarizes the principal usages represented in recent literature.

Usage	Meaning	Domain
ReLOAD in Reload-Mamba (Chan et al., 16 Jun 2026)	Reload mechanism for anti-dilution refinement in multi-class semantic segmentation	Computer vision
ReLOAD (Jayakumar et al., 22 Sep 2025)	Relayed-Loop Optically Amplified Deflection	Optical scanning
ReLOAD (Moskovitz et al., 2023)	Reinforcement Learning with Optimistic Ascent-Descent	Constrained RL
ReLOAD (Chaudhary et al., 17 Jul 2025)	Reinforcement Learning with Offline Reward Annotation via Distillation	Offline RL
RELOAD (Moghadam et al., 2021)	Reinforcement-learning–driven load testing agent	Software testing
RELOAD (Lee et al., 16 Apr 2026)	Learned query optimizer with knowledge retention and transfer	Database systems
Reload (Newatia et al., 12 Apr 2026)	Partially-blind unlearning framework	Privacy / machine unlearning
reload cost (Gözüpek et al., 2016)	Traversal cost induced by color changes on edges	Graph algorithms

This multiplicity is not merely nominal. Each usage defines a different object: a decoder refinement gate, an optical cavity, an optimistic primal–dual update rule, an intrinsic reward annotator, a workload-generation policy, a replay-and-meta-learning wrapper for query planning, an unlearning pipeline, or a cost functional on edge-colored graphs.

2. Reload-Mamba and anti-dilution semantic segmentation

In semantic segmentation, ReLOAD is the anti-dilution mechanism introduced by “Reload-Mamba: Hierarchical Anti-Dilution State-Space Modeling for Multi-Class Semantic Segmentation” (Chan et al., 16 Jun 2026). The paper starts from the observation that Mamba-based state space models provide linear-time long-range dependency modeling for high-resolution dense prediction, but that sequential state-space propagation can attenuate boundary-sensitive and detail-sensitive responses. This effect is termed propagation-induced response dilution and is especially harmful for multi-class semantic segmentation, where object boundaries, thin structures, and small categories have small spatial support.

Reload-Mamba addresses this with three segmentation-specific designs. The first is a boundary-supervised local detail prior trained with ground-truth boundary masks. Given ground-truth labels $G$, the boundary mask is defined over a $3\times 3$ neighborhood, and lower-resolution masks are obtained by max-pooling. The predicted prior $P^{(l)}$ gates local decoder features through

$$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$

with $\lambda_{\text{low}} = 0.3$ (Chan et al., 16 Jun 2026). The second is a class-uncertainty-aware Reload Gate that incorporates per-pixel normalized class entropy

$$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$

derived from a pre-reload auxiliary head. The gate uses the original decoder feature $D_l$, the propagation-induced change $D_l - M^{(l)}$, and $U^{(l)}$, modulates the result by the prior $P^{(l)}$, and reinjects local detail via

$3\times 3$0

The third is a hierarchical multi-level Reload mechanism applied at decoder levels $3\times 3$1, $3\times 3$2, and $3\times 3$3, followed by top-down fusion (Chan et al., 16 Jun 2026).

Architecturally, Reload-Mamba uses a ConvNeXt-Tiny encoder, a U-shaped multi-scale decoder, four-directional Mamba scanning with no weight sharing across directions, pixel-wise directional attention, and an edge prior derived from Sobel gradients of the raw image (Chan et al., 16 Jun 2026). Controlled ablations isolate each component. On ADE20K, a direct single-level anti-dilution baseline reaches 45.7% mIoU, adding the boundary-supervised prior yields 46.3%, adding the class-uncertainty-aware gate yields 47.0%, and adding hierarchical multi-level Reload yields 47.9%, for a cumulative +2.2 mIoU over the direct-port baseline (Chan et al., 16 Jun 2026). Reported final benchmarks are 47.9% single-scale and 48.9% multi-scale mIoU on ADE20K, 83.2% single-scale mIoU on Cityscapes, and 87.8% mIoU on PASCAL VOC 2012 val under the stated ResNet-101 + COCO pre-training protocol (Chan et al., 16 Jun 2026).

3. Relayed-Loop Optically Amplified Deflection

In optics, ReLOAD expands to Relayed-Loop Optically Amplified Deflection (Jayakumar et al., 22 Sep 2025). It is an optical architecture that amplifies the scan angle, and therefore the effective optical invariant, of a laser deflector by sending the same beam through the deflector multiple times in a controlled optical loop. If $3\times 3$4 is the single-pass deflection angle, the net output angle is

$3\times 3$5

where $3\times 3$6 is the number of passes (Jayakumar et al., 22 Sep 2025).

The implementation described in the paper places the deflector at a pupil plane of an 8-$3\times 3$7 cavity made of two 4-$3\times 3$8 afocal relays. A loop control mirror tilted in a direction orthogonal to the active scan axis shifts the beam in a loop axis while preserving imaging conjugation onto the deflector. The beam re-encounters the same pupil with the same beam diameter, so each pass adds the same angular increment. This multiplies the effective optical invariant in the scanning coordinate and thereby the number of resolvable spots (Jayakumar et al., 22 Sep 2025).

The demonstrated system uses a KTN electro-optic deflector. For the reported EO-ReLOAD implementation, $3\times 3$9 mm, $P^{(l)}$0 nm, $P^{(l)}$1 mrad, and $P^{(l)}$2 mrad, which gives $P^{(l)}$3; the authors selected $P^{(l)}$4 to limit losses and aberrations (Jayakumar et al., 22 Sep 2025). With $P^{(l)}$5, $P^{(l)}$6 mm, $P^{(l)}$7 nm, and $P^{(l)}$8 mrad, the ideal diffraction-limited estimate is $P^{(l)}$9 resolvable spots. Experimentally, a single-pass KTN deflector at 100 kHz drive and 592 V$$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$0 produced 23 resolvable spots, whereas 8× EO-ReLOAD produced 132 resolvable spots. The paper attributes the gap between full 8× angle amplification and only ~6× improvement in usable resolvable spots to spot growth, vignetting at KTN edges, and telecentric errors in stock relay lenses (Jayakumar et al., 22 Sep 2025).

The system-level demonstration integrates EO-ReLOAD into a laser scanning reflectance microscope and reports 10 kHz frame rate imaging, 1 MHz line scan rate, and $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$1s step times. The measured step response is 1.5 $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$2s for a motion corresponding to about 50 resolvable spots in the 8× output. Prototype throughput is 11% after all 8 passes, and spot intensity can drop by up to 76% from center to edge, which the paper presents as a primary trade-off for higher $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$3 (Jayakumar et al., 22 Sep 2025).

4. ReLOAD in reinforcement learning and adaptive optimization

In constrained reinforcement learning, ReLOAD denotes Reinforcement Learning with Optimistic Ascent-Descent (Moskovitz et al., 2023). The paper considers discounted CMDPs with occupancy measure $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$4 and Lagrange multipliers $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$5, and writes the saddle-point objective as

$$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$6

Its motivation is that standard primal–dual mirror descent guarantees average-iterate convergence but not last-iterate convergence; policies can oscillate between satisfying constraints and maximizing reward. ReLOAD replaces ordinary descent–ascent with optimistic updates using $$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$7, and the paper proves last-iterate convergence in the convex occupancy-measure formulation. It also reports empirical effectiveness on discrete MDPs and continuous control and introduces a benchmark of challenging constrained RL problems (Moskovitz et al., 2023).

In offline RL, ReLOAD denotes Reinforcement Learning with Offline Reward Annotation via Distillation (Chaudhary et al., 17 Jul 2025). The method assumes a reward-free offline dataset

$$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$8

and a small expert transition dataset

$$X^{(l)} = D_l \odot P^{(l)} + \lambda_{\text{low}}\, D_l \odot (1 - P^{(l)}),$$9

A fixed target network $\lambda_{\text{low}} = 0.3$0 and trainable predictor $\lambda_{\text{low}} = 0.3$1 are used in a Random Network Distillation construction over transitions $\lambda_{\text{low}} = 0.3$2. The predictor is trained on $\lambda_{\text{low}} = 0.3$3, and reward on the offline data is defined by the negative prediction error

$\lambda_{\text{low}} = 0.3$4

This inverts the usual RND novelty semantics: low error means expert-like transition, high error means non-expert transition. The relabeled dataset is then used by a standard offline RL learner, specifically IQL in the experiments. On D4RL, the paper reports performance competitive with reward-annotated methods (Chaudhary et al., 17 Jul 2025).

In software performance engineering, RELOAD is a self-adaptive reinforcement-learning–driven load testing agent (Moghadam et al., 2021). It treats workload generation as a model-free RL problem, uses discrete performance states derived from average response time and error rate, and defines actions as transaction-specific load increases of one third:

$\lambda_{\text{low}} = 0.3$5

The reward is

$\lambda_{\text{low}} = 0.3$6

The implementation uses Q-learning with $\lambda_{\text{low}} = 0.3$7-greedy exploration and a DQN variant. On a WooCommerce-based e-commerce application, RELOAD reports 30–34% test cost saving relative to the standard baseline and 17–20% relative to random testing after convergence; in transfer learning experiments it reports 25% and 13% savings, respectively (Moghadam et al., 2021).

A distinct, non-acronymic usage appears in PWR core reload optimization, where PPO is shown statistically superior to Genetic Algorithm, Parallel Simulated Annealing, and Tabu Search for equilibrium-cycle loading pattern optimization under an LCOE-based objective with constraint penalties (Seurin et al., 2024). This usage does not define a framework named ReLOAD, but it places “reload” in the specific reactor-engineering sense of core loading pattern design.

5. Database optimization and forget set-free unlearning

In database systems, RELOAD is “A Robust and Efficient Learned Query Optimizer for Database Systems” (Lee et al., 16 Apr 2026). It is a reinforcement-learning–based learned query optimizer designed to wrap and augment existing learned optimizers such as Balsa and LIMAO on PostgreSQL and SQL Server. RELOAD focuses on robustness, defined in terms of reducing query-level performance regressions, and efficiency, defined in terms of faster convergence to expert-level plan quality. Its two main mechanisms are knowledge retention, implemented as a specialized prioritized experience replay over join-rooted subplans, and knowledge transfer, implemented as a MAML-based initialization across task partitions derived from query complexity, operator count, estimated cost, or estimated rows (Lee et al., 16 Apr 2026). On standard benchmarks including JOB, TPC-DS, and SSB, the paper reports up to 2.4x higher robustness and 3.1x greater efficiency than state-of-the-art RL-based query optimization techniques (Lee et al., 16 Apr 2026).

In privacy-preserving machine learning, Reload is a partially-blind unlearning framework that removes the influence of training subsets without explicit access to the forget set (Newatia et al., 12 Apr 2026). The paper formalizes the retain set $\lambda_{\text{low}} = 0.3$8, forget set $\lambda_{\text{low}} = 0.3$9, the full-data model $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$0, and the ideal retrained model

$$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$1

Its auxiliary object for conventional models is the final-epoch full-data gradient $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$2, and it uses the identity

$$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$3

to support forget-set-free updates (Newatia et al., 12 Apr 2026). Operationally, the method combines gradient optimization with structured weight sparsification and selective reinitialization. On LLMs, the paper reports that Reload can unlearn entities using <0.025% of the retain set and <7% of model weights in <8 minutes on Llama2-7B, and in the corrective setting it achieves unlearning when only 10% of corrupted data is identified (Newatia et al., 12 Apr 2026).

6. Reload cost in graph algorithms

In graph theory, reload cost is not an acronym but a cost model on edge-colored graphs (Gözüpek et al., 2016). If a path traverses a vertex through two incident edges of different colors, it incurs a traversal cost determined only by that color pair. This induces two global measures: reload cost, which sums traversal costs over all path occurrences, and changeover cost, which does not depend on the amount of commodity or number of paths traversing a vertex (Gözüpek et al., 2016). “Edge Coloring with Minimum Reload/Changeover Costs” studies the optimization problem in which the edge coloring itself is chosen so as to minimize these costs; it gives strong hardness results and polynomial-time solvable cases on trees, bounded-degree trees, and graphs whose block structure is tree-like (Gözüpek et al., 2016).

The same cost model underlies cycle-cover and factor problems. “Minimum Reload Cost Cycle Cover in Complete Graphs” studies 2-factors in complete graphs with equitable or nearly equitable 2-edge-colorings and proves that, except possibly for certain small cases, the minimum reload cost is zero because a monochromatic cycle cover exists; for complete graphs with at least 13 vertices and a nearly equitable 2-edge-coloring, the optimal value is $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$4 (Büyükçolak et al., 2017). “Minimum Reload Cost Graph Factors” generalizes from 2-factors to $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$5-factors, proves NP-hardness even when $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$6, $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$7, and $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$8, shows planar hardness for $$U^{(l)}_i = -\frac{1}{\log C}\sum_{c=1}^{C} p^{(l)}_{c,i}\log p^{(l)}_{c,i},$$9, gives a polynomial-time algorithm when $D_l$0, and provides an FPT algorithm parameterized by $D_l$1 and $D_l$2 (Baste et al., 2018).

A plausible implication is that “reload” in this graph-theoretic line consistently refers to a local transition penalty at a vertex, rather than to restoration or reinitialization. That sense is conceptually distinct from the acronymic ML and optics usages, even though the spelling is identical.

7. Reload-style cache attacks and defenses

In computer architecture, “reload” appears in attack names rather than as a standalone framework. “RELOAD+REFRESH: Abusing Cache Replacement Policies to Perform Stealthy Cache Attacks” introduces a RELOAD-style last-level cache side-channel that exploits Intel’s deterministic replacement policy to monitor a victim’s use of a shared cache line without evicting that line from the LLC and with negligible additional LLC misses on the victim (Briongos et al., 2019). The attack depends on reverse-engineering a Quad-Age LRU model, arranging the target as the eviction candidate, then using a controlled miss plus a REFRESH phase to reconstruct the age configuration. On AES, it requires 110,000 samples per 4-byte group, close to FLUSH+RELOAD at 108,000; on RSA it reports 94.9% true positives with 38.75% false positives; and it produces a victim miss distribution close to the no-attack baseline (Briongos et al., 2019).

Mitigation work has focused on Flush+Reload and related primitives. “Lookout for Zombies: Mitigating Flush+Reload Attack on Shared Caches by Monitoring Invalidated Lines” proposes Zombie-Based Mitigation (ZBM), which marks Flush-Caused Invalidation lines as Zombies and treats hits on Zombie lines as misses to eliminate timing leakage. The design requires 4-bits per cache line in its extended form, retains OS-based page sharing, requires no OS or ISA changes, and does not incur slowdown for benign applications (Saileshwar et al., 2019).

The same naming pattern appears in Prefetch+Reload, introduced as part of “Adversarial Prefetch: New Cross-Core Cache Side Channel Attacks” (Guo et al., 2021). That paper identifies two vulnerabilities in Intel PREFETCHW: it can execute on read-only data, and its execution time leaks coherence state. Based on these, it constructs Prefetch+Reload and Prefetch+Prefetch, which work with both inclusive and non-inclusive LLCs. As covert channels, Prefetch+Reload and Prefetch+Prefetch reach 782 KB/s and 822 KB/s, respectively, using only one shared cache line, and in transient execution attacks they leak about 2 times as many secret bytes as Flush+Reload (Guo et al., 2021).

These microarchitectural usages are related by attack structure rather than by acronym expansion. “Reload” denotes a measurement phase that infers prior accesses from cache-state timing, and later work either refines that phase, replaces the eviction mechanism, or attempts to remove the timing signal entirely.