pith. machine review for the scientific record. sign in

arxiv: 2604.18715 · v1 · submitted 2026-04-20 · 💻 cs.CL · cs.AI

Recognition: unknown

Characterizing AlphaEarth Embedding Geometry for Agentic Environmental Reasoning

Mashrekur Rahman , Samuel J. Barrett , Christina Last
Authors on Pith no claims yet

Pith reviewed 2026-05-10 05:09 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords AlphaEarth embeddingsmanifold geometrynon-Euclidean manifoldagentic reasoningenvironmental reasoningembedding retrievaltangent space rotationretrieval coherence
0
0 comments X

The pith

AlphaEarth embeddings occupy a twisting non-Euclidean manifold where local retrieval outperforms vector arithmetic for environmental reasoning.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper measures the geometry of 64-dimensional land-surface embeddings across 12 million Continental US locations from 2017 to 2023. It reports that the space collapses to an effective dimension of roughly 13 with local intrinsic dimension near 10, while tangent spaces rotate sharply and show almost no alignment with global directions. These properties make arithmetic on embedding vectors unreliable for answering land-related questions. The authors respond by building an agentic system that decomposes queries into retrieval steps over a database of the same embeddings and demonstrate that the retrieval component accounts for most of the gain in answer quality compared with a language model relying only on its parametric knowledge.

Core claim

The embeddings form a non-Euclidean manifold whose effective dimensionality is 13.3 by participation ratio and whose local intrinsic dimensionality is approximately 10. Tangent spaces rotate more than 60 degrees at 84 percent of locations, local-global alignment reaches only 0.17, and concept directions shift across the manifold according to linear probes. Compositional vector arithmetic therefore yields poor precision while retrieval produces physically coherent results whose quality is predicted by local geometry with R-squared of 0.32. An agentic system equipped with nine specialized tools that operate over a FAISS-indexed embedding database therefore achieves higher-quality responses to

What carries the argument

The nine-tool agentic system that decomposes environmental queries into reasoning chains over a FAISS-indexed embedding database, guided by measurements of manifold dimensionality, tangent rotation, and retrieval coherence.

Load-bearing premise

The geometric properties measured on Continental US 2017-2023 data and the performance gains from the nine-tool agentic system will generalize to other regions, time periods, or foundation models without being confounded by tool design or query selection.

What would settle it

Re-running the five-condition ablation on embeddings from a different continent or different years and finding that the retrieval-augmented agent no longer outperforms the parametric baseline.

Figures

Figures reproduced from arXiv: 2604.18715 by Christina Last, Mashrekur Rahman, Samuel J. Barrett.

Figure 1
Figure 1. Figure 1: Overview of the analysis and system architecture. (1) Prior work (Rahman 2026) established a dimension dictionary mapping AlphaEarth’s 64 embedding dimensions to environmental variables and built a FAISS-indexed retrieval pipeline over 12.1 million CONUS samples. (2) Phase 1 characterizes the manifold geometry through global covariance analysis, intrinsic dimensionality estimation, local PCA, and multi-sca… view at source ↗
Figure 2
Figure 2. Figure 2: Geometric structure of the AlphaEarth embedding space. (a) Cumulative variance explained by principal components, with the participation ratio of 13.3 annotating the effective dimensional￾ity. Markers indicate the number of components needed for 80%, 90%, and 95% of total variance. (b) Eigenvector weights for the top two principal components (PC1: moisture–vegetation, PC2: tem￾perature), colored by Rahman … view at source ↗
Figure 3
Figure 3. Figure 3: Intrinsic dimensionality of the AlphaEarth embedding manifold. (a) Three-dimensional PCA projection of 200,000 embedding vectors colored by local intrinsic dimensionality (Levina–Bickel MLE, 𝑘 = 20). (b) Same projection colored by elevation, showing that high-dimensionality regions correspond to topographically complex terrain. (c) Spatial distribution of local intrinsic dimensionality across CONUS (mean =… view at source ↗
Figure 4
Figure 4. Figure 4: Local geometry and tangent space analysis at 10,000 probe locations (𝑘 = 100 neighbors). (a) Alignment between local and global PC1 (| cos 𝜃|) across CONUS; warm colors indicate high alignment. (b) Tangent space instability measured as the angle between adjacent tangent spaces; 84% of locations exceed 60°. (c) Locally dominant environmental category at each probe, showing that temperature dominates 36% of … view at source ↗
Figure 5
Figure 5. Figure 5: Multi-scale geometric analysis at neighborhood sizes 𝑘 ∈ {20, 100, 500, 2000}. (a) Local– global alignment for PC1 (moisture–vegetation) and PC2 (temperature) as a function of 𝑘; alignment increases slowly but remains below 0.3 even at 𝑘 = 2,000. (b) Tangent angle and local participation ratio as functions of 𝑘. (c) Spatial map of PC1 alignment at 𝑘 = 20 (mean = 0.150). (d) Spatial map of PC1 alignment at … view at source ↗
Figure 6
Figure 6. Figure 6: Linear probes for concept directions across spatial scales. (a) Probe predictive accuracy (𝑅 2 ) at global, regional, and local scales for precipitation; local probes achieve the highest 𝑅 2 but with substantial variance across locations. (b) Distribution of | cos 𝜃| between local and global probe directions for precipitation, with median = 0.14 approaching the random baseline of 0.10; concept directions r… view at source ↗
Figure 7
Figure 7. Figure 7: Retrieval coherence and regional geometric profiles. (a) Spatial distribution of FAISS retrieval coherence across CONUS, measured as normalized spread of environmental variables among 𝑘 = 10 nearest neighbors; warmer colors indicate less coherent retrieval. Regional bounding boxes delineate the six subregions. (b) Local intrinsic dimensionality versus retrieval spread at 10,000 probe locations, with binned… view at source ↗
Figure 8
Figure 8. Figure 8: Regional structure of the AlphaEarth embedding manifold. (a) FAISS retrieval coherence across CONUS (normalized spread of environmental variables among 𝑘 = 10 nearest neighbors) with the six analysis subregions overlaid. (b–g) Three-dimensional PCA projections of the embedding space, with points belonging to each subregion highlighted in color and the remaining CONUS samples shown in gray. Each panel is an… view at source ↗
Figure 9
Figure 9. Figure 9: Ablation study results (Claude Sonnet 4.5 system, Gemma-3-27B judge). (a) Overall weighted scores by ablation condition, with the Rahman (2026) - (Paper 1) baseline (𝜇 = 3.74) shown as a dashed reference line. Error bars indicate one standard deviation. (b) Weighted scores by query tier and condition. Tier 2 (multi-step comparison) produces the highest scores across all agentic conditions; the deterministi… view at source ↗
read the original abstract

Earth observation foundation models encode land surface information into dense embedding vectors, yet the geometric structure of these representations and its implications for downstream reasoning remain underexplored. We characterize the manifold geometry of Google AlphaEarth's 64-dimensional embeddings across 12.1 million Continental United States samples (2017--2023) and develop an agentic system that leverages this geometric understanding for environmental reasoning. The manifold is non-Euclidean: effective dimensionality is 13.3 (participation ratio) from 64 raw dimensions, with local intrinsic dimensionality of approximately 10. Tangent spaces rotate substantially, with 84\% of locations exceeding 60\textdegree{} and local-global alignment (mean$|\cos\theta| = 0.17$) approaching the random baseline of 0.125. Supervised linear probes indicate that concept directions rotate across the manifold, and compositional vector arithmetic using both PCA-derived and probe-derived directions yields poor precision. Retrieval instead produces physically coherent results, with local geometry predicting retrieval coherence ($R^2 = 0.32$). Building on this characterization, we introduce an agentic system with nine specialized tools that decomposes environmental queries into reasoning chains over a FAISS-indexed embedding database. A five-condition ablation (120 queries, three complexity tiers) shows that embedding retrieval dominates response quality ($\mu = 3.79 \pm 0.90$ vs.\ $3.03 \pm 0.77$ parametric-only; scale 1--5), with peak performance on multi-step comparisons ($\mu = 4.28 \pm 0.43$). A cross-model benchmark show that geometric tools reduce Sonnet 4.5's score by 0.12 points but improve Opus 4.6's by 0.07, with Opus achieving higher geometric grounding (3.38 vs.\ 2.64), suggesting that the value of geometric characterization scales with the reasoning capability of the consuming model.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript characterizes the manifold geometry of Google AlphaEarth's 64-dimensional embeddings from 12.1 million Continental US samples (2017-2023). It reports an effective dimensionality of 13.3 via participation ratio, local intrinsic dimensionality of ~10, substantial tangent space rotations (84% >60°), and low local-global alignment (mean |cos θ| = 0.17). Vector arithmetic is shown to be imprecise, while retrieval coherence correlates with local geometry (R² = 0.32). The authors then present an agentic system with nine geometry-informed tools using FAISS retrieval, demonstrating via ablation on 120 queries that retrieval-augmented responses score higher (3.79 ± 0.90) than parametric-only (3.03 ± 0.77) on a 1-5 scale, with variations across models.

Significance. If the results hold, the work provides concrete empirical metrics on the non-Euclidean structure of Earth-observation foundation-model embeddings and illustrates how retrieval can outperform pure parametric reasoning in environmental query tasks. The cross-model benchmark and ablation design offer a useful template for evaluating geometry-aware agentic systems, though stronger isolation of the geometry contribution would increase impact.

major comments (3)
  1. [Ablation experiments] The five-condition ablation (abstract and associated results) compares the full nine-tool agentic system to a parametric-only baseline but lacks a control using plain FAISS kNN retrieval without the geometry-derived specializations such as coherence prediction or local tangent tools. This omission means the performance delta (μ = 3.79 vs. 3.03) cannot be confidently attributed to the geometric characterization rather than retrieval augmentation in general.
  2. [Geometric characterization results] The reported participation ratio of 13.3 and local intrinsic dimensionality of ~10 are central to the non-Euclidean manifold claim, yet the manuscript provides insufficient detail on the precise estimators, any post-hoc data filtering, or uncertainty quantification applied to the 12.1 million samples.
  3. [Retrieval coherence analysis] The R² = 0.32 correlation between local geometry and retrieval coherence is presented as evidence supporting tool design, but the paper does not test this relationship causally inside the agentic loop (e.g., via an ablation that disables geometry-informed components while retaining retrieval).
minor comments (2)
  1. [Abstract] Abstract: 'A cross-model benchmark show' should read 'shows'.
  2. [Discussion] The generalization discussion is brief; adding a short paragraph on expected behavior outside the Continental US 2017-2023 domain would improve clarity without altering the core claims.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed feedback, which highlights important opportunities to strengthen the attribution of results and the reproducibility of our geometric analyses. We address each major comment below and commit to revisions that directly respond to the concerns raised.

read point-by-point responses

  1. Referee: The five-condition ablation (abstract and associated results) compares the full nine-tool agentic system to a parametric-only baseline but lacks a control using plain FAISS kNN retrieval without the geometry-derived specializations such as coherence prediction or local tangent tools. This omission means the performance delta (μ = 3.79 vs. 3.03) cannot be confidently attributed to the geometric characterization rather than retrieval augmentation in general.

    Authors: We agree that the current ablation does not fully isolate the contribution of the geometry-informed tools from generic retrieval augmentation. The five conditions focus on variations in reasoning strategy and model choice but omit a plain FAISS kNN baseline. In the revised manuscript we will add this control condition to the ablation, enabling a clearer decomposition of performance gains attributable to the geometric specializations. revision: yes

  2. Referee: The reported participation ratio of 13.3 and local intrinsic dimensionality of ~10 are central to the non-Euclidean manifold claim, yet the manuscript provides insufficient detail on the precise estimators, any post-hoc data filtering, or uncertainty quantification applied to the 12.1 million samples.

    Authors: We acknowledge that additional methodological detail is required for full reproducibility. The revised methods section will explicitly describe the participation-ratio formula (trace of squared eigenvalues over sum of squared eigenvalues), the maximum-likelihood estimator and neighborhood size used for local intrinsic dimensionality, the absence of post-hoc filtering beyond the initial continental-US sampling, and uncertainty quantification via bootstrap resampling across spatial subsamples. revision: yes

  3. Referee: The R² = 0.32 correlation between local geometry and retrieval coherence is presented as evidence supporting tool design, but the paper does not test this relationship causally inside the agentic loop (e.g., via an ablation that disables geometry-informed components while retaining retrieval).

    Authors: The referee is correct that the reported correlation is observational and does not demonstrate causality within the agentic system. While the correlation guided tool selection, we did not ablate the geometry-derived components while keeping retrieval fixed. We will add this ablation in the revision, comparing the full nine-tool system against a retrieval-only variant with geometry tools disabled, to quantify the incremental benefit. revision: yes

Circularity Check

0 steps flagged

No significant circularity in geometric characterization or agentic ablation

full rationale

The paper measures manifold properties (participation ratio 13.3, local ID ~10, tangent rotations, alignment 0.17) via standard participation-ratio and local-PCA techniques on the 12.1M-sample dataset; these quantities are computed directly from the embeddings and do not presuppose the downstream agentic results. The reported R^2=0.32 link between local geometry and retrieval coherence is an empirical correlation obtained after the geometry is fixed, not a definitional reduction. The nine-tool agentic system is constructed after the geometry analysis, and the five-condition ablation (120 queries) reports a measured performance delta (3.79 vs 3.03) rather than a quantity forced by construction or by self-citation. No load-bearing step reduces to its own inputs, no uniqueness theorem is imported from the authors' prior work, and no ansatz is smuggled via citation. The derivation chain therefore remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The paper relies on standard manifold learning assumptions and the pre-existing AlphaEarth model; no new free parameters or invented entities are introduced in the abstract.

axioms (1)
  • domain assumption Embedding vectors lie on a differentiable manifold whose local geometry can be estimated via participation ratio and tangent space PCA
    Invoked for all dimensionality and rotation measurements.

pith-pipeline@v0.9.0 · 5660 in / 1413 out tokens · 65251 ms · 2026-05-10T05:09:34.077434+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

37 extracted references · 10 canonical work pages · 2 internal anchors

  1. [1]

    Physically Interpretable

    Rahman, Mashrekur , journal=. Physically Interpretable. 2026 , note=

    2026

  2. [2]

    and others , journal=

    Brown, Christopher F. and others , journal=

  3. [3]

    Nature , volume =

    Tollefson, Jeff , title =. Nature , volume =. 2025 , doi =

    2025

  4. [4]

    arXiv (2025)

    Xiao, Aoran and Xuan, Weihao and Wang, Junjue and Huang, Jiaxing and Tao, Dacheng and Lu, Shijian and Yokoya, Naoto , title =. arXiv preprint arXiv:2410.16602 , year =

    work page arXiv

  5. [5]

    arXiv preprint arXiv:2405.04285 , year=

    On the Foundations of Earth and Climate Foundation Models , author=. arXiv preprint arXiv:2405.04285 , year=

    work page arXiv

  6. [6]

    and Lucic, Ana and Stanley, Megan and Allen, Anna and Brandstetter, Johannes and Garvan, Patrick and Riechert, Maik and Weyn, Jonathan A

    Bodnar, Cristian and Bruinsma, Wessel P. and Lucic, Ana and Stanley, Megan and Allen, Anna and Brandstetter, Johannes and Garvan, Patrick and Riechert, Maik and Weyn, Jonathan A. and Dong, Haiyu and Gupta, Jayesh K. and Thambiratnam, Kit and Archibald, Alexander T. and Wu, Chun-Chieh and Heider, Elizabeth and Welling, Max and Turner, Richard E. and Perdik...

    2025

  7. [7]

    On the opportunities and challenges of foundation models for geospatial artificial intelligence,

    On the opportunities and challenges of foundation models for geospatial artificial intelligence , author=. arXiv preprint arXiv:2304.06798 , year=

    work page arXiv

  8. [8]

    , journal=

    Murakami, K. , journal=. Within- and Cross-Regional Crop Classification for Cool Climate Upland Agriculture Using

  9. [9]

    arXiv preprint arXiv:2510.09894 , year=

    Beyond AlphaEarth: toward human-centered spatial representation via POI-guided contrastive learning , author=. arXiv preprint arXiv:2510.09894 , year=

    work page arXiv

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    Maximum Likelihood Estimation of Intrinsic Dimension , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    International Conference on Learning Representations (ICLR) Workshop , year=

    Efficient Estimation of Word Representations in Vector Space , author=. International Conference on Learning Representations (ICLR) Workshop , year=

  12. [12]

    Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    Linguistic Regularities in Continuous Space Word Representations , author=. Proceedings of the 2013 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies , pages=

    2013

  13. [13]

    Retrieval-Augmented Generation for Knowledge-Intensive

    Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume=

  14. [14]

    Billion-Scale Similarity Search with

    Johnson, Jeff and Douze, Matthijs and J. Billion-Scale Similarity Search with. IEEE Transactions on Big Data , volume=. 2019 , doi=

    2019

  15. [15]

    and Stoica, Ion , booktitle=

    Zheng, Lianmin and Chiang, Wei-Lin and Sheng, Ying and Zhuang, Siyuan and Wu, Zhanghao and Zhuang, Yonghao and Lin, Zi and Li, Zhuohan and Li, Dacheng and Xing, Eric and Zhang, Hao and Gonzalez, Joseph E. and Stoica, Ion , booktitle=. Judging

  16. [16]

    International Journal of Digital Earth , volume=

    GPT, large language models (LLMs) and generative artificial intelligence (GAI) models in geospatial science: a systematic review , author=. International Journal of Digital Earth , volume=. 2024 , publisher=

    2024

  17. [17]

    Zhang, Yifan and Wei, Cheng and Wu, Shangyou and He, Zhengting and Yu, Wenhao , journal=

  18. [18]

    and Ye, X

    Xu, Yue and Kibria, Golam and Peeta, Srinivas and Wang, G. and Ye, X. and Yang, Y. and Ding, Y. , booktitle=. Agentic

  19. [19]

    Big Earth Data , pages=

    An LLM-based multi-agent system for remote sensing analysis , author=. Big Earth Data , pages=. 2026 , publisher=

    2026

  20. [20]

    Empowering LLM Agents with Geospatial Awareness: Toward Grounded Reasoning for Wildfire Response

    Empowering LLM Agents with Geospatial Awareness: Toward Grounded Reasoning for Wildfire Response , author=. arXiv preprint arXiv:2510.12061 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  21. [21]

    International Conference on Learning Representations (ICLR) , year=

    Earth-Agent: Unlocking the Full Landscape of Earth Observation with Agents , author=. International Conference on Learning Representations (ICLR) , year=

  22. [22]

    Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle=

  23. [23]

    Advances in Neural Information Processing Systems , volume=

    Toolformer: Language Models Can Teach Themselves to Use Tools , author=. Advances in Neural Information Processing Systems , volume=

  24. [24]

    Nature Communications , volume=

    Spatial Validation Reveals Poor Predictive Performance of Large-Scale Ecological Mapping Models , author=. Nature Communications , volume=

  25. [25]

    Ecography , volume=

    Cross-Validation Strategies for Data with Temporal, Spatial, Hierarchical, or Phylogenetic Structure , author=. Ecography , volume=. 2017 , doi=

    2017

  26. [26]

    arXiv preprint arXiv:2503.19786 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    2024 , url=

    Dartmouth Chat: Institutional. 2024 , url=

    2024

  28. [28]

    2017 , doi=

    Gorelick, Noel and Hancher, Matt and Dixon, Mike and Ilyushchenko, Simon and Thau, David and Moore, Rebecca , journal=. 2017 , doi=

    2017

  29. [29]

    How Contextual Are Contextualized Word Representations?

    Ethayarajh, Kawin , booktitle=. How Contextual Are Contextualized Word Representations?

  30. [30]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=

    Representation Learning: A Review and New Perspectives , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume=. 2013 , doi=

    2013

  31. [31]

    Jakubik, Johannes and Roy, Sujit and Phillips, C. E. and Fraccaro, Paolo and Godwin, Denys and Zadrozny, Bianca and Szwarcman, Daniela and Gomes, Carlos and Nyirjesy, Gabby and Edwards, Blair and others , title =. arXiv preprint arXiv:2310.18660 , year =

    work page arXiv

  32. [32]

    Reed, Colorado J. and Gupta, Ritwik and Li, Shufan and Brockman, Sarah and Funk, Christopher and Clipp, Brian and Keutzer, Kurt and Candido, Salvatore and Uyttendaele, Matt and Darrell, Trevor , title =. arXiv preprint arXiv:2212.14532 , year =

    work page arXiv

  33. [33]

    2024 , url=

    Clay Foundation Model , author=. 2024 , url=

    2024

  34. [34]

    arXiv preprint arXiv:2601.13134 , year=

    Earth Embeddings as Products: Taxonomy, Ecosystem, and Standardized Access , author=. arXiv preprint arXiv:2601.13134 , year=

    work page arXiv

  35. [35]

    International Conference on Learning Representations , year=

    All-but-the-Top: Simple and Effective Postprocessing for Word Representations , author=. International Conference on Learning Representations , year=

  36. [36]

    bioRxiv , year=

    A Theory of Multineuronal Dimensionality, Dynamics and Measurement , author=. bioRxiv , year=

  37. [37]

    Measuring the Intrinsic Dimension of

    Rao, Arjun and Ru. Measuring the Intrinsic Dimension of. arXiv preprint arXiv:2511.02101 , year=

    work page arXiv