ASTREA: Spacecraft Thermal Autonomy

• ASTREA is a hybrid agentic system for spacecraft thermal control that combines a resource-constrained large language model with a Soft Actor-Critic controller.
• It is often confused with similar-sounding methods in exoplanet spectroscopy, stellar photometry, quantum error correction, and vision-language modeling, necessitating careful disambiguation.
• Empirical results show improved ground thermal stability though latency in orbital environments emphasizes the need for decoupled timelines and explicit latency budgeting.

Searching arXiv for "ASTREA" and closely related variants to ground the article in the cited papers. ASTREA most precisely denotes Agentic System for Thermal Regulation and Embedded Adaptation, a hybrid agentic autonomy system for spacecraft thermal control that combines a resource-constrained LLM with a Soft Actor-Critic controller on flight-heritage hardware (Mousist, 16 Sep 2025). In the arXiv literature, however, the same string and close orthographic variants—AESTRA, Astraea, and Astrea—also label unrelated methods in extreme-precision radial velocities, stellar rotation-period inference, real-time quantum error correction, and mixture-of-experts vision-language modeling (Liang et al., 2023, Liang et al., 11 Jun 2026, Yuxi et al., 2020, Alavisamani et al., 2024, Yang et al., 12 Mar 2025). The term therefore has a strong disambiguation component: in some contexts it is a specific acronym, whereas in others it is a naming confusion or a distinct proper name.

1. Nomenclature and disambiguation

A common source of confusion is that ASTREA is not a single research program spanning multiple domains. In the spacecraft-autonomy literature, ASTREA is explicitly expanded as Agentic System for Thermal Regulation and Embedded Adaptation (Mousist, 16 Sep 2025). By contrast, the radial-velocity method is AESTRA, expanded as Auto-Encoding STellar Radial-velocity and Activity, and the associated paper explicitly states that “ASTREA” is a naming confusion or typo (Liang et al., 2023). The 2026 solar extension likewise notes that Æstra/AESTRA is “sometimes referred to as ASTREA in the literature,” while still treating them as the same generative spectral modeling framework (Liang et al., 11 Jun 2026). The 2020 stellar-rotation tool is Astraea, “often misspelled ‘ASTREA’” (Yuxi et al., 2020). In quantum decoding and multimodal modeling, Astrea is a separate proper name, not a typo (Alavisamani et al., 2024, Yang et al., 12 Mar 2025).

Name in literature	Domain	Core role
ASTREA	Spacecraft autonomy	Agentic thermal control with LLM supervision and SAC control
AESTRA / Æstra	Exoplanet spectroscopy	Spectrum-level RV estimation and activity disentangling
Astraea	Stellar photometry	Long rotation-period prediction from short light curves
Astrea / Astrea-G	Quantum error correction	Real-time MWPM decoding framework
Astrea	Vision-language modeling	MoE VLM with progressive pre-alignment

This multiplicity matters methodologically. A reader encountering “ASTREA” in exoplanet spectroscopy is usually being pointed to AESTRA, whereas in onboard autonomy it denotes a genuinely different acronym and architecture. This suggests that citation by arXiv id is essential whenever the name appears in technical discussion.

2. ASTREA as orbital thermal autonomy

ASTREA is a hybrid agentic system designed for orbital thermal autonomy on space-qualified platforms under tight compute and power constraints. Its architecture is explicitly asynchronous. A real-time RL-Agent based on Soft Actor-Critic (SAC) continuously monitors thermal state and adjusts CPU core frequencies and power states to maximize utilization while avoiding threshold violations, while an LLM-Agent periodically analyzes episode-level summaries and recommends updates to the SAC entropy coefficient $\alpha$ in order to modulate exploration and exploitation (Mousist, 16 Sep 2025).

The system is deployed on a flight-qualified 64-bit ARM SoC with 16 Cortex-A72 cores, dynamic frequency 1.0–2.0 GHz, and 16 GB LPDDR4 memory, with no GPU/NPU. Core 0 is reserved for orchestration and agentic services, and the remaining 15 cores are controlled under load. The LLM is a Qwen2.5-1.54B variant with 4-bit quantization, inferred using Llama.cpp and orchestrated with the OpenAI Agents SDK via an OpenAI-compatible local server. The supervisory loop uses a 60-minute window on ground and a 15-minute window on the ISS, while measured LLM inference latency is approximately 40 seconds to over 8 minutes per prompt, with preliminary instances around 10 minutes. The thermal safety bound is 60°C, and prompt logic treats steps $\ge 59^\circ\mathrm{C}$ as near-limit behavior (Mousist, 16 Sep 2025).

At the control-interface level, the RL agent pushes three episode-summary quantities into an asynchronous queue: total iterations, count of iterations where peak temperature from any sensor is within $1^\circ\mathrm{C}$ of the critical threshold, and average thermal gradient during the episode. The LLM aggregates these summaries over the supervisory window and returns a tool-selected recommendation among Increase exploration ($\alpha \in [0.4,0.8]$), Moderate exploration ($\alpha \in [0.2,0.4]$), Decrease exploration ($\alpha \in [0.05,0.2]$), Keep alpha, or Reset alpha ($\alpha=0.2$). The LLM remains outside the real-time loop; it neither vetoes actions nor performs reward shaping in the reported study (Mousist, 16 Sep 2025).

Empirically, ASTREA improves ground performance but degrades certain on-orbit metrics because of delay mismatch. In the first 4 hours of ground experiments, average episode duration increased by 67.2%, from 47.17 ± 18.15 steps to 78.83 ± 11.33, while thermal violations fell by 58.5%, from 39.33 ± 9.29 to 16.33 ± 2.08. Over 24 hours, average episode duration was 135.24 ± 32.94 for baseline and 142.29 ± 8.06 for agentic control, while thermal violations fell from 88.67 ± 20.50 to 51.33 ± 4.04; average CPU usage changed minimally. On the ISS, average episode duration became shorter under agentic supervision—224.26 ± 163.81 versus 146.60 ± 71.04—although thermal violations still decreased from 204 ± 201.72 to 184.33 ± 69.47, and CPU usage rose slightly from 18.84 ± 13.01\% to 19.72 ± 17.43\% (Mousist, 16 Sep 2025).

The central lesson is architectural rather than merely empirical. ASTREA shows that LLM-guided supervision can improve thermal stability when the supervisory timescale is slower than the dominant plant dynamics, but the same design can become ill-timed in Low Earth Orbit when inference latency approaches the timescale of sun/eclipse thermal transitions. The paper therefore advocates decoupled timelines, queue-based communication, explicit latency budgeting, and preservation of RL autonomy within a hard safety envelope (Mousist, 16 Sep 2025).

3. AESTRA in precise radial-velocity estimation

In exoplanet spectroscopy, the relevant system is AESTRA, not ASTREA. AESTRA is a self-supervised deep-learning framework for precise radial velocity measurement in the presence of stellar activity. It addresses the fact that spots, plages, granulation, and oscillations reshape spectral line profiles and induce apparent radial-velocity “jitter” at the meter-per-second level, masking the centimeter-per-second Doppler signatures sought in Earth-analog searches (Liang et al., 2023).

The model combines two branches. The first is a spectrum autoencoder that learns a rest-frame stellar spectrum plus low-dimensional activity structure. Given observed residuals $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$, the encoder $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$ uses three 1D convolutional layers, an attention layer, and a three-layer MLP; the paper sets $S=3$. The decoder $\ge 59^\circ\mathrm{C}$0 is a three-layer MLP, and a trainable rest-frame template $\ge 59^\circ\mathrm{C}$1 yields

$\ge 59^\circ\mathrm{C}$2

The second branch is an RV estimator $\ge 59^\circ\mathrm{C}$3 composed of two 1D convolutional layers with PReLU nonlinearities, max pooling, a 64 × 20 feature arrangement with a softmax along the wavelength axis, flattening to 1280, and an MLP with layers $\ge 59^\circ\mathrm{C}$4 ending in a scalar velocity estimate (Liang et al., 2023).

The interface between branches is a Doppler shift operator implemented by interpolation. Using the small-velocity relation $\ge 59^\circ\mathrm{C}$5 and the forward mapping $\ge 59^\circ\mathrm{C}$6, AESTRA evaluates

$\ge 59^\circ\mathrm{C}$7

with cubic spline interpolation to align the rest-frame reconstruction to observed wavelengths. The loss is

$\ge 59^\circ\mathrm{C}$8

where $\ge 59^\circ\mathrm{C}$9 is weighted spectral reconstruction error, $1^\circ\mathrm{C}$0 enforces recovery of injected augmentation offsets $1^\circ\mathrm{C}$1, and $1^\circ\mathrm{C}$2 is a Ridge penalty on the decoded activity spectrum with $1^\circ\mathrm{C}$3 and $1^\circ\mathrm{C}$4 cycled $1^\circ\mathrm{C}$5 over 1000 iterations and then reset. Training is two-phase: first pretrain $1^\circ\mathrm{C}$6 until $1^\circ\mathrm{C}$7, then jointly train all modules (Liang et al., 2023).

AESTRA’s strategy for suppressing activity-induced spurious shifts has two parts. First, augmentation-based self-supervision forces the RV estimator to respond to global wavelength stretch rather than line-shape changes. Second, after training, it removes an activity-dependent zero-point bias in latent space by Gaussian smoothing:

$1^\circ\mathrm{C}$8

and defines corrected velocities

$1^\circ\mathrm{C}$9

Here $\alpha \in [0.4,0.8]$0 is set to the average distance to the 10 nearest neighbors, and in practice the smoothing uses the closest 5\% of samples with a minimum of 10 neighbors (Liang et al., 2023).

The reported simulated performance is strong. In a case with 1000 spectra, 3.38 m s$\alpha \in [0.4,0.8]$1 apparent scatter, activity-induced noise at 3 m s$\alpha \in [0.4,0.8]$2, photon noise near 0.3 m s$\alpha \in [0.4,0.8]$3, and a planetary signal of $\alpha \in [0.4,0.8]$4 m s$\alpha \in [0.4,0.8]$5, traditional detrending based on CCF bisector span, FWHM, and depth regression yielded 0.98 m s$\alpha \in [0.4,0.8]$6 residual scatter, whereas AESTRA $\alpha \in [0.4,0.8]$7 yielded 0.46 m s$\alpha \in [0.4,0.8]$8, close to the “best possible” reference 0.39 m s$\alpha \in [0.4,0.8]$9. A circular Keplerian recovery with 32 walkers, 10k steps, and 2k burn-in recovered $\alpha \in [0.2,0.4]$0 days and $\alpha \in [0.2,0.4]$1 m s$\alpha \in [0.2,0.4]$2. Across the paper’s cases, AESTRA detects signals near 0.1 m s$\alpha \in [0.2,0.4]$3 under 3 m s$\alpha \in [0.2,0.4]$4 activity jitter and 0.3 m s$\alpha \in [0.2,0.4]$5 photon noise per spectrum (Liang et al., 2023).

The main caveat is that the method was validated on simulations, not yet on real stellar spectra in that original study. The paper also notes sample-size sensitivity, calibration sensitivity, possible degeneracy between $\alpha \in [0.2,0.4]$6 and $\alpha \in [0.2,0.4]$7, and slight underestimation of $\alpha \in [0.2,0.4]$8 when $\alpha \in [0.2,0.4]$9 m s$\alpha \in [0.05,0.2]$0 because planetary variation can be treated as “noise” during latent smoothing (Liang et al., 2023).

4. AESTRA II and generative spectral modeling of the Sun

AESTRA II extends the AESTRA program from simulated spectra to real NEID Sun-as-a-star observations by replacing the earlier spectrum-rest-frame reconstruction with a broader generative decomposition of tellurics, continuum variability, and stellar line-shape variability (Liang et al., 11 Jun 2026). This continuation is important because it makes explicit that the AESTRA line of work is not only an RV estimator, but a spectrum-level generative framework for extreme-precision radial velocities.

The data pipeline begins with publicly available NEID Solar Feed Level-1/2 products. After quality cuts removing high airmass $\alpha \in [0.05,0.2]$1, cloud passages with irradiance drops $\alpha \in [0.05,0.2]$2, low-irradiance or low-count days, high intraday RV RMS $\alpha \in [0.05,0.2]$3 m s$\alpha \in [0.05,0.2]$4, and documented instrument anomalies, 72,449 spectra remained; approximately 30,000 exposures spanning 1294 days and 521 observing days were then used. The analysis restricted the spectra to 4300–6230 Å, excluded two orders around strong tellurics near 5875–5990 Å, retained 42 orders, and merged them into cleaned spectra covering 4307.5–6231.7 Å with 170,923 wavelength bins. A local “glitch” variance estimated in six-pixel windows was added to the pipeline variances, yielding a heteroskedastic Gaussian noise model (Liang et al., 11 Jun 2026).

The per-exposure generative model is

$\alpha \in [0.05,0.2]$5

or equivalently $\alpha \in [0.05,0.2]$6. Here $\alpha \in [0.05,0.2]$7 is Earth-frame telluric absorption, $\alpha \in [0.05,0.2]$8 is smooth continuum modulation, $\alpha \in [0.05,0.2]$9 is a time-independent empirical stellar template, and $\alpha=0.2$0 is time-variable stellar line-shape distortion. The telluric branch uses one learned eigenspectrum with an exposure-dependent depth and a trainable LSF-like kernel. The continuum branch uses five eigenspectra convolved by a fixed $\alpha=0.2$1-pixel Gaussian smoothing kernel. The stellar branch uses three eigenspectra, high-pass filtered and penalized in continuum bins to enforce line-localized variability (Liang et al., 11 Jun 2026).

Training minimizes a weighted reconstruction loss plus morphology regularizers:

$\alpha=0.2$2

with $\alpha=0.2$3, $\alpha=0.2$4, and $\alpha=0.2$5. The learned telluric component reproduces narrow Earth-frame features with line depths correlated at $\alpha=0.2$6 with the NEID precipitable-water-vapor diagnostic, indicating that the model isolates real atmospheric variability without external templates (Liang et al., 11 Jun 2026).

After removing only the learned telluric and continuum components, AESTRA II measures an apparent bulk Doppler shift $\alpha=0.2$7 by template fitting to the average spectrum, shifts each spectrum by $\alpha=0.2$8, and trains an attentive autoencoder with a 20-dimensional latent space to encode residual stellar variability. The first principal component of the latent representation correlates strongly and monotonically with the NEID Ca II H&K index at $\alpha=0.2$9, despite the Ca H&K region being excluded from training. The retained apparent RV series is then decomposed into activity and planet components using an activity estimator $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$0, where $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$1 is a two-layer MLP with hidden sizes 128 and 8, and a circular-orbit planet model

$$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$2

Candidate periods and phases are initialized from an RV-level Lomb–Scargle search, and amplitudes are optimized in four period blocks: $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$3 d, $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$4–$$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$5 d, $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$6–$$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$7 d, and $$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$8 d (Liang et al., 11 Jun 2026).

Detection uses

$$r_{\mathrm{obs}} = y_{\mathrm{obs}} - b_{\mathrm{obs}}$$9

with $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$0 and threshold $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$1, calibrated on 500 single-planet injections to yield zero spurious detections. Over periods 2.5–400 days and semi-amplitudes $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$2 m s$f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$3, AESTRA recovered 238 of 500 injected planets, including 13 with $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$4 m s$f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$5. The traditional CCF-indicator baseline recovered 9 planets and none below $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$6 m s$f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$7. Recovered AESTRA amplitudes had median relative error $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$8, with 84th percentile 27.1\% and 95th percentile 43.1\% (Liang et al., 11 Jun 2026).

The result is not merely an incremental gain over the 2023 formulation. It shows that the AESTRA program can be reinterpreted as a generative latent-variable model for spectrum-level disentangling of stellar activity, micro-tellurics, and continuum systematics, with a detection threshold calibrated directly by injection–recovery at fixed false-positive rate (Liang et al., 11 Jun 2026).

5. Astraea for stellar rotation-period inference

Astraea is an open-source Random Forest pipeline for predicting long stellar rotation periods from short-duration light curves, especially the $f_\theta(r_{\mathrm{obs}})\to s\in\mathbb{R}^S$9-day segments typical of TESS (Yuxi et al., 2020). The motivation is directly connected to radial-velocity science: stellar rotation informs activity modeling, RV mitigation, and high-energy irradiation history, yet many stars—particularly field M dwarfs—rotate more slowly than the TESS single-sector baseline.

The pipeline has two stages. A classifier predicts whether a star’s rotation period is “measurable,” meaning consistent with detectable photometric modulation as defined by the McQuillan (2014) catalog. A regressor then predicts the rotation period for stars classified as measurable. Inputs include light-curve summary statistics—$S=3$0, 8-hour flicker, and maximum Lomb–Scargle peak height—together with Gaia DR2 quantities and derived kinematics, including $S=3$1, effective temperature, radius, luminosity, parallax, photometric flux-over-error, distance intervals, galactic latitude, $S=3$2, and $S=3$3. The classifier uses LG_peaks, $S=3$4, $S=3$5, phot_g_mean_flux_over_error, $S=3$6, $S=3$7, parallax, while the regressor uses $S=3$8, $S=3$9, $\ge 59^\circ\mathrm{C}$00, lum_val, flicker, radius_val, $\ge 59^\circ\mathrm{C}$01, phot_g_mean_flux_over_error, $\ge 59^\circ\mathrm{C}$02, $\ge 59^\circ\mathrm{C}$03 (Yuxi et al., 2020).

The Kepler training data comprise 133,030 main-sequence targets from McQuillan (2014), with measured periods for 34,030 stars over 0.2–70 days, augmented by 4,637 stars from Santos (2019) and García (2014) with periods up to $\ge 59^\circ\mathrm{C}$04 days, including 70 stars with $\ge 59^\circ\mathrm{C}$05 days. Full four-year Kepler light curves were split into roughly 54 segments of 27 days, producing $\ge 59^\circ\mathrm{C}$06 million training instances from about 34,000 stars. The classifier threshold was selected by ROC analysis to maximize TPR − FPR, achieving 98\% accuracy at threshold 0.4. Hyperparameter search for the regressor considered $\ge 59^\circ\mathrm{C}$07, max_depth $\ge 59^\circ\mathrm{C}$08, and max_features $\ge 59^\circ\mathrm{C}$09, with the chosen setting $\ge 59^\circ\mathrm{C}$10, max_depth $\ge 59^\circ\mathrm{C}$11, and max_features $\ge 59^\circ\mathrm{C}$12 because it minimized average $\ge 59^\circ\mathrm{C}$13 (Yuxi et al., 2020).

Performance depends strongly on domain. On Kepler 4-year light curves, Astraea predicts rotation periods with $\ge 59^\circ\mathrm{C}$14 uncertainty overall and $\ge 59^\circ\mathrm{C}$15 for periods $\ge 59^\circ\mathrm{C}$16 days. When trained on 27-day Kepler segments, it predicts periods up to $\ge 59^\circ\mathrm{C}$17 days with $\ge 59^\circ\mathrm{C}$18 uncertainty overall and $\ge 59^\circ\mathrm{C}$19 for periods $\ge 59^\circ\mathrm{C}$20 days. On real TESS data, however, performance degrades: for the 195 stars observed by both Kepler and TESS, the model predicts periods with $\ge 59^\circ\mathrm{C}$21 uncertainty. The paper attributes this to bandpass differences, higher white-noise floor, larger pixels, contamination, and field-dependent kinematic shifts (Yuxi et al., 2020).

This establishes Astraea as a supervised, tabular-plus-time-series alternative to direct period detection. Traditional methods such as autocorrelation, Lomb–Scargle, and Gaussian-process modeling generally require at least one full rotation cycle; Astraea instead learns a nonlinear mapping from short-segment variability proxies and stellar parameters to $\ge 59^\circ\mathrm{C}$22. The principal limitations are that Random Forests do not extrapolate beyond the training domain, fast rotators may include synchronized binaries, and low-S/N short segments—especially in TESS—degrade the reliability of $\ge 59^\circ\mathrm{C}$23 and flicker (Yuxi et al., 2020).

6. Astrea and Astrea-G in real-time quantum decoding

In quantum error correction, Astrea and Astrea-G denote a real-time decoding framework for the surface code, described in the context of the Promatch predecoder (Alavisamani et al., 2024). Here the problem is not spectroscopy or autonomy, but hardware decoding of parity-bit syndromes within a strict $\ge 59^\circ\mathrm{C}$24 cycle-time budget.

Astrea is a hardware real-time Minimum-Weight Perfect Matching (RT-MWPM) decoder that performs exact MWPM for low-Hamming-weight syndromes by brute-force search over all perfect matchings up to HW $\ge 59^\circ\mathrm{C}$25; the paper notes 945 matchings at HW = 10. It achieves RT-MWPM up to $\ge 59^\circ\mathrm{C}$26 in approximately 456 ns, but the brute-force approach does not scale beyond about $\ge 59^\circ\mathrm{C}$27 because the number of candidate matchings grows exponentially with Hamming weight and code distance. Astrea-G is the greedy extension that reaches $\ge 59^\circ\mathrm{C}$28 in $\ge 59^\circ\mathrm{C}$29 by pruning the MWPM graph below a probability threshold and then performing greedy near-exhaustive search, but its accuracy deteriorates at larger distances; at $\ge 59^\circ\mathrm{C}$30 its logical error rate is 43× higher than MWPM (Alavisamani et al., 2024).

The Promatch paper introduces an adaptive predecoder that reduces high-HW syndromes before dispatching them to Astrea. The Hamming weight is

$\ge 59^\circ\mathrm{C}$31

and the logical error rate is estimated as

$\ge 59^\circ\mathrm{C}$32

Promatch operates locally and greedily, prioritizing matches that do not create new singletons, because singleton creation forces long error chains with probability approximately $\ge 59^\circ\mathrm{C}$33 for chain length $\ge 59^\circ\mathrm{C}$34. Its prioritized matching steps are: isolated pairs of degree-1 nodes; neighboring flipped bits that avoid singleton creation; singleton-to-flipped-bit shortest paths when needed; and, as a last resort, risky neighbor matches that may create singletons. The decoder stops prematching once the remaining syndrome falls within Astrea’s exact RT capability (Alavisamani et al., 2024).

The hardware implementation uses a Kintex UltraScale+ FPGA at 250 MHz. The edge-processing pipeline uses approximately 3\% LUTs and 1\% FFs. Edge tables occupy 3.6 KB at $\ge 59^\circ\mathrm{C}$35 and 6 KB at $\ge 59^\circ\mathrm{C}$36; path tables occupy 129 KB and 345 KB, respectively. For high-HW syndromes (HW $\ge 59^\circ\mathrm{C}$37), predecoding latency peaks at 824 ns for $\ge 59^\circ\mathrm{C}$38 and 928 ns for $\ge 59^\circ\mathrm{C}$39, while end-to-end Promatch + Astrea latency peaks at 904 ns and 960 ns, with average about 524 ns. The probability of exceeding $\ge 59^\circ\mathrm{C}$40 at $\ge 59^\circ\mathrm{C}$41 is reported as $\ge 59^\circ\mathrm{C}$42 (Alavisamani et al., 2024).

At physical error rate $\ge 59^\circ\mathrm{C}$43, ideal MWPM achieves LER $\ge 59^\circ\mathrm{C}$44 at $\ge 59^\circ\mathrm{C}$45 and $\ge 59^\circ\mathrm{C}$46 at $\ge 59^\circ\mathrm{C}$47. Astrea-G alone yields $\ge 59^\circ\mathrm{C}$48 at $\ge 59^\circ\mathrm{C}$49 and $\ge 59^\circ\mathrm{C}$50 at $\ge 59^\circ\mathrm{C}$51. Promatch + Astrea yields $\ge 59^\circ\mathrm{C}$52 at $\ge 59^\circ\mathrm{C}$53 and $\ge 59^\circ\mathrm{C}$54 at $\ge 59^\circ\mathrm{C}$55. The parallel design Promatch || Astrea-G reaches MWPM-equivalent LER at $\ge 59^\circ\mathrm{C}$56, namely $\ge 59^\circ\mathrm{C}$57 (Alavisamani et al., 2024).

The quantum-decoding meaning of Astrea is thus highly specific: it is a real-time MWPM architecture whose exactness is bounded by syndrome Hamming weight, and whose extension to larger code distances depends on predecoding and parallel greedy search. It is unrelated to the spacecraft ASTREA acronym despite the near-identical spelling.

7. Astrea as a mixture-of-experts vision-LLM

In multimodal machine learning, Astrea is a Mixture-of-Experts (MoE) vision-LLM designed to address task heterogeneity and expert load imbalance by combining four specialized visual experts—detection, segmentation, classification, and captioning—within a shared latent space (Yang et al., 12 Mar 2025). The paper frames the problem as one of representational conflict: localization and segmentation require geometric precision, while classification and captioning require global semantic alignment.

The model integrates Grounding-DINO-1.5 for localization, SAM2 for segmentation, RAM++ for classification, and SigLIP ViT-SO400M/14@384 for captioning or global semantics. Their outputs are projected to a common hidden size $\ge 59^\circ\mathrm{C}$58 and fused by sparse top-$\ge 59^\circ\mathrm{C}$59 routing. If the router produces scores $\ge 59^\circ\mathrm{C}$60 over experts and each expert applies an adapter $\ge 59^\circ\mathrm{C}$61, the mixture is

$\ge 59^\circ\mathrm{C}$62

With top-$\ge 59^\circ\mathrm{C}$63 gating, the selected-expert distribution is

$\ge 59^\circ\mathrm{C}$64

and non-selected experts are zeroed. The main backbones are Vicuna-1.5-13B and Hermes2-Yi-34B (Yang et al., 12 Mar 2025).

Astrea’s distinctive mechanism is progressive pre-alignment. The experts are aligned into the shared VLM space in a coarse-to-fine sequence: Caption, then Classification, then Detection, then Segmentation. Each stage is trained and then frozen, while a feature cache preserves prior representations for later stages. Alignment is enforced with an InfoNCE-style contrastive objective

$\ge 59^\circ\mathrm{C}$65

and knowledge continuity is maintained by probabilistically activated stochastic residual connections, expressible as

$\ge 59^\circ\mathrm{C}$66

The full model combines generative and momentum-contrastive objectives, with

$\ge 59^\circ\mathrm{C}$67

Training uses LLAVA-Next multi-resolution visual tokenization with 576 visual tokens per image, AdamW with $\ge 59^\circ\mathrm{C}$68, $\ge 59^\circ\mathrm{C}$69, weight decay 0.05, peak learning rate $\ge 59^\circ\mathrm{C}$70, and 15\% linear warm-up, on 48×8 A800 GPUs (Yang et al., 12 Mar 2025).

The empirical claim is broad benchmark improvement. Across 12 image benchmarks and multiple video suites, Astrea reports an average +4.7\% gain over state of the art. Representative Astrea-34B scores include MMB 83.6, SEED 76.9, GQA 66.7, SQA$\ge 59^\circ\mathrm{C}$71 88.2, MMMU$\ge 59^\circ\mathrm{C}$72 51.5, MathVista$\ge 59^\circ\mathrm{C}$73 55.4, ChartQA 78.2, OCRBench 63.0, and DocVQA 80.6. For video, Astrea-34B attains EgoSchema 66.5, Perception-Test 60.5, MVBench 64.2, and VideoMME 62.5/64.8. Ablations show that top-$\ge 59^\circ\mathrm{C}$74 is the best trade-off, and that removing either residual connections or contrastive learning reduces metrics (Yang et al., 12 Mar 2025).

Its limitations are structural rather than incidental. Performance depends on the quality of external specialist models; it is sensitive to curriculum order and hyperparameters; it relies on sparsity and calibration rather than an explicit load-balancing regularizer; and further scaling to additional experts such as OCR-specific or 3D specialists would require more elaborate routing and pre-alignment. This suggests that “Astrea” in the VLM literature denotes not just a model instance, but a particular design philosophy for heterogeneous-expert harmonization (Yang et al., 12 Mar 2025).