---
title: 'ASTREA: Spacecraft Thermal Autonomy'
url: https://www.emergentmind.com/topics/astrea
type: topic
---

# ASTREA: Spacecraft Thermal Autonomy

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 large language model with a Soft Actor-Critic controller on flight-heritage hardware [2509.13380]. 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 [2311.18326] [2606.13574] [2008.03351] [2404.03136] [2503.09445]. 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** [2509.13380]. 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 [2311.18326]. 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 [2606.13574]. The 2020 stellar-rotation tool is **Astraea**, “often misspelled ‘ASTREA’” [2008.03351]. In quantum decoding and multimodal modeling, **Astrea** is a separate proper name, not a typo [2404.03136] [2503.09445].

| 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 [2509.13380].

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 [2509.13380].

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 [2509.13380].

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\%** [2509.13380].

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 [2509.13380].

## 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 [2311.18326].

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 $g_\phi(s)\to y_{\mathrm{act}}$ is a **three-layer MLP**, and a trainable rest-frame template $b_{\mathrm{rest}}$ yields
$$
y_{\mathrm{rest}} = y_{\mathrm{act}} + b_{\mathrm{rest}}.
$$
The second branch is an **RV estimator** $h_\psi(r_{\mathrm{obs}})\to v_{\mathrm{encode}}$ 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 $(128,64,32)$** ending in a scalar velocity estimate [2311.18326].

The interface between branches is a Doppler shift operator implemented by interpolation. Using the small-velocity relation $v \approx c(\Delta\lambda/\lambda)$ and the forward mapping $\lambda_{\mathrm{obs}}=\lambda_{\mathrm{rest}}(1+v/c)$, AESTRA evaluates
$$
T_v[S](\lambda_{\mathrm{obs}})=S\!\left(\frac{\lambda_{\mathrm{obs}}}{1+v/c}\right),
$$
with **cubic spline interpolation** to align the rest-frame reconstruction to observed wavelengths. The loss is
$$
L_{\mathrm{total}} = L_{\mathrm{fid}} + L_{\mathrm{RV}} + L_{\mathrm{reg}},
$$
where $L_{\mathrm{fid}}$ is weighted spectral reconstruction error, $L_{\mathrm{RV}}$ enforces recovery of injected augmentation offsets $v_{\mathrm{offset}}\sim U(-3,3)\,\mathrm{m\,s^{-1}}$, and $L_{\mathrm{reg}}$ is a Ridge penalty on the decoded activity spectrum with **$\sigma_y=0.1$** and **$k_{\mathrm{reg}}$ cycled $0\to1$ over 1000 iterations and then reset**. Training is two-phase: first pretrain $h_\psi$ until **$L_{\mathrm{RV}}\lesssim 1$**, then jointly train all modules [2311.18326].

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:
$$
\langle v\rangle_i=\frac{\sum_{j\neq i}w_{ij}v_{\mathrm{encode},j}}{\sum_{j\neq i}w_{ij}},
\qquad
w_{ij}=\exp\!\left(-\frac{\|s_i-s_j\|^2}{2\sigma_R^2}\right),
$$
and defines corrected velocities
$$
v_{\mathrm{correct},i}=v_{\mathrm{encode},i}-\langle v\rangle_i.
$$
Here $\sigma_R$ 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** [2311.18326].

The reported simulated performance is strong. In a case with **1000 spectra**, **3.38 m s$^{-1}$** apparent scatter, **activity-induced noise at 3 m s$^{-1}$**, **photon noise near 0.3 m s$^{-1}$**, and a planetary signal of **$K=0.10$ m s$^{-1}$**, traditional detrending based on **CCF bisector span, FWHM, and depth regression** yielded **0.98 m s$^{-1}$** residual scatter, whereas **AESTRA $v_{\mathrm{correct}}$** yielded **0.46 m s$^{-1}$**, close to the “best possible” reference **0.39 m s$^{-1}$**. A circular Keplerian recovery with **32 walkers**, **10k steps**, and **2k burn-in** recovered **$P = 100.0 \pm 0.27$ days** and **$K = 0.12 \pm 0.01$ m s$^{-1}$**. Across the paper’s cases, AESTRA detects signals near **0.1 m s$^{-1}$** under **3 m s$^{-1}$** activity jitter and **0.3 m s$^{-1}$** photon noise per spectrum [2311.18326].

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 $y_{\mathrm{act}}$ and $b_{\mathrm{rest}}$, and slight underestimation of $K$ when **$K \gtrsim 1$ m s$^{-1}$** because planetary variation can be treated as “noise” during latent smoothing [2311.18326].

## 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 [2606.13574]. 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 $>2.5$**, cloud passages with **irradiance drops $\ge 5\%$**, low-irradiance or low-count days, high intraday RV RMS **$>1.5$ m s$^{-1}$**, 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 [2606.13574].

The per-exposure generative model is
$$
y_{\mathrm{model},i}(\lambda)
=
[1-y_{t,i}(\lambda)]\,[1+y_{c,i}(\lambda)]
\odot
[y_*(\lambda)+\Delta y_{*,i}(\lambda)]
+
b_i,
$$
or equivalently $y(\lambda,t)\approx C(\lambda,t)\times S(\lambda,t)\times T(\lambda,t)+b(t)$. Here $y_t$ is Earth-frame telluric absorption, $y_c$ is smooth continuum modulation, $y_*$ is a time-independent empirical stellar template, and $\Delta y_*$ 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 **$\sim200$-pixel Gaussian smoothing kernel**. The stellar branch uses **three eigenspectra**, high-pass filtered and penalized in continuum bins to enforce line-localized variability [2606.13574].

Training minimizes a weighted reconstruction loss plus morphology regularizers:
$$
L = L_{\mathrm{fid}} + \alpha_t L_t + \alpha_c L_c + \alpha_* L_*,
$$
with **$\alpha_t = 0.02$**, **$\alpha_c = 1$**, and **$\alpha_* = 10$**. The learned telluric component reproduces narrow Earth-frame features with line depths correlated at **$r > 0.999$** with the NEID precipitable-water-vapor diagnostic, indicating that the model isolates real atmospheric variability without external templates [2606.13574].

After removing only the learned telluric and continuum components, AESTRA II measures an apparent bulk Doppler shift $v_{\mathrm{app}}$ by template fitting to the average spectrum, shifts each spectrum by $-v_{\mathrm{app}}$, 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 **$r = 0.889$**, 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 $v_{\mathrm{act},i}=f_\theta(z_i)$, where $f_\theta$ is a **two-layer MLP** with hidden sizes **128** and **8**, and a circular-orbit planet model
$$
v_{\mathrm{planets},i}=\sum_{k=1}^{N_p}K_k\sin(2\pi t_i/P_k+\phi_k).
$$
Candidate periods and phases are initialized from an RV-level Lomb–Scargle search, and amplitudes are optimized in four period blocks: **$P<10$ d**, **$10$–$100$ d**, **$100$–$250$ d**, and **$>250$ d** [2606.13574].

Detection uses
$$
Q_{\mathrm{detect}} = K_{\mathrm{rec}} - \alpha P_{\mathrm{resid}},
$$
with **$\alpha = 3.5$** and threshold **$Q_{\mathrm{detect}} > 0.185$**, calibrated on **500 single-planet injections** to yield **zero spurious detections**. Over periods **2.5–400 days** and semi-amplitudes **$K\in[0.1,0.7]$ m s$^{-1}$**, AESTRA recovered **238 of 500** injected planets, including **13 with $K<0.3$ m s$^{-1}$**. The traditional CCF-indicator baseline recovered **9** planets and **none below $K=0.5$ m s$^{-1}$**. Recovered AESTRA amplitudes had median relative error **$\approx 13.6\%$**, with **84th percentile 27.1\%** and **95th percentile 43.1\%** [2606.13574].

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 [2606.13574].

## 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 **$\sim27$-day** segments typical of **TESS** [2008.03351]. 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—**$R_{\mathrm{var}}$**, **8-hour flicker**, and **maximum Lomb–Scargle peak height**—together with **Gaia DR2** quantities and derived kinematics, including **$G_{BP}-G$**, **effective temperature**, **radius**, **luminosity**, **parallax**, **photometric flux-over-error**, distance intervals, **galactic latitude**, **$v_{\mathrm{tan}}$**, and **$v_b$**. The classifier uses **LG\_peaks, $R_{\mathrm{var}}$, $bp\_g$, phot\_g\_mean\_flux\_over\_error, $r\_lo$, $r\_hi$, parallax**, while the regressor uses **$R_{\mathrm{var}}$, $T_{\mathrm{eff}}$, $bp\_g$, lum\_val, flicker, radius\_val, $v_{\mathrm{tan}}$, phot\_g\_mean\_flux\_over\_error, $b$, $v_b$** [2008.03351].

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 **$\sim150$ days**, including **70** stars with **$P_{\mathrm{rot}} > 70$ days**. Full four-year Kepler light curves were split into roughly **54** segments of **27 days**, producing **$>1.8$ 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 **$n_{\mathrm{estimators}}\in[1,100]$**, **max\_depth $\in[1,150]$**, and **max\_features $\in[1,10]$**, with the chosen setting **$n_{\mathrm{estimators}}=20$**, **max\_depth $=50$**, and **max\_features $=6$** because it minimized average $\chi^2$ [2008.03351].

Performance depends strongly on domain. On **Kepler 4-year** light curves, Astraea predicts rotation periods with **$\sim13\%$** uncertainty overall and **$\sim9\%$** for periods **$>30$ days**. When trained on **27-day Kepler segments**, it predicts periods up to **$\sim150$ days** with **$\sim9\%$** uncertainty overall and **$\sim5\%$** for periods **$>30$ days**. On real **TESS** data, however, performance degrades: for the **195** stars observed by both Kepler and TESS, the model predicts periods with **$\sim55\%$** uncertainty. The paper attributes this to bandpass differences, higher white-noise floor, larger pixels, contamination, and field-dependent kinematic shifts [2008.03351].

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 $P_{\mathrm{rot}}$. 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 $R_{\mathrm{var}}$ and flicker [2008.03351].

## 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 [2404.03136]. Here the problem is not spectroscopy or autonomy, but hardware decoding of parity-bit syndromes within a strict **$\le 1\,\mu\mathrm{s}$** 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 $\le 10$**; the paper notes **945 matchings at HW = 10**. It achieves RT-MWPM up to **$d=7$** in approximately **456 ns**, but the brute-force approach does not scale beyond about **$d\approx 7$** because the number of candidate matchings grows exponentially with Hamming weight and code distance. **Astrea-G** is the greedy extension that reaches **$d\le 9$** in **$1\,\mu\mathrm{s}$** by pruning the MWPM graph below a probability threshold and then performing greedy near-exhaustive search, but its accuracy deteriorates at larger distances; at **$d=13$** its logical error rate is **43× higher than MWPM** [2404.03136].

The Promatch paper introduces an adaptive predecoder that reduces high-HW syndromes before dispatching them to Astrea. The Hamming weight is
$$
w_H=\sum_i s_i,
$$
and the logical error rate is estimated as
$$
\mathrm{LER}=\sum_k P_o(k)\times P_f(k).
$$
Promatch operates locally and greedily, prioritizing matches that do **not create new singletons**, because singleton creation forces long error chains with probability approximately **$p^L$** for chain length $L\ge 2$. 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 [2404.03136].

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 **$d=11$** and **6 KB** at **$d=13$**; path tables occupy **129 KB** and **345 KB**, respectively. For high-HW syndromes (**HW $\ge 10$**), predecoding latency peaks at **824 ns** for **$d=11$** and **928 ns** for **$d=13$**, while end-to-end **Promatch + Astrea** latency peaks at **904 ns** and **960 ns**, with average about **524 ns**. The probability of exceeding **$1\,\mu\mathrm{s}$** at **$d=13$** is reported as **$\sim1.5\times10^{-17}$** [2404.03136].

At physical error rate **$p=10^{-4}$**, ideal MWPM achieves LER **$1.8\times10^{-13}$** at **$d=11$** and **$3.4\times10^{-15}$** at **$d=13$**. **Astrea-G alone** yields **$4.5\times10^{-13}$** at **$d=11$** and **$1.4\times10^{-13}$** at **$d=13$**. **Promatch + Astrea** yields **$4.5\times10^{-13}$** at **$d=11$** and **$2.6\times10^{-14}$** at **$d=13$**. The parallel design **Promatch || Astrea-G** reaches **MWPM-equivalent LER** at **$d=13$**, namely **$3.4\times10^{-15}$** [2404.03136].

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-language model

In multimodal machine learning, **Astrea** is a **Mixture-of-Experts (MoE) vision-language model** 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 [2503.09445]. 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 $D$ and fused by sparse top-$k$ routing. If the router produces scores $g_e(x)$ over experts and each expert applies an adapter $f_e(\cdot)$, the mixture is
$$
y=\sum_{e=1}^{E} g_e(x)\,f_e(x).
$$
With top-$k$ gating, the selected-expert distribution is
$$
\mathcal{P}_i=\frac{\exp(\omega_i)}{\sum_{j\in\mathrm{Top\text{-}k}}\exp(\omega_j)},
\qquad
f=\sum_{i=1}^{k}\mathcal{P}_i\cdot\mathcal{F}_i(f_i),
$$
and non-selected experts are zeroed. The main backbones are **Vicuna-1.5-13B** and **Hermes2-Yi-34B** [2503.09445].

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
$$
\mathcal{L}_{\mathrm{CL}}
=
-\log
\frac{\exp(\operatorname{sim}(z_i,z_j)/\tau)}
{\sum_k \exp(\operatorname{sim}(z_i,z_k)/\tau)},
$$
and knowledge continuity is maintained by **probabilistically activated stochastic residual connections**, expressible as
$$
y=x+\beta\cdot f(x),\qquad \beta\sim\mathrm{Bernoulli}(p).
$$
The full model combines generative and momentum-contrastive objectives, with
$$
\mathcal{L}_{\mathrm{total}}=\mathcal{L}_{\mathrm{reg}}+\mu\cdot\mathcal{L}_{\mathrm{moco}}.
$$
Training uses **LLAVA-Next multi-resolution** visual tokenization with **576 visual tokens per image**, **AdamW** with **$\beta_1=0.9$**, **$\beta_2=0.98$**, **weight decay 0.05**, **peak learning rate $10^{-4}$**, and **15\% linear warm-up**, on **48×8 A800 GPUs** [2503.09445].

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$^I$ 88.2**, **MMMU$^V$ 51.5**, **MathVista$^M$ 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-$k=3$** is the best trade-off, and that removing either residual connections or contrastive learning reduces metrics [2503.09445].

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 [2503.09445].

Source: https://www.emergentmind.com/topics/astrea