Pith. sign in

REVIEW 4 major objections 8 minor 58 references

TTCov curates training data by matching the deployment test distribution, expressed as a Knowledge Atlas of atomic driving propositions, and achieves near-parity with (and beyond) a manually curated set on a hard closed-loop driving benchma

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-01 20:33 UTC pith:DQ67ZWY3

load-bearing objection A genuinely new idea — curating the training set from unlabeled test data before training — is executed cleanly, but the headline performance gains over baselines are within noise and the comparison lacks a control that would isolate the mechanism. the 4 major comments →

arxiv 2607.22697 v1 pith:DQ67ZWY3 submitted 2026-07-18 cs.AI cs.CVcs.LG

Test-Time Coverage: Test-Conditioned Data Curation for Deployment-Aware Learning

Nadine Chang , Maying Shen , Shizhe Diao , Jialiang Wang , Jingde Chen , Thomas Breuel , Pavlo Molchanov , Rafid Mahmood
show 1 more author
Jose M. Alvarez
This is my paper
classification cs.AI cs.CVcs.LG
keywords test-time coveragedata curationdeployment-aware learningatomic propositionsknowledge atlasdistribution matchingautonomous drivingKL divergence
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper tackles the problem of choosing which training examples to keep when a deployed AI system will face a specific, known deployment distribution rather than a generic one. It argues that instead of scoring training samples on diversity or uncertainty, the curation objective should directly ask whether the subset covers the concepts the deployment environment will exercise, and with the same frequencies. To make that concrete, TTCov builds a "Knowledge Atlas" from the deployment test data plus open-world driving knowledge, then selects the budgeted training subset whose own Atlas distribution is closest to it in KL divergence. The paper shows that this test-conditioned subset yields better closed-loop driving scores than training-side curation baselines and can be extended city by city without re-curating earlier data. If true, the implication is that unlabeled deployment logs are an underused signal that should drive data curation before training, not only model adaptation at inference.

Core claim

TTCov's central claim is that deployment-aware data curation can be decomposed into coverage and distribution and solved before training. Coverage is represented by an Atlas: a deduplicated set of atomic propositions extracted by an LLM from open driving knowledge and from unlabeled deployment videos. Distribution is represented by the K-Atlas, the normalized frequency of those propositions across the deployment set. The paper claims that greedily selecting training samples to minimize the KL divergence between the selected subset's AP distribution and the K-Atlas produces a training set that is closer to the deployment distribution than standard curation baselines, and that on a challenging

What carries the argument

The Atlas is a set of atomic propositions (APs), single-clause declarative sentences about driving scenarios, agents, conditions, and behaviors (e.g., 'pedestrian crosses crosswalk'). APs are extracted from LLM-generated captions of deployment videos, unified across paraphrases by an LLM, and seeded with open-world driving ontologies. The K-Atlas is the frequency-weighted histogram of APs over deployment samples, which serves as the curation target. The selection algorithm greedily adds the candidate training sample that most reduces KL divergence between the running selected-set AP distribution and the K-Atlas target.

Load-bearing premise

The load-bearing premise is that the LLM-based extraction and the available deployment test set together capture every deployment-relevant scenario class; any class that is missing from both cannot be targeted, and curation will remain blind to it.

What would settle it

Hold out a scenario class that is prevalent in the live deployment but absent from the test set and from the open-world seed Atlas, then compare TTCov against a baseline on logs dominated by that class. A large EPDMS drop for TTCov would confirm that curation is only as good as the Atlas coverage; alternately, measuring the marginal EPDMS gain from adding just that class's APs to the Atlas would quantify how much of the improvement is driven by the coverage component versus the distribution matching.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • At the standard budget, test-conditioned curation reaches near-parity with a manually curated oracle (24.42 vs 24.70 EPDMS), suggesting the Atlas target captures much of what human curation filters for.
  • At larger budgets the curated subset surpasses the oracle (26.40 at 1.5x), so the method does not just recover human judgment; it extends it.
  • When the deployment set changes (new city), TTCov keeps previously selected data and adds only new relevant samples, improving scores on both old and new cities without full re-curation.
  • The distribution-matching objective correlates with downstream performance: lower KL divergence to the K-Atlas corresponds to higher EPDMS across metrics.
  • Combining K-Atlas matching with a long-tail reweighting of rare APs improves EPDMS further (26.15), indicating the framework is compatible with standard long-tail techniques.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • A testable extension beyond driving: any domain with unlabeled deployment logs and an LLM captioning channel could use the same Atlas/K-Atlas pipeline to curate fine-tuning data, e.g., robotics or document processing.
  • The paper's own limitation implies an upper bound: if a safety-critical scenario is absent from both the test set and the open-world seed, TTCov cannot select for it. Editors could test this by measuring performance on a held-out scenario class that was deliberately excluded from Atlas construction.
  • The strong result with repeat-factor sampling suggests the proportional K-Atlas may underweight rare APs; a production system might combine TTCov with explicit rare-event over-sampling and evaluate the trade-off against the KL objective.
  • TTCov is a data-level counterpart to test-time adaptation; if the correlation between distribution matching and performance holds, one could compare the two strategies directly on the same deployment stream.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 8 minor

Summary. The paper proposes TTCov, a data-curation method that uses unlabeled deployment/test data before training. It constructs an 'Atlas' of atomic propositions (APs) by prompting an LLM/MLLM over open driving knowledge and test videos, then forms a K-Atlas frequency distribution over APs from the test set. A budgeted training subset is selected by greedily minimizing the KL divergence between the selected set's AP distribution and the K-Atlas target. Experiments on Navhard with Latent Transfuser compare TTCov against Coreset, SSE, and Random across 0.5x-1.5x budgets, reporting EPDMS, distribution-matching metrics, and city-to-city expansion. The paper claims that TTCov achieves higher deployment coverage, closer K-Atlas matching, and stronger downstream E2E driving performance than baselines, with parity to the manually curated Navtrain oracle at 1x budget.

Significance. If the empirical claims were fully supported, TTCov would be a valuable contribution: it is one of the few curation methods that directly injects deployment-side information into data selection without test-time weight updates, and the Atlas/K-Atlas representation is interpretable and actionable. The paper is transparent in providing prompts (Fig. 11), a full sub-metric table with standard deviations (Table 8), and ablations of the greedy optimizer and LLM extraction thresholds, which aid reproducibility. However, the current evidence does not establish the central superiority claim: pairwise EPDMS differences are within sampling noise, no test-conditioned control baseline isolates the Atlas/KL mechanism, and the distribution-matching metrics partly report success on the very objective being optimized. The method is promising but the empirical case is incomplete.

major comments (4)
  1. [Table 8 / Table 2 (Sec. 5.1)] The headline 'consistent performance gains' is not supported statistically. With three seeds per condition, TTCov's 1x EPDMS (24.42±0.66) differs from SSE (23.45±0.53) by 0.97; a two-sample t-test gives p≈0.11, and vs Coreset (23.63±1.19) p≈0.36. Across all budgets, no pairwise difference reaches p<0.05, and at 0.5x TTCov (20.62±0.52) is numerically worse than Coreset (20.77±0.40), contradicting 'consistent performance gains over all baselines'. Please report confidence intervals or significance tests, increase the number of seeds, or temper the claims to observed trends.
  2. [Eq. (1)-(2), Algorithm 1, Table 2] The baseline comparison is structurally asymmetric. Coreset and SSE are test-blind; TTCov's objective is built directly from the K-Atlas of the test set (Eq. 1), and Algorithm 1 greedily matches that target. Any method given test-side access—e.g., selecting training samples nearest to test embeddings, or matching the test feature histogram—would be expected to improve over test-blind curation. A test-conditioned control is needed to attribute the Table 2 gains to the Atlas/KL mechanism rather than to generic transductive benefit. Without such a control, the causal claim that TTCov outperforms baselines is not established.
  3. [Table 4, Sec. 3.2, Fig. 4] The distribution-matching and coverage metrics are partly circular. The KL, JS, Hellinger, and cosine values in Table 4 compare the selected set's K-Atlas to the same target p* that the greedy procedure directly minimizes (Eq. 2); Fig. 4 is a diagnostic of optimization success, not independent evidence. Likewise, Table 1's NN/MMD metrics use embeddings produced by the same caption/embedding pipeline that generates the APs, so these metrics overlap with the selection objective. To support claims of 'closest distributional match', please report an external or held-out evaluation of AP coverage, or at least frame Table 4 as an objective-fit check.
  4. [Table 3, Sec. 5.1] The city-expansion claim that 'TTCov significantly outperforms all baselines' is reported without standard deviations, confidence intervals, or number of seeds. Round 2 City 2 shows a 6.9-point gap (38.11 vs 31.18), but without variance information the significance cannot be assessed. The claim of avoiding 'catastrophic forgetting' is also only supported by a single EPDMS number for City 1 in Round 2. Please report repeated runs or treat these results as preliminary.
minor comments (8)
  1. [Table 2 vs Table 8] The oracle Navtrain EPDMS is reported as 24.70 in Table 2 but 24.49±1.41 in Table 8. Please reconcile these numbers.
  2. [Eq. (2), Sec. 3.2] Typo: 'the goal fo TTCov' should be 'the goal of TTCov'. Also, Algorithm 1's title says 'KL coverage' but the objective is KL distribution matching; consider renaming for clarity.
  3. [Table 4] The 'Cov.' column is not defined in the caption or text. Please define it or remove it.
  4. [Fig. 10] Captions read '(1) Round 1 (2) Round 1'; likely should be Round 1 and Round 2. Also, no caption explains the right panel.
  5. [Sec. 7] The limitation that TTCov cannot capture scenarios absent from the test set or open-world knowledge is acknowledged in Sec. 7 but should be reflected in the abstract and conclusion, where 'deployment-relevant coverage' and 'seamless adaptability' are stated without this caveat.
  6. [Sec. 4.2, Table 1] The text says TTCov yields 'significantly more' selected points and 'lower MMD', but no statistical comparison or error bars are reported for these coverage metrics.
  7. [Table 6, Sec. B.1] The Atlas threshold ablation reports single runs; without repeated runs the robustness claim is not quantitatively supported.
  8. [Sec. G] The code release states code 'will be released alongside the final version'; for a reproducibility-focused review, please provide an anonymous link or detailed pseudocode for the Atlas expansion and AP matching steps in a supplementary appendix.

Circularity Check

1 steps flagged

Distribution-matching metrics are tautological relative to the K-Atlas objective; the headline EPDMS result is an independent planner benchmark and is not circular.

specific steps
  1. self definitional [Eq. (1), Eq. (2), Section 5.1, Table 4]
    "S ⋆ := arg min S⊂C,|S|=B DKL(p⋆ ∥ ˆpS ). (2) ... By targeting our K-Atlas test distribution comprised of relevant APs, TTCov curates a training set through an optimizer that minimizes the distributional distance between the selected data and the target K-Atlas."

    The K-Atlas target p* in Eq. (1) is defined as the AP frequency distribution of the test samples. Eq. (2) then selects the curated set S by minimizing D_KL(p* || p_hat_S). Table 4 reports KL divergence, JS distance, Hellinger distance, and cosine similarity between the selected set and this same p*. Minimizing the exact objective that is later reported as an evaluation measure guarantees a low value; TTCov's smaller KL/JS/Hellinger/cosine numbers compared to methods optimizing different criteria are a property of the optimizer, not independent evidence that the selected distribution matches deployment. The downstream EPDMS score is computed by an external planner benchmark and does not reduce to p*.

full rationale

The paper's central derivation is not circular at the level of the downstream EPDMS claim: training subsets are selected by matching a test-derived K-Atlas, models are trained, and performance is measured on the Navsim/Navhard planner benchmark, which is external to the curation objective. That part is self-contained and falsifiable. However, the paper's 'closer K-Atlas matching' evidence is partially circular: Algorithm 1 and Eq. (2) minimize D_KL(p* || p_hat_S), and Table 4/Figure 4 then report distances to the same p* as evidence of superiority. Those numbers are forced by the objective. The coverage metrics (NN@0.15, MMD) are not literally the optimization target but are made easier by test-side access, and the paper lacks a test-conditioned control that would isolate the Atlas/KL mechanism from the generic benefit of transductive test information. This is an experimental-design concern rather than a derivation-level circularity. There is also no load-bearing reliance on self-citations: the cited same-author works are motivational or baselines, not the mechanism justifying the result. Overall, the distribution-matching results are partially circular, but the central EPDMS comparison remains independent, so the paper does not warrant a high circularity score.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

TTCov introduces no new physical entities. The Atlas and K-Atlas are conceptual/semantic constructs built from LLM outputs, not independently postulated objects. The free parameters are all hyperparameters, not fitted to downstream EPDMS, and the ablation shows robustness to the main threshold. The key assumptions are about LLM reliability and the representativeness of the test/seed knowledge.

free parameters (3)
  • AP matching cosine threshold = 0.85
    Used in Faiss-based AP-to-Atlas matching; ablated in Table 6 with EPDMS stable across 0.83–0.87, so not a driver of results.
  • RFS threshold t (ablation only) = 1e-2
    Used only in the repeat-factor-sampling ablation (Sec. D.2); not part of main results.
  • Over-coverage penalty rho (ablation only) = 1.0
    Ablated in Table 5; rho=1 (no penalty) is used for main results.
axioms (4)
  • domain assumption LLM-based AP extraction faithfully captures deployment-relevant concepts from video captions
    The entire Atlas and K-Atlas are built from Gemini 2.5 Pro captions and Qwen embeddings; there is no human validation of the extracted APs. Invoked throughout Sec. 3.1.
  • domain assumption The test set (Navhard) is representative of the deployment distribution
    TTCov is evaluated on the same test set used to build the K-Atlas; the method cannot handle deployment conditions absent from both the test set and the LLM seed knowledge (Sec. 7).
  • domain assumption Matching the K-Atlas AP distribution leads to better end-to-end driving performance
    The optimization objective (Eq. 2) assumes AP-frequency matching is a good proxy for planner quality; the paper supports this with correlational evidence but does not prove causality.
  • standard math Greedy sequential minimization approximates the global optimum of the KL objective
    Greedy is a standard heuristic for submodular-ish objectives; no optimality guarantee is given (Algorithm 1).

pith-pipeline@v1.3.0-alltime-deepseek · 22615 in / 9113 out tokens · 98179 ms · 2026-08-01T20:33:34.956903+00:00 · methodology

0 comments
read the original abstract

Deployed AI systems are often trained from broad candidate data pools, necessitating data curation towards the deployment test distribution. However, standard data curation methods score training-side criteria rather than directly optimizing deployment match. We introduce TTCov (Test-Time Coverage), a data-level test-conditioned curation method that uses test-side information before training instead of updating model weights at inference. TTCov decomposes deployment-conditioned curation into coverage and distribution. To represent coverage, it builds a task Atlas, a collection of LLM-based atomic propositions (APs) describing deployment-relevant concepts, seeded from open task knowledge and expanded with unmatched APs extracted from unlabeled deployment samples. To represent distribution, it instantiates the matched deployment APs with their frequencies, yielding a Knowledge Atlas (K-Atlas) that operationalizes the deployment distribution as a curation target. TTCov then selects a budgeted training set whose deployment APs distribution approximates this target. We apply TTCov towards autonomous driving (AD), keeping adaptation off the inference path while selecting data with greater deployment-relevant coverage, closer K-Atlas matching, and stronger downstream end-to-end driving performance than data-curation baselines, including seamless adaptability to novel domains via city-to-city expansion.

Figures

Figures reproduced from arXiv: 2607.22697 by Jialiang Wang, Jingde Chen, Jose M. Alvarez, Maying Shen, Nadine Chang, Pavlo Molchanov, Rafid Mahmood, Shizhe Diao, Thomas Breuel.

Figure 1
Figure 1. Figure 1: The Atlas consists of a collection of LLM-based atomic propositions (APs) describing [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: TTCov Overview. Top: The AP extraction process transforms complex descriptions into [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: TTCov’s selected data coverage over first test city. We show the histogram of the number [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: KL div to K-Atlas distribution [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visual samples of TTCov’s selected data with diverse relevant APs from our K-Atlas, [PITH_FULL_IMAGE:figures/full_fig_p009_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Samples of complex triplets in knowledge graph. [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Samples of unnecessary modifiers, which we removed in our final de-noising phrase [PITH_FULL_IMAGE:figures/full_fig_p015_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Samples of decomposed APs prior to de-noising and unification from LLM generated [PITH_FULL_IMAGE:figures/full_fig_p015_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: TTCov’s selected data coverage over the 2nd test city. We show the histogram of the [PITH_FULL_IMAGE:figures/full_fig_p016_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: The geography distribution of selected data during the first and second round of city [PITH_FULL_IMAGE:figures/full_fig_p016_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: All relevant prompts used throughout TTCov’s process. [PITH_FULL_IMAGE:figures/full_fig_p017_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Additional visualization examples of TTCov selected data. [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗

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

58 extracted references · 11 linked inside Pith

  1. [1]

    Semd- edup: Data-efficient learning at web-scale through semantic deduplication.arXiv preprint arXiv:2303.09540, 2023

    Amro Abbas, Kushal Tirumala, Dániel Simig, Surya Ganguli, and Ari S Morcos. Semd- edup: Data-efficient learning at web-scale through semantic deduplication.arXiv preprint arXiv:2303.09540, 2023

    Pith/arXiv arXiv 2023

  2. [2]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  3. [3]

    Deep batch active learning by diverse, uncertain gradient lower bounds.arXiv preprint arXiv:1906.03671, 2019

    Jordan T Ash, Chicheng Zhang, Akshay Krishnamurthy, John Langford, and Alekh Agar- wal. Deep batch active learning by diverse, uncertain gradient lower bounds.arXiv preprint arXiv:1906.03671, 2019

    Pith/arXiv arXiv 1906

  4. [4]

    Manifold regularization: A geometric framework for learning from labeled and unlabeled examples.JMLR, 7:2399–2434, 2006

    Mikhail Belkin, Partha Niyogi, and Vikas Sindhwani. Manifold regularization: A geometric framework for learning from labeled and unlabeled examples.JMLR, 7:2399–2434, 2006

    2006

  5. [5]

    Bennett and Ayhan Demiriz

    Kristin P. Bennett and Ayhan Demiriz. Semi-supervised support vector machines. InNeurIPS, 1998

    1998

  6. [6]

    Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom

    Holger Caesar, Varun Bankiti, Alex H. Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuScenes: A multimodal dataset for autonomous driving. InCVPR, pages 11621–11631, 2020

    2020

  7. [7]

    Pseudo-simulation for autonomous driving.arXiv preprint arXiv:2506.04218, 2025

    Wei Cao, Marcel Hallgarten, Tianyu Li, Daniel Dauner, Xunjiang Gu, Caojun Wang, Yakov Miron, Marco Aiello, Hongyang Li, Igor Gilitschenski, et al. Pseudo-simulation for autonomous driving.arXiv preprint arXiv:2506.04218, 2025

    arXiv 2025

  8. [8]

    Nadine Chang, Maying Shen, Jialiang Wang, Rafid Mahmood, and Jose M. Alvarez. Position: Stop reactively patching your model every time and start proactive test-driven ai development. InICML, 2026

    2026

  9. [9]

    Vadv2: End-to-end vectorized autonomous driving via probabilistic planning.arXiv preprint arXiv:2402.13243, 2024

    Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning.arXiv preprint arXiv:2402.13243, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    TransFuser: Imitation with transformer-based sensor fusion for autonomous driving.TPAMI, 45(11):12878–12895, 2023

    Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. TransFuser: Imitation with transformer-based sensor fusion for autonomous driving.TPAMI, 45(11):12878–12895, 2023

    2023

  11. [11]

    End-to-end driving via conditional imitation learning

    Felipe Codevilla, Matthias Müller, Antonio López, Vladlen Koltun, and Alexey Dosovitskiy. End-to-end driving via conditional imitation learning. InICRA, pages 4693–4700. IEEE, 2018

    2018

  12. [12]

    Selection via proxy: Efficient data selection for deep learning

    Cody Coleman, Christopher Yeh, Stephen Mussmann, Baharan Mirzasoleiman, Peter Bailis, Percy Liang, Jure Leskovec, and Matei Zaharia. Selection via proxy: Efficient data selection for deep learning. InICLR, 2020. 10

    2020

  13. [13]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Gheorghe Comanici, Eric Bieber, Mike Schaekermann, Ice Pasupat, Noveen Sachdeva, Inderjit Dhillon, Marcel Blistein, Ori Ram, Dan Zhang, Evan Rosen, et al. Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities.arXiv preprint arXiv:2507.06261, 2025

    Pith/arXiv arXiv 2025

  14. [14]

    Openscene: The largest up-to-date 3d occupancy prediction bench- mark in autonomous driving.https://github.com/OpenDriveLab/OpenScene, 2023

    OpenScene Contributors. Openscene: The largest up-to-date 3d occupancy prediction bench- mark in autonomous driving.https://github.com/OpenDriveLab/OpenScene, 2023

    2023

  15. [15]

    Michael J. Coren. Tesla has 780 million miles of driving data, and adds another million every 10 hours. Quartz, 2016. Published May 28, 2016. Accessed: 2026-05-06

    2016

  16. [16]

    Parting with miscon- ceptions about learning-based vehicle motion planning

    Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with miscon- ceptions about learning-based vehicle motion planning. InCoRL, 2023

    2023

  17. [17]

    NA VSIM: Data-driven non-reactive autonomous vehicle simulation and benchmarking

    Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. NA VSIM: Data-driven non-reactive autonomous vehicle simulation and benchmarking. InNeurIPS, volume 37, 2024

    2024

  18. [18]

    Tolga Dimlioglu, Nadine Chang, Maying Shen, Rafid Mahmood, and Jose M. Alvarez. Scaling- aware data selection for end-to-end autonomous driving systems. InCVPR, 2026. Accepted

    2026

  19. [19]

    The Faiss library.TBD, pages 1–17, 2025

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre- Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. The Faiss library.TBD, pages 1–17, 2025

    2025

  20. [20]

    Scaling laws for data filtering–data curation cannot be compute agnostic

    Sachin Goyal, Pratyush Maini, Zachary C Lipton, Aditi Raghunathan, and J Zico Kolter. Scaling laws for data filtering–data curation cannot be compute agnostic. InCVPR, pages 22702–22711, 2024

    2024

  21. [21]

    Training AI for self-driving vehicles: The challenge of scale

    Adam Grzywaczewski. Training AI for self-driving vehicles: The challenge of scale. NVIDIA Technical Blog, Oct. 2017. Accessed: 2026-05-06

    2017

  22. [22]

    Lvis: A dataset for large vocabulary instance segmentation

    Agrim Gupta, Piotr Dollar, and Ross Girshick. Lvis: A dataset for large vocabulary instance segmentation. InCVPR, pages 5356–5364, 2019

    2019

  23. [23]

    Test-time training on nearest neighbors for large language models

    Moritz Hardt and Yu Sun. Test-time training on nearest neighbors for large language models. InICLR, 2024

    2024

  24. [24]

    Vad: Vectorized scene representation for efficient autonomous driving

    Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. InICCV, pages 8340–8350, October 2023

    2023

  25. [25]

    Transductive inference for text classification using support vector machines

    Thorsten Joachims. Transductive inference for text classification using support vector machines. InICML, 1999

    1999

  26. [26]

    Multi-class active learning for image classification

    Ajay J Joshi, Fatih Porikli, and Nikolaos Papanikolopoulos. Multi-class active learning for image classification. InCVPR, pages 2372–2379. IEEE, 2009

    2009

  27. [27]

    Adadedup: Adaptive hybrid data pruning for efficient large-scale object detection training.arXiv preprint arXiv:2507.00049, 2025

    Feiyang Kang, Nadine Chang, Maying Shen, Marc T Law, Rafid Mahmood, Ruoxi Jia, and Jose M Alvarez. Adadedup: Adaptive hybrid data pruning for efficient large-scale object detection training.arXiv preprint arXiv:2507.00049, 2025

    Pith/arXiv arXiv 2025

  28. [28]

    Lewis and Jason Catlett

    David D. Lewis and Jason Catlett. Heterogeneous uncertainty sampling for supervised learning. InICML, pages 148–156. Elsevier, 1994

    1994

  29. [29]

    Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation.arXiv preprint arXiv:2406.06978, 2024

    Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi-target hydra-distillation.arXiv preprint arXiv:2406.06978, 2024

    Pith/arXiv arXiv 2024

  30. [30]

    Ztrs: Zero-imitation end-to-end autonomous driving with trajectory scoring.arXiv preprint arXiv:2510.24108, 2025

    Zhenxin Li, Wenhao Yao, Zi Wang, Xinglong Sun, Jingde Chen, Nadine Chang, Maying Shen, Jingyu Song, Zuxuan Wu, Shiyi Lan, et al. Ztrs: Zero-imitation end-to-end autonomous driving with trajectory scoring.arXiv preprint arXiv:2510.24108, 2025. 11

    Pith/arXiv arXiv 2025

  31. [31]

    Is ego status all you need for open-loop end-to-end autonomous driving? InCVPR, pages 14864–14873, 2024

    Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? InCVPR, pages 14864–14873, 2024

    2024

  32. [32]

    KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d.TPAMI, 45(3):3292–3310, 2023

    Yiyi Liao, Jun Xie, and Andreas Geiger. KITTI-360: A novel dataset and benchmarks for urban scene understanding in 2d and 3d.TPAMI, 45(3):3292–3310, 2023

    2023

  33. [33]

    Mingyu Liu, Ekim Yurtsever, Jonathan Fossaert, Xingcheng Zhou, Walter Zimmer, Yuning Cui, Bare Luka Zagar, and Alois C. Knoll. A survey on autonomous driving datasets: Statistics, annotation quality, and a future outlook.TIV, 9(11):7138–7164, 2024

    2024

  34. [34]

    TTT++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808–21820, 2021

    Yuejiang Liu, Parth Kothari, Bastien van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In NeurIPS, volume 34, pages 21808–21820, 2021

    2021

  35. [35]

    Kanatsoulis, and Sanmi Koyejo

    Belinda Mo, Kyssen Yu, Joshua Kazdan, Proud Mpala, Lisa Yu, Charilaos I. Kanatsoulis, and Sanmi Koyejo. KGGen: Extracting knowledge graphs from plain text with language models. In NeurIPS, 2025

    2025

  36. [36]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. InICML, volume 162, pages 16888–16905, 2022

    2022

  37. [37]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. InICLR, 2023

    2023

  38. [38]

    Survey of large multimodal model datasets, application categories and taxonomy.arXiv preprint arXiv:2412.17759, 2024

    Priyaranjan Pattnayak, Hitesh Laxmichand Patel, Bhargava Kumar, Amit Agarwal, Ishan Banerjee, Srikant Panda, and Tejaswini Kumar. Survey of large multimodal model datasets, application categories and taxonomy.arXiv preprint arXiv:2412.17759, 2024

    Pith/arXiv arXiv 2024

  39. [39]

    Deep learning on a data diet: Finding important examples early in training

    Mansheej Paul, Surya Ganguli, and Gintare Karolina Dziugaite. Deep learning on a data diet: Finding important examples early in training. InNeurIPS, volume 34, pages 20596–20607, 2021

    2021

  40. [40]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agar- wal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, Gretchen Krueger, and Ilya Sutskever. Learning transferable visual models from natural language supervision. InICML, 2021

    2021

  41. [41]

    Gupta, Xiaojiang Chen, and Xin Wang

    Pengzhen Ren, Yun Xiao, Xiaojun Chang, Po-Yao Huang, Zhihui Li, Brij B. Gupta, Xiaojiang Chen, and Xin Wang. A survey of deep active learning.CSUR, 54(9):180:1–180:40, 2021

    2021

  42. [42]

    LAION-5B: An open large-scale dataset for training next generation image-text models

    Christoph Schuhmann, Romain Beaumont, Richard Vencu, Cade Gordon, Ross Wightman, Mehdi Cherti, Theo Coombes, Aarush Katta, Clayton Mullis, Mitchell Wortsman, Patrick Schramowski, Srivatsa Kundurthy, Katherine Crowson, Ludwig Schmidt, Robert Kaczmarczyk, and Jenia Jitsev. LAION-5B: An open large-scale dataset for training next generation image-text models....

    2022

  43. [43]

    Active learning for convolutional neural networks: A core-set approach

    Ozan Sener and Silvio Savarese. Active learning for convolutional neural networks: A core-set approach. InICLR, 2018

    2018

  44. [44]

    Sse: Multimodal semantic data selection and enrichment for industrial-scale data assimilation

    Maying Shen, Nadine Chang, Sifei Liu, and Jose M Alvarez. Sse: Multimodal semantic data selection and enrichment for industrial-scale data assimilation. InKDD, pages 2525–2535, 2025

    2025

  45. [45]

    Centaur: Robust end-to-end autonomous driving with test-time training

    Chonghao Sima, Kashyap Chitta, Zhiding Yu, Shiyi Lan, Ping Luo, Andreas Geiger, Hongyang Li, and Jose M Alvarez. Centaur: Robust end-to-end autonomous driving with test-time training. arXiv preprint arXiv:2503.11650, 2025

    Pith/arXiv arXiv 2025

  46. [46]

    Fairdedup: Detecting and mitigating vision-language fairness disparities in semantic dataset deduplication

    Eric Slyman, Stefan Lee, Scott Cohen, and Kushal Kafle. Fairdedup: Detecting and mitigating vision-language fairness disparities in semantic dataset deduplication. InCVPR, pages 13905– 13916, 2024. 12

    2024

  47. [47]

    Prediction-oriented bayesian active learning

    Freddie Bickford Smith, Andreas Kirsch, Sebastian Farquhar, Yarin Gal, Adam Foster, and Tom Rainforth. Prediction-oriented bayesian active learning. InAISTATS, pages 7331–7348, 2023

    2023

  48. [48]

    Ben Sorscher, Robert Geirhos, Shashank Shekhar, Surya Ganguli, and Ari S. Morcos. Beyond neural scaling laws: Beating power law scaling via data pruning. InNeurIPS, volume 35, pages 19523–19536, 2022

    2022

  49. [49]

    Scalability in perception for autonomous driving: Waymo open dataset

    Pei Sun, Henrik Kretzschmar, Xerxes Dotiwalla, Aurelien Chouard, Vijaysai Patnaik, Paul Tsui, James Guo, Yin Zhou, Yuning Chai, Benjamin Caine, Vijay Vasudevan, Wei Han, Jiquan Ngiam, Hang Zhao, Aleksei Timofeev, Scott Ettinger, Maxim Krivokon, Amy Gao, Aditya Joshi, Yu Zhang, Jonathon Shlens, Zhifeng Chen, and Dragomir Anguelov. Scalability in perception...

    2020

  50. [50]

    Test-time training with self-supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. InICML, pages 9229–9248. PMLR, 2020

    2020

  51. [51]

    Mariya Toneva, Alessandro Sordoni, Remi Tachet des Combes, Adam Trischler, Yoshua Bengio, and Geoffrey J. Gordon. An empirical study of example forgetting during deep neural network learning. InICLR, 2019

    2019

  52. [52]

    Springer science & business media, 2013

    Vladimir Vapnik.The nature of statistical learning theory. Springer science & business media, 2013

    2013

  53. [53]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. InICLR, 2021

    2021

  54. [54]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. InCVPR, pages 7201–7211, 2022

    2022

  55. [55]

    Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, et al. Qwen3 technical report.arXiv preprint arXiv:2505.09388, 2025

    Pith/arXiv arXiv 2025

  56. [56]

    MEMO: Test time robustness via adaptation and augmentation

    Marvin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. InNeurIPS, 2022

    2022

  57. [57]

    Learning with local and global consistency

    Dengyong Zhou, Olivier Bousquet, Thomas Lal, Jason Weston, and Bernhard Schölkopf. Learning with local and global consistency. InNeurIPS, 2004

    2004

  58. [58]

    a man":

    Xiaojin Zhu, Zoubin Ghahramani, and John Lafferty. Semi-supervised learning using Gaussian fields and harmonic functions. InICML, 2003. 13 A Broader Impact TTCov improves autonomous driving safety by enabling interpretable, deployment-conditioned data curation. By guiding selection in an explicit Atlas of atomic propositions, TTCov enables adaptability to...

    2003