{"id":"7d36478a-0556-4a23-9438-cbcec8a615a7","arxiv_id":"2607.13655","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"This paper introduces activation rate, feature coverage, syntactic distance, and semantic distance as objective metrics for logic-rule explanations of RL and multi-agent RL policies.","lead":"The paper turns reinforcement-learning policies into human-readable logic rules, then adds four metrics to measure how well those rules explain behavior and how they change over training or across agents. The metrics are meant to give an objective, user-free signal about what an RL agent has learned, beyond its reward curve.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Activation rate is computed on the same traces used to induce the ILASP theory, so it measures in-sample fit rather than faithful policy alignment; RQ1/RQ4 conclusions conflate approximation fidelity with agent policy properties.","rationale":"The reader's weakest assumption—that H is a faithful and stable surrogate for the neural policy—is exactly the load-bearing point. The activation rate is the metric used to support the strongest empirical claims (RQ1: action-specific learning dynamics, RQ4: transferability), and its definition in §4.3.1 does not specify that the traces used for metric computation are disjoint from those used for ILASP induction. If they are the same, the metric is a training-accuracy score, which is not an objective measurement of the agent's policy, but of the fit of the symbolic approximation. The RWARE result, α(H)≈0.30, shows that the theory is far from covering observed behavior, so the interpretation of low activation as 'agent uncertainty' or 'specialization' is not justified without separating approximation error from policy structure. A held-out evaluation and a control baseline would settle this cleanly. If the held-out α remains similar to the training α and significantly exceeds the control, then the metric would be validated as a faithful quantitative tool. If not, the paper's central contribution of 'objective, user-independent' metrics would be seriously weakened, though the framework could still be useful as a diagnostic of symbolic approximation quality. The paper does provide reproducible code and clearly defined metric formulas, which is positive, but the missing independent validation is a real correctness risk. I agree with the reader's conditional verdict and do not see a need to move to REJECT, because the concern is directly addressable with additional experiments.","tokens_in":27296,"tokens_out":5430,"duration_ms":61245,"concrete_test":"For each checkpoint in RWARE and Simple Adversary, split the collected traces into disjoint train and test sets at the episode level. Learn H from train traces; compute α(a) and α(H) on both train and test traces using Eq. 7–8. Also compute α on test traces for a control theory: e.g., a random theory of the same size/head distribution, and a theory that always predicts the global most frequent action. If test α is substantially lower than train α at any checkpoint, or if the control theory achieves comparable test α, then the activation rate is not measuring faithful policy alignment; it is measuring in-sample fit. The RQ1/RQ4 interpretations should be revised accordingly.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the proposed metrics provide objective, user-independent measurements of RL/MARL policies via ILP-learned ASP hypotheses. The most load-bearing assumption is that the learned hypothesis H (§4.2) is a faithful and stable surrogate for the neural policy, so that metrics on H can be read as properties of the agent. The weakest point is the activation rate (Def. 1, Eq. 7–8; §4.3.1), which is computed over the same execution traces used to construct the ILP examples E in §4.2. Under Eq. 3, ILASP is asked to find H such that for every observed state–action pair the chosen action atom is in the answer set and the other action atoms are not. Thus α(a) measures the fraction of training examples the induced theory happens to cover—essentially a train-set fidelity score—not an independently measured alignment with the agent. In RWARE (§5.3, Fig. 6a) α(H) saturates at ~0.30, i.e., the theory fails to explain 70% of the very traces it was induced from. The paper interprets this residual as evidence about action-specific confidence/specialization, but it could equally reflect ILASP's inability to fit the policy within the given search space and feature map. Without held-out traces or a baseline control, the claims in RQ1 and RQ4 (learning dynamics, transferability) rest on an unvalidated proxy. This is not an internal inconsistency of the metric, but it is a correctness risk for the headline contribution of objective policy quantification.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper introduces a framework for evaluating symbolic, logic-based explanations of RL and MARL policies. It uses ILASP to induce ASP rule sets from execution traces and then defines four metrics: activation rate, feature coverage, syntactic distance, and semantic distance. These are intended to quantify action-specific policy confidence, feature relevance, policy evolution during training, inter-agent agreement, and transferability. The experiments cover the Intersection single-agent domain, the RWARE cooperative MARL domain, and the Simple Adversary contrastive MARL domain, and each is organized around four research questions. The central claim is that the metrics provide objective, user-independent measurements of policy properties that go beyond return curves.","tokens_in":27711,"tokens_out":5402,"duration_ms":63305,"significance":"If validated, this would be a useful contribution to explainable RL: the metrics are formally defined, the code is released, and the multi-domain study gives concrete evidence that symbolic approximations can reveal action-level and agent-level structure not visible in global returns. The out-of-distribution activation-rate experiments in §5.6 are a particularly good step toward testing generalization of the learned rules. However, the paper's stronger interpretative claims—that activation rate measures agent confidence, or that semantic distance measures behavioral convergence—depend on the ILASP hypothesis being a faithful surrogate of the neural policy. The current evidence is largely in-sample, and one metric definition is ambiguous. These issues are fixable but need to be addressed before the central claims are fully supported.","major_comments":[{"comment":"Activation rate is computed on the same execution traces used to construct the ILP examples. Since ILASP is asked to find H that covers exactly those state–action pairs, α(a) is a train-set coverage score of the fitted symbolic model, not an independent measurement of how faithfully H represents the neural policy. The RWARE result in Fig. 6a (α(H) ≈ 0.30 at convergence) is interpreted in §5.3 as evidence about action-specific confidence, but an equally plausible reading is that ILASP's search space and feature map cannot fit the policy on 70% of the training traces. RQ1 and RQ4 therefore need a validation step: held-out traces, a random-policy or random-rule baseline, or a precision/recall decomposition. The OOD experiments in §5.6 partly address transfer but do not validate the training-phase interpretation.","section":"§4.2 / §4.3.1, Eqs. (3), (7)–(8)"},{"comment":"AS_H(c) is not defined precisely. If, as in ASP, the context C = F_F(s) is part of the program, the answer sets contain all context atoms, so the Jaccard index in Eq. (11) is inflated by the shared context and may be largely insensitive to behavioral differences. If only action atoms are intended, that restriction must be stated and used consistently in the experiments (Fig. 7b–d). Without this, the semantic-distance conclusions about inter-agent convergence and specialization rest on an ambiguous quantity.","section":"§4.3.4, Eq. (11)"}],"minor_comments":[{"comment":"The notation AS_{r_k} is used but never formally defined; it should be clarified whether this is the set of answer sets of H in which rule r_k is applicable/fires. The definition should be explicit to make Eq. (7) reproducible.","section":"§4.3.1, Eq. (7)"},{"comment":"The 'activation rate of the whole theory' is described as a sum of per-action rates. This is not the overall fraction of matched actions unless actions occur equally often; the sum weights actions equally. The paper should define α(H) explicitly (e.g., as a weighted average) and state which quantity is plotted in Fig. 6a.","section":"§4.3.1, Eq. (8) and text after it"},{"comment":"The 'objective, user-independent' claim is qualified by the fact that the whole pipeline depends on the user-defined feature map F_F and on the chosen value discretizations (Dist sets, Dir sets). This should be acknowledged explicitly in the limitations section, because different feature maps can lead to different metric values.","section":"§4.1–§4.3"},{"comment":"The interpretation that activation for move_up/move_down goes to zero because the adversary 'rarely moves up or down' conflates low action frequency with low rule fidelity. The denominator in Eq. (8) conditions on the action being performed, so a zero value means that whenever the action was performed the rule did not fire, which is a different statement. The text should be reworded to avoid this confusion.","section":"§5.3, Fig. 6c"},{"comment":"Several claims, especially in the abstract and Section 6, state that the metrics are 'planning-oriented' without demonstrating a connection to planning or sequential decision-making. Either define this term or soften the claim.","section":"General"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is within scope for JAIR and the proposed metric suite is potentially valuable, but the interpretation of the metrics as objectively describing the neural policy needs stronger validation. The in-sample circularity concern and the semantic-distance ambiguity are the two load-bearing issues. I would encourage the editor to ask the authors to add held-out evaluations and to tighten Definition 5."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper does something real. It packages four metrics for evaluating ILP-learned ASP policy approximations in RL and MARL—activation rate, feature coverage, syntactic distance, semantic distance—and shows they surface dynamics that return curves miss, like action-specific learning trajectories and inter-agent specialization. That addresses a known gap in XRL, where shared evaluation metrics are scarce. The experiments are extensive for a methods paper: three domains, multiple seeds, confidence intervals, transfer runs, permutation tests. Credit where due: this is more empirical grounding than most XRL papers bother with.\n\nThe soft spots are real, though. The activation rate is computed on the same traces used to build the ILP examples. So it measures in-sample fit of the induced theory, not independent alignment with the agent's policy. In RWARE it saturates around 0.30, meaning the theory fails to explain 70% of the very traces it was learned from. Interpreting that as action confidence or specialization conflates approximation fidelity with agent behavior. The RQ4 transfer experiments and RQ3 permutation tests do provide some independent grounding, so it's not hopelessly circular, but the core RQ1/RQ2 conclusions rest on this shaky measure. Fixable: compute activation on held-out trajectories, ideally fresh rollouts at each checkpoint.\n\nThere are two smaller issues. The semantic distance definition (Eq. 11) is ambiguous: AS_H(c) might include context atoms, and if so the Jaccard is trivially high regardless of action alignment. Needs explicit projection onto action atoms. And the paper compares against feature-importance baselines only qualitatively, not with measured differences; the example rules are from the best seed, which risks cherry-picking. These are minor beside the circularity.\n\nOne more thing: the \"first evaluation framework\" claim is plausible but not rigorously established—they cite the relevant surveys but don't systematically prove absence. That's a literature-claim weakness, not a technical one.\n\nOverall, the components are standard—Jaccard variants and frequency counts—but the packaged framework applied to training dynamics and MARL is genuinely new and likely useful. The paper deserves a serious referee. It should go to peer review, not be desk rejected. The circularity is addressable, and if fixed, this becomes a solid methods contribution. I'd bring it to reading group for the discussion of evaluation proxies in XRL.","headline":"A useful packaged set of metrics for ILP-based policy explanations, but the headline objectivity claim rests on an in-sample circularity that needs a held-out fix.","tokens_in":28153,"tokens_out":1771,"would_cite":true,"duration_ms":23839,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["68T05","68N17","68T27"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper treats the explainability of a reinforcement-learning agent as a measurable property of the logical rules learned from its behavior, not as a matter of user preference, and proposes four objective metrics to quantify that propert","keywords":["explainable reinforcement learning","inductive logic programming","answer set programming","policy evaluation metrics","multi-agent reinforcement learning","feature importance","policy transfer","semantic distance"],"falsifier":"Compute the activation rate of a converged rule set on both the traces it was learned from and on a fresh set of traces collected from the same agent in the same environment; if the rate drops sharply out of the training traces while staying high on them, the metric is measuring trace memorization, not policy alignment, and the framework's core diagnostic would be false.","tokens_in":27197,"feed_emoji":"🤖","tokens_out":8971,"duration_ms":85767,"temperature":0.7,"pith_summary":"The authors want explainable reinforcement learning to have shared, objective measurements instead of user studies; they claim this is the first evaluation framework aimed at quantitatively explaining multiple aspects of RL and multi-agent RL policies. Their method translates a trained agent's decisions into a compact set of Answer Set Programming rules, then defines four metrics on those rules: activation rate (how often a rule fires in agreement with the agent's action), feature coverage (how much each feature is used), syntactic distance (how differently two rule sets are built), and semantic distance (how differently they behave on identical observations). Experiments across three domains — driving through an intersection, a cooperative warehouse task, and an adversarial pursuit task — show the metrics exposing action-level learning dynamics, coordination and specialization between agents, and rules that transfer across harder variants. If the claim holds, policy explanation becomes a repeatable measurement rather than a subjective judgment, complementing the return curve and user opinion.","feed_headline":"Four metrics expose what a black-box agent's logic actually does","feed_subtitle":"Learned from the agent's own decisions, they reveal per-action confidence, coordination, and gaps returns hide.","key_machinery":"The load-bearing object is the logical policy approximation: an Answer Set Programming rule set (a rule-based logic formalism in which rules derive actions from observed features), learned from state-action traces, turning a black-box policy into an inspectable artifact. The four metrics are defined on it: activation rate counts how often rules reproduce observed actions; feature coverage counts feature mentions normalized by total body size; syntactic distance is an intersection-over-union distance between rule-body sets; semantic distance is the same overlap on the conclusions produced for identical contexts. This is what lets a qualitative judgment (is this explanation any good?) be repla","core_discovery":"Central claim: a trained agent's decisions can be summarized as a finite rule set, and four numbers computed on that rule set surface what return curves hide. Activation rate = how often a rule fires when the agent actually performed that action, read as per-action reliability. Feature coverage = the share of rule bodies mentioning a given feature, an action-level analogue of feature importance. Syntactic distance = one minus the intersection-over-union overlap between rule bodies for the same action across two policies. Semantic distance = the same overlap computed on the conclusions the two rule sets draw from identical observations, capturing behavioral rather than structural difference.","pith_inferences":["Editorial inference: the same numbers could drive data collection — traces where activation is low are precisely the states where the current rule set is wrong, so they are candidates for more exploration or for curriculum design.","Editorial inference: because all four metrics are computed on a user-chosen feature map, they double as a quantitative audit of that map: features that never appear in rules or whose permutation barely changes activation are candidates for removal, suggesting a principled way to co-design explanation and representation.","Editorial inference: if semantic distance keeps moving after return plateaus, it offers a testable stopping criterion or a stability monitor for deployment, since a policy can look converged in reward while still reshaping its behavior."],"forward_implications":["Activation rate works as a diagnostic that separates actions the agent has learned confidently from actions whose rules barely fire, catching instability or over-specialization that a flat return curve misses.","Feature coverage gives action-level feature relevance and lets a practitioner spot features that are effectively unused, supporting principled pruning of the feature set.","Syntactic and semantic distances track policy evolution during training and between agents, revealing coordination, specialization, and continued policy change even after mean return has plateaued.","Activation rate computed in altered environments measures how far learned rules generalize and points to which action-level skills are worth transferring.","Together the metrics position the logical rule set as a debugging artifact for the whole training loop, not just a final explanation."],"fun_headline_variants":["Four metrics decode the logic inside a black-box agent","ILP rules plus four metrics reveal RL policies' hidden logic","Measuring explainability: four metrics for RL rule sets","Activation, coverage, distances: quantitative RL explainability","From black-box to clear-cut: four metrics explain RL"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The framework assumes the shortest rule set induced from observed decisions is a faithful, stable surrogate for the agent's policy — the assumption enters when traces become learning examples and when the rule set is read as the explanation — and in the warehouse experiments the rules reproduce only about 30% of decisions, so low activation there may reflect approximation failure rather than agent uncertainty.","fun_headline_variants_meta":{"raw":{"variants":["Four metrics decode the logic inside a black-box agent","ILP rules plus four metrics reveal RL policies' hidden logic","Measuring explainability: four metrics for RL rule sets","Activation, coverage, distances: quantitative RL explainability","From black-box to clear-cut: four metrics explain RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000866,"raw_usage":{"total_tokens":3616,"prompt_tokens":793,"completion_tokens":2823,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":537,"completion_tokens_details":{"reasoning_tokens":2743}},"tokens_in":537,"tokens_out":2823,"duration_ms":18740,"temperature":1.0,"reasoning_tokens":2743,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T04:31:53.096046+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute the activation rate of a converged rule set on both the traces it was learned from and on a fresh set of traces collected from the same agent in the same environment; if the rate drops sharply out of the training traces while staying high on them, the metric is measuring trace memorization, not policy alignment, and the framework's core diagnostic would be false.","supporting_citations":[],"review_version":1}