{"id":"6f1183cc-c941-4d1e-afcb-52d798c6de69","arxiv_id":"2504.13368","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"IDRL iteratively filters an offline dataset using learned visitation ratios and then runs weighted behavior cloning, outperforming several prior offline RL methods on D4RL and corrupted demonstrations.","lead":"A new offline reinforcement learning method, IDRL, repeatedly prunes low-quality transitions from a fixed dataset and relearns on the cleaner data. If correct, it would make offline RL more reliable on messy, real-world datasets where expert demonstrations are rare.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 4 does not apply to Algorithm 1: filtering by w_k>0 yields the original data restricted to a support, not the optimized visitation distribution d*_k, so the monotonic-improvement/curriculum claim is unsupported.","rationale":"I read the paper as making two intertwined claims: (1) IDRL empirically beats strong Primal- and Dual-RL baselines on all datasets; (2) IDRL is theoretically principled because iterated filtering moves the reference distribution toward the optimal discriminator weight. The reader's CONDITIONAL verdict focuses mainly on approximation error in the learned action ratio and on overclaims in the abstract. My independent stress-test identifies a more fundamental, internal mismatch: even with exact ratios, the implemented algorithm does not realize the distribution replacement analyzed in Theorem 4. Algorithm 1 filters by support (w_k>0), so the next iteration's behavior distribution is the original data distribution conditioned on the support of d*_k, not d*_k itself. Theorem 4 and its proof equate the filtered dataset with d*_{k+1}, which is a different object; the monotonicity conclusion therefore does not follow for the actual procedure. This is not an approximation or empirical disagreement—it is an inconsistency between the stated theory and the algorithm. The empirical results in Table 1 may still support a weaker 'matches or outperforms on most tasks' claim, and the reader was right to condition acceptance; I would keep the CONDITIONAL verdict. The condition should now include either modifying the algorithm to resample according to w_k (making Theorem 4 applicable) or providing a new analysis of the support-filtering procedure. I partially agree with the reader: they located the risk in the learned-ratio approximation, whereas the more load-bearing issue is the theory-algorithm mismatch, which is independent of function approximation error.","tokens_in":21327,"tokens_out":14787,"duration_ms":134299,"concrete_test":"Run the §3.2 gridworld toy with logging: after iteration 1, compute the normalized empirical distribution p_2(s,a) of D_2 and compare it with the learned ratio w_1(s,a) (i.e., d*_1) using total variation distance. If TV(p_2, d*_1)>0, the algorithm is not replacing the behavior distribution by the optimized visitation distribution, contradicting the premise of Theorem 4. Then run iteration 2 twice, once with D_2 as in Algorithm 1 and once with D_2 resampled according to w_1; Theorem 4's monotonicity should hold only for the resampled variant, isolating the gap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Algorithm 1 (line 12) constructs D_{k+1} by keeping transitions with w_k(s,a)>0. The next Dual-RL iteration therefore uses behavior distribution p_{k+1}(s,a) ∝ p_k(s,a) · 1_{w_k(s,a)>0}. This is not the optimized visitation distribution d*_k whose ratio w_k = d*_k/p_k was just learned: within the retained support, suboptimal transitions keep their original frequency. Section 3.2 and Theorem 4 instead analyze the idealized sequence d*_{k+1} = argmax_d E_d[r] - α D_f(d || d*_k), i.e., they assume the reference distribution is replaced by d*_k. The proof of Theorem 4 defines V(D_k)=E_{d*_k}[r] and concludes V(D_{k+1})≥V(D_k), but the actual filtered dataset's expected reward E_{p_{k+1}}[r] need not be monotone: zero-weight transitions can carry high reward that the regularized optimum deliberately avoids, and positive-weight transitions can remain suboptimal. Unless D_{k+1} is resampled according to w_k, the 'curriculum of improved visitation distribution ratios' and the monotonic improvement guarantee do not hold for the implemented procedure. This gap undercuts the paper's claim that IDRL provably breaks the regularization barrier; the theory concerns a sequence of distributions the algorithm never realizes.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Iterative Dual Reinforcement Learning (IDRL), an offline RL method that views the problem through optimal-discriminator-weighted imitation learning. The authors observe that weighted behavior cloning with a discriminator ratio trained on the offline dataset plus an additional expert dataset performs strongly. They argue that semi-gradient Dual-RL methods learn only an action-level distribution ratio rather than the full state-action visitation ratio, and they propose a two-stage correction: first estimate the action ratio by semi-gradient Dual-RL, then solve an OPE-style objective to recover the state ratio. IDRL then iteratively removes transitions assigned zero weight by the learned state-action ratio and reruns Dual-RL on the filtered dataset. The paper provides theoretical statements about the behavior-cloning bound and monotonic improvement across iterations, a grid-world illustration, and experiments on D4RL and corrupted-demonstration datasets.","tokens_in":21609,"tokens_out":5614,"duration_ms":51191,"significance":"If the central claims hold, IDRL would be a meaningful contribution to offline RL: it identifies a real fixed-point issue in semi-gradient Dual-RL, proposes a decomposition that avoids OOD action queries, and introduces a sparse-filtering mechanism for iterative dataset refinement. The paper's strengths include the explicit derivation of the correction objective, the careful distinction between action-level and state-action visitation ratios, the toy experiment that visually demonstrates the difference, and a fairly broad empirical evaluation including corrupted demonstrations. However, the main theoretical guarantee—Theorem 4—is stated for an idealized sequence of distributions that Algorithm 1 does not actually implement, so the claim that IDRL provably breaks the regularization barrier is not currently supported. The empirical results are promising but the abstract overstates them relative to the paper's own Table 1.","major_comments":[{"comment":"Theorem 4 does not apply to the implemented algorithm. Algorithm 1 (line 12) constructs D_{k+1} by keeping transitions with w_k(s,a)>0, so the next Dual-RL iteration uses the support-restricted behavior distribution p_{k+1}(s,a) ∝ p_k(s,a)·1_{w_k(s,a)>0}. This is not the optimized visitation distribution d*_k whose ratio w_k=d*_k/p_k was learned: within the retained support, the relative frequencies of transitions are unchanged. The proof of Theorem 4 instead considers d*_{k+1}=argmax_d E_d[r]-αD_f(d||d*_k) and defines V(D_k)=E_{d*_k}[r], concluding V(D_{k+1})≥V(D_k). Thus the monotonic-improvement/curriculum guarantee concerns a sequence of distributions the algorithm never realizes. The paper itself hedges by saying the filtering can be 'approximately viewed' as replacing the behavior distribution, but Theorem 4 is then stated and used as an unconditional guarantee. This gap undercuts the central theoretical claim that IDRL provably breaks the regularization barrier. The authors should either prove a monotonicity statement for the actual filtered-data sequence E_{p_{k+1}}[r]≥E_{p_k}[r] or explicitly and consistently reframe Theorem 4 as a statement about an idealized iteration and remove the claim that the implemented procedure is provably improving.","section":"§3.2, Theorem 4; Algorithm 1"},{"comment":"The claim that W*(s) provides an 'unbiased estimate' of (f')^{-1}(E_{a~μ}[w*(a|s)(TU(s,a)-U(s))]) is not accurate as stated. The result in Lemma 1 identifies W*(s) as the minimizer of a convex expected objective; the argmin of a finite-sample empirical version of that objective is generally biased because the argmin does not commute with expectation under the nonlinear map (f')^{-1}. What is unbiased is the gradient of the empirical objective with respect to W. Since Eqs. (10)-(11) are implemented with minibatches, the paper's later claim that the correction stage recovers the true state-action visitation ratio 'in an unbiased way' is not justified by the stated theorems. Please either correct the terminology and provide a finite-sample analysis, or soften the claim to 'unbiased gradient estimates' and state the resulting bias-variance caveat.","section":"§3.1, Theorem 2 and Eqs. (10)-(11)"},{"comment":"The abstract's claim that IDRL 'beats strong Primal-RL and Dual-RL baselines in terms of both performance and stability, on all datasets' is contradicted by the paper's own Table 1. In the antmaze-large-diverse row, ReBRAC reports 54.4±25.1 while IDRL reports 54.2±3.8, so IDRL does not beat that baseline on that dataset. The table caption itself says IDRL 'matches or outperforms' baselines 'on almost all tasks.' The abstract and the main-text summary in Section 4.1 should be revised to match the actual results, removing the 'all datasets' overstatement.","section":"Abstract; Table 1"},{"comment":"The proof of Theorem 3 invokes the imitation-learning bound of Li et al. (2024) by treating D_{k+1} as the 'expert' dataset and D_k \\ D_{k+1} as the supplementary dataset, but the required premise is not verified. D_{k+1} consists of transitions kept by the learned weight w_k>0, which are not necessarily expert demonstrations, and the ratio w*_{k+1}(s) used in the bound is not the expert/behavior ratio assumed by Li et al. Additionally, the proof text writes the supplementary set as 'D_{k+1}−D_k', which is empty because D_{k+1}⊆D_k, and should be D_k \\ D_{k+1}. Please clarify the assumptions under which the cited bound applies and correct the set notation.","section":"Theorem 3 proof"}],"minor_comments":[{"comment":"Line 10 says to update Uψ1 and Wψ2 by Eqs. (11) and (12), but Eq. (12) is the formula for the ratio w(s,a), not an update objective. The W update should reference Eq. (10) and the U update Eq. (11).","section":"Algorithm 1, line 10"},{"comment":"There is a typo in the abstract: 'addtional' should be 'additional'.","section":"Abstract"},{"comment":"In the introduction, 'uncentainty estimation' should be 'uncertainty estimation', and 'faciliate' should be 'facilitate'.","section":"Section 1"},{"comment":"Table 3 lists λ values for Mujoco and Antmaze tasks but not for the Kitchen tasks, despite Kitchen results appearing in Table 1. Please report the λ values used for Kitchen.","section":"Appendix C, Table 3"},{"comment":"The caption of Figure 3 says 'this table shows the results' although the results are displayed in a figure; this appears to be a copyediting error.","section":"Figure 3 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper is an arXiv posting of an ICLR 2025 conference paper. If this is being considered as a journal submission, the authors should clarify what new material is added beyond the conference version, since the present text does not contain such a statement. The main issue is that the flagship theoretical guarantee (Theorem 4) is about an idealized distribution sequence and does not match Algorithm 1; this is fixable by reframing or by adding a real analysis of the support-filtering operation, but it is not merely a presentation issue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"I'll give you the short version first: this paper has a real algorithmic idea, and the empirical work is worth taking seriously, but the headline theory overstates what is actually proved. The two-stage correction—learn the action ratio, then solve an OPE problem to get the state-action ratio—is a clean fix for a known semi-gradient distortion, and the iterative support filtering is a sensible way to focus imitation on plausible transitions. The D4RL results and the corrupted-demo experiments show consistent gains with low variance across seven seeds, and the ablation shows both components matter. That alone makes it worth reading.\n\nThe soft spots are mostly in the claims, not the method. The biggest one: Theorem 4 proves monotonic improvement for an idealized sequence where the reference distribution is replaced by d*_k, but Algorithm 1 filters by w_k > 0. That gives you the original data restricted to a support, not a resampling from d*_k. So the next iteration's behavior distribution is p_k conditioned on the support, and the theorem's V(D_{k+1}) doesn't track the actual filtered dataset. The paper says \"approximately\" in the text, but the theorem is stated as a property of the algorithm. That is a genuine gap, and the monotonic-improvement claim should be presented as a heuristic.\n\nAlso, the abstract says \"beats ... on all datasets,\" but Table 1 shows several tasks where IDRL is not the best (e.g., walker2d-medium-replay, halfcheetah-medium-expert). The text later says \"almost all,\" so the abstract is just overreaching. The \"unbiased estimator\" language in Lemma 1/Theorem 2 is only correct in population: the sample argmin is not an unbiased estimate of (f')^{-1}(E[g]) because of the nonlinearity. Minor. And Theorem 3's proof writes D_{k+1}-D_k where it should be D_k-D_{k+1}. Likely a typo, but confusing.\n\nNo code is shipped, which makes the empirical claims harder to verify. I'd ask for code.\n\nWho is this for? People working on DICE-style offline RL. It gives a practical recipe and a useful negative observation about semi-gradient updates. I'd send it to review—it deserves referee time, but the theory needs revision or honest relabeling as motivation, not proof.","headline":"IDRL is a genuinely useful offline RL algorithm with a clean correction for semi-gradient Dual-RL, but its monotonic-improvement theorem doesn't apply to the implemented filtering step.","tokens_in":22157,"tokens_out":4293,"would_cite":true,"duration_ms":38910,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that offline RL is best solved as imitation learning with a corrected, iteratively refined visitation ratio, and that this single mechanism suffices to beat the strongest primal- and dual-RL baselines on every dataset…","keywords":["offline reinforcement learning","dual reinforcement learning","visitation distribution ratio","discriminator-weighted behavior cloning","semi-gradient fixed point","iterative dataset filtering","D4RL benchmark","corrupted demonstrations"],"falsifier":"Construct a synthetic MDP with a known behavior policy $\\mu$ and a known expert policy $\\pi_E$, so the true ratios $\\pi_E(a|s)/\\mu(a|s)$ and $d_E(s,a)/d_D(s,a)$ can be computed exactly. Run IDRL's stage one and its two-stage correction on a dataset drawn from $\\mu$, and compare the recovered weights with the true ratios at every state-action pair. If the recovered action ratio is not pointwise close to the true ratio, or if the positive-weight support excludes state-action pairs the true optimal policy visits, then the fixed-point correction and the filtering guarantee that rest on it are empirically false. The same test, applied to the corrupted-demonstration setting, would check whether IDRL's filtered support matches the support of the hidden expert transitions rather than a superset of them.","tokens_in":21029,"feed_emoji":"🎯","tokens_out":14537,"duration_ms":111985,"temperature":0.7,"pith_summary":"The paper's claim is that offline reinforcement learning — learning a policy from a fixed dataset without further interaction — is best understood as a form of imitation learning: assign every logged transition a weight equal to how much more likely the optimal policy is to take that action than the data-collecting policy was, then clone the data under those weights. It shows that dual RL, the family of methods that tries to learn exactly this visitation ratio, has been estimating the wrong object: a semi-gradient update silently changes the fixed point so that what is learned is an action ratio, which assigns positive weight at states the optimal policy would never visit. The proposed method, IDRL, corrects the ratio in two stages and then exploits the fact that the optimal ratio is sparse — transitions with zero weight are suboptimal and can be removed — by iteratively filtering the dataset and repeating the estimation on the cleaned set. The paper proves that the filtered datasets improve monotonically in attainable return and reports that IDRL matches or beats the strongest primal- and dual-RL baselines on D4RL benchmarks and on corrupted demonstrations where expert transitions are as rare as one percent of the data. If the central claim is right, the practical payoff is a parameter-light offline RL algorithm that needs no expert data, no generative policy model, and no environment interaction, and that gets more stable as it runs more iterations.","feed_headline":"Self-filtering data lifts offline RL past its strongest baselines","feed_subtitle":"IDRL corrects a hidden ratio error in dual RL, then keeps only the transitions worth imitating — no expert data needed.","key_machinery":"The load-bearing object is the corrected state-action visitation distribution ratio $w^*(s,a) = w^*(s) \\cdot w^*(a|s)$ of Eq. 12. It is assembled in two stages: stage one runs semi-gradient Dual-RL (Eqs. 3-4) to obtain the action ratio $w^*(a|s) = \\pi^*(a|s)/\\mu(a|s)$; stage two feeds that ratio into a Fenchel–Rockafellar dual of a Bellman-flow-constrained off-policy evaluation problem (Eqs. 7-9), yielding convex, unbiased objectives (Eqs. 10-11) for the state weight $W^*(s)$ and the value function $U^*(s)$, with $w^*(s) = \\max(0, (f')^{-1}(\\mathbb{E}_{a\\sim\\mu}[w^*(a|s)(T U^*(s,a) - U^*(s))]))$. The second mechanism is the iterative filter: after each pass the dataset is replaced by its positive-weight support, $D_{k+1} = \\{(s,a,r,s') \\in D_k : w_k(s,a) > 0\\}$, which the paper interprets as replacing the behavior visitation distribution with the optimized one; Theorem 4 guarantees monotone improvement in the filtered datasets' attainable return, and Theorem 3 turns that into a behavior-cloning performance bound by viewing iteration $k$ as weighted-BC on an expert set $D_{k+1}$ with a supplementary set $D_k \\setminus D_{k+1}$.","core_discovery":"IDRL's central claim is that the optimal discriminator weight — the ratio of the optimal policy's visitation distribution to the behavior policy's — can be recovered from a static dataset alone, without expert data, provided two errors in current Dual-RL methods are corrected. Proposition 1 establishes that semi-gradient Dual-RL converges to the action distribution ratio $w^*(a|s) = \\pi^*(a|s)/\\mu(a|s)$ rather than the state-action visitation ratio $w^*(s,a) = d^*(s,a)/d_D(s,a)$; the action-only ratio cannot tell whether a state is worth visiting and, under function approximation, generalizes badly to suboptimal states. The correction treats the action ratio as known and recovers the missing state factor by solving a Bellman-flow-constrained off-policy evaluation problem (Theorem 1), with Lemma 1 and Theorem 2 providing unbiased sample-based objectives (Eqs. 10-11); the product $w^*(s,a) = w^*(s) \\cdot w^*(a|s)$ is the corrected state-action ratio. Because this corrected ratio is sparse, the regularized optimal policy's support defines a cleaned dataset, and IDRL iterates: filter, re-estimate, filter again. Theorem 3 bounds the weighted-BC suboptimality of the resulting curriculum by the usual imitation gap plus a term that shrinks as the filtered dataset approaches the expert distribution, and Theorem 4 guarantees $V(D_{k+1}) \\geq V(D_k)$ — each iteration's dataset is no worse than the previous one in attainable return.","pith_inferences":["A direct diagnostic follows from the paper's own decomposition: train an independent behavior model $\\hat{\\mu}(a|s)$ on the dataset and check whether $w^*(a|s) \\cdot \\hat{\\mu}(a|s)$ sums to one over actions per state; systematic deviation would flag fixed-point drift in the semi-gradient stage before it contaminates the filter — a test the paper does not run.","The hard-thresholding filter ($w>0$) could be softened into a smooth reweighting with a small floor for low-weight transitions; this would likely reduce the small-data generalization problems the authors list as a limitation, at the cost of a slightly weaker sparsity guarantee.","Nothing in the correction is specific to offline RL: the same action-ratio-to-visitation-ratio fix applies to off-policy evaluation, so the two-stage OPE correction may improve DualDICE-style stationary-ratio estimators in settings the paper never touches.","The monotone-improvement theorem suggests a self-tuning stopping rule — iterate until $V(D_{k+1}) - V(D_k)$ falls within noise — which would remove the manually chosen iteration count $M$ from the algorithm."],"forward_implications":["Any Dual-RL method that trains its value head with a semi-gradient update — IQL, SQL, and their relatives — is learning an action ratio, not a visitation ratio; IDRL's two-stage correction can be grafted onto them as a policy-extraction upgrade.","The sparsity result gives a principled transition-level filtering rule (keep the positive-weight support) that is strictly finer than trajectory-level filtering such as X%-BC, which keeps or discards whole trajectories.","The oracle experiment that motivates the paper — expert-augmented discriminator-weighted BC beating classical offline RL — becomes achievable without any expert data, since iteration replaces the missing expert distribution.","Because Theorem 4 guarantees $V(D_{k+1}) \\geq V(D_k)$, practitioners can pick the iteration count by watching the filtered dataset stabilize instead of tuning a regularization strength down, which the paper argues has no monotonicity guarantee and risks overestimation.","On corrupted demonstrations, iterative filtering should recover the rare high-performing transitions that reweighting-based methods miss, which is exactly the regime the paper tests at 1%, 5%, and 10% expert ratios."],"supporting_citations":[{"why":"Supplies the Fenchel–Rockafellar duality and off-policy evaluation construction behind the two-stage state-ratio correction (Eqs. 7-9).","marker":"Nachum & Dai, 2020"},{"why":"Establishes the Dual-RL unification and the lambda-parameterized surrogate objective (Eq. 3) that IDRL's semi-gradient analysis starts from.","marker":"Sikchi et al., 2023b"},{"why":"Shows the regularized optimal visitation ratio is sparse; the zero-weight property IDRL's dataset filter exploits, and the SQL baseline that inherits the action-ratio error.","marker":"Xu et al., 2023"},{"why":"O-DICE, the strongest Dual-RL baseline, which targets the correct state-action ratio via orthogonal gradients but requires a hard-to-tune eta that IDRL avoids.","marker":"Mao et al., 2024a"},{"why":"Advantage-weighted regression, the weighted-BC extraction step (Eq. 6) that turns the learned ratio into a policy.","marker":"Peng et al., 2019"},{"why":"Discriminator-weighted offline imitation learning, the oracle procedure whose strong results motivate the optimal-discriminator view.","marker":"Xu et al., 2022b"},{"why":"Supplies the imitation-gap bound that IDRL's Theorem 3 builds on for the iterative filtering analysis.","marker":"Li et al., 2024"},{"why":"The D4RL benchmark suite, the main evaluation setting for both the D4RL and corrupted-demonstration experiments.","marker":"Fu et al., 2020"},{"why":"ReBRAC, the strongest Primal-RL baseline that IDRL must outscore; this comparison establishes the in-sample advantage claim.","marker":"Tarasov et al., 2024"},{"why":"IQL, a semi-gradient Dual-RL baseline whose weighted-BC extraction demonstrates the action-ratio error IDRL corrects.","marker":"Kostrikov et al., 2021b"}],"fun_headline_variants":["IDRL corrects dual-RL ratio, filters data to beat baselines","Self-filtering offline RL: IDRL needs no expert demonstrations","Iterative dual RL: prune zero-weight data, improve each round","IDRL recovers optimal discriminator weight from static offline data","Offline RL gains: IDRL iteratively cleans dataset, outperforms prior"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument collapses if the action-level ratio learned in stage one is not exactly $\\pi^*(a|s)/\\mu(a|s)$ — if the semi-gradient fixed point is only approximately the implicit policy ratio, or if finitely many gradient steps on Eqs. (10)-(11) leave $W^*(s)$ inaccurate, then the product $w^*(s) \\cdot w^*(a|s)$ inherits both errors and the dataset filter can discard the very transitions the optimal policy needs.","fun_headline_variants_meta":{"raw":{"variants":["IDRL corrects dual-RL ratio, filters data to beat baselines","Self-filtering offline RL: IDRL needs no expert demonstrations","Iterative dual RL: prune zero-weight data, improve each round","IDRL recovers optimal discriminator weight from static offline data","Offline RL gains: IDRL iteratively cleans dataset, outperforms prior"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000213,"raw_usage":{"total_tokens":1503,"prompt_tokens":1107,"completion_tokens":396,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":723,"completion_tokens_details":{"reasoning_tokens":303}},"tokens_in":723,"tokens_out":396,"duration_ms":4522,"temperature":1.0,"reasoning_tokens":303,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:10:07.725803+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Construct a synthetic MDP with a known behavior policy $\\mu$ and a known expert policy $\\pi_E$, so the true ratios $\\pi_E(a|s)/\\mu(a|s)$ and $d_E(s,a)/d_D(s,a)$ can be computed exactly. Run IDRL's stage one and its two-stage correction on a dataset drawn from $\\mu$, and compare the recovered weights with the true ratios at every state-action pair. If the recovered action ratio is not pointwise close to the true ratio, or if the positive-weight support excludes state-action pairs the true optimal policy visits, then the fixed-point correction and the filtering guarantee that rest on it are empirically false. The same test, applied to the corrupted-demonstration setting, would check whether IDRL's filtered support matches the support of the hidden expert transitions rather than a superset of them.","supporting_citations":[],"review_version":1}