| { |
| "openreview_id": "9uENnRAcSl", |
| "arxiv": "2603.20538", |
| "claims": [ |
| { |
| "claim_index": 1, |
| "registered_claim": "Behavior cloning with quantized actions and log-loss is proven to achieve sample complexity matching known lower bounds, up to the quantization error term (Theorem 2, Section 3.2).", |
| "source_status": "confirmed", |
| "source_page": "1; 7; 11-12", |
| "anchor": "Abstract; Theorem 2; Theorems 8-9", |
| "note": "The paper states optimal sample complexity for log-loss behavior cloning. The matching is for the statistical sample term; the quantization contribution is separately additive and depends on stability assumptions.", |
| "claim": 1 |
| }, |
| { |
| "claim_index": 2, |
| "registered_claim": "Under Probabilistic Incremental Input-to-State Stability (P-IISS) of the dynamics and Relaxed Total Variation Continuity (RTVC) of the expert policy, the regret bound has only polynomial (not exponential) dependence on the horizon H with respect to quantization error epsilon_q (Theorem 3, Definition 3, Definition 4, Section 3.1-3.2).", |
| "source_status": "confirmed", |
| "source_page": "6-7", |
| "anchor": "Definitions 3-4; Theorem 3", |
| "note": "The theorem proves polynomial horizon dependence under global P-IISS and RTVC, for the specified binning quantizer and policy-class regularity conditions.", |
| "claim": 2 |
| }, |
| { |
| "claim_index": 3, |
| "registered_claim": "Theorem 6 shows that without a smoothness assumption on the quantizer, non-smooth quantizers can incur regret of order H*Omega(1) even though their in-distribution one-step error is only O(epsilon_q) (Theorem 6, Section 4.1).", |
| "source_status": "confirmed", |
| "source_page": "9-10", |
| "anchor": "Theorem 6", |
| "note": "The deterministic construction has average expert-distribution one-step quantization error O(ε_q) but deployed regret H·Ω(1). The theorem also gives a weaker stochastic lower bound.", |
| "claim": 3 |
| }, |
| { |
| "claim_index": 4, |
| "registered_claim": "Theorem 7 proves that model-based data augmentation improves the horizon dependence to H*[sqrt(log|Pi|/n) + epsilon_q] without requiring the policy smoothness (RTVC) assumption (Theorem 7, Section 4.2).", |
| "source_status": "misstated", |
| "source_page": 10, |
| "anchor": "Theorem 7", |
| "note": "The theorem removes RTVC but its statistical term is H·sqrt((log(|Π|/δ)+log(|M|/δ))/n), not H·sqrt(log|Π|/n). The registered claim omits transition-model realizability and the model-class complexity log|M|.", |
| "claim": 4 |
| }, |
| { |
| "claim_index": 5, |
| "registered_claim": "Information-theoretic lower bounds (Theorems 8-9) establish that regret must scale at least as H*(1/n + epsilon_q) for deterministic experts and H*(sqrt(1/n) + epsilon_q) for stochastic experts, matching the achievable upper bounds (Section 5, Theorems 8-9).", |
| "source_status": "confirmed", |
| "source_page": 11, |
| "anchor": "Theorems 8-9", |
| "note": "The deterministic expected lower bound is H(1/n+ε_q), and the stochastic high-probability lower bound is H(sqrt(1/n)+ε_q) under the theorem's allowance for a suboptimal expert.", |
| "claim": 5 |
| }, |
| { |
| "claim_index": 6, |
| "registered_claim": "Empirically, binning quantizers are shown to preserve policy smoothness better than learned quantizers, while deterministic experts more often violate the RTVC requirement needed for the sharp regret bound (Section 4.1).", |
| "source_status": "misstated", |
| "source_page": "9-10", |
| "anchor": "Proposition 5; Section 4.1 discussion", |
| "note": "The paper contains no empirical comparison establishing this claim. Proposition 5 is a theoretical sufficient result for binning, and the text argues that generic learned quantizers need not preserve the structure; remarks about empirical practice cite external work.", |
| "claim": 6 |
| } |
| ] |
| } |
