Pith. sign in

REVIEW 2 major objections 6 minor 3 cited by

RoboSSM claims that replacing the Transformer backbone with a state-space model lets an in-context imitation learning policy maintain or improve success as test-time demonstrations grow to 16 times the training count, while a Transformer ba

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

2026-08-04 15:20 UTC pith:B6IDQXZ3

load-bearing objection The paper shows a real and useful result — SSM-based ICIL extrapolates to much longer prompts than a Transformer baseline — but the unspecified test-time beta-scaling hyperparameter muddies the exact source of the advantage. the 2 major comments →

arxiv 2509.19658 v2 pith:B6IDQXZ3 submitted 2025-09-24 cs.RO cs.AI

RoboSSM: Scalable In-context Imitation Learning via State-Space Models

Youngju Yoo , Jiaheng Hu , Yifeng Zhu , Bo Liu , Qiang Liu , Roberto Mart\'in-Mart\'in , Peter Stone This is my paper
classification cs.RO cs.AI
keywords in-context imitation learningstate-space modelsLonghornprompt-length extrapolationrobot manipulationLIBEROfew-shot learninglinear-time inference
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.

The paper introduces RoboSSM, an in-context imitation learning method that replaces the Transformer backbone with the Longhorn state-space model. It aims to show that a robot policy can be prompted at test time with many more demonstrations than it saw during training, and that success rate stays flat or improves even when the prompt is up to 16 times longer. If true, users could add demonstrations at deployment to improve performance on new tasks without any parameter updates or retraining. The paper also shows RoboSSM outperforms multi-task learning baselines on unseen tasks and handles temporally dilated demonstrations.

Core claim

On the paper's terms, the central discovery is that Longhorn's recurrent key-value memory, derived from an online regression objective, gives RoboSSM prompt-length extrapolation in in-context imitation learning. Trained with N_train=2 demonstrations, RoboSSM keeps or improves success as N_test grows to 32, attaining its best LIBERO-Object performance at N_test=32. ICRT, a Transformer method with a LLaMA2 backbone, degrades sharply once N_test exceeds N_train and collapses at N_test=32. The same pattern holds when the prompt is stretched by repeating frames (time dilation up to factor 16) rather than adding demonstrations. The paper interprets these results as evidence that SSMs are a scalabl

What carries the argument

Longhorn, a state-space model whose hidden state s_t is a d×m memory updated by s_t = A_t ⊙ s_{t-1} + B_t, with A_t and B_t functions of the input key-value pair. The update is the solution to an online convex programming problem balancing old-state retention against fitting the new observation; a sigmoid-gated vector β_t sets the trade-off. At test time RoboSSM scales β_t by γ ∈ (0,1], which the paper uses to bias the model toward attending to accumulated prompt context. This linear recurrence gives O(L+T_q) inference and lets the memory integrate many demonstrations without the quadratic cost and length-mismatch failure of attention.

Load-bearing premise

The central extrapolation claim assumes that the reported RoboSSM success rates, including the best-at-32 result, do not depend on the test-time gamma multiplier being tuned per task suite on the test tasks; the paper sweeps gamma but does not state which value produced the headline curves.

What would settle it

Run the Figure 3 extrapolation experiment with the test-time beta-scaling multiplier fixed to γ=1 across all suites and report success rates for N_test=2..32; if RoboSSM's success rate no longer rises beyond N_train and instead declines like ICRT, the claimed advantage is an artifact of test-set gamma selection rather than an architectural extrapolation property.

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

Share X LinkedIn Reddit HN

If this is right

  • If the central claim holds, deploying RoboSSM with more demonstrations than used in training should not hurt, and may help, success on unseen tasks.
  • Long-horizon or variable-speed demonstration prompts can be used at test time without retraining, since time dilation up to 16x preserves performance.
  • In-context imitation learning with SSMs can scale to long prompts at linear-time inference cost, making it practical for embodied agents with limited compute.
  • Prompting a trained RoboSSM policy with demonstrations of a new task becomes a way to adapt without parameter updates, supporting continual adaptation as new tasks arrive.

Where Pith is reading between the lines

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

  • The paper does not state whether the headline extrapolation curves use γ=1 or a test-selected γ from the sweep; a decisive test is to fix γ=1 and see whether the monotonic improvement at N_test=32 remains. If it does not, part of the claimed advantage is fitted to the evaluation set.
  • The β-scaling dependence suggests a simpler alternative design: a learned or schedule-based forgetting rate that lets the state integrate longer prompts, which could be tested by varying the training objective to expose the model to longer prompt lengths.
  • Because the paper's core input excludes language instructions, the mechanism is purely visual-proprioceptive. An extension would be to combine language with long prompts; if the advantage persists, the two conditioning signals may yield even better unseen-task generalization.
  • The LIBERO results are simulated; transferring the same setup to a physical robot with noisy teleoperated demonstrations would test whether the 16x extrapolation survives real-world distribution shift.

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

2 major / 6 minor

Summary. RoboSSM replaces the Transformer backbone of in-context imitation learning (ICIL) with the Longhorn state-space model, aiming to handle prompts substantially longer than those seen in training while maintaining linear-time inference. The paper evaluates on multiple LIBERO suites with disjoint train/test task splits, comparing against ICRT and multi-task baselines. The central claims are that RoboSSM maintains or improves success as the number of test-time demonstrations grows beyond the training prompt length (up to N_test=32 with N_train=2, i.e., 16x longer), that it remains robust under temporal dilation of demonstrations, and that it achieves competitive or better in-distribution performance than Transformer-based ICRT. The method includes a test-time beta-scaling multiplier gamma for the Longhorn state update.

Significance. If the central length-extrapolation result holds, the paper provides a valuable empirical demonstration that SSMs can be a scalable backbone for ICIL, addressing a known weakness of Transformer-based policies. The paper's strengths include multi-suite LIBERO evaluation with six seeds, released code, a runtime analysis confirming nearly linear inference scaling, and an explicit study of the beta-scaling mechanism. The gamma-disclosure issue described below is the main load-bearing concern; the rest is largely presentation and statistical reporting. The result is not circular: it is an empirical comparison, and Longhorn is evaluated in a new domain rather than assumed to work.

major comments (2)
  1. [Sec. III.A.3, Sec. IV-E, Figs. 3-6] The test-time beta-scaling multiplier gamma in (0,1] is introduced in Sec. III.A.3, but the value used for the headline extrapolation results is never stated. Sec. IV-E sweeps gamma in [0.4, 1.0] only for N_test=8, N_train=2, while Figs. 3-6 report the main 'prompts 16x longer' claim without disclosing gamma. If a test-selected gamma was used, part of the claimed advantage over ICRT at N_test=32 could be attributable to oracle tuning rather than the SSM's intrinsic length extrapolation. Please state the gamma value used in every reported experiment; if gamma=1 was used in the headline curves, say so explicitly; if gamma<1 was used, provide the gamma=1 results and describe how gamma was selected without using the test tasks.
  2. [Table I vs. Sec. IV-C] In Table I, ICRT w/o lang achieves 66.3±3.8 on LIBERO-Object while RoboSSM w/o lang achieves 57.9±3.5, yet Sec. IV-C states that RoboSSM 'consistently achieves higher success rates than ICRT across most scenarios.' Clarify whether Table I and Fig. 7 use identical protocols (N_train, N_test=4 and the same task split), and explain the apparent discrepancy. If the 'consistently higher' claim is meant only for the Study and Living Room scenes, the text should be revised to avoid overclaiming.
minor comments (6)
  1. [Figs. 3-6] The main comparison figures do not show error bars or confidence intervals, despite success rates being averaged over 6 seeds. Please add shaded regions or error bars, or report standard deviations in the captions, so the significance of the N_test > N_train trend can be assessed.
  2. [Fig. 6] Temporal dilation by repeating observation embeddings is an artificial proxy for long-horizon tasks. The caption should state explicitly that this simulates temporal variability rather than demonstrating generalization to genuinely novel long-horizon task structures.
  3. [Sec. IV-A.2] The text says the two backbones have 'a similar number of parameters' but no parameter counts are reported. Please include the actual numbers for ICRT and RoboSSM.
  4. [Sec. IV-G] The runtime analysis reports only LIBERO-Object and does not describe how many runs or which statistic (mean/median) is shown. Specify the hardware and the variance across runs.
  5. [Sec. IV-F / Ref. [37]] The t-SNE visualization cites reference [37], which is 'Gradients are not all you need' rather than the original t-SNE paper (van der Maaten & Hinton, 2008). The citation should be corrected.
  6. [Fig. 1 caption] The caption contains an obvious duplication: 'Longhorn Longhorn'. Also, the introduction refers to 'beta-scaling ablations' before the beta-scaling mechanism is defined; a forward pointer to Sec. IV-E would help.

Circularity Check

0 steps flagged

No significant circularity; central claim is empirical and self-contained, with one minor overlapping-author citation and an unspecified test-time beta-scaling knob that is a correctness risk, not a circular reduction.

full rationale

The paper's central claim—that RoboSSM maintains or improves success as the number of test-time demonstrations grows beyond training length—is an empirical comparison on LIBERO, not a derivation from an equation that reduces to its inputs. The main potential circularity is the self-citation of Longhorn [23], which shares authors with this paper. However, RoboSSM evaluates Longhorn on a new domain (ICIL on LIBERO) and does not assume the conclusion; the extrapolation is demonstrated experimentally in Figures 3–6. This is therefore a non-load-bearing self-citation. A second concern is the test-time beta-scaling multiplier γ introduced in Sec. III.A.3 and swept in Sec. IV-E (γ∈[0.4,1.0] for N_test=8, N_train=2). The paper never states which γ produced the headline Figures 3–6, including the peak at N_test=32 after training with N_train=2. If a per-suite, test-selected γ was used, part of the claimed long-prompt advantage would be fitted to the evaluation set. But this is a missing-specification / external-validity issue, not demonstrated circularity: no equation equates the prediction to the fitted parameter, and the paper does not rename the fit as a prediction in its derivation. No self-definitional, uniqueness-imported, ansatz-smuggled, or renaming patterns are present. Overall score reflects the minor self-citation and the unresolved gamma-specification risk while affirming that the central empirical result is not circular by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 4 axioms · 0 invented entities

No new physical or architectural entities are introduced; Longhorn is prior work. The main free parameter is the test-time beta-scaling multiplier gamma, whose role in the headline results is under-specified. The axioms are standard benchmark and design assumptions, with the temporal-dilation proxy being the most fragile.

free parameters (1)
  • gamma (test-time beta-scaling multiplier) = not reported; swept in [0.4, 1.0] on test suites
    Introduced in Section III-A.3 and Section IV-E to trade off reliance on previous state vs. new input; if selected per suite on test data, it is a free parameter fit to the evaluation set.
axioms (4)
  • domain assumption Longhorn recurrence (Eqs. 2-5) is a suitable in-context learner for robot demonstrations.
    Adopted from self-cited Longhorn [23]; no derivation or validation internal to this paper beyond the presented empirical results.
  • ad hoc to paper Excluding actions and language from the input forces the policy to infer the task from observation prompts.
    Design choice in Section III-A.1; if actions were needed for disambiguation, the interpretation of the benchmark results would change.
  • ad hoc to paper Frame repetition creates valid long-horizon test scenarios.
    Section IV-B.2; repeated observation embeddings simulate temporal dilation, not new skills, objects, or task structure.
  • domain assumption LIBERO success rate on unseen tasks is a meaningful proxy for ICIL generalization.
    Section IV-A; LIBERO is an accepted benchmark, but it is simulated and only 2 unseen tasks per suite are used.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of RoboSSM: Scalable In-context Imitation Learning via State-Space Models." pith.science (2026). https://pith.science/paper/B6IDQXZ3

@misc{pith2026250919658,
  author       = {Pith},
  title        = {Pith review of: RoboSSM: Scalable In-context Imitation Learning via State-Space Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/B6IDQXZ3}},
  note         = {Machine review of arXiv:2509.19658}
}
Share X LinkedIn Reddit HN
read the original abstract

In-context imitation learning (ICIL) enables robots to learn tasks from prompts consisting of just a handful of demonstrations. By eliminating the need for parameter updates at deployment time, this paradigm supports few-shot adaptation to novel tasks. However, recent ICIL methods rely on Transformers, which have computational limitations and tend to underperform when handling longer prompts than those seen during training. In this work, we introduce RoboSSM, a scalable recipe for in-context imitation learning based on state-space models (SSM). Specifically, RoboSSM replaces Transformers with Longhorn -- a state-of-the-art SSM that provides linear-time inference and strong extrapolation capabilities, making it well-suited for long-context prompts. Through diverse experiments on the LIBERO benchmark, we demonstrate the effectiveness of applying SSMs to ICIL, achieving improved generalization to both unseen and long-horizon tasks than Transformer-based ICIL methods by handling longer contexts at test-time. These results show for the first time that SSMs are an efficient and scalable backbone for ICIL. Our code is available at https://github.com/youngjuY/RoboSSM.

Figures

Figures reproduced from arXiv: 2509.19658 by Bo Liu, Jiaheng Hu, Peter Stone, Qiang Liu, Roberto Mart\'in-Mart\'in, Yifeng Zhu, Youngju Yoo.

Figure 1
Figure 1. Figure 1: Overview of RoboSSM. Training (left): Longhorn receives Ntrain trajectories from Ptrain and a query trajectory for the same task. Inference (right): Given Ntest trajectories from Ptest containing unseen tasks, the model predicts actions and updates the environment iteratively from the initial observation embedding. these methods still struggle to perform completely unseen tasks without additional data coll… view at source ↗
Figure 2
Figure 2. Figure 2: We evaluate RoboSSM on challenging manipulation tasks from the LIBERO benchmark. In LIBERO-90, the Study and Living Room suites each comprise 4 different scenes, while the Kitchen suite comprises 10 different scenes. 1) Input Encoding: RoboSSM encodes multimodal ob￾servations at each time step of a demonstration. Visual data are processed by convolutional neural networks (CNNs), and proprioceptive data are… view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of RoboSSM and ICRT across test-time demonstrations (Ntest), with both models trained with Ntrain=2. ICRT’s performance drops sharply once Ntest>Ntrain [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Results with a fixed test-time prompt Ntest = 8 across models trained with different numbers of demonstrations (Ntrain) [PITH_FULL_IMAGE:figures/full_fig_p004_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Results with a fixed test-time prompt Ntest = 32 across models trained with different numbers of demonstrations (Ntrain). ICRT yields zero success across all Ntrain [PITH_FULL_IMAGE:figures/full_fig_p004_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Comparison of RoboSSM and ICRT when evaluated on test-time prompts with temporally dilated demonstrations [PITH_FULL_IMAGE:figures/full_fig_p005_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Results of RoboSSM and ICRT on in-context imitation learning, where the prompt consists of the same number of demonstrations during training and testing (Ntrain = Ntest). We design our experiments to answer the following re￾search questions: • Q1: Can RoboSSM extrapolate to prompts composed of demonstrations that are longer than those used in training? • Q2: How many training demonstrations are required fo… view at source ↗
Figure 8
Figure 8. Figure 8: Effectiveness of test-time β-scaling across task suites. length α · T, where T is the original trajectory length and α ∈ {1, 2, 4, 8, 16} is the dilation factor. Although models are trained with the original prompt length (α = 1), we evaluate their robustness under extended test-time prompts, where |Ptest| = α · |Ptrain|, with α up to 16 [PITH_FULL_IMAGE:figures/full_fig_p006_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Inference runtime on LIBERO-Object for ICRT and RoboSSM. in equation (4). RoboSSM generates trajectories that stay close to the prompt demonstration clusters, indicating sta￾ble long-context extrapolation. In contrast, ICRT produces trajectories that drift far outside the prompt manifold, with the deviation becoming more pronounced as the number of prompt demonstrations increases. G. Runtime Analysis [PI… view at source ↗
Figure 9
Figure 9. Figure 9: Visualization of latent representations with Ntest = 8 (left) and Ntest = 16 (right). Blue corresponds to prompt demonstrations and red to the predicted trajectory. D. Comparison to Multi-Task Learning We compare RoboSSM against multi-task learning base￾lines, MTL-TF and MTL-SSM. During training, we set Ntrain = 4, and for evaluation, Ntest = 0 for MTL baselines, while Ntest = 4 for RoboSSM and ICRT. To en… view at source ↗

discussion (0)

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

Forward citations

Cited by 3 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. DSSP: Diffusion State Space Policy with Full-History Encoding

    cs.RO 2026-05 conditional novelty 7.0

    DSSP is a history-conditioned diffusion state space policy that uses SSMs to encode full observation streams with an auxiliary dynamics objective and hierarchical fusion, achieving SOTA results with reduced model size...

  2. From Abstraction to Instantiation: Learning Behavioral Representation for Vision-Language-Action Model

    cs.CV 2026-05 unverdicted novelty 5.0

    BehaviorVLA introduces a symmetric encoder-decoder architecture with causal Mamba and phase conditioning to learn unified long-horizon behavioral representations for improved generalization in VLA models.

  3. From Abstraction to Instantiation: Learning Behavioral Representation for Vision-Language-Action Model

    cs.CV 2026-05 unverdicted novelty 4.0

    BehaviorVLA learns long-horizon behavioral representations via causal Mamba encoder and phase-conditioned decoder, reporting SOTA results of 58% on RoboTwin 2.0, 98% on LIBERO, 4.36 on CALVIN, and matching OpenVLA-OFT...

Reference graph

Works this paper leans on

37 extracted references · 20 linked inside Pith · cited by 2 Pith papers

  1. [1]

    A survey of robot learning from demonstration,

    B. D. Argall, S. Chernova, M. Veloso, and B. Browning, “A survey of robot learning from demonstration,”Robotics and autonomous systems, vol. 57, no. 5, pp. 469–483, 2009

    2009

  2. [2]

    Robot program- ming by demonstration,

    A. Billard, S. Calinon, R. Dillmann, and S. Schaal, “Robot program- ming by demonstration,” inSpringer handbook of robotics. Springer, 2008, pp. 1371–1394

    2008

  3. [3]

    Octo: An open-source generalist robot policy,

    O. M. Team, D. Ghosh, H. Walke, K. Pertsch, K. Black, O. Mees, S. Dasari, J. Hejna, T. Kreiman, C. Xu,et al., “Octo: An open-source generalist robot policy,”arXiv preprint arXiv:2405.12213, 2024

    Pith/arXiv arXiv 2024

  4. [4]

    Open- vla: An open-source vision-language-action model,

    M. J. Kim, K. Pertsch, S. Karamcheti, T. Xiao, A. Balakrishna, S. Nair, R. Rafailov, E. Foster, G. Lam, P. Sanketi,et al., “Open- vla: An open-source vision-language-action model,”arXiv preprint arXiv:2406.09246, 2024

    Pith/arXiv arXiv 2024

  5. [5]

    Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,

    A. O’Neill, A. Rehman, A. Maddukuri, A. Gupta, A. Padalkar, A. Lee, A. Pooley, A. Gupta, A. Mandlekar, A. Jain,et al., “Open x-embodiment: Robotic learning datasets and rt-x models: Open x- embodiment collaboration 0,” in2024 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2024, pp. 6892–6903

    2024

  6. [6]

    Droid: A large-scale in-the-wild robot manipulation dataset,

    A. Khazatsky, K. Pertsch, S. Nair, A. Balakrishna, S. Dasari, S. Karam- cheti, S. Nasiriany, M. K. Srirama, L. Y . Chen, K. Ellis,et al., “Droid: A large-scale in-the-wild robot manipulation dataset,”arXiv preprint arXiv:2403.12945, 2024

    Pith/arXiv arXiv 2024

  7. [7]

    Fine-tuning vision-language- action models: Optimizing speed and success,

    M. J. Kim, C. Finn, and P. Liang, “Fine-tuning vision-language- action models: Optimizing speed and success,”arXiv preprint arXiv:2502.19645, 2025

    Pith/arXiv arXiv 2025

  8. [8]

    Flare: Achieving masterful and adaptive robot policies with large-scale reinforcement learning fine- tuning,

    J. Hu, R. Hendrix, A. Farhadi, A. Kembhavi, R. Mart ´ın-Mart´ın, P. Stone, K.-H. Zeng, and K. Ehsani, “Flare: Achieving masterful and adaptive robot policies with large-scale reinforcement learning fine- tuning,”arXiv preprint arXiv:2409.16578, 2024

    Pith/arXiv arXiv 2024

  9. [9]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat,et al., “Gpt-4 technical report,”arXiv preprint arXiv:2303.08774, 2023

    Pith/arXiv arXiv 2023

  10. [10]

    Llama 2: Open foundation and fine-tuned chat models,

    H. Touvron, L. Martin, K. Stone, P. Albert, A. Almahairi, Y . Babaei, N. Bashlykov, S. Batra, P. Bhargava, S. Bhosale,et al., “Llama 2: Open foundation and fine-tuned chat models,”arXiv preprint arXiv:2307.09288, 2023

    Pith/arXiv arXiv 2023

  11. [11]

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities,

    G. Comanici, E. Bieber, M. Schaekermann, I. Pasupat, N. Sachdeva, I. Dhillon, M. Blistein, O. Ram, D. Zhang, E. 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

  12. [12]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell,et al., “Language models are few-shot learners,”Advances in neural information pro- cessing systems, vol. 33, pp. 1877–1901, 2020

    1901

  13. [13]

    One-shot imitation learning,

    Y . Duan, M. Andrychowicz, B. Stadie, O. Jonathan Ho, J. Schnei- der, I. Sutskever, P. Abbeel, and W. Zaremba, “One-shot imitation learning,”Advances in neural information processing systems, vol. 30, 2017

    2017

  14. [14]

    Towards more generalizable one-shot visual imitation learning,

    Z. Mandi, F. Liu, K. Lee, and P. Abbeel, “Towards more generalizable one-shot visual imitation learning,” in2022 International conference on robotics and automation (ICRA). IEEE, 2022, pp. 2434–2444

    2022

  15. [15]

    Demonstrate once, imitate immediately (dome): Learning visual servoing for one- shot imitation learning,

    E. Valassakis, G. Papagiannis, N. Di Palo, and E. Johns, “Demonstrate once, imitate immediately (dome): Learning visual servoing for one- shot imitation learning,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2022, pp. 8614– 8621

    2022

  16. [16]

    Prompting decision transformer for few-shot policy generalization,

    M. Xu, Y . Shen, S. Zhang, Y . Lu, D. Zhao, J. Tenenbaum, and C. Gan, “Prompting decision transformer for few-shot policy generalization,” ininternational conference on machine learning. PMLR, 2022, pp. 24 631–24 645

    2022

  17. [17]

    In-context imitation learning via next-token prediction,

    L. Fu, H. Huang, G. Datta, L. Y . Chen, W. C.-H. Panitch, F. Liu, H. Li, and K. Goldberg, “In-context imitation learning via next-token prediction,”arXiv preprint arXiv:2408.15980, 2024

    Pith/arXiv arXiv 2024

  18. [18]

    Action tokenizer matters in in-context imitation learning,

    A. D. Vuong, M. N. Vu, D. An, and I. Reid, “Action tokenizer matters in in-context imitation learning,”arXiv preprint arXiv:2503.01206, 2025

    arXiv 2025

  19. [19]

    Keypoint action tokens enable in-context imitation learning in robotics,

    N. Di Palo and E. Johns, “Keypoint action tokens enable in-context imitation learning in robotics,”arXiv preprint arXiv:2403.19578, 2024

    Pith/arXiv arXiv 2024

  20. [20]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems, vol. 30, 2017

    2017

  21. [21]

    Transformers can achieve length generalization but not robustly,

    Y . Zhou, U. Alon, X. Chen, X. Wang, R. Agarwal, and D. Zhou, “Transformers can achieve length generalization but not robustly,” arXiv preprint arXiv:2402.09371, 2024

    Pith/arXiv arXiv 2024

  22. [22]

    Flashattention: Fast and memory-efficient exact attention with io-awareness,

    T. Dao, D. Fu, S. Ermon, A. Rudra, and C. R ´e, “Flashattention: Fast and memory-efficient exact attention with io-awareness,”Advances in neural information processing systems, vol. 35, pp. 16 344–16 359, 2022

    2022

  23. [23]

    Longhorn: State space models are amortized online learners,

    B. Liu, R. Wang, L. Wu, Y . Feng, P. Stone, and Q. Liu, “Longhorn: State space models are amortized online learners,”arXiv preprint arXiv:2407.14207, 2024

    Pith/arXiv arXiv 2024

  24. [24]

    Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,

    B. Liu, Y . Zhu, C. Gao, Y . Feng, Q. Liu, Y . Zhu, and P. Stone, “Libero: Benchmarking knowledge transfer for lifelong robot learn- ing,”Advances in Neural Information Processing Systems, vol. 36, pp. 44 776–44 791, 2023

    2023

  25. [25]

    Deeply aggrevated: Differentiable imitation learning for sequential prediction,

    W. Sun, A. Venkatraman, G. J. Gordon, B. Boots, and J. A. Bagnell, “Deeply aggrevated: Differentiable imitation learning for sequential prediction,” inInternational conference on machine learning. PMLR, 2017, pp. 3309–3318

    2017

  26. [26]

    Imitation learning for human pose prediction,

    B. Wang, E. Adeli, H.-k. Chiu, D.-A. Huang, and J. C. Niebles, “Imitation learning for human pose prediction,” inProceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 7124–7133

    2019

  27. [27]

    A framework for behavioural cloning

    M. Bain and C. Sammut, “A framework for behavioural cloning.” in Machine intelligence 15, 1995, pp. 103–129

    1995

  28. [28]

    Behavioral cloning from obser- vation,

    F. Torabi, G. Warnell, and P. Stone, “Behavioral cloning from obser- vation,”arXiv preprint arXiv:1805.01954, 2018

    Pith/arXiv arXiv 2018

  29. [29]

    Efficiently modeling long sequences with structured state spaces,

    A. Gu, K. Goel, and C. R ´e, “Efficiently modeling long sequences with structured state spaces,”arXiv preprint arXiv:2111.00396, 2021

    Pith/arXiv arXiv 2021

  30. [30]

    Hungry hungry hippos: Towards language modeling with state space models,

    D. Y . Fu, T. Dao, K. K. Saab, A. W. Thomas, A. Rudra, and C. R ´e, “Hungry hungry hippos: Towards language modeling with state space models,”arXiv preprint arXiv:2212.14052, 2022

    Pith/arXiv arXiv 2022

  31. [31]

    Simpli- fied state space layers for sequence modeling,

    J. T. Smith, A. Warrington, and S. W. Linderman, “Simpli- fied state space layers for sequence modeling,”arXiv preprint arXiv:2208.04933, 2022

    Pith/arXiv arXiv 2022

  32. [32]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,”arXiv preprint arXiv:2312.00752, 2023

    Pith/arXiv arXiv 2023

  33. [33]

    Structured state space models for in-context reinforcement learning,

    C. Lu, Y . Schroecker, A. Gu, E. Parisotto, J. Foerster, S. Singh, and F. Behbahani, “Structured state space models for in-context reinforcement learning,”Advances in Neural Information Processing Systems, vol. 36, pp. 47 016–47 031, 2023

    2023

  34. [34]

    Mail: Improving imita- tion learning with mamba,

    X. Jia, Q. Wang, A. Donat, B. Xing, G. Li, H. Zhou, O. Celik, D. Blessing, R. Lioutikov, and G. Neumann, “Mail: Improving imita- tion learning with mamba,”arXiv preprint arXiv:2406.08234, 2024

    Pith/arXiv arXiv 2024

  35. [35]

    Online convex programming and generalized in- finitesimal gradient ascent,

    M. Zinkevich, “Online convex programming and generalized in- finitesimal gradient ascent,” inProceedings of the 20th international conference on machine learning (icml-03), 2003, pp. 928–936

    2003

  36. [36]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101, 2017

    Pith/arXiv arXiv 2017

  37. [37]

    Gradi- ents are not all you need,

    L. Metz, C. D. Freeman, S. S. Schoenholz, and T. Kachman, “Gradi- ents are not all you need,”arXiv preprint arXiv:2111.05803, 2021

    Pith/arXiv arXiv 2021

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.