pith. sign in

arxiv: 2606.20967 · v1 · pith:6DH2Q4MBnew · submitted 2026-06-18 · 💻 cs.LG · cs.SY· eess.SY

Formalizing Task-Space Complexity for Zero-Shot Generalization

Jung-Hoon Cho , Heling Zhang , Siqi Du , Roy Dong , Cathy Wu This is my paper

Pith reviewed 2026-06-26 17:26 UTC · model grok-4.3

classification 💻 cs.LG cs.SYeess.SY
keywords zero-shot generalizationsigned divergencetask-space complexitycontextual dynamical systemspolicy selectionset coverreinforcement learning control
0
0 comments X

The pith

Signed divergence upper bounds the generalization gap from source to target contexts and induces ε-tolerance sets for policy classes.

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

The paper introduces a signed divergence as a performance-based directional measure of dissimilarity between contexts in dynamical systems. This quantity upper bounds the gap in how well a policy performs when moving from trained source contexts to new target contexts. It defines ε-tolerance sets around each source that certify generalization within a chosen error level, and it formalizes task-space complexity as the smallest number of sources needed to keep every possible target within ε. Under a local smoothness condition on performance, the tolerance sets admit inner and outer ball certificates along with volume bounds that depend on the specific instance. Source selection then reduces to a set-cover problem whose greedy solution carries the usual approximation guarantee and requires fewer policies than uniform or random baselines in the reported control examples.

Core claim

The signed divergence is introduced as a performance-centric directional task dissimilarity that upper bounds the generalization gap from a source context to a target context; it induces ε-tolerance sets that certify when a source policy class generalizes, and it yields the concrete notion of task-space complexity as the minimum number of source contexts needed so that every target context incurs at most ε generalization gap.

What carries the argument

The signed divergence, a directional performance-difference measure between contexts that bounds generalization gaps and generates tolerance sets.

If this is right

  • ε-tolerance sets supply explicit certificates that a given source policy class will generalize to any target inside the set.
  • Task-space complexity equals the size of the smallest source collection whose tolerance sets cover the entire target space at level ε.
  • Under the smoothness assumption the tolerance sets admit certified inner and outer balls together with instance-dependent volume bounds.
  • Selecting a minimal source set reduces to a set-cover problem on which the greedy algorithm inherits the standard H(n) approximation guarantee.
  • In the Mass-Spring-Damper LQR and nonlinear CartPole examples the greedy choice achieves the target ε-coverage with strictly fewer policies than uniform or random selection.

Where Pith is reading between the lines

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

  • The same signed-divergence construction could be used to decide how many simulation contexts are required before transferring a policy to hardware.
  • Volume bounds derived from the smoothness assumption might be turned into scaling predictions for the number of training tasks needed as state dimension increases.
  • The set-cover reduction suggests that existing approximation algorithms for covering problems could be imported directly into task-selection pipelines for reinforcement learning.

Load-bearing premise

Performance varies smoothly enough with small changes in context to support certified inner and outer balls around the tolerance sets.

What would settle it

A counter-example pair of source and target contexts where the measured performance gap exceeds the signed-divergence upper bound, or an instance where the local smoothness condition is violated and the claimed volume bounds on task-space complexity fail to hold.

Figures

Figures reproduced from arXiv: 2606.20967 by Cathy Wu, Heling Zhang, Jung-Hoon Cho, Roy Dong, Siqi Du.

Figure 1
Figure 1. Figure 1: Coverage vs. training computation. (a) Mass-Spring-Damper with LQR policies; (b) CartPole with PPO policies. uniform grid baseline is its dependence on a pre-selected grid resolution. A coarse grid may fail to place policies in critical regions, while a fine grid is inefficient. Greedy approach avoids this hyper￾parameter tuning by adaptively placing policies based on the local performance landscape [PITH… view at source ↗
Figure 2
Figure 2. Figure 2: Diagrams of two dynamical systems. Defining the state vector x(t) and input u(t) as x(t) = x(t) ˙x(t) ⊤ and u(t) = 0 F(t) ⊤ , the state-space equations are x˙(t) = A(k,m,c)x(t) + B(k,m,c)u(t) with A(k,m,c) =  0 1 − k m − c m  and B(k,m,c) = [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗
read the original abstract

Policies must operate across diverse conditions, yet a single policy is often conservative while fully adaptive schemes can be complex. We study zero-shot generalization in contextual dynamical systems and introduce a performance-centric, directional task dissimilarity--the signed divergence--that upper bounds the generalization gap from a source context to a target context. The signed divergence induces $\varepsilon$-tolerance sets that certify when a source policy class generalizes, and it yields a concrete notion of task-space complexity: the minimum number of source contexts needed so that every target context incurs at most $\varepsilon$ generalization gap. Under a mild local smoothness assumption on performance, the induced tolerance sets admit certified inner/outer balls and instance-dependent volume bounds on task-space complexity. In the finite-oracle setting, source selection reduces to set cover; a greedy strategy inherits the standard $H(n)$ approximation guarantee. Using a Mass-Spring-Damper system with linear-quadratic regulator (LQR) controllers and a nonlinear CartPole system with deep reinforcement learning controllers, we show that greedy selection achieves the same $\varepsilon$-coverage with fewer policies than uniform or random baselines. Our approach delivers a performance-based task similarity measure and practical certificates for building generalizable control with simple policies.

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

2 major / 2 minor

Summary. The paper introduces signed divergence, a directional performance-based dissimilarity between contexts in dynamical systems, which is shown to upper-bound the generalization gap from source to target contexts. This induces ε-tolerance sets that certify when a source policy class generalizes zero-shot, and defines task-space complexity as the smallest number of source contexts guaranteeing ε-coverage of all targets. Under a local smoothness assumption on the performance map, the tolerance sets admit inner/outer ball certificates and instance-dependent volume bounds. Source selection reduces to minimum set cover (with the standard greedy H(n) approximation), and experiments on an LQR-controlled mass-spring-damper system and a DRL-controlled CartPole system report that greedy selection achieves the same ε-coverage with fewer policies than uniform or random baselines.

Significance. If the local smoothness assumption is satisfied and the signed-divergence bound is non-vacuous, the work supplies a concrete, performance-centric notion of task similarity together with certificates for zero-shot generalization and a set-cover reduction that justifies greedy source selection. The empirical demonstration that greedy covers targets with fewer policies than baselines is a practical strength; the set-cover connection is a clean application of a standard result.

major comments (2)
  1. [Abstract, §4] Abstract and §4 (theorems on tolerance sets and volume bounds): the certified inner/outer balls and instance-dependent volume bounds on task-space complexity are derived only after invoking the local smoothness assumption on the performance map (context → expected return). No diagnostic is supplied in the LQR or CartPole experiments (gradient norms, local Lipschitz estimates, or counter-example search) to confirm the assumption holds even locally; if violated, the certificates and volume bounds become invalid while the reported coverage numbers remain unchanged.
  2. [§2] §2 (definition of signed divergence): the claim that signed divergence 'upper bounds the generalization gap' appears to follow directly from the definition of the measure rather than from independent premises about the dynamics or policy class; this makes the bounding relation definitional. The paper should clarify whether any non-trivial derivation occurs beyond the definition itself.
minor comments (2)
  1. [§2] Notation for ε-tolerance sets and task-space complexity should be introduced with a single consistent symbol in §2 before being used in the theorems.
  2. [§5] The experimental figures would benefit from error bars or multiple random seeds to show variability in the greedy vs. baseline coverage curves.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on our manuscript. We address each major comment point by point below.

read point-by-point responses

  1. Referee: [Abstract, §4] Abstract and §4 (theorems on tolerance sets and volume bounds): the certified inner/outer balls and instance-dependent volume bounds on task-space complexity are derived only after invoking the local smoothness assumption on the performance map (context → expected return). No diagnostic is supplied in the LQR or CartPole experiments (gradient norms, local Lipschitz estimates, or counter-example search) to confirm the assumption holds even locally; if violated, the certificates and volume bounds become invalid while the reported coverage numbers remain unchanged.

    Authors: We agree that the local smoothness assumption is required for the inner/outer ball certificates and volume bounds, and that the experiments do not include diagnostics (such as local Lipschitz estimates) to verify it. The reported ε-coverage numbers are computed directly from the signed divergence and remain valid independently of the assumption. In the revised manuscript we will add a discussion paragraph in the experimental section that (i) explicitly states the independence of the coverage results from the smoothness assumption and (ii) outlines practical verification methods (e.g., finite-difference estimates of local Lipschitz constants on the performance map). This is a partial revision because we add clarification and guidance rather than new empirical diagnostics. revision: partial

  2. Referee: [§2] §2 (definition of signed divergence): the claim that signed divergence 'upper bounds the generalization gap' appears to follow directly from the definition of the measure rather than from independent premises about the dynamics or policy class; this makes the bounding relation definitional. The paper should clarify whether any non-trivial derivation occurs beyond the definition itself.

    Authors: The referee is correct: the inequality D_s(θ, θ') ≥ J(θ') − J(θ) holds by construction of the signed divergence and does not rely on additional premises about the dynamics or policy class. No non-trivial derivation beyond the definition is claimed or performed. The subsequent technical contributions—inducing ε-tolerance sets, obtaining ball certificates under smoothness, and reducing source selection to set cover—are non-definitional. We will revise the text in §2 to state explicitly that the bounding relation is definitional while emphasizing the downstream utility of the measure. This change will be incorporated in the revision. revision: yes

Circularity Check

1 steps flagged

Signed divergence introduced to upper-bound generalization gap by definition; ε-tolerance sets and task-space complexity follow tautologically

specific steps
  1. self definitional [Abstract]
    "introduce a performance-centric, directional task dissimilarity--the signed divergence--that upper bounds the generalization gap from a source context to a target context. The signed divergence induces ε-tolerance sets that certify when a source policy class generalizes, and it yields a concrete notion of task-space complexity: the minimum number of source contexts needed so that every target context incurs at most ε generalization gap."

    The signed divergence is introduced precisely as the dissimilarity measure that upper-bounds the gap and induces the certifying sets; the bounding relation and certification therefore hold by the definition of the new object rather than being derived from prior independent premises about the performance map.

full rationale

The paper's core objects (signed divergence, ε-tolerance sets, task-space complexity) are defined such that the claimed bounding and certification properties hold by construction. The abstract explicitly introduces the signed divergence as the object that 'upper bounds the generalization gap' and 'induces ε-tolerance sets that certify' generalization. Under the local smoothness assumption the volume bounds are then derived, but the primary performance interpretation reduces to the definitional choice rather than an independent derivation. The set-cover reduction in the finite-oracle case is standard and non-circular. No self-citations or imported uniqueness theorems appear load-bearing in the provided text. This yields partial circularity (score 6) but leaves the empirical coverage results and smoothness invocation as separate issues.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 3 invented entities

The framework rests on the definition of signed divergence to produce the bound and on the local smoothness assumption to obtain geometric certificates; no free parameters beyond the application-chosen tolerance ε are identified in the abstract.

free parameters (1)
  • ε
    Tolerance threshold for acceptable generalization gap; chosen per application rather than fitted to data in the abstract.
axioms (1)
  • domain assumption mild local smoothness assumption on performance
    Invoked to derive certified inner/outer balls and instance-dependent volume bounds on task-space complexity.
invented entities (3)
  • signed divergence no independent evidence
    purpose: Directional task dissimilarity that upper-bounds the generalization gap
    Newly introduced performance-centric measure.
  • ε-tolerance sets no independent evidence
    purpose: Certify generalization of source policy classes
    Induced directly by the signed divergence.
  • task-space complexity no independent evidence
    purpose: Minimum number of source contexts guaranteeing ε-coverage
    Defined via the minimum set cover induced by the tolerance sets.

pith-pipeline@v0.9.1-grok · 5756 in / 1559 out tokens · 37183 ms · 2026-06-26T17:26:42.312297+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

77 extracted references · 2 canonical work pages

  1. [1]

    IEEE Transactions on automatic Control , volume=

    Guaranteed margins for LQG regulators , author=. IEEE Transactions on automatic Control , volume=. 2003 , publisher=

    2003

  2. [2]

    Journal of Machine Learning Research , volume=

    End-to-end training of deep visuomotor policies , author=. Journal of Machine Learning Research , volume=

  3. [3]

    2018 , publisher=

    High-dimensional probability: An introduction with applications in data science , author=. 2018 , publisher=

    2018

  4. [4]

    and Veness, Joel and Bellemare, Marc G

    Human-level control through deep reinforcement learning , volume =. Nature , author =. 2015 , pages =. doi:10.1038/nature14236 , language =

    work page doi:10.1038/nature14236 2015

  5. [5]

    Nature , volume=

    Autonomous navigation of stratospheric balloons using reinforcement learning , author=. Nature , volume=. 2020 , publisher=

    2020

  6. [6]

    and Barto, Andrew G

    Sutton, Richard S. and Barto, Andrew G. , year =. Reinforcement learning: an introduction , isbn =

  7. [7]

    Annual Reviews in Control , volume=

    Reinforcement learning for control: Performance, stability, and deep approximators , author=. Annual Reviews in Control , volume=. 2018 , publisher=

    2018

  8. [8]

    Advances in Neural Information Processing Systems , volume=

    Model-Based Transfer Learning for Contextual Reinforcement Learning , author=. Advances in Neural Information Processing Systems , volume=

  9. [9]

    arXiv preprint arXiv:2312.09436 , year=

    Temporal Transfer Learning for Traffic Optimization with Coarse-grained Advisory Autonomy , author=. arXiv preprint arXiv:2312.09436 , year=

    Pith/arXiv arXiv

  10. [10]

    Advances in Neural Information Processing Systems , volume=

    The impact of task underspecification in evaluating deep reinforcement learning , author=. Advances in Neural Information Processing Systems , volume=

  11. [11]

    IEEE Transactions on Automation Science and Engineering , volume=

    Unified automatic control of vehicular systems with reinforcement learning , author=. IEEE Transactions on Automation Science and Engineering , volume=. 2022 , publisher=

    2022

  12. [12]

    2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , pages=

    Reinforcement learning for mixed autonomy intersections , author=. 2021 IEEE International Intelligent Transportation Systems Conference (ITSC) , pages=. 2021 , organization=

    2021

  13. [13]

    arXiv preprint arXiv:1502.02259 , year=

    Contextual markov decision processes , author=. arXiv preprint arXiv:1502.02259 , year=

    Pith/arXiv arXiv

  14. [14]

    The Thirteenth International Conference on Learning Representations , year=

    IntersectionZoo: Eco-driving for Benchmarking Multi-Agent Contextual Reinforcement Learning , author=. The Thirteenth International Conference on Learning Representations , year=

  15. [15]

    2023 , journal=

    Contextualize Me--The Case for Context in Reinforcement Learning , author=. 2023 , journal=

    2023

  16. [16]

    2018 21st International Conference on Intelligent Transportation Systems (ITSC) , pages=

    Zero-shot deep reinforcement learning driving policy transfer for autonomous vehicles based on robust control , author=. 2018 21st International Conference on Intelligent Transportation Systems (ITSC) , pages=. 2018 , organization=

    2018

  17. [17]

    Robotics Research: The 18th International Symposium ISRR , pages=

    Adapt: zero-shot adaptive policy transfer for stochastic dynamical systems , author=. Robotics Research: The 18th International Symposium ISRR , pages=. 2019 , organization=

    2019

  18. [18]

    IEEE Robotics and Automation Letters , volume=

    Multi-task reinforcement learning with attention-based mixture of experts , author=. IEEE Robotics and Automation Letters , volume=. 2023 , publisher=

    2023

  19. [19]

    The Twelfth International Conference on Learning Representations , year=

    Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts , author=. The Twelfth International Conference on Learning Representations , year=

  20. [20]

    IEEE Transactions on Robotics , volume=

    Flow: A modular learning framework for mixed autonomy traffic , author=. IEEE Transactions on Robotics , volume=. 2021 , publisher=

    2021

  21. [21]

    Algorithmic Learning Theory , pages=

    Markov decision processes with continuous side information , author=. Algorithmic Learning Theory , pages=. 2018 , organization=

    2018

  22. [22]

    Proceedings of the 28th International Conference on Machine Learning (ICML-11) , pages=

    Learning with whom to share in multi-task feature learning , author=. Proceedings of the 28th International Conference on Machine Learning (ICML-11) , pages=

  23. [23]

    Advances in neural information processing systems , volume=

    Distral: Robust multitask reinforcement learning , author=. Advances in neural information processing systems , volume=

  24. [24]

    International conference on machine learning , pages=

    Which tasks should be learned together in multi-task learning? , author=. International conference on machine learning , pages=. 2020 , organization=

    2020

  25. [25]

    nature , volume=

    Mastering the game of go without human knowledge , author=. nature , volume=. 2017 , publisher=

    2017

  26. [26]

    The International Journal of Robotics Research , volume=

    Reinforcement learning in robotics: A survey , author=. The International Journal of Robotics Research , volume=. 2013 , publisher=

    2013

  27. [27]

    2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=

    Generalizing cooperative eco-driving via multi-residual task learning , author=. 2024 IEEE International Conference on Robotics and Automation (ICRA) , pages=. 2024 , organization=

    2024

  28. [28]

    , author=

    Metrics for Finite Markov Decision Processes. , author=. UAI , volume=

  29. [29]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Scalable methods for computing state similarity in deterministic markov decision processes , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

  30. [30]

    International Conference on Learning Representations , year=

    Contrastive Behavioral Similarity Embeddings for Generalization in Reinforcement Learning , author=. International Conference on Learning Representations , year=

  31. [31]

    Proceedings of the 24th international conference on Machine learning , pages=

    Cross-domain transfer for reinforcement learning , author=. Proceedings of the 24th international conference on Machine learning , pages=

  32. [32]

    International Conference on Learning Representations , year=

    Learning Invariant Feature Spaces to Transfer Skills with Reinforcement Learning , author=. International Conference on Learning Representations , year=

  33. [33]

    28th AAAI Conference on Artificial Intelligence, AAAI 2014 , pages=

    An automated measure of mdp similarity for transfer in reinforcement learning , author=. 28th AAAI Conference on Artificial Intelligence, AAAI 2014 , pages=. 2014 , organization=

    2014

  34. [34]

    2012 , publisher=

    Robust adaptive control , author=. 2012 , publisher=

    2012

  35. [35]

    arXiv preprint arXiv:1707.06347 , year=

    Proximal policy optimization algorithms , author=. arXiv preprint arXiv:1707.06347 , year=

    Pith/arXiv arXiv

  36. [36]

    International Conference on Machine Learning , pages=

    Darla: Improving zero-shot transfer in reinforcement learning , author=. International Conference on Machine Learning , pages=. 2017 , organization=

    2017

  37. [37]

    Journal of Artificial Intelligence Research , volume=

    A survey of zero-shot generalisation in deep reinforcement learning , author=. Journal of Artificial Intelligence Research , volume=

  38. [38]

    Neural computation , volume=

    Robust reinforcement learning , author=. Neural computation , volume=. 2005 , publisher=

    2005

  39. [39]

    International conference on machine learning , pages=

    Robust adversarial reinforcement learning , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  40. [40]

    International conference on machine learning , pages=

    Policy gradient method for robust reinforcement learning , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  41. [41]

    Proceedings of the 24th international conference on Machine learning , pages=

    Multi-task reinforcement learning: a hierarchical bayesian approach , author=. Proceedings of the 24th international conference on Machine learning , pages=

  42. [42]

    Electronics , volume=

    A survey of multi-task deep reinforcement learning , author=. Electronics , volume=. 2020 , publisher=

    2020

  43. [43]

    1996 , publisher=

    Robust adaptive control , author=. 1996 , publisher=

    1996

  44. [44]

    2014 , publisher=

    Adaptive filtering prediction and control , author=. 2014 , publisher=

    2014

  45. [45]

    1998 , publisher=

    Essentials of robust control , author=. 1998 , publisher=

    1998

  46. [46]

    Foundations of Computational Mathematics , volume=

    A theoretical and empirical comparison of gradient approximations in derivative-free optimization , author=. Foundations of Computational Mathematics , volume=. 2022 , publisher=

    2022

  47. [47]

    Johns Hopkins apl technical digest , volume=

    An overview of the simultaneous perturbation method for efficient optimization , author=. Johns Hopkins apl technical digest , volume=

  48. [48]

    IEEE Transactions on aerospace and electronic systems , volume=

    Implementation of the simultaneous perturbation algorithm for stochastic optimization , author=. IEEE Transactions on aerospace and electronic systems , volume=. 2002 , publisher=

    2002

  49. [49]

    IEEE transactions on automatic control , volume=

    Multivariate stochastic approximation using a simultaneous perturbation gradient approximation , author=. IEEE transactions on automatic control , volume=. 1992 , publisher=

    1992

  50. [50]

    International conference on machine learning , year =

    Learning Policy Committees for Effective Personalization in MDPs with Diverse Tasks , author=. International conference on machine learning , year =

  51. [51]

    ACM Transactions on Algorithms (TALG) , volume=

    On the set multicover problem in geometric settings , author=. ACM Transactions on Algorithms (TALG) , volume=. 2012 , publisher=

    2012

  52. [52]

    Balls in Rk do not cut all subsets of k + 2 points , journal =

    R.M Dudley , abstract =. Balls in Rk do not cut all subsets of k + 2 points , journal =. 1979 , issn =. doi:https://doi.org/10.1016/0001-8708(79)90047-1 , url =

    work page doi:10.1016/0001-8708(79)90047-1 1979

  53. [53]

    Annales de l'Institut Fourier , pages =

    Assouad, Patrick , title =. Annales de l'Institut Fourier , pages =. 1983 , doi =

    1983

  54. [54]

    Automatica , volume=

    Research on gain scheduling , author=. Automatica , volume=. 2000 , publisher=

    2000

  55. [55]

    2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=

    Domain randomization for transferring deep neural networks from simulation to the real world , author=. 2017 IEEE/RSJ international conference on intelligent robots and systems (IROS) , pages=. 2017 , organization=

    2017

  56. [56]

    International conference on machine learning , pages=

    Model-agnostic meta-learning for fast adaptation of deep networks , author=. International conference on machine learning , pages=. 2017 , organization=

    2017

  57. [57]

    Machine learning , volume=

    Multitask learning , author=. Machine learning , volume=. 1997 , publisher=

    1997

  58. [58]

    IEEE transactions on knowledge and data engineering , volume=

    A survey on multi-task learning , author=. IEEE transactions on knowledge and data engineering , volume=. 2021 , publisher=

    2021

  59. [59]

    2001 , publisher=

    Approximation algorithms , author=. 2001 , publisher=

    2001

  60. [60]

    Journal of Machine Learning Research , year =

    Antonin Raffin and Ashley Hill and Adam Gleave and Anssi Kanervisto and Maximilian Ernestus and Noah Dormann , title =. Journal of Machine Learning Research , year =

  61. [61]

    Information Processing Letters , volume=

    Hitting sets when the VC-dimension is small , author=. Information Processing Letters , volume=. 2005 , publisher=

    2005

  62. [62]

    1972 , publisher=

    Reducibility among combinatorial problems , author=. 1972 , publisher=

    1972

  63. [63]

    Advances in Neural Information Processing Systems , volume=

    When is generalizable reinforcement learning tractable? , author=. Advances in Neural Information Processing Systems , volume=

  64. [64]

    arXiv preprint arXiv:2503.07988 , year=

    Provable zero-shot generalization in offline reinforcement learning , author=. arXiv preprint arXiv:2503.07988 , year=

    arXiv

  65. [65]

    arXiv preprint arXiv:2505.18447 , year=

    Pessimism Principle Can Be Effective: Towards a Framework for Zero-Shot Transfer Reinforcement Learning , author=. arXiv preprint arXiv:2505.18447 , year=

    arXiv

  66. [66]

    Proceedings of the 24th annual Conference On Learning Theory , pages=

    Contextual bandits with similarity information , author=. Proceedings of the 24th annual Conference On Learning Theory , pages=. 2011 , organization=

    2011

  67. [67]

    Conference on Lifelong Learning Agents , pages=

    Statistical Context Detection for Deep Lifelong Reinforcement Learning , author=. Conference on Lifelong Learning Agents , pages=. 2025 , organization=

    2025

  68. [68]

    Cognitive psychology , volume=

    What makes some problems really hard: Explorations in the problem space of difficulty , author=. Cognitive psychology , volume=. 1990 , publisher=

    1990

  69. [69]

    arXiv preprint arXiv:2509.00125 , year=

    Know When to Explore: Difficulty-Aware Certainty as a Guide for LLM Reinforcement Learning , author=. arXiv preprint arXiv:2509.00125 , year=

    arXiv

  70. [70]

    9th International Conference on Education and Information Systems, Technologies and Applications (EISTA 2011) , year=

    Task complexity and design science , author=. 9th International Conference on Education and Information Systems, Technologies and Applications (EISTA 2011) , year=

    2011

  71. [71]

    Procedia Computer Science , volume=

    The concept of problem complexity , author=. Procedia Computer Science , volume=. 2014 , publisher=

    2014

  72. [72]

    Nature communications , volume=

    Task complexity interacts with state-space uncertainty in the arbitration between model-based and model-free learning , author=. Nature communications , volume=. 2019 , publisher=

    2019

  73. [73]

    Part 1: Finite controller coverings , author=

    Multiple model adaptive control. Part 1: Finite controller coverings , author=. International Journal of Robust and Nonlinear Control: IFAC-Affiliated Journal , volume=. 2000 , publisher=

    2000

  74. [74]

    Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

    Taskonomy: Disentangling task transfer learning , author=. Proceedings of the IEEE conference on computer vision and pattern recognition , pages=

  75. [75]

    ICLR , year =

    Emilio Parisotto and Jimmy Ba and Ruslan Salakhutdinov , title =. ICLR , year =

  76. [76]

    Learning for Dynamics and Control , pages=

    Suboptimal coverings for continuous spaces of control tasks , author=. Learning for Dynamics and Control , pages=. 2021 , organization=

    2021

  77. [77]

    2022 International Conference on Robotics and Automation (ICRA) , pages=

    Synergistic scheduling of learning and allocation of tasks in human-robot teams , author=. 2022 International Conference on Robotics and Automation (ICRA) , pages=. 2022 , organization=

    2022