Cartridges in Modular Systems and Applications

• Cartridges are defined as modular subassemblies with a stable interface that isolate complex functions, enabling seamless integration into larger systems.
• They are applied across domains, including machine learning (KV-cache adaptation), radio astronomy (cryogenic receivers), biosensing (analytical modules), and even ammunition classification.
• Their design supports efficient maintenance, streamlined upgrades, and novel adaptations like stealth-bias detection and latent-memory compression in advanced applications.

Searching arXiv for the supplied papers to ground the article with current records. Across the cited literature, cartridges are bounded modules inserted into larger systems to supply a specialized function while preserving the host system’s main architecture. The term is used for a small KV-cache prefix adapter or a trainable KV cache in LLMs, a self-contained receiver unit in cryogenic radio astronomy front ends, a plug-in field-emission cathode assembly in an RF electron injector, a single-use paper-based analytical cartridge in biosensing, a glass primary pharmaceutical container, and an ammunition-specific descriptor recorded alongside firearm acoustics metadata (Talaei et al., 1 Jul 2026, Eyuboglu et al., 6 Jun 2025, Baryshev et al., 2015, Qiu et al., 2017, Jang et al., 2024, Gurny et al., 16 Jun 2026). This range of usage suggests a common engineering pattern: the cartridge is a localized, replaceable, or reusable subassembly that concentrates system-specific functionality behind a stable interface.

1. Cartridge as a modular systems concept

In physical instrumentation, a cartridge is typically a mechanically bounded unit that can be inserted, aligned, replaced, or serviced without redesigning the full host apparatus. The ALMA Band 9 receiver units are explicitly called “cartridges” and are installed in the telescope’s front end as complete cryogenic heterodyne receiver modules (Baryshev et al., 2015). In the miniature X-band injector, the cathode cartridge is the replaceable assembly inserted into the back wall of the RF cavity and designed so the emitter can be installed, conditioned, replaced, or recycled without redesigning the injector (Qiu et al., 2017). In the SMA upgrade, aging receiver cartridges are replaced by new dual-polarized wideband receiver cartridges within a redesigned cryostat (Grimes et al., 2024). In biosensing, the paper-based analytical cartridge is the disposable reaction module that couples wet chemistry to a reusable FET readout (Jang et al., 2024).

In machine learning, the term is extended from hardware to latent-state infrastructure. A cartridge becomes a KV-cache prefix or compact learned KV block that is loaded directly into an LLM’s attention stack, allowing the model to behave as if it had processed a much larger document context (Eyuboglu et al., 6 Jun 2025). In the D2D auditing framework, the cartridge is again a KV-cache prefix adapter, but its purpose is not long-context retrieval; it is to compress the behavioral difference between a suspected model and its base so that stealth preferential bias becomes visible in generated text (Talaei et al., 1 Jul 2026).

The semantic breadth of the term therefore does not indicate conceptual vagueness. Rather, it reflects a recurring design principle: the cartridge isolates a difficult function—heterodyne reception, field emission, assay chemistry, document memory, or behavior-shift concentration—inside a portable module with a constrained interface.

2. Learned KV-cache cartridges for long-context inference

In LLM research, a cartridge is a small, trainable KV cache that stands in for a long corpus (Eyuboglu et al., 6 Jun 2025). The central problem is that standard in-context learning over long corpora is memory-hungry because the KV cache scales with input length. The long-context baseline prepends the entire corpus $C$ to every query, whereas the cartridge approach trains a compact cache $Z$ offline and reuses it across all subsequent queries about that corpus. The learned object is written as $Z \in \mathbb{R}^{L \times p \times d \times 2}$, with memory equivalent to a prefix of $p$ tokens (Eyuboglu et al., 6 Jun 2025).

The original training recipe is Self-Study, which replaces naive next-token training on the corpus with synthetic conversations plus context distillation. The model chunks the corpus into subcorpora, generates synthetic dialogues from generic but diverse seed prompts, and trains the cartridge to minimize KL divergence between the teacher distribution $F(\cdot \mid C \oplus q)$ and the cartridge-augmented student distribution $F_Z(\cdot \mid q)$ (Eyuboglu et al., 6 Jun 2025). The distinction from summarization is explicit in subsequent work: token-space summarization is described as highly lossy, whereas latent-space compaction preserves the model’s own memory format (Zweiger et al., 18 Feb 2026).

The reported systems gains are substantial. Self-Study cartridges match ICL performance while using 38.6x less memory and enabling 26.4x higher throughput (Eyuboglu et al., 6 Jun 2025). The same paper reports that LLaMA 70B needs 84 GB of memory at 16-bit precision to answer a single question over a 128k-token context, and that on a single H100, LLaMA-8B throughput drops by $77\times$ from 1k to 120k tokens (Eyuboglu et al., 6 Jun 2025). Cartridges also extend effective context length: on MTOB, a model with a 128k context window can use a cartridge built from a 484k-token textbook (Eyuboglu et al., 6 Jun 2025).

Mechanistic work refines this picture. A later analysis argues that cartridge keys act as stable, shareable retrieval routers and that most learned compression occurs in the value vectors (Diaz, 23 Aug 2025). In that study, value rotations are often an order of magnitude larger than key rotations, and swapping learned keys across tasks causes only mild degradation—about 4–5% on Llama models and about 7% on Qwen3—while still outperforming the no-cartridge baseline (Diaz, 23 Aug 2025). The same paper proposes Sampled Chunk Initialization (SCI) as an alternative to First-$k$ Token Initialization and reports faster convergence, with convergence defined as reaching perplexity 1.10 and a paired $t$-test showing $p<0.05$ (Diaz, 23 Aug 2025).

The question of composability is not settled in a single universal sense. One paper reports that independently trained Cartridges can be concatenated at inference time without retraining (Eyuboglu et al., 6 Jun 2025). By contrast, Cartridges at Scale (CAS) argues that cartridges trained in isolation are non-compositional over large document collections and that naively mixing them can collapse performance to near chance unless training explicitly includes distractor co-visibility (Hardalov et al., 3 Jun 2026). This suggests that cartridge composition is regime-dependent: it can work in some inference settings, but scalable multi-cartridge systems require dedicated training procedures.

A further systems development is Attention Matching, which replaces slow end-to-end optimization with per-layer, per-head matching of attention output and attention mass (Zweiger et al., 18 Feb 2026). It decomposes compaction into key selection, bias fitting by NNLS, and value fitting by ordinary least squares, achieving up to $Z$0 compaction in seconds to minutes and roughly 2 orders of magnitude speedup over cartridge training in the main comparison (Zweiger et al., 18 Feb 2026). The literature therefore distinguishes between cartridges as a representational idea and any single training algorithm used to realize them.

3. Cartridge distillation for stealth-bias detection

The 2026 D2D framework uses the word cartridge in a more adversarially oriented sense: a cartridge is a small KV-cache prefix adapter trained with frozen base-model weights to expose hidden bias (Talaei et al., 1 Jul 2026). It is “a form of prefix tuning” that optimizes learned key-value states prepended at each transformer layer, with far fewer trainable parameters than the full model (Talaei et al., 1 Jul 2026). The operational setting assumes a suspected model and its known base model. The cartridge is trained so that the adapted base matches the suspected model’s output distribution on data unrelated to the bias topic, with the objective of minimizing forward KL divergence (Talaei et al., 1 Jul 2026).

The motivating threat model is a stealth preferential bias that resides in the soft logit distribution and may be invisible to text-based inspection. D2D’s central claim is that cartridge distillation learns an approximation to that distributional shift, but because the cartridge is capacity-limited, it cannot reproduce every detail of the shift. Instead, it captures the dominant, coherent bias signal and drops diffuse masking residuals. The result is that a bias originally visible only in logits becomes visible in generated text, where behavioral auditors such as Petri or AuditBench can detect it (Talaei et al., 1 Jul 2026).

The paper makes a strong theoretical point: the capacity bottleneck is not merely a limitation but the mechanism of exposure. If the adapter is too small, it cannot capture even the real bias; if it is too large, it starts learning the same masking residuals as the original stealth model. Detection is predicted to peak at an intermediate size near the intrinsic rank of the bias, yielding an inverted-U curve (Talaei et al., 1 Jul 2026). A common misconception would be to treat more adapter capacity as monotonically better; the D2D analysis rejects that view.

Under a local quadratic approximation of the KL objective, D2D becomes a Fisher-weighted low-rank projection problem on the logit shift between suspected and base models (Talaei et al., 1 Jul 2026). The logit shift is decomposed into a low-rank bias component and a high-rank residual component, and the cartridge learns the top Fisher-weighted singular directions. The paper’s Bias Concentration Ratio formalizes the possibility that removing masking residuals can make the recovered bias stronger than in the original stealth model, so that $Z$1 (Talaei et al., 1 Jul 2026). In this formulation, the cartridge is neither a compression artifact nor a deployment-time memory object; it is a deliberately capacity-bounded behavioral amplifier for auditing.

4. Receiver cartridges in radio and submillimeter astronomy

In radio astronomy, a cartridge is a self-contained cryogenic receiver unit with standardized interfaces. The ALMA Band 9 receiver cartridges occupy the telescope front-end cryostat and cover 602–720 GHz (Baryshev et al., 2015). Each cartridge accepts the telescope beam, separates the two linear polarizations with a polarizing grid, injects the local oscillator through beam splitters, down-converts each polarization with an independent double-sideband SIS mixer, and amplifies the resulting 4–12 GHz IF chain with cryogenic and room-temperature HEMT amplifiers (Baryshev et al., 2015). The optical system is fully integrated within the cartridge, implemented as precision-machined aluminum blocks with no shimming or adjustable alignment and requiring about 40 μm relative positioning accuracy (Baryshev et al., 2015).

The ALMA work emphasizes both modular production and performance uniformity. Laboratory characterization was reported for 73 Band 9 cartridges, all meeting the ALMA specifications of SSB noise temperature < 335 K over 80% of the band and SSB noise temperature < 500 K at any frequency (Baryshev et al., 2015). The project used both AlO$Z$2 and AlN SIS barrier technologies, with AlN performing better at the high-frequency end of Band 9 and at higher IF frequencies (Baryshev et al., 2015). On-sky commissioning further showed stable phases, stable amplitudes, and line-rich science-verification spectra (Baryshev et al., 2015).

The wSMA upgrade uses the same term in a different but related architectural setting. The original SMA configuration used four single-polarized double-sideband receiver cartridges per antenna—230 GHz (RxA), 240 GHz (RxB), 345 GHz (RxA), and 400 GHz (RxB)—inside one cryostat (Grimes et al., 2024). The new system replaces these with a two-cartridge cryostat, each cartridge being dual-polarized: wSMA-Low: 194–286 GHz and wSMA-High: 264–376 GHz (Grimes et al., 2024). Each new cartridge contains a profiled corrugated feed, an ortho-mode transducer (OMT), local oscillator couplers, two double-sideband SIS mixers, and up to four IF output chains (Grimes et al., 2024).

The upgrade is not merely a refresh of aging hardware. It is explicitly tied to improvements in instantaneous bandwidth, sensitivity, polarization alignment, and dual-frequency flexibility (Grimes et al., 2024). Estimated Stage I performance gives typical $Z$3 values of 147 K for the 1.3 mm band and 484 K for the 850 μm band, versus 222 K and 702 K in the current system (Grimes et al., 2024). Selector-wheel optics—open aperture, plane mirror, wire polarizing grid, and dichroic beamsplitter—enable same-frequency dual-polarization observing, dual-frequency observing, and future tri-band configurations (Grimes et al., 2024). In this context, the cartridge is not merely a holder for a mixer; it is the receiver architecture’s fundamental modular unit.

5. Cartridges as reaction modules and emission modules

In point-of-care biosensing, the cartridge is a single-use paper-based analytical cartridge built around a porous sensing membrane (PSM) integrated with an ITO sensing electrode and electrically coupled to a commercial n-type MOSFET (CD4007UB) (Jang et al., 2024). Its layered construction includes an acrylic substrate, double-sided tape spacers, and an air gap (~50 μm) between the PSM and ITO surface (Jang et al., 2024). Dried reagents on the PSM include cholesterol esterase, cholesterol oxidase, peroxidase, surfactants, 10% BSA, and optimized buffers (Jang et al., 2024). The assay injects 20 μL of plasma, the cartridge generates protons (H$Z$4) through the biochemical cascade, and the FET transfer curve is measured repeatedly for 5 minutes—more specifically, 49 cycles at 7 s per cycle, for a total of 343 s (Jang et al., 2024).

A crucial feature is that the cartridge eliminates the need for conventional FET surface functionalization. The enzymes and assay chemistry are pre-dried in the paper matrix, while the ITO surface requires only BSA blocking rather than complex immobilization chemistry (Jang et al., 2024). The cartridge cost is reported as < \$0.15 per test (Jang et al., 2024). For cholesterol testing, the DL pipeline uses a heatmap representation of the kinetic transfer-curve sequence, and the best-performing subset is $Z$5 to $Z$6 and time = 91 s to 119 s, reducing assay operating time to < 2.5 minutes (Jang et al., 2024). Blind comparison against a CLIA-certified clinical laboratory achieved $Z$7 for all three batches and maximum CV = 6.46% (Jang et al., 2024). The paper explicitly warns against viewing the cartridge as just a sample holder: it is the entire wet-chemistry assay module.

In accelerator hardware, the cartridge is a plug-in field-emission cathode cartridge carrying a thin-film (N)UNCD emitter for a miniature single-cell X-band electron injector (Qiu et al., 2017). The injector is a conventional single-cell quarter-wave RF gun with a copper body, stainless steel flanges, and standard 1.33-inch ConFlat flanges for vacuum connections and cathode insertion (Qiu et al., 2017). The cathode plug comprises a base with a venting hole, a stainless-steel cartridge attached with a set screw, and a spring-loaded groove for RF contact (Qiu et al., 2017). The cartridge is mirror polished, coated with Mo, and then receives the (N)UNCD thin film by microwave-plasma-assisted CVD (Qiu et al., 2017).

When installed, the cathode plug defines an internal insertion length $Z$8 mm, setting the cavity resonance to 9.17 GHz (Qiu et al., 2017). The cartridge diameter is 4.4 mm, and the device produced pulsed beams with about 1 mA/cm² at up to 28 MV/m, beam energy around 10 keV, and beam diameter as small as 1 mm (Qiu et al., 2017). A principal engineering advantage is that the cartridge requires no laser system and no heater, unlike photocathodes or thermionic emitters (Qiu et al., 2017). Here again, the cartridge localizes a difficult function—electron emission in a high-gradient RF cavity—inside a replaceable module.

6. Cartridges as containers and ammunition descriptors

The term also denotes objects whose significance is material, not merely modular. In pharmaceutical manufacturing, a glass cartridge is a primary container whose integrity must survive conveyor transport before filling (Boso et al., 2020). The integrated DEM–FEM study models many upright cartridges on an accumulation table, representing each cylindrical cartridge in DEM by an equivalent sphere with the same mass and material, then using FEM on representative collision clusters to compute local stresses (Boso et al., 2020). The combined Hertz–Reissner contact model is reported to match direct FE collision results better than Hertz or Reissner alone (Boso et al., 2020).

The transport process is shown to involve repeated shocks of varying intensity. The DEM study identifies an optimal conveyor configuration of $Z$9 cm, $Z \in \mathbb{R}^{L \times p \times d \times 2}$0 s, and belt speed $Z \in \mathbb{R}^{L \times p \times d \times 2}$1 cm/s, lower than the real line speed of 30 cm/s (Boso et al., 2020). In the FEM stage, confinement from neighboring cartridges amplifies loading; for the 1–3 pair with confinement, normal contact force reaches about 55 N at the bottom and about 80 N at the top, while maximum principal stress reaches about 140 MPa at the bottom of cartridge 1 (Boso et al., 2020). The paper’s fracture interpretation is especially important: the maximum tensile stress appears on the interior side of the wall, so a crack may originate inside the cartridge and propagate outward (Boso et al., 2020).

In firearm acoustics, the relevant object is the ammunition cartridge rather than a packaging or instrument module. The Certus Caliber Classification Gunshot Dataset (C3GD) contains 8015 audio files from 28 firearms across 16 calibers, and its metadata includes exact cartridges used in addition to calibers, platforms, microphone locations, microphone details, and testing events (Gurny et al., 16 Jun 2026). The paper argues that caliber-based classification is more practical than firearm make/model classification, but it also emphasizes that caliber and cartridge are related yet not perfectly interchangeable descriptors (Gurny et al., 16 Jun 2026).

That distinction matters acoustically. The dataset discussion highlights ambiguities in prior work, such as whether a 12 gauge used slugs or shot, or whether 9x19mm ammunition was subsonic, supersonic, or overpressure (Gurny et al., 16 Jun 2026). It also distinguishes muzzle blast from the ballistic shockwave produced by supersonic ammunition, noting that subsonic projectiles lack the shockwave (Gurny et al., 16 Jun 2026). In this usage, “cartridge” names the ammunition configuration that partly determines the observed waveform; it is therefore not a module inserted into a host system but a ballistic and metadata-bearing entity in its own right.

Taken together, these literatures show that the word cartridge has become a technically precise but domain-specific term. In some fields it denotes a replaceable subsystem; in others, a latent memory prefix; in others still, a container or an ammunition specification. What unifies these uses is not the material form of the object but its role as a bounded carrier of function, state, or identity within a larger operational framework.