AS-Bridge: Bridging Heterogeneous Modalities

• AS-Bridge is a collection of frameworks that insert an intermediate bridging module to reconcile mismatches between heterogeneous modalities, such as LSST/Euclid in astronomy and speech enhancement/ASR in language processing.
• In astronomy, it uses a bidirectional Brownian Bridge diffusion model to probabilistically map LSST and Euclid observations, enabling reconstruction of missing data and rare-event detection.
• For speech applications, AS-Bridge integrates spectral similarity-based post-processing and unsupervised connectors (wav2vec 2.0 to PT5) to mediate between noisy inputs and robust recognition outputs.

Searching arXiv for the relevant AS-Bridge papers to ground the article in the current literature. AS-Bridge is a reused designation for several bridge-based research programs that mediate between heterogeneous modalities, pretrained systems, or survey domains rather than replacing them outright. In current arXiv usage, the term denotes at least three distinct frameworks: a bidirectional Brownian-Bridge diffusion model connecting LSST and Euclid observations in observational cosmology (Zhang et al., 12 Mar 2026), a lightweight post-processing module inserted between frozen speech-enhancement and automatic-speech-recognition systems for robust ASR (Wang et al., 2024), and an unsupervised speech-to-text connector that links wav2vec 2.0 and phoneme-level T5 for spoken language understanding (Shi et al., 2022). Across these settings, the shared principle is the construction of an intermediate stochastic, spectral, or symbolic bridge that compensates for modality mismatch while preserving uncertainty or downstream utility.

1. Nomenclature and recurring design pattern

The designation “AS-Bridge” is not attached to a single canonical architecture. In observational cosmology, it is explicitly expanded as A(stronomical)S(urvey)-Bridge and refers to a bidirectional generative model between LSST and Euclid (Zhang et al., 12 Mar 2026). In robust speech recognition, “AS-Bridge” denotes a bridge module that evaluates signal-level information between noisy and enhanced speech and then applies observation addition before a frozen ASR backend (Wang et al., 2024). In spoken language understanding, Jiatong Shi et al. describe AS-Bridge as a three-stage stack—wav2vec2, unsupervised ASR, and phoneme-level T5—that connects speech and textual pre-trained models without paired speech-text supervision (Shi et al., 2022).

A concise summary of the three uses is given below.

Usage	Bridge object	Reported function
Astronomical surveys	Brownian Bridge diffusion between LSST and Euclid endpoints	Probabilistic prediction of missing observations and inter-survey rare-event detection
Robust ASR	Lightweight bridge module producing an OA coefficient $S' \in [0.6,1]$	Integration of diverse pre-trained SE and ASR models
Speech-text SLU	Unsupervised ASR connector between wav2vec 2.0 and PT5	Reconciliation of speech and textual pre-trained models

The common pattern is an intermediate mechanism that is neither a trivial adapter nor a full end-to-end retraining strategy. This suggests a broader methodological role for “bridge” constructions: they provide controlled transfer across systems whose native representations are misaligned.

2. AS-Bridge in astronomical survey translation

In the astronomy formulation, AS-Bridge addresses joint analysis between the ground-based LSST at the Vera C. Rubin Observatory and the space-based Euclid mission, whose observations differ in depth, resolution, wavelength, point-spread function, and scanning cadence (Zhang et al., 12 Mar 2026). LSST provides six optical bands $(u,g,r,i,z,y)$, while Euclid provides a broad visible band and near-infrared $(Y,J,H)$ bands. The paper frames the survey–survey mapping as ill-posed and inherently stochastic because both observations are noisy, partial projections of an unobservable “true sky” process $\Phi$.

The generative model treats paired LSST and Euclid images of the same object as endpoints of a Brownian Bridge in high-dimensional pixel space. Survey observations are written as

$$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$

AS-Bridge then models both $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$ and $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$ with a bridge process whose conditional marginal has the closed form

$$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$

Rather than the classical Brownian-Bridge diffusion objective that combines drift and denoising, the model uses an $\epsilon$-prediction loss,

$$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$

which the paper states is formally equivalent to a proper likelihood-inspired objective with timestep weight $(u,g,r,i,z,y)$0, avoiding vanishing gradients at bridge endpoints (Zhang et al., 12 Mar 2026).

The network is a time-conditional U-Net-style score network similar to Palette’s conditional diffusion UNet. Training uses simulated paired LSST $(u,g,r,i,z,y)$1 and Euclid VIS images generated with SLSim and SkyPy. From an astrophysical catalog over $(u,g,r,i,z,y)$2, the dataset contains 115 000 regular galaxies and 5 000 galaxy–galaxy strong-lens systems rendered into $(u,g,r,i,z,y)$3 cutouts; training uses 110 000 regular galaxies, while the remainder plus all lenses are reserved for evaluation. The reported hyperparameters are $(u,g,r,i,z,y)$4 diffusion steps, linear schedule $(u,g,r,i,z,y)$5, Adam with learning rate $(u,g,r,i,z,y)$6, and batch size $(u,g,r,i,z,y)$7 (Zhang et al., 12 Mar 2026).

Reported results use Continuous Ranked Probability Score for probabilistic reconstruction and FPR@1%TPR, FPR@5%TPR, and AUPR for anomaly detection. On CRPS, the best AS-Bridge variant, “BB $(u,g,r,i,z,y)$8-prediction,” achieves $(u,g,r,i,z,y)$9 for LSST$(Y,J,H)$0Euclid / Euclid$(Y,J,H)$1LSST, compared with Palette at $(Y,J,H)$2, Joint diffusion at $(Y,J,H)$3, SPADE at $(Y,J,H)$4, OASIS at $(Y,J,H)$5, and Pix2Pix at $(Y,J,H)$6 (Zhang et al., 12 Mar 2026). For rare-event detection, AS-Bridge reports FPR@1%TPR $(Y,J,H)$7, FPR@5%TPR $(Y,J,H)$8, and AUPR $(Y,J,H)$9, versus Deco-Diff at $\Phi$0, $\Phi$1, and $\Phi$2, and CFM at $\Phi$3, $\Phi$4, and $\Phi$5 (Zhang et al., 12 Mar 2026).

Scientifically, the paper positions these results as enabling faithful probabilistic predictions of missing survey observations, cross-survey deblending, and anomaly detection through reconstruction inconsistency. A central point is that AS-Bridge is explicitly bidirectional and probabilistic; it is not a deterministic image-to-image translator.

3. AS-Bridge as a bridge between speech enhancement and ASR

In robust speech recognition, AS-Bridge is a post-processing method designed for the common setting in which a speech-enhancement model $\Phi$6 and an automatic-speech-recognition system $\Phi$7 are pre-trained independently (Wang et al., 2024). The motivating observation is that directly feeding enhanced waveforms into ASR can degrade recognition because neural speech enhancement introduces artifacts such as musical noise that the ASR model was never exposed to. The method is therefore intended for cases where jointly retraining or fine-tuning SE+ASR is impossible, including closed-source ASR systems or prohibitive training costs.

The architecture is a four-stage pipeline:

1. $\Phi$8,
2. $\Phi$9,
3. observation addition,

$$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$0

1. $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$1 (Wang et al., 2024).

The bridge operates on magnitude STFT spectrograms $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$2 and $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$3. For each frame $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$4, it computes cosine similarity along the frequency axis,

$$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$5

stacks these similarities into $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$6, and applies a single linear layer,

$$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$7

The output is then clipped into a valid OA coefficient,

$$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$8

with $$x_{\mathrm{LSST}} = O_{\mathrm{LSST}}(\Phi) + \epsilon_{\mathrm{LSST}}, \qquad x_{\mathrm{Euclid}} = O_{\mathrm{Euclid}}(\Phi) + \epsilon_{\mathrm{Euclid}}.$$9 in all experiments (Wang et al., 2024).

Training uses only unlabeled audio: clean speech $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$0 with label $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$1, pure noise $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$2 with label $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$3, and enhanced speech $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$4 from a fixed SE model. The objective is mean-squared error,

$p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$5

No ASR transcripts or ASR-related loss are used. The implementation details reported are: frozen SE models CMGAN, MP-SENet, DEMUCS, and SEMamba; frozen ASR models Whisper-base and Whisper-large-v3; sample rate 16 kHz; Hann-window STFT with 400-sample window, 100-sample hop, and FFT size 512; SGD with learning rate $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$6, momentum $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$7, and batch size $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$8; training on LibriSpeech-train360 plus 80% DNS-challenge noise (Wang et al., 2024).

The quantitative results focus on WER. On LibriSpeech-testclean contaminated by five random DNS noises at SNRs $p(x_{\mathrm{Euclid}} \mid x_{\mathrm{LSST}})$9 dB, Whisper-base yields noisy WER $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$0, while Whisper-base + CMGAN enhanced speech yields $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$1. Adding SNR-level OA reduces WER to $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$2, and adding the proposed Bridge reduces it to $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$3, corresponding to $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$4 relative improvement over noisy input and $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$5 relative improvement over enhanced input (Wang et al., 2024). On Aurora-4, Bridge+CMGAN gives $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$6 for Whisper-base and $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$7 for Whisper-large, compared with noisy WER of $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$8 and $p(x_{\mathrm{LSST}} \mid x_{\mathrm{Euclid}})$9. On VoiceBank-DEMAND, Bridge+SEMamba gives $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$0 for Whisper-base and $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$1 for Whisper-large, compared with noisy WER of $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$2 and $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$3, described as a $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$4 relative WER reduction over noisy input (Wang et al., 2024).

A common misconception is that the method depends on transcript supervision or ASR fine-tuning. The reported formulation explicitly avoids both: the bridge is trained purely on clean, noise, and enhanced unlabeled audio, and the SE and ASR components remain frozen.

4. AS-Bridge in unsupervised speech-to-text representation bridging

A third use of the term appears in spoken language understanding, where AS-Bridge denotes a framework that uses unsupervised ASR as a connector between speech self-supervised models and textual pre-trained models (Shi et al., 2022). The motivation is the modality mismatch between speech signals and text tokens, which earlier work often addressed with more complex designs. The proposed connector is an unsupervised speech-to-semantic pre-trained model for multiple SLU tasks.

The architecture has three components. The speech encoder $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$5 is a pre-trained wav2vec 2.0 model trained on LibriLight, producing frame-level features $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$6. On top of $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$7, a simplified wav2vec-U 2.0 acts as an unsupervised ASR generator $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$8 with discriminator $$x_t \mid (x_0,x_T) \sim \mathcal N\bigl((1-m_t)x_0 + m_t x_T,\; \delta_t I\bigr), \qquad \delta_t = m_t(1-m_t).$$9, producing pseudo-phoneme distributions $\epsilon$0 without paired speech-text data. The textual encoder is a phoneme-level T5, PT5, pre-trained on phonemized text. It consumes discrete phoneme sequences obtained from $\epsilon$1 and emits contextual embeddings (Shi et al., 2022).

The final multimodal representation is formed by simple upsampling and concatenation rather than a dedicated learned projection network: $\epsilon$2 where $\epsilon$3 and $\epsilon$4. The paper explicitly states that no separate projection network is required; alignment is achieved simply by upsampling and concatenating streams (Shi et al., 2022).

The unsupervised ASR objective is a GAN-style adversarial loss with three regularizers: $\epsilon$5 where $\epsilon$6 is gradient penalty, $\epsilon$7 enforces segment smoothness, and $\epsilon$8 encourages phoneme-type diversity. The reported data for UASR pre-training are Librispeech 960 h as $\epsilon$9 and Librispeech LM text of approximately 800 M words as $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$0 (Shi et al., 2022).

Results are reported in both a freeze-only setting and a fine-tuning setting. In the freeze setting, wav2vec2, $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$1, $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$2, and PT5 are frozen; only a small task-specific head is trained. Relative to wav2vec2 alone, adding UASR improves phone recognition from $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$3 to $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$4 PER, while the full $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$5UASR $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$6PT5 system gives $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$7 PER, $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$8 ASR error, $$\mathcal L_\epsilon = E\bigl[\|\epsilon - \epsilon_\theta(x_t,x_T,t)\|^2\bigr],$$9 intent classification accuracy, $(u,g,r,i,z,y)$00 BLEU for speech translation, and $(u,g,r,i,z,y)$01 speaker identification (Shi et al., 2022). In the fine-tuning setting, the full AS-Bridge configuration, “UASR + PT5,” reports $(u,g,r,i,z,y)$02 intent classification, $(u,g,r,i,z,y)$03 slot filling, $(u,g,r,i,z,y)$04 emotion recognition, and $(u,g,r,i,z,y)$05 BLEU for speech translation, outperforming no connector, UASR without a text model, and a K-means connector with PT5 on most listed tasks (Shi et al., 2022).

On spoken question answering over NMSQA, the paper reports that “+UASR + PT5 (AS-Bridge)” reaches AOS/FF1 of $(u,g,r,i,z,y)$06 on SQuAD and $(u,g,r,i,z,y)$07 on out-of-domain NEWSQA+QuAC, compared with prior SOTA Longformer at $(u,g,r,i,z,y)$08 on SQuAD (Shi et al., 2022). The paper’s ablation analysis further states that unsupervised ASR alone improves phonetic representations but gives minimal or slightly negative effects on downstream semantics, whereas the full AS-Bridge pipeline gives the strongest SLU performance.

Here again, the bridge is not supervised ASR. Its connector is explicitly unsupervised and is used to reconcile modalities so that a textual model can slot into a speech pipeline.

5. Methodological commonalities and relation to adjacent bridge-based work

Despite their domain differences, the three AS-Bridge formulations share several structural commitments. First, each introduces an intermediate object that is narrower than full joint retraining but richer than direct feature passing: a stochastic Brownian-Bridge path in astronomy, a scalar observation-addition coefficient derived from spectral similarity in robust ASR, and a pseudo-phoneme sequence plus text-encoder stream in SLU (Zhang et al., 12 Mar 2026, Wang et al., 2024, Shi et al., 2022). Second, each bridge is designed around a specific mismatch: survey modality mismatch, SE–ASR artifact mismatch, or speech–text modality mismatch. Third, each formulation treats uncertainty or ambiguity as intrinsic rather than incidental: probabilistic image translation in astronomy, partial retention of noisy observations in OA-based ASR, and discrete intermediate phonemic structure in speech-text transfer.

Adjacent arXiv work broadens this perspective. “Efficient Generative Modeling beyond Memoryless Diffusion via Adjoint Schrödinger Bridge Matching” frames generative modeling as optimal coupling construction followed by backward bridge matching under a non-memoryless regime, yielding straighter and more efficient sampling paths (Shin et al., 17 Feb 2026). “Semantic Bridging Domains: Pseudo-Source as Test-Time Connector” treats a pseudo-source as a semantic bridge connecting source and target, rectifies the pseudo-source with universal semantics, and then aligns the target domain via stepwise semantic alignment (Yang et al., 4 Mar 2026). This suggests that “bridge” is increasingly used as a general methodological motif for dealing with domain shift, imperfect coupling, or asymmetric observability.

The differences are equally important. The astronomy model is generative and bidirectional. The robust ASR model is a lightweight post-processing module that preserves frozen black-box components. The SLU model is representational and relies on unsupervised symbolic mediation. Accordingly, “AS-Bridge” should not be read as a single transferable recipe; it is a family resemblance centered on intermediate alignment mechanisms.

6. Limitations, interpretive cautions, and outlook

Each AS-Bridge variant comes with domain-specific limitations. In robust ASR, the bridge only uses spectral cosine similarity, so other artifact types such as phase distortion are not explicitly modeled; the clipping floor $(u,g,r,i,z,y)$09 is manually chosen; and the approach still relies on fixed SE+ASR black boxes, even though the paper notes that a small amount of ASR-supervised fine-tuning could help further (Wang et al., 2024). In the SLU setting, the paper does not present the connector as universally beneficial in isolation: its own ablation states that unsupervised ASR alone gives clear gains in phone error rate but minimal or slightly negative effects on downstream semantics (Shi et al., 2022). In astronomy, the framework is trained on realistic simulations rather than public real LSST and Euclid releases, and the paper identifies transfer learning to real data as a key next step (Zhang et al., 12 Mar 2026).

Several recurrent misconceptions can therefore be rejected. AS-Bridge is not a synonym for end-to-end fusion; in the robust ASR formulation, the SE and ASR modules are frozen (Wang et al., 2024). It is not necessarily deterministic; in the astronomy formulation, the mapping is explicitly stochastic because both surveys are noisy partial projections of an underlying $(u,g,r,i,z,y)$10 (Zhang et al., 12 Mar 2026). It is not equivalent to supervised speech recognition; in the SLU formulation, the connector is unsupervised ASR trained without paired speech-text data (Shi et al., 2022).

The outlook differs by field but follows a common trajectory. The astronomy paper proposes extensions to other modalities such as the Roman Space Telescope and multi-epoch time series (Zhang et al., 12 Mar 2026). The robust ASR paper proposes richer bridge features, a frame-wise $(u,g,r,i,z,y)$11 rather than a single global scalar, and application to downstream tasks such as speaker verification and emotion recognition (Wang et al., 2024). The SLU paper shows that a simple connector can suffice to bridge speech and textual pre-trained models across five SLU tasks and spoken QA, suggesting further work on modality mediation without paired supervision (Shi et al., 2022).

Taken together, AS-Bridge names a set of architectures built around the same operational insight: when two systems, modalities, or domains are mismatched, a deliberately constructed bridge can be more effective than either naive interfacing or wholesale retraining.