{
 "day": "fri",
 "date": "2026-07-09",
 "radar": [
  {
   "id": "http://arxiv.org/abs/2607.06641v1",
   "title": "Healthier LLMs: Retrieval-Augmented Generation for Public Health Question Answering",
   "summary": "Large language models (LLMs) achieve promising results on medical question answering benchmarks, yet their use in public health is constrained by hallucinations and the rapid evolution of official guidance. Retrieval-Augmented Generation (RAG) mitigates these risks by grounding responses in an explicitly maintained corpus, but end-to-end performance depends critically on retrieval configuration and on evaluation beyond multiple-choice formats. We extend PubHealthBench, a question answering (QA) benchmark of 7,929 questions derived from UK Government public health guidance, into a retrieval-augmented setting and systematically evaluate retrieval and generation choices. We compare dense, sparse, and hybrid retrieval across multiple embedding models and corpus variants, and show that hybrid retrieval consistently improves recall and ranking quality, with chunk length and topic interacting with ranking performance. Providing retrieved context substantially increases multiple-choice accuracy across a diverse set of LLMs, enabling smaller open-weight models to match or outperform larger models used without retrieval, with gains primarily driven by retrieval quality and careful context selection. To assess realistic free-form answering, we introduce a rubric-based LLM-as-a-judge covering faithfulness, completeness, clarity, and factual consistency, and validate it against dual human annotations. Judge-human agreement is strongest for faithfulness and completeness, while factual consistency and clarity are less reliably reproduced, motivating caution when interpreting those dimensions at scale. Overall, our results highlight retrieval as a primary lever for reliable public health QA and provide practical guidance for building and evaluating RAG systems grounded in official guidance.",
   "published": "2026-07-07T14:47:42Z",
   "authors": [
    "Felix Feldman",
    "Joshua Harris",
    "Timothy Laurence",
    "Leo Loman",
    "Ollie Higgins",
    "Fan Grayson"
   ],
   "why": "The paper's systematic evaluation of RAG configurations, retrieval quality, and faithfulness aligns directly with the researcher's focus on RAG pipelines and evaluation methodologies."
  },
  {
   "id": "http://arxiv.org/abs/2607.06544v1",
   "title": "Rethinking Indic AI from a Lens of Cultural Heritage Preservation",
   "summary": "As Artificial Intelligence (AI) makes inroads into different parts of the Indian subcontinent, there is significant interest in studying how AI impacts the linguistic and cultural foundations of this civilization. AI is seen as a ''double-edged sword'' where on the one hand, it can enable access and inclusion for a large population, on the other, it can homogenize worldviews and exclude underrepresented languages and worldviews. In this paper, we try to characterize this problem by addressing the extensive characteristic nature of Indian linguistics and the way they closely connect to cultural practices and worldview. We then perform a longitudinal survey of how Natural Language Processing (NLP) techniques have evolved in this space, tracing the historical development of Indic NLP, covering key milestones, methodological shifts, and resource creation efforts. In addition, the paper also examines the structural and sociolinguistic characteristics of Indian languages, such as rich morphology, complex scripts and grammar rules, diglossia, and large dialectal variation, and explains how these create unique challenges for building AI foundation models. We then discuss the growing role of Indic foundation models and analyze how these models address these long-standing resource and representation gaps. Finally, we propose a research direction called 'Culture Sensing', which re-imagines AI based on hermeneutic reasoning. Culture Sensing aims to address open problems such as ensuring equitable performance across low-resource languages and producing outputs that are culturally meaningful. By bringing together past work, current techniques, and emerging trends, this paper outlines research directions that can guide the next phase of Indic NLP and contribute to the development of more robust and inclusive Indic foundation models.",
   "published": "2026-07-07T17:51:10Z",
   "authors": [
    "Aparna Madva",
    "Sharath Srivatsa",
    "Srinath Srinivasa",
    "Tulika Saha"
   ],
   "why": "The paper addresses the intersection of linguistic diversity, cultural preservation, and the challenges of building AI for complex, underrepresented languages, which aligns with the researcher's focus on Indigenous language sovereignty and low-resource NLP."
  },
  {
   "id": "http://arxiv.org/abs/2607.06327v1",
   "title": "Estimating Uncertainty from Reasoning: A Large-Scale Study of Multi- and Crosslingual MCQA Performance in LLMs",
   "summary": "Uncertainty estimation (UE) enables LLM-powered systems to recognize when to abstain, yet existing research has predominantly focused on English. We present the first large-scale evaluation of UE methods across 22 languages, spanning high-, mid-, and low-resource settings. Using two human-curated Q\\&A datasets, we compare open and closed box UE methods (nine in total) across different model sizes and architectures while eliciting long-form reasoning, avoiding LLM-as-a-judge and embedding-based scoring, which can introduce evaluation noise. We report three main actionable findings. First, we find that prompting models to reason in English while keeping questions in low-resource languages substantially improves UE performance, suggesting that comprehension of low-resource languages is largely intact, and that the reliability bottleneck lies in generation rather than understanding. Second, prompting models to reason in English closes the UE performance gap between low and high-resource languages, demonstrating that generation language matters more than the question language. Third, the choice of UE method should depend on model scale: at smaller scales, open-box probability-based methods outperform alternatives; at larger scales, closed-box self-verbalized uncertainty becomes superior. Finally, we provide an analysis of threshold selection for selective prediction, offering guidance on calibrating abstention in multilingual settings.",
   "published": "2026-07-07T14:25:09Z",
   "authors": [
    "Andrea Alfarano",
    "Andrea Bacciu",
    "Saab Mansour",
    "Amin Mantrach",
    "Marcello Federico"
   ],
   "why": "The paper's findings on how reasoning in English improves performance in low-resource languages directly informs the researcher's work on data-efficient adaptation and reliability in underserved languages."
  },
  {
   "id": "http://arxiv.org/abs/2607.06289v1",
   "title": "From Sinhala to Dhivehi: Cross-Lingual Transfer Learning for Low-Resource Speech Recognition",
   "summary": "Dhivehi, the national language of the Maldives, is currently under-resourced for automatic speech recognition (ASR) and other NLP tasks. This study investigates whether cross-lingual transfer learning from Sinhala, a linguistically related, relatively well-resourced Insular Indo-Aryan language, can improve Dhivehi ASR. We conduct seventeen experiments across five transfer learning paradigms: Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a control using Turkish as an unrelated language. The strongest system, continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM, achieves 12.89% WER and 2.70% CER, outperforming the Dhivehi-only baseline by 13.50% WER and 3.02% CER. However, the adaptation strategy and decoding configuration are equally critical for a successful transfer learning experiment. We conduct seventeen controlled experiments spanning five transfer learning paradigms: Dhivehi-only baselines, sequential fine-tuning, multilingual fine-tuning, continual pre-training, and a control experiment using Turkish as an unrelated language. The strongest system, continual pre-training on Sinhala followed by fine-tuning on Dhivehi with KenLM, achieves 12.89% WER and 2.70% CER, outperforming the Dhivehi-only baseline by 13.50% WER and 3.02% CER. The Turkish control experiment confirms that observed improvements stem from linguistic relatedness; adaptation strategy and decoding configuration are also critical.",
   "published": "2026-07-07T13:57:54Z",
   "authors": [
    "Lukmal Ilyas",
    "Nevidu Jayatilleke"
   ],
   "why": "The paper focuses on cross-lingual transfer learning for low-resource languages, which aligns with the researcher's interest in data-efficient adaptation for underserved languages."
  },
  {
   "id": "http://arxiv.org/abs/2607.05992v1",
   "title": "PluraMath: Extending Mathematical Reasoning Evaluation Beyond High-Resource Languages",
   "summary": "Mathematical reasoning has become a central task for evaluating and tuning reasoning Large Language Models (LLMs), yet existing benchmarks remain heavily biased toward high-resource languages, with English and Chinese dominating both pre-training corpora and evaluation suites. The recently released PolyMath (Wang et al., 2025) dataset represents a significant step forward, yet its coverage is still limited to 18 only high-resource languages. To address this gap, we introduce PluraMath, an extension of PolyMath to 18 additional {underrepresented languages spanning 6 language families -- ranging from mid-resource to extreme low-resource settings. We constructed the dataset through a human-curated pipeline, where native speakers thoroughly validated pre-computed translations. Using PluraMath, we then benchmark 27 reasoning LLMs across four model scales -- small, mid-size, large, and closed-source ensembles -- probing the multilingual mathematical reasoning capabilities of state-of-the-art models under diverse linguistic conditions. Our fine-grained analysis confirms a persistent gap in mathematical reasoning performance between high-resource and underrepresented languages, with stronger results largely associated with better instruction-following ability. We fully open-source our dataset, data acquisition pipeline, and evaluation framework, with the goal of lowering the barrier to multilingual benchmark development for underrepresented communities.",
   "published": "2026-07-07T08:25:29Z",
   "authors": [
    "Daryna Dementieva",
    "Nikolay Babakov",
    "Kathy H\u00e4mmerl",
    "Ilseyar Alimova",
    "Jind\u0159ich Libovick\u00fd",
    "Shu Okabe"
   ],
   "why": "The paper addresses the gap in mathematical reasoning for underrepresented and low-resource languages, which aligns with the researcher's focus on low-resource NLP and multilingual evaluation."
  },
  {
   "id": "http://arxiv.org/abs/2607.05849v1",
   "title": "CoPiT: Cognitive Pivot Translation for Digraphic Low-Resource Mongolian in the Traditional Script",
   "summary": "Low-resource languages remain challenging for machine translation, and Mongolian is a representative case. As a digraphic language, Mongolian is written in both Cyrillic and Traditional scripts, which exhibit a severe imbalance in data availability. While the Cyrillic script is relatively well-resourced, the Traditional script remains extremely data-scarce and orthographically ambiguous, leading to substantial performance degradation in direct translation. We propose CoPiT, a cognitively motivated pivot-based translation pipeline that exploits this internal resource hierarchy by routing translation through the Cyrillic script. The pipeline explicitly resolves script-induced ambiguity in the Traditional script before translation, enabling more stable and accurate meaning transfer. Across multiple backbone models and target languages, CoPiT consistently outperforms direct translation, achieving substantial absolute BLEU improvements together with consistent 1.5-1.6x COMET gains. These gains allow strong open-source models to match or outperform GPT-4.1 under comparable evaluation settings. Beyond inference-time improvements, CoPiT enables the construction of synthetic parallel data directly from Traditional-script text, mitigating data scarcity in realistic low-resource scenarios. We release a new multi-script parallel dataset covering Mongolian in both scripts alongside English, Korean, and Russian. All datasets and code are publicly available at https://anonymous.4open.science/r/anonymous_project-76C7.",
   "published": "2026-07-07T05:12:13Z",
   "authors": [
    "Burte Bayarsaikhan",
    "Serynn Kim",
    "Buru Chang"
   ],
   "why": "The paper addresses low-resource, digraphic language translation and script-induced ambiguity, which aligns with the researcher's focus on underserved orthographies and data-efficient adaptation."
  }
 ],
 "geoai": {
  "paper": {
   "id": "http://arxiv.org/abs/2607.03949v1",
   "title": "TESSERA v2: Scaling Pixel-wise Earth Foundation Models",
   "summary": "Pixel-wise Earth-observation (EO) foundation models are now achieving state-of-the-art performance via generated spatial embeddings. However, how these models scale and how best to spend a pretraining budget remain poorly understood. We present the largest controlled scaling study for EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks. We find that pretraining loss barely predicts downstream performance (|Pearson r| < 0.2), so selecting models by loss wastes a large share of the compute. We also find that, as the training budget grows, the encoder and the data should grow together while the projector stays fixed, which gives a simple rule for allocating compute. Using this rule, we train a family of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill them into compact students for embeddings-as-data deployment. The 21-million-parameter distilled TESSERA v2-1B-M in aggregate outperforms all open and proprietary models tested, some of which are orders of magnitude larger. These students produce Matryoshka representations that are inexpensive to serve: a 16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of the storage. Upon completion of training we plan to release v2 global embeddings covering 2017-2025. Together, these results give a concrete, empirically grounded recipe for scaling pixel-wise EO foundation models: train large encoders, select by downstream performance, and distil into flexible student models. All code will be released at https://github.com/ucam-eo/tessera.",
   "published": "2026-07-04T16:52:34Z",
   "authors": [
    "Zhengpeng Feng",
    "Sadiq Jaffer",
    "Ira Shokar",
    "Jovana Knezevic",
    "Mark Elvers",
    "Clement Atzberger"
   ]
  },
  "reason": "This paper directly addresses the scaling and pretraining strategies of pixel-wise Earth-observation foundation models, which is the most central and high-leverage topic in your stated research interests.",
  "text": "Preprint.\n\nTESSERA V 2: S CALING P IXEL - WISE E ARTH\nF OUNDATION M ODELS\nZhengpeng Feng1 Sadiq Jaffer1 Ira Shokar2 Jovana Knezevic1 Mark Elvers1\nClement Atzberger3 Robin Young1 Aneesh Naik1 Niall Robinson2 Andrew Blake1\nDavid Coomes1 Anil Madhavapeddy1 Srinivasan Keshav1\u2217\n1\n2\n3\nUniversity of Cambridge\nNVIDIA\ndClimate Labs\n\narXiv:2607.03949v1 [cs.CV] 4 Jul 2026\n\nA BSTRACT\nPixel-wise Earth-observation (EO) foundation models are now achieving state-ofthe-art performance via generated spatial embeddings. However, how these models\nscale and how best to spend a pretraining budget remain poorly understood. We\npresent the largest controlled scaling study for EO to date: 395 training runs on\n1,024 GH200 superchips within a fixed pixel-wise BARLOW T WINS family, each\nevaluated on 15 downstream tasks. We find that pretraining loss barely predicts\ndownstream performance (|Pearson r| < 0.2), so selecting models by loss wastes\na large share of the compute. We also find that, as the training budget grows,\nthe encoder and the data should grow together while the projector stays fixed,\nwhich gives a simple rule for allocating compute. Using this rule, we train a\nfamily of pixel-wise models (0.5B and 1B, with a 2B model in training) and distill\nthem into compact students for embeddings-as-data deployment. The 21-millionparameter distilled TESSERA v2-1B-M in aggregate outperforms all open and\nproprietary models tested, some of which are orders of magnitude larger. These\nstudents produce M ATRYOSHKA representations that are inexpensive to serve: a\n16-dimensional prefix keeps 92% of the full 128-dimensional performance at 1/8 of\nthe storage. Upon completion of training we plan to release v2 global embeddings\ncovering 2017-2025. Together, these results give a concrete, empirically grounded\nrecipe for scaling pixel-wise EO foundation models: train large encoders, select by\ndownstream performance, and distil into flexible student models. All code will be\nreleased at https://github.com/ucam-eo/tessera.\n\n1\n\nI NTRODUCTION\n\nThe main bottleneck in using Earth observation (EO) for downstream tasks is preparing the data:\nradiometric calibration, cloud and shadow masking, cross-sensor harmonisation, and expensive\ncomputation over raw imagery. Beyond this preparation, the ground-truth labels these tasks need are\noften scarce and tied to specific regions and seasons (Metcalfe et al., 2025; Hou et al., 2026). The\ndata itself is awkward in ways natural images are not: Sentinel-2 and Sentinel-1 observe a given\npixel at different, irregular cadences, cloud removes much of the optical record (Figure 1a), and the\ncloud-free composites most models train on remove the phenological dynamics that downstream\ntasks depend on (Zeng et al., 2020; Xiao et al., 2025). Lowering these barriers requires a reusable\nrepresentation layer for global surface state, delivered as spatially mapped data.\nEarth embeddings are vector representations of specific places and times that compress and fuse\nmulti-source observations (Klemmer et al., 2025; Fang et al., 2026). Provided at global scale, in the\n\u2018embeddings-as-data\u2019 approach, these representations free analysts from the onerous acquisition of\nheterogeneous raw data, the need for remote-sensing expertise, and from user-side GPU compute.\nWith task-specific heads, they have been shown to match, and often exceed, task-specific models.\nAn ideal embedding product should satisfy at least the following five desiderata. It should be analysisready, giving users a geospatial data layer they can work on directly, without processing raw imagery\nor running GPU inference. It should be transferrable, so that the representation carries across tasks\n\u2217\n\nCorresponding author.\n\n1\n\n\fPreprint.\n\nFigure 1: Overview of TESSERA v2. (a) Sparse, irregular Sentinel-2/Sentinel-1 sampling at one\nlocation. (b) 395-run downstream-driven scaling sweep and compute-optimal fits; inset: pretraining\nloss vs. downstream score. (c) M ATRYOSHKA distillation of the 1 B teacher into N/S/M/L students,\nd \u2208 {16, 32, 64, 128}. (d) Artefact removal and inter-annual stability vs. v1. (e) Deployment\nvia G EOT ESSERA. (f) Leading composite performance across the 29-task full suite (15 shared\nA LPHA E ARTH suite tasks + 14 held-out datasets). Radars show per-task scores on the shared\n(left) and held-out (right) tasks; bars give the 29-task composite (TESSERA v2-1B-M, shades\nd \u2208 {16, 32, 64, 128}) and encoder size. TESSERA v2 leads at 0.611, clearly ahead of all three\nbaselines.\nand workflows (classification, regression, change detection, high-resolution mapping, etc.) rather\nthan being tuned to only one benchmark or region (Klemmer et al., 2025; Lyu et al., 2026). It\nshould be reproducible, letting users inspect, compare, and extend the released artifacts rather than\ndownload opaque outputs. It should be economical in the resources that limit EO adoption, namely\nlabels, training compute, inference compute, storage, I/O, and engineering time. Finally, it should\noffer adaptivity, an explicit characterisation of the trade-off between accuracy and computational\n(or economic) cost, so that users are not locked into a single model size or a single embedding\ndimensionality.\nMeasured against these criteria, every existing system leaves a gap. Remote-sensing foundation\nmodels (Cong et al., 2022; Reed et al., 2023; Fuller et al., 2023; Guo et al., 2024; Tseng et al.,\n2024; 2025; Astruc et al., 2025) learn strong representations but release models rather than products,\nso the preprocessing and inference burden stays with the user (Fang et al., 2026). Specialised\nembedding archives are useful within their scope: ESD (Chen et al., 2026) compresses 25 years of\nLandsat/MODIS reflectance into quantised 30 m embeddings, though its resolution, sensor scope,\nand demonstrated evaluation breadth stop short of the specification above. A LPHA E ARTH (Brown\net al., 2025) is analysis-ready at global scale but only open-output: the embeddings are public while\nthe training procedure and model weights are not (Hou et al., 2026), which limits inspection and\nextension. TESSERA v1 (Feng et al., 2026) is open, pixel-wise, and label-efficient, yet it ships one\nfixed 128-dimensional specification, so every user inherits the same storage and I/O budget regardless\nof their deployment constraints. Each of these meets part of the specification. None is an open,\npixel-wise, analysis-ready family in which users pick the model size and embedding dimension that\nfit their budget.\nThese product constraints reach back into training. If an embedding field is produced once, served\nglobally, and reused by many downstream users, then pretraining compute becomes part of the product\nbudget rather than an isolated modelling expense. The relevant question is not simply whether a\n\n2\n\n\fPreprint.\n\nlarger backbone lowers a self-supervised loss, but which allocation of compute produces the most\nuseful representation per unit cost.\nIn the language and vision domains, that allocation comes from scaling laws fit to the pretraining\nloss (Kaplan et al., 2020; Hoffmann et al., 2022; Zhai et al., 2022). The few existing EO scaling studies\neither entangle architecture with capacity or fit power laws to the loss and defer the downstream\ncomparison (Dionelis et al., 2025; Wickrema et al., 2025). In EO the loss is a doubtful target, because\ncloud and orbital sampling dominate input variance and a redundancy-reduction objective can be\nminimised through invariances that carry no downstream value.\nWe therefore ran a downstream-driven scaling study: 395 controlled pretraining runs of pixel-wise\nBARLOW T WINS (Zbontar et al., 2021) encoders (the backbone family of TESSERA v1),1 each\nevaluated on 15 downstream tasks. The result is a different scaling picture from loss-driven practice.\nPretraining loss turns out to be a weak proxy for downstream utility, and selecting models by loss\nwastes roughly 254% of compute relative to downstream-driven selection (F1). The downstreamoptimal allocation instead puts additional budget into encoder capacity and training data while the\nprojector stays essentially fixed (F2). This yields a simple production rule: train a large encoder on\nmatched data, then recover deployment efficiency through distillation. The sweep is expensive, but\nit is a one-off measurement of the production function of this model family: once the downstreamcalibrated allocation is known, larger teachers can be trained by rule rather than by repeated loss-driven\nsearch.\nTESSERA v2 follows this approach. We train a large pixel-wise Sentinel-1/2 teacher as a representation distribution rather than a deployed artifact, and distil it into a family of compact students\n(N/S/M/L). The students produce precomputed annual embeddings served through G EOT ESSERA, so\nusers train lightweight heads on analysis-ready data instead of running a backbone. M ATRYOSHKA\nprefixes expose 16-, 32-, 64-, and 128-dimensional views of the same embedding, which turns storage\nand I/O into a user-side knob. Across a 29-task suite (the 15-task A LPHA E ARTH suite plus 14 further\nheld-out datasets), the distilled students lead every open and proprietary embedding product we\ncompare, the smallest prefixes keep most of the full-dimensional score, and the embeddings shed the\nSentinel acquisition artefacts visible in v1 (Figure 1d). This nested representation must be learned\nthrough distillation rather than by adding prefix losses during self-supervised pretraining.\nIn summary, our contributions are:\n1. Downstream-driven scaling for product-grade EO embeddings. Across 395 runs evaluated on\n15 downstream tasks, we show that the pretraining loss is a poor selection target (F1) and fit a\ncompute-allocation rule (F2): additional pretraining budget should go to encoder capacity and\ntraining data, not projector size.\n2. A deployable pixel-wise embedding product family. Guided by this rule, we train large teachers\nand distil them into compact N/S/M/L students whose annual Sentinel-1/2 embeddings are served\nas analysis-ready data through G EOT ESSERA, with the best composite score and mean rank on the\n15-task A LPHA E ARTH suite at a deployment cost two orders of magnitude below the teacher\u2019s.\n3. Storage-adaptive M ATRYOSHKA embeddings. Each student exposes prefixes at d \u2208\n{16, 32, 64, 128} from one embedding, an accuracy/storage knob that needs no retraining. The\nd=16 prefix keeps \u223c92% of the d=128 score at 1/8 of the storage.\n\n2\n\nR ELATED WORK\n\nThe present work builds on TESSERA v1 (Feng et al., 2026), which adapts BARLOW T WINS\nredundancy reduction to cloud-corrupted EO time series following Lisaius et al. (2024). We reuse its\nd-pixel formulation and pretraining recipe as the fixed model family for the scaling study.\nEO foundation models. Remote-sensing foundation models pretrain spatial backbones on singletime, often cloud-filtered patches with contrastive or masked-image objectives (Ma\u00f1as et al., 2021;\nGuo et al., 2024; Cong et al., 2022; Reed et al., 2023; Tang et al., 2023; Li et al., 2024; Noman et al.,\n2024; Wang et al., 2022; Wanyan et al., 2024), and have grown into broader families (Sun et al., 2023;\n1\nThe sweep ran on 1,024 NVIDIA GH200 superchips McIntosh-Smith et al. (2024), each pairing a Grace\nCPU with one H100 GPU (96 GB HBM3).\n\n3\n\n\fPreprint.\n\nWang et al., 2023; Bastani et al., 2023; Schmude et al., 2024; Szwarcman et al., 2025; Mendieta\net al., 2023; Han et al., 2024; Wu et al., 2025; Luo et al., 2024; Zhu et al., 2025; Zhang et al., 2025).\nMulti-sensor and multi-resolution models fuse optical and SAR or ingest many sensors at once (Fuller\net al., 2023; Yao et al., 2023; Xiong et al., 2024; Astruc et al., 2025; Perron et al., 2026; Tseng\net al., 2025). P RESTO (Tseng et al., 2024) processes per-pixel time series, and MOSAIKS (Rolf\net al., 2021) explored lightweight universal features. Two assumptions recur across these designs:\nnatural-image scaling intuition is inherited by analogy, and the model is deployed through per-task\nfine-tuning of the backbone, which is compute- and label-intensive for EO users. We are not aware of\na controlled study of where additional compute should go given irregular revisits, cloud occlusion,\nand label scarcity. Section 3 provides one for the pixel-wise BARLOW T WINS family.\nEmbedding products. A second line publishes precomputed, analysis-ready embeddings instead\nof a backbone: TESSERA v1 releases global annual 10 m pixel-wise int8 embeddings with the\nG EOT ESSERA retrieval library (Feng et al., 2026; Madhavapeddy et al., 2026), A LPHA E ARTH\nprovides global annual 10 m embedding fields from many instruments (Brown et al., 2025), and\nESD (Chen et al., 2026) compresses 25 years of Landsat/MODIS reflectance into quantized 30 m\nembeddings. The shared limitation is a fixed embedding specification: one dimension, one storage\nand I/O budget for every user, and no coordinate ordering that would let a user truncate to a smaller\ndimension without retraining or loss. TESSERA v2 keeps the paradigm and adds two degrees of\nfreedom: students at four sizes, produced by scaling-law-guided distillation, and M ATRYOSHKA\nprefixes d \u2208 {16, 32, 64, 128} from a single embedding without retraining. Section 5 explains why\nthe ordering that makes these prefixes usable requires distillation rather than self-supervision alone.\nEvaluating embeddings across such budgets has begun to attract dedicated benchmarks (Vinge et al.,\n2025).\nScaling laws. Empirical scaling laws have shaped recent language and vision work (Kaplan et al.,\n2020; Hoffmann et al., 2022; Zhai et al., 2022), and are typically fit to a self-supervised pretraining\nloss on the assumption that the loss proxies downstream quality. This assumption can fail: in language\nmodelling the mapping from pretraining loss to downstream performance is sometimes noisy or\nnon-monotone, so a lower loss need not yield a better task model (Hu et al., 2025). The two EO\nstudies we are aware of do not test it directly. Dionelis et al. (2025) sweep architecture, size, and\ndata on PhilEO Bench but entangle architecture with capacity over a coarse grid, and Wickrema et al.\n(2025) fit peta-pixel power laws to the validation loss in a data-limited regime that, as they note, is\nconfounded by under-trained large models, and defer the loss-vs-downstream comparison to future\nwork. Neither isolates encoder, projector, and data under matched compute. We isolate all three and\nfit the compute allocation against task performance directly, evaluating every run on 15 downstream\ntasks.\nDistillation and nested embeddings. Knowledge distillation (Hinton et al., 2015) transfers representations from a high-capacity teacher to a compact student. M ATRYOSHKA representation\nlearning (Kusupati et al., 2022) produces nested embeddings whose prefixes work at multiple\ndimensionalities. The two are rarely combined in EO, and rarely with attention to whether the\nself-supervised objective can support nested coordinates at all. Our analysis shows that naive M A TRYOSHKA-BARLOW T WINS fails because redundancy-reduction objectives identify subspaces only\nup to rotation. Distillation against a fixed teacher supplies the ordering signal that self-supervision\nlacks.\n\n3\n\nD OWNSTREAM - DRIVEN SCALING LAWS\n\nThis section answers the allocation question: within a fixed architectural family, how should pretraining compute be split between encoder size, projector size, and training data? The experiments\nbelow describe the family we sweep, pixel-wise Sentinel-1/2 encoders pretrained with BARLOW\nT WINS (Zbontar et al., 2021). Note that we do not advance them as universal EO scaling laws.\n3.1\n\nS TUDY DESIGN\n\nArchitecture choice. Performance differences attributed to size or data are only reasonable when\nthe architecture is fixed. We therefore fix the architecture before scaling, sweeping seven structural\n4\n\n\fPreprint.\n\naxes (encoder microarchitecture, projector form, Sentinel-1/Sentinel-2 fusion, temporal aggregation,\ncloud handling, sequence length, and at what point M ATRYOSHKA nesting is introduced) one at a\ntime and selecting by aggregated downstream score. The selected configuration is held fixed for the\nentire scaling sweep.\nControlled sweep. With architecture fixed, we pretrain 395 models on 1,024 H100 GPUs in an\niso-FLOP-style grid over encoder size Nenc (16 widths, 7\u2013278 M), projector size Nproj (four widths),\nand training data D (0.03\u20139,984 M d-pixels). At each compute level, the compute-optimal size is the\nvertex of a quadratic fit in log N . The compute axis is\n\u0001\nC = 12 D Nenc Lref + Nproj ,\nLref = 240,\n(1)\nwhere the constant 12 = 6 \u00d7 2 is the textbook 6N D factor (two FLOPs per multiply\u2013add, threefold\nforward-plus-backward) times the two BARLOW T WINS augmentation views per step, and Lref is a\nnominal annual sequence length common to all runs.\nDownstream evaluation. Every pretrained model is evaluated on 15 A LPHA E ARTH suite tasks\ndrawn from 10 source datasets (classification, segmentation, change detection, regression), with\nchance-adjusted metrics: (balanced accuracy \u2212 1/K)/(1 \u2212 1/K) for classification and max(0, R2 )\nfor regression. The per-task scores average into one composite downstream score, the y-axis of\nFigure 2. On the same suite, TESSERA v1 scores 0.541 and A LPHA E ARTH 0.560. Both are\nbaselines in Figure 2a\u2013c. Both baselines are full-budget production systems, whereas the sweep\ngrid deliberately spans many small, data-limited configurations in order to trace out the compute\nfrontier; most individual runs therefore fall below the baselines, while the upper envelope of the\nsweep approaches them. For every run we also record the converged BARLOW T WINS loss on\nheld-out d-pixels, normalised so that runs with different projector widths are comparable.\n\nFigure 2: Downstream-driven scaling laws (395 runs; composite score over 15 tasks). (a) Pretraining\nloss vs. downstream score, coloured by model size. (b, c) Iso-FLOP parabolas for encoder and\nprojector; vertices mark compute-optimal sizes per compute bucket. (d) Power-law fits to the\n\u22c6\n\u22c6\nvertices, Nenc\n\u221d C 0.36 , D\u22c6 \u221d C 0.63 , Nproj\n\u221d C 0.00 ; dashed line is the loss-optimal encoder fit.\n(e1\u2013e3) Bootstrap and leave-one-bucket-out stability of the three exponents.\n\n5\n\n\fPreprint.\n\n3.2\n\nF INDING 1: THE PRETRAINING LOSS IS NOT A GOOD PREDICTOR OF DOWNSTREAM TASK\nPERFORMANCE\n\nAcross the 395 runs, the converged BARLOW T WINS loss and the composite downstream score\nare nearly independent (Figure 2a; Pearson r = \u22120.18, Spearman \u03c1 = \u22120.16). Clouds and orbital\nsampling dominate the per-sample variance without being downstream-relevant, and a redundancyreduction objective can be minimised through invariances orthogonal to the physical processes\ndownstream tasks depend on. The consequence is quantitative, not just statistical: fitting separate\npower laws through loss-selected and score-selected bucket peaks, loss-based selection needs roughly\n254% more compute to reach the same downstream score. EO scaling laws must therefore be fit\nagainst downstream metrics. F1 prices the alternative.\n3.3\n\nF INDING 2: ENCODER SIZE AND DATA REQUIREMENTS SCALE WITH COMPUTE ; THE\nPROJECTOR DOES NOT\n\nWe group runs into nine iso-FLOP buckets and fit, within each, a quadratic in log N whose vertex is the\ncompute-optimal size at that budget. Encoder vertices shift right as compute grows (Figure 2b), while\nprojector vertices stack along a vertical line (Figure 2c). Power-law fits through the vertices (Figure 2d)\n\u22c6\n\u22c6\ngive Nenc\n\u221d C 0.36 (95% CI [+0.29, +0.45]), D\u22c6 \u221d C 0.63 ([+0.53, +0.70]), and Nproj\n\u221d C 0.00\n([\u22120.03, +0.08]). Two observations. First, 0.36 + 0.63 \u2248 1 (precisely 0.99), recovering the C \u221d N\u00b7D\nbalance of Hoffmann et al. (2022) without it being assumed by the fitting procedure, an internal\nconsistency check on the sweep. Second, the projector exponent is statistically indistinguishable\nfrom zero: the compute-optimal projector size does not grow with compute. Encoder capacity, not\nprojector capacity, is the load-bearing scaling axis.\nFrom findings to a recipe. Since encoder capacity and data are the axes that absorb compute and the\nprojector is compute-invariant, the compute-optimal use of a large budget is one oversized encoder\ntrained on correspondingly more data, with the projector held near its optimum as a disposable\ntraining scaffold. Deployment flexibility, in model size and embedding dimension, is then recovered\nthrough distillation. Section 4 trains the large teacher at the upper end of the swept regime. Section 5\ndistils it into a compact student family.\n\n4\n\nA PIXEL - WISE TEMPORAL TEACHER\n\nThis section focuses on pretraining using a single large encoder. Figure 3 summarises the design.\nRelative to TESSERA v1 (Feng et al., 2026), TESSERA v2 adds multi-scale temporal pretraining,\nadaptive full-observation inference, a unified all-Transformer architecture with cross-modal fusion,\nand scaling-law-guided distillation into M ATRYOSHKA students (Section 5). We discuss these\ninnovations in more detail next.\nInputs. A d-pixel at location (i, j) is the time series of all Sentinel-1 and Sentinel-2 observations at\nthat 10 m pixel over one year, with a binary mask marking valid timesteps (cloud-free for Sentinel-2,\npresent for Sentinel-1) (Feng et al., 2026). This preserves the full temporal phenology while tolerating\ncloud occlusion and/or irregular revisits.\nArchitecture. The teacher is a 1 B dual-branch pixel-wise encoder (Figure 3a). Each modality\nbranch linearly embeds its valid observations, adds a sinusoidal day-of-year positional encoding,\nruns a four-layer Transformer, and aggregates the variable-length sequence by learned attention\npooling. A two-layer fusion Transformer then combines the two modality tokens into one embedding\nt \u2208 RdT , dT = 768, followed by a final affine-free LayerNorm (no learnable scale or shift). A\nbatch-normalised projector is also used, but only during pretraining. dT deliberately exceeds the\nstudent\u2019s 128: this leaves room for compression, and the prefix heads of Section 5 map student\nprefixes into the teacher\u2019s space. The encoder is larger than any point in the Section 3 sweep (which\ntops out at 278 M); following the recipe, we extrapolate its C 0.36 encoder law to the 1 B teacher,\nwhose training compute is measured on the same axis (Equation (1)).\nObjective and training. Similar to TESSERA v1, we train using BARLOW T WINS over two\ntemporally subsampled views, with an additional mix-up consistency regulariser (Bandara et al.,\n6\n\n\fPreprint.\n\nFigure 3: TESSERA v2 architecture. (a) Pretraining: two views per d-pixel at random length\nL \u2208 {8, 16}, per-modality Transformers with day-of-year encoding, cross-modal fusion, and BARLOW\nT WINS + mix-up. (b) Distillation: M ATRYOSHKA prefix heads at d \u2208 {16, 32, 64, 128} reconstruct\nthe frozen teacher embedding. (c) Inference: each pixel\u2019s k valid observations are packed into the\nsmallest bucket B \u22c6 \u2265 k, with residual slots filled by midpoint resampling.\n2023) and global shuffling of d-pixels across tiles. Unlike v1\u2019s single fixed sample length, at every\nstep the view length is a random L \u223c Uniform{8, 16}, so each view is a sparse random subsample of\nthe year that forces the encoder to recover annual phenology from few observations. At inference\nwe instead pass all valid observations (Section 4): the encoder is given strictly more evidence about\nthe same phenology, each placed at its true day-of-year and aggregated by length-agnostic attention\npooling, rather than a longer out-of-distribution input. Whole modalities are also dropped with\nsmall probability, which doubles as the training signal for the inference case of a pixel with no valid\nobservations in one modality (e.g. a persistently clouded pixel with zero Sentinel-2).\nThe teacher is pretrained for a single epoch over \u223c4.2 billion d-pixels at a global batch size of\n131,072 on 512 GPUs under FSDP in bf16.\nThe teacher as a distillation target. Given its large size, we expect most users cannot run a 1 B\npixel-wise encoder over global Sentinel-1/2 even once, let alone on a recurring basis. The scaling\nlaws say where to spend pretraining compute: on a large encoder (Section 3). What users can afford to\nserve is something much smaller, set by the recurring cost of global inference. The teacher is therefore\nthe right model to train but the wrong one to deploy, so we treat the frozen teacher T\u03b8 : Pi,j 7\u2192 t as a\nfixed representation distribution (the distribution of target embeddings it induces over d-pixels) and\nobtain every deployed artifact by distilling against it (Section 5).\nAdaptive bucket sampling at inference. The number of valid observations k varies from a handful\nin heavily clouded regions to roughly a hundred in clean ones, yet batched inference wants fixedlength inputs. Where TESSERA v1 sampled a fixed L = 40 timesteps\u2014discarding observations when\nk > L and bluntly duplicating when k < L\u2014TESSERA v2 packs each pixel into the smallest bucket\nfrom the ladder {16, 32, 48, . . .} that fits all k observations, filling the residual slots by midpoint\nresampling across the year (Figure 3c). No observation is discarded, and a batch still partitions into a\nfew fixed-length groups that run in parallel.\n\n7\n\n\fPreprint.\n\n5\n\nD ISTILLING A DEPLOYABLE STUDENT FAMILY\n\nWe expect users to use our embedding as data: consuming precomputed annual pixel embeddings\nwith lightweight task heads, without running a backbone (Feng et al., 2026; Brown et al., 2025). The\nteacher, however, is too computationally expensive to support this, since running a complex teacher\nfor all land areas on earth would be extremely expensive. Instead, we adopt the well-known idea of\ndistillation to reduce the cost of inference.\nInference cost for one global, annual, 10 m Sentinel-1/2 pass follows a near-linear power law in\nencoder parameters, GPUy(N ) \u2248 0.041 N 1.03 H100-years for N in millions. The 1 B teacher costs\nroughly 50 H100-years per global pass. In contrast, the students cost 0.04 to 2 H100-years, two\norders of magnitude lower. Note that, even with lower inference costs, embedding users still need\nto download and pay for storage in proportion to embedding dimension. We therefore provide\nM ATRYOSHKA prefixes to reduce this cost.\nStudent family. We distil the teacher into four students of the same architectural form at different\ncapacities: TESSERA v2-1B-L (44 M , for provider-side global inference), M (21 M , a balanced\ndefault), S (7 M , low-resource), and N (1 M , edge and on-device). Each emits a 128-dimensional\nembedding and differs only in backbone width and depth. Distillation uses \u223c200 million d-pixels on\n64 H100 GPUs.\nM ATRYOSHKA distillation objective. For each prefix length k \u2208 K = {16, 32, 64, 128} we attach\na linear head hk : Rk \u2192 RdT , used only during distillation, and train the student against the frozen\nteacher embedding t with\nX\u0010\n\u0001\u0011\nLDIST (s, t) =\n1 \u2212 cos hk (s1:k ), t .\n(2)\nk\u2208K\n\nEach prefix is supervised to reconstruct the full teacher embedding, which makes it a rate\u2013distortion\ncode for t rather than a copy of the teacher\u2019s first k coordinates. At inference the user takes any prefix\ns1:k directly. Because the teacher is frozen, we run it over the distillation corpus once and cache the\ntargets, so distillation compute is spent almost entirely on the student.\nNesting through distillation, not pretraining. We found that the obvious alternative, adding\nprefix-BARLOW T WINS losses during self-supervised pretraining unexpectedly fails. This is because\nthe redundancy-reduction objective identifies a representation subspace only up to rotation, so prefix\nlosses break the coordinate symmetry through gradient imbalance rather than through a signal about\naxes\u2019 information content. Distillation against a fixed target, however, supplies this important signal.\n5.1\n\nB ENCHMARK RESULTS\n\nWe evaluate in two stages. First, on the 15-task A LPHA E ARTH suite (classification, segmentation,\nregression, and change detection) we compare against the two directly comparable embeddingas-data systems, A LPHA E ARTH (Brown et al., 2025) and TESSERA v1 (Feng et al., 2026), plus\nP RESTO (Tseng et al., 2024), OlmoEarth (Herzog et al., 2026), MOSAIKS (Rolf et al., 2021), and\na suite of RSFMs and generic backbones (Fuller et al., 2023; Guo et al., 2024; Reed et al., 2023;\nTseng et al., 2025; Astruc et al., 2025; Dosovitskiy et al., 2021). Second, we take the four strongest\nof these on the A LPHA E ARTH suite (TESSERA v2, TESSERA v1, A LPHA E ARTH, and OlmoEarth)\nand run a held-out evaluation on 14 further datasets that played no part in development, covering\nclassification, segmentation, regression, and distribution prediction; the two stages together form a\n29-task full suite. Every task uses a fixed lightweight head: a two-layer MLP for pixel-wise tasks and\na simple CNN (< 2M parameters) for patch-level tasks.\nHeadline results. On the full 29-task suite TESSERA v2-1B-M has the best composite score of any\nsystem, 0.611, against 0.576 for TESSERA v1, 0.574 for A LPHA E ARTH, and 0.562 for OlmoEarthL, and it is the smallest of the four (Figure 1f). On the 15 A LPHA E ARTH suite tasks alone (Figure 4c)\nTESSERA v2-1B-L leads at 0.584 (M, S, N at 0.581, 0.570, 0.558), above A LPHA E ARTH (0.560)\nand TESSERA v1 (0.541); it also has the best mean rank (2.4 at d=128 vs. 3.8 and 4.9; Figure 4b)\nand is best or second-best on nearly all tasks. The ordering depends on the task set: A LPHA E ARTH\n8\n\n\fPreprint.\n\nFigure 4: A LPHA E ARTH suite results (15 tasks) and held-out generalisation (14 datasets). (a) Geographic density of all downstream labels. (b) Per-task heatmap with mean rank. (c) Composite\nscore per model; the four markers per TESSERA v2 row are d \u2208 {16, 32, 64, 128}. (d) Score vs.\nprefix dimension: d=16 retains \u223c92% of the d=128 composite at 1/8 of the storage. (e) Score vs.\nglobal annual storage, relative to A LPHA E ARTH at d=64 int8. (f) Label efficiency on the 14 held-out\ndatasets (vegetation, urban) at 1/30/100% of the labels: \u2206 score vs. the best non-v2 baseline.\nedges TESSERA v1 on its own suite but falls behind it across the full 29 tasks, while TESSERA v2\nleads on both.\nGraceful degradation across students. Across the four students, d=16 performs at \u223c92% of\nthe d=128 composite at 1/8 of the storage, with d=32 and d=64 at \u223c97% and \u223c99% (Figure 4d).\nPerformance is best on land-cover and change-detection tasks and worst on fine-class-count and\nregression tasks, for which d=64 is the default. Regarding storage use, every student is on the\nPareto frontier of score versus storage and encoder parameters (Figure 4e): at 0.25\u00d7 the storage of\nA LPHA E ARTH, TESSERA v2-1B-S already matches it.\nHeld-out generalisation. The 14 held-out datasets split into a vegetation group (tree species, crop\nand parcel segmentation, biomass) and an urban group (land use, road density, population, GDP,\nnighttime lights, pollution, surface temperature, demographics). Across 1/30/100% label budgets\nTESSERA v2 beats the best non-v2 baseline on most of them, with the widest margins in the\nlow-label regime (Figure 4f).\nEmbedding perceptual quality and temporal stability. v1 embeddings often show along-track\nstriping and tile-seam discontinuities aligned with Sentinel-2/Sentinel-1 acquisition geometry. In\ncontrast, v2 embeddings have far fewer artefacts while preserving geographic structure. We also note\nthat, as long as land cover is stable, consecutive-year cosine distances for the same pixel are markedly\nlower for v2 than v1. Moreover, after a temporary disturbance v2 returns to its prior baseline where\nv1 stays noisy (Figure 1d).\n\n6\n\nD ISCUSSION\n\nIn language and vision models, the pretraining loss tracks downstream quality closely enough that\nscaling laws can be fit directly to it. EO breaks this assumption: selecting models by the loss wastes\n\n9\n\n\fPreprint.\n\nroughly a factor of three in compute (F1). Future EO scaling studies should therefore budget for\ndownstream evaluation rather than trust in the loss.\nOnce selection is downstream-task driven, F2 gives the compute budget allocation rule we use,\nwhich is to spend the budget on encoder capacity and matched data, holding the projector fixed,\nand recovering deployable encoders by distillation. Treating the teacher as a frozen representation\ndistribution then lets a single pretraining run amortise across four student sizes and, within each, four\nembedding dimensions.\nDistillation, not pretraining, is what makes the embedding dimension a usable control \"knob.\"\nSelf-supervision fixes only the subspace the embedding spans, leaving its coordinates unordered;\ndistillation against the frozen teacher imposes an order on them using a M ATRYOSHKA loss target, so\nthat each prefix is a usable lower-dimensional code.\nLimitations. Our work suffers from a few limitations. First, the scaling laws are empirical and apply\nonly to pixel-wise Sentinel-1/2 encoders, one self-supervised objective, and one 15-task evaluation\nsuite. Hence, our analysis of budget allocation is specific to this set of tasks. Second, the use of an\nexpensive teacher model requires substantial computation. Finally, our benchmarks are drawn from\nwell-studied regions, so generalisation to under-represented climates and unseen seasons remains to\nbe evaluated.\n\n7\n\nC ONCLUSION\n\nWe present the design, implementation, and analysis of the second generation of Tessera foundation\nmodel. Our principal research question is how to allocate a compute budget across the encoder and\nprojector. A 395-run downstream-driven scaling study found that pretraining loss is a weak predictor\nof downstream performance, encoder size and data scale together with compute, and projector size\ndoes not. Following this rule, we trained a 1 B pixel-wise teacher and distilled it into embeddingas-data students that lead the 15-task A LPHA E ARTH suite; the d=16 M ATRYOSHKA prefix keeps\n\u223c92% of the d=128 score at 1/8 of the storage. Finally, we found that M ATRYOSHKA-style nested\nembeddings cannot be learned by naive self-supervised prefix losses. Instead, we use distillation to\nturn prefix learning into supervised ordered compression, a technique we hope carries over to other\nEO embedding products.\nR EPRODUCIBILITY S TATEMENT\nWe will release training code, the controlled scaling-study sweeps, distilled pixel-wise student\ncheckpoints, the AlphaEarth-suite evaluation harness, and a frozen G EOT ESSERA-style embeddingas-data product. Hyperparameters, dataset splits, and compute estimates are documented alongside\nthe public code release.\nE THICS S TATEMENT\nTESSERA v2 is a generic representation model for publicly available Sentinel-1/2 imagery. We\nfollow standard responsible-release practices for publicly available foundation models.\nACKNOWLEDGMENT\nThe authors acknowledge the use of resources provided by the Isambard-AI National AI Research\nResource (AIRR). Isambard-AI is operated by the University of Bristol and is funded by the UK\nGovernment\u2019s Department for Science, Innovation and Technology (DSIT) via UK Research and\nInnovation; and the Science and Technology Facilities Council [ST/AIRR/I-A-I/1023].\n\nR EFERENCES\nGuillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. Omnisat: Self-supervised\nmodality fusion for earth observation. In European Conference on Computer Vision, pp. 409\u2013427.\nSpringer, 2024.\n\n10\n\n\fPreprint.\n\nGuillaume Astruc, Nicolas Gonthier, Clement Mallet, and Loic Landrieu. AnySat: One earth\nobservation model for many resolutions, scales, and modalities, 2025. URL https://arxiv.\norg/abs/2412.14123.\nWele Gedara Chaminda Bandara, Celso M. de Melo, and Vishal M. Patel. Guarding Barlow Twins\nagainst overfitting with mixed samples. arXiv preprint arXiv:2312.02151, 2023.\nFavyen Bastani, Piper Wolters, Ritwik Gupta, Joe Ferdinando, and Aniruddha Kembhavi. SatlasPretrain: A large-scale dataset for remote sensing image understanding. In Proceedings of the\nIEEE/CVF International Conference on Computer Vision (ICCV), 2023.\nChristopher F. Brown, Michal R. Kazmierski, Valerie J. Pasquarella, William J. Rucklidge, Masha\nSamsikova, Chenhui Zhang, Evan Shelhamer, Estefania Lahera, Olivia Wiles, Simon Ilyushchenko,\nNoel Gorelick, Lihui Lydia Zhang, Sophia Alj, Emily Schechter, Sean Askay, Oliver Guinan,\nRebecca Moore, Alexis Boukouvalas, and Pushmeet Kohli. AlphaEarth foundations: An embedding\nfield model for accurate and efficient global mapping from sparse label data, 2025. URL https:\n//arxiv.org/abs/2507.22291.\nShuang Chen, Jie Wang, Shuai Yuan, Jiayang Li, Yu Xia, Yuanhong Liao, Junbo Wei, Jincheng\nYuan, Xiaoqing Xu, Xiaolin Zhu, Peng Zhu, Hongsheng Zhang, Yuyu Zhou, Haohuan Fu, Huabing\nHuang, Bin Chen, Fan Dai, and Peng Gong. Democratizing planetary-scale analysis: An ultralightweight Earth embedding database for accurate and flexible global land monitoring. Earth\nSystem Science Data Discussions, 2026. doi: 10.5194/essd-2026-57. URL https://essd.\ncopernicus.org/preprints/essd-2026-57/. Preprint.\nYezhen Cong, Samar Khanna, Chenlin Meng, Patrick Liu, Erik Rozi, Yutong He, Marshall Burke,\nDavid B. Lobell, and Stefano Ermon. SatMAE: Pre-training transformers for temporal and multispectral satellite imagery. In Advances in Neural Information Processing Systems (NeurIPS),\n2022.\nNikolaos Dionelis, Riccardo Musto, Jente Bosmans, Simone Sarti, Giancarlo Paoletti, Peter Naylor,\nValerio Marsocci, S\u00e9bastien Lef\u00e8vre, Bertrand Le Saux, and Nicolas Long\u00e9p\u00e9. Scaling laws for\ngeospatial foundation models: A case study on PhilEO bench. arXiv preprint arXiv:2506.14765,\n2025.\nAlexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas\nUnterthiner, et al. An image is worth 16x16 words: Transformers for image recognition at scale.\nIn International Conference on Learning Representations (ICLR), 2021.\nHeng Fang, Adam J. Stewart, Isaac Corley, Xiao Xiang Zhu, and Hossein Azizpour. Earth embeddings\nas products: Taxonomy, ecosystem, and standardized access. arXiv preprint arXiv:2601.13134,\n2026.\nZhengpeng Feng, Clement Atzberger, Sadiq Jaffer, Jovana Knezevic, Silja Sormunen, Robin Young,\nMadeline C Lisaius, Markus Immitzer, Toby Jackson, James Ball, et al. Tessera: Temporal\nembeddings of surface spectra for earth representation and analysis. In Proceedings of the\nIEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 34818\u201334831, 2026.\nAnthony Fuller, Koreen Millard, and James Green. CROMA: Remote sensing representations\nwith contrastive radar-optical masked autoencoders. Advances in Neural Information Processing\nSystems (NeurIPS), 36:5506\u20135538, 2023.\nXin Guo, Jiangwei Lao, Bo Dang, Yingying Zhang, Lei Yu, Lixiang Ru, Liheng Zhong, Ziyuan\nHuang, Kang Wu, Dingxiang Hu, Huimei He, Jian Wang, Jingdong Chen, Ming Yang, Yongjun\nZhang, and Yansheng Li. SkySense: A multi-modal remote sensing foundation model towards\nuniversal interpretation for earth observation imagery. In Proceedings of the IEEE/CVF Conference\non Computer Vision and Pattern Recognition (CVPR), pp. 27662\u201327673, 2024. doi: 10.1109/\nCVPR52733.2024.02613.\nBoran Han, Shuai Zhang, Xingjian Shi, and Markus Reichstein. Bridging remote sensors with\nmultisensor geospatial foundation models. In Proceedings of the IEEE/CVF Conference on\nComputer Vision and Pattern Recognition (CVPR), pp. 27852\u201327862, 2024. doi: 10.1109/\nCVPR52733.2024.02631.\n11\n\n\fPreprint.\n\nHenry Herzog, Favyen Bastani, Yawen Zhang, Gabriel Tseng, Joseph Redmon, Hadrien Sablon,\nRyan Park, Jacob Morrison, Alexandra Buraczynski, Karen Farley, et al. Olmoearth: Stable latent\nimage modeling for multimodal earth observation. In Proceedings of the IEEE/CVF Conference\non Computer Vision and Pattern Recognition, pp. 34806\u201334817, 2026.\nGeoffrey Hinton, Oriol Vinyals, and Jeff Dean. Distilling the knowledge in a neural network. In\nNeurIPS Deep Learning and Representation Learning Workshop, 2015.\nJordan Hoffmann, Sebastian Borgeaud, Arthur Mensch, et al. Training compute-optimal large\nlanguage models. In Advances in Neural Information Processing Systems (NeurIPS), 2022.\nShuyang Hou, Haoyue Jiao, Ziqi Liu, Lutong Xie, Guanyu Chen, Shaowen Wu, Zhangyan Xu,\nZengjie Wang, Shaoqing Tang, Yaxian Qing, Jianyuan Liang, Xuefeng Guan, and Huayi Wu.\nAlphaearth foundations (aef) in earth observation: A systematic review of applications and practices.\nPreprints.org, 2026. doi: 10.20944/preprints202605.0981.v1.\nMichael Hu et al. Scaling laws are unreliable for downstream tasks: A reality check. In Findings of\nthe Association for Computational Linguistics: EMNLP 2025, 2025. arXiv:2507.00885.\nJared Kaplan, Sam McCandlish, Tom Henighan, Tom B. Brown, Benjamin Chess, Rewon Child,\nScott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. Scaling laws for neural language models.\narXiv preprint arXiv:2001.08361, 2020.\nKonstantin Klemmer, Esther Rolf, Marc Russwurm, Gustau Camps-Valls, Mikolaj Czerkawski,\nStefano Ermon, Alistair Francis, Nathan Jacobs, Hannah Rae Kerner, Lester Mackey, et al. Earth\nembeddings: Towards ai-centric representations of our planet. EarthArXiv preprint, 2025. doi:\n10.31223/X5HX9S.\nAditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham Kakade, Prateek Jain, and Ali Farhadi.\nMatryoshka representation learning. In Advances in Neural Information Processing Systems\n(NeurIPS), 2022.\nXuyang Li, Danfeng Hong, and Jocelyn Chanussot. S2MAE: A spatial-spectral pretraining foundation\nmodel for spectral remote sensing data. In Proceedings of the IEEE/CVF Conference on Computer\nVision and Pattern Recognition (CVPR), pp. 24088\u201324097, 2024.\nMadeline C. Lisaius, Andrew Blake, Srinivasan Keshav, and Clement Atzberger. Using Barlow\nTwins to create representations from cloud-corrupted remote sensing time series. IEEE Journal of\nSelected Topics in Applied Earth Observations and Remote Sensing, 17:13162\u201313168, 2024.\nJunyuan Liu, Xinglei Wang, Zichao Zeng, Jiazhuang Feng, Quan Qin, Ilya Ilyankou, Guangsheng\nDong, and Tao Cheng. CITYREP: A unified benchmark for"
 },
 "zeitgeist": [
  {
   "title": "OpenAI launches GPT-5.6 Sol, Terra, and Luna",
   "url": "https://openai.com/index/previewing-gpt-5-6-sol/",
   "summary": "OpenAI made its GPT-5.6 model family \u2014 Sol (flagship, for long-horizon agentic coding/cybersecurity/biology work), Terra (mid-tier, GPT-5.5-class performance at half price), and Luna (fast/cheap) \u2014 broadly available across ChatGPT, the API, and Codex on July 9, 2026. The launch followed a 12-day voluntary U.S. government review window; Sol also debuted on Cerebras wafer-scale hardware for select customers at up to 750 tokens/sec.",
   "published": "2026-07-09"
  },
  {
   "title": "SpaceXAI and Cursor launch Grok 4.5",
   "url": "https://www.bloomberg.com/news/articles/2026-07-08/spacexai-cursor-unveil-grok-ai-model-for-legal-finance-tasks",
   "summary": "SpaceXAI (xAI, absorbed into SpaceX in February 2026) released Grok 4.5, its first model built jointly with Cursor following SpaceX's pending $60B acquisition of the coding-tool startup. Musk pitched it as \"Opus-class\" but faster and cheaper, targeting coding, finance, and legal work; it's priced at $2/$6 per million input/output tokens and is not yet available in the EU.",
   "published": "2026-07-08"
  },
  {
   "title": "Anthropic finds an internal \"J-Space\" workspace inside Claude",
   "url": "https://www.anthropic.com/news",
   "summary": "Anthropic published interpretability research describing a privileged internal neural workspace in Claude, dubbed \"J-Space,\" identified via a Jacobian-based method (\"J-lens\"), that resembles global workspace theory's account of human conscious access. The company is explicit that this shows only functional \"access consciousness,\" not phenomenal consciousness, and frames the finding as a safety tool for detecting hidden reasoning like covert test-awareness or fabricated data. Google DeepMind's Neel Nanda independently reproduced the core result on an unrelated open-weight model (Qwen 3.6 27B).",
   "published": "2026-07-06"
  },
  {
   "title": "Claude Fable 5 redeployed globally after US export controls lifted",
   "url": "https://www.anthropic.com/news/redeploying-fable-5",
   "summary": "The US government had imposed export controls on Claude Fable 5 and Mythos 5 on June 12, 2026 after Amazon researchers showed Fable 5's safeguards could be bypassed to produce exploit code, forcing Anthropic to suspend both models worldwide. Controls were lifted June 30, and Fable 5 returned to Claude.ai, the API, Claude Code, and Claude Cowork on July 1, with Anthropic committing to new proactive security-detection and reporting obligations.",
   "published": "2026-07-01"
  },
  {
   "title": "FTC proposes policy statement on AI \"accuracy\" and ideological steering",
   "url": "https://www.ftc.gov/news-events/news/press-releases/2026/07/ftc-seeks-public-comment-policy-statement-addressing-ai-accuracy",
   "summary": "The FTC voted 2-0 to issue a proposed policy statement on the \"Suppression of Accuracy in Artificial Intelligence Systems,\" addressing whether AI firms that steer model outputs toward undisclosed ideological objectives violate Section 5 of the FTC Act. Issued under Executive Order 14365, it also targets conflicts between state laws mandating output alterations and federal law; public comment is open through July 31, 2026.",
   "published": "2026-07-01"
  },
  {
   "title": "Claude Cowork expands to web and mobile",
   "url": "https://techcrunch.com/2026/07/07/the-coding-agent-wars-are-spilling-into-the-rest-of-the-office-claude-cowork/",
   "summary": "Anthropic extended Claude Cowork beyond desktop to web and mobile, starting with Max-plan users, adding remote/cloud-run sessions, synced files, and scheduled tasks that continue running with no device online. The move follows a broader industry push (alongside OpenAI's Codex and others) to extend coding-agent-style tooling into general office work.",
   "published": "2026-07-07"
  }
 ]
}