{"id":"df143bd2-cb96-4377-becd-a280354ec1c3","arxiv_id":"2501.19256","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"low","formal_verification":"none","parameter_count":0,"one_line_summary":"For evaluating explanations in reinforcement learning, this paper recommends measuring observable human behavior, such as predicting the agent's next action or goal, instead of relying only on subjective ratings.","lead":"This paper argues that explainable reinforcement learning should be evaluated with objective behavioral metrics, such as whether people can predict an agent's next action or goal, rather than only subjective self-reports. It catalogs these metrics and illustrates them in a simple gridworld, aiming to make XRL user studies more reproducible and comparable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Proposed objective metrics largely measure policy mimicry, not actionability; the construct-validity link is unestablished and should gate the central recommendation.","rationale":"The reader identifies the load-bearing premise as the validity of the proposed metrics as proxies for actionability. I agree this is the central weakness, but I sharpen it: the issue is not merely that the metrics lack empirical validation; several metrics are conceptually better characterized as measures of policy fidelity or user mental-model accuracy than of actionability as defined by the paper. Next-action, goal, and sub-goal prediction ask users to reproduce the agent's decisions; counterfactual policy asks users to predict state changes that alter the policy. These are observable behaviors, but the paper does not argue why reproducing the agent's policy should enable debugging or improve teaming outcomes. Task completion is more directly behavioral, but the paper notes it can be achieved without explanation, so its sensitivity to explanation quality is unclear without a baseline. The paper is honest about these limitations in Section 3, and as a position paper it makes a reasonable case for objective metrics. However, the central recommendation—that researchers adopt these metrics as a methodological standard—depends on the construct-validity link, and that link is the least secure part of the argument. The reader's CONDITIONAL verdict is therefore appropriate; my concern does not change that verdict, but it specifies where additional evidence is most needed.","tokens_in":7342,"tokens_out":2750,"duration_ms":32971,"concrete_test":"Run a within-subjects user study in the mini-world environment with two explanation conditions plus a no-explanation baseline. For each participant, measure: (a) next-action prediction accuracy, (b) goal prediction accuracy, (c) task completion score in a teaming setting, and (d) debugging success, defined as correctly identifying and fixing a planted policy bug. Compute the correlation between prediction accuracies and debugging success, and compare task completion across the three conditions. If prediction accuracy is high while debugging success is at chance, or if task completion does not differ from the no-explanation baseline, the proposed metrics do not operationalize actionability as defined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that objective human metrics should be used because subjective metrics only capture opinions, whereas objective metrics capture an explanation's actionability, defined in Section 1 as \"the ability of explanations to affect or change human behaviour.\" For this claim to hold, the proposed metrics must actually track actionability in debugging and teaming contexts. That link is not established, and for several metrics it is conceptually doubtful. Next Action Prediction, Goal Prediction, Sub-Goal Prediction, and Counterfactual Policy (Section 2.2) all ask participants to reproduce or anticipate the agent's policy choices: they measure whether users can mimic the agent's decision-making, not whether explanations change behavior in a way that improves task outcomes. A user who can predict the agent's next action may still be unable to identify a policy bug or take corrective action; prediction accuracy is a fidelity measure, not a behavioral-effectiveness measure. Task Completion (Section 2.3) is closer to actionability, but the paper itself notes in Table 2 that it \"can be achieved without an explanation,\" so without a controlled no-explanation baseline it cannot isolate the effect of the explanation. Section 3 acknowledges that \"in practice, our proposed metrics may be instantiated in a wide range of ways and their effectiveness in capturing explanation success will vary,\" but it does not provide evidence, from a user study or otherwise, that any of the metrics correlate with downstream debugging or teaming performance. Absent this criterion validity, the paper's recommendation risks replacing one unvalidated evaluation practice with another that is more reproducible but not demonstrably more actionability-relevant.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper is a position/vision paper arguing that evaluation of explainable reinforcement learning (XRL) should include objective human-subject metrics based on observable behavior, and that subjective self-reports alone are insufficient. It defines an explanation's actionability as its ability to affect or change human behavior, and it curates a set of objective metrics for two application contexts: debugging (next action prediction, goal prediction, sub-goal prediction, counterfactual policy, time taken) and human-agent teaming (task completion, inter-agent conflict, time taken). The paper illustrates each metric through a small grid-world environment ('mini-world') and discusses how subjective and objective metrics can complement each other, while also recommending baselines and standardized benchmarks for future work.","tokens_in":7545,"tokens_out":4036,"duration_ms":37958,"significance":"If the central claim is accepted, the paper could help shift XRL evaluation practice from self-reported opinions toward measurable behavioral outcomes, improving reproducibility and comparability across studies. Its strengths are its clear definitions, its concrete tabular comparison of metric pros and cons, its worked mini-world examples, and its transparent acknowledgment of limitations, including the need for 'No Explanation' baselines and the variability of metric effectiveness. However, the central claim that the proposed metrics are valid measures of actionability and that they make XRL research 'epistemically grounded' is not demonstrated: no user study, benchmark, or formal derivation supports the construct-validity link. The paper is best read as a proposal rather than an established result, and the gap between the actionability definition and several of the proposed metrics is a load-bearing concern that needs to be addressed.","major_comments":[{"comment":"The debugging metrics Next Action Prediction, Goal Prediction, Sub-Goal Prediction, and Counterfactual Policy all ask participants to predict or reproduce the agent's policy choices. These measure the participant's fidelity of mental simulation of the policy, not whether an explanation changes behavior in a way that improves debugging or teaming outcomes. This sits uneasily with the Section 1 definition of actionability as 'the ability of explanations to affect or change human behaviour': a user who can predict the agent's next action may still fail to identify a policy bug or to take corrective action. Either the definition of actionability must be broadened to include policy comprehension, or the paper must provide an argument or evidence connecting prediction accuracy to downstream task performance.","section":"Section 2.2"},{"comment":"Task Completion is the metric closest to behavioral effectiveness, but the paper itself notes in Table 2 that it 'can be achieved without an explanation.' The Section 3 recommendation of a 'No Explanation' baseline is therefore essential, but the paper does not explain how the metric should be analyzed relative to that baseline, nor does it discuss ceiling or floor effects. A user study that compares task completion with and without explanations, and that reports effect sizes and confidence intervals, is needed to support the claim that this metric captures explanation actionability; as written, the methodology is underspecified.","section":"Section 2.3 and Table 2"},{"comment":"The abstract and Section 1 assert that objective human metrics enable 'reproducible, comparable, and epistemically grounded research' and 'better reproducibility, incremental work, and more rigorous practices.' Yet Section 3 admits that 'in practice, our proposed metrics may be instantiated in a wide range of ways and their effectiveness in capturing explanation success will vary.' This admission undercuts the strength of the central claim. Since no user study, benchmark, or formal analysis is presented, the claim is a plausible position rather than a demonstrated result; the paper should either qualify the claim accordingly or include a concrete validation agenda specifying the controlled comparisons needed to establish each metric's validity.","section":"Section 3"},{"comment":"The paper states that 'we are unaware of an objective method for determining an agent's \"true\" sub-goal' and footnote 2 notes that the effectiveness of existing approaches 'is unknown.' This means sub-goal labels must be handpicked and annotated by researchers, which introduces experimenter judgment into the supposedly objective measurement. The paper should clarify how the metric remains objective when the ground-truth sub-goal is subjectively assigned, or explicitly classify this metric as relying on a researcher-defined ground truth with its own validation burden.","section":"Section 2.2, Sub-Goal Prediction"}],"minor_comments":[{"comment":"Typo: '(i.e., the soldier in the red circle).' should not have a stray comma after the closing parenthesis; the sentence continues with 'but for Section 2.3'.","section":"Section 2.1"},{"comment":"The phrase 'passabletiles' is missing a space between 'passable' and 'tiles'.","section":"Figure 1 caption"},{"comment":"The metric is called 'Task Competition' in Table 2 but 'Task Completion' in Section 2.3; the terminology should be made consistent.","section":"Table 2 and Section 2.3"},{"comment":"The sentence 'users can mentally simulate an state to check' contains a grammar error; it should read 'a state'.","section":"Section 2.2, Counterfactual Policy"},{"comment":"The reference format '[17, Table 1.]' includes an unnecessary period inside the bracket; it should be '[17, Table 1]'.","section":"Section 1"},{"comment":"The term 'agentic systems' is used without definition; since the paper addresses a general AI audience, it would be helpful to define it or use a more standard term such as 'autonomous agents.'","section":"Section 1"}],"recommendation":"major_revision","confidential_remarks":"This is a position/vision paper, so the absence of a user study is not by itself disqualifying. However, the central construct-validity gap between the definition of actionability and the proposed metrics must be resolved before publication. The self-citations are used as illustrative examples and are not load-bearing for the core argument, so I see no circularity problem. I recommend major revision rather than rejection because the framework is potentially useful and the paper is transparent about its limitations; the authors need to either substantially qualify the epistemic claims or provide a concrete validation agenda that would make those claims testable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a position paper, not an empirical study, and its value is the curation of objective metrics plus the debugging/teaming taxonomy. The central recommendation—prefer objective behavioral metrics over subjective ratings—is sensible, but the link from these metrics to 'actionability' is assumed rather than demonstrated.\n\nThe paper does several things well. It defines actionability clearly, separates debugging from teaming, and gives concrete examples of six metrics with pros and cons. The tables are useful reference material. The authors are honest about limitations: they note that task completion can be achieved without an explanation, recommend a no-explanation baseline, and point to cognitive psychology for experimental design. That's more self-awareness than most position papers show.\n\nThe soft spot is real, and the stress-test note captures it. Next-action, goal, sub-goal, and counterfactual prediction ask participants to reproduce the agent's choices. Those are fidelity measures—they test whether a user can mimic the policy, not whether they can spot a bug or correct the agent's behavior. Task completion and inter-agent conflict are closer to actionability, but without a baseline they are confounded; the paper acknowledges this. The missing piece is criterion validity: no user study or benchmark shows that better scores on these metrics correspond to better debugging or teaming outcomes. The claim that objective metrics are 'more epistemically grounded' is asserted, not shown. That matters because the whole point of the paper is to change evaluation practice.\n\nThat said, for a position paper, the absence of a user study isn't fatal. The contribution is conceptual, and the authors don't oversell the metrics as validated—they explicitly call the paper a first step. The mini-world is illustrative, not a released benchmark; that's fine for this purpose.\n\nI'd like to see the final version engage explicitly with the mimicry concern and propose one experiment that could discriminate between policy mimicry and actionability. But as is, this is a solid, readable position piece that deserves a serious referee. The target reader is anyone designing XRL user studies; it would be a good starting point for a methods section.","headline":"A useful taxonomy of objective XRL metrics, but the actionability link is asserted rather than shown; still deserves a serious referee.","tokens_in":8101,"tokens_out":3424,"would_cite":true,"duration_ms":30887,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Explainable reinforcement learning should be judged by whether humans can act on an explanation, not just by what they say about it.","keywords":["Explainable reinforcement learning","Human-subjects evaluation","Objective metrics","Subjective metrics","Actionability","Human-agent teaming","Debugging","User study"],"falsifier":"Run a between-subjects study in mini-world with two explanation mechanisms, one optimised for next-action prediction and one for task completion. If participants using the better-prediction mechanism still fail to spot a deliberately planted agent fault, or complete debugging tasks no better than a no-explanation control, then next-action prediction does not measure actionability.","tokens_in":7126,"feed_emoji":"🎯","tokens_out":7757,"duration_ms":63607,"temperature":0.7,"pith_summary":"This paper argues that self-reported subjective metrics—confidence, comprehension, preference—cannot tell researchers whether an explanation actually helps a human act on an agent's behaviour. To fix this, it proposes evaluating explanations through objective, observable human behaviour, defining 'actionability' as the capacity of an explanation to change or guide what a person does. It curates two families of metrics: debugging metrics (next-action, goal, and sub-goal prediction; counterfactual state selection; time taken) and teaming metrics (task completion, inter-agent conflict, time taken), illustrating each in a grid-based mini-world. If adopted, the field would move from opinion polls to measurable behavioural outcomes, making studies more reproducible and comparable. The paper frames this as a first step, not a settled protocol.","feed_headline":"Grade explainable RL by observable actions, not self-reports","feed_subtitle":"A metric suite for debugging and teaming measures whether explanations actually change what users do.","key_machinery":"The central object is the concept of actionability, defined as the ability of an explanation to affect or change human behaviour, operationalized through a catalogue of objective behavioural metrics. In debugging, the metrics are next-action prediction (can the user select the agent's action $\\pi(s)$ from the action set?), goal prediction, sub-goal prediction, counterfactual policy selection (which state change $s'$ makes $\\pi(s) \\neq \\pi(s')$ or $\\pi(s')=A$?), and time taken; in teaming, task completion, inter-agent conflict counts, and time taken. The mini-world grid environment serves as a shared illustration for how each metric would be administered, and the metrics are structured around two application goals—debugging agent behaviour before or after deployment and supporting human-agent teaming—so that the choice of metric is tied to the goal and audience of the explanation.","core_discovery":"The central claim is that an explanation's quality in XRL should be assessed by its actionability—whether it changes or enables human behaviour—measured through objective tasks rather than subjective self-report. For debugging, users should be able to predict an agent's next action, overall goal, or intermediate sub-goal, or to identify counterfactual state changes that would alter the agent's decision; for teaming, the measures are task completion, the number of inter-agent conflicts, and time taken. These metrics are presented as complements, not replacements, to subjective measures: subjective reports can track preferences and confidence, while objective metrics track practical effectiveness. The paper also recommends a 'No Explanation' control condition and argues that the choice of environment and experimental controls, informed by cognitive psychology, determines whether the metrics measure what they claim.","pith_inferences":["Beyond the paper's explicit claims, the metric list suggests a cost-fidelity ordering: next-action prediction is cheap but narrow, task completion is behaviourally rich but confounded, so practical protocols would pair a low-cost prediction metric with an outcome metric.","The 'No Explanation' baseline, if standardised, would allow effect-size meta-analysis across XRL studies, a step the paper calls for but does not formalise.","A testable extension is a mini-world comparison where objective and subjective rankings of explanation mechanisms diverge; such a divergence would directly demonstrate the paper's central distinction between actionability and opinion.","The debugging/teaming split leaves other explanation goals, such as trust calibration or regulatory accountability, without operationalised metrics; extending the catalogue to those goals would require new behavioural tasks."],"forward_implications":["Researchers evaluating XRL explanations would replace or supplement Likert-style self-reports with behavioural tasks such as next-action, goal, and sub-goal prediction, making results more directly comparable across studies.","Debugging and teaming require different metrics; a metric that validates an explanation for fault investigation will not necessarily validate it for human-agent coordination, so papers must state which goal they target.","Every human-subject study should include a 'No Explanation' baseline, preventing results from being attributed to explanations when users could have succeeded from task description alone.","Subjective and objective metrics are complementary: subjective measures capture preference and confidence, while objective measures capture actionability, and their correlation can reveal when users' beliefs about an explanation diverge from its actual effect.","Because metric effectiveness depends on experimental setup, the field needs standardised benchmarks and shared tasks, ideally in more realistic environments than toy gridworlds, to make cross-paper comparison meaningful."],"supporting_citations":[{"why":"documents the field's overwhelming preference for fidelity and performance metrics and the neglect of human-centred evaluation, framing the gap the paper addresses.","marker":"[17]"},{"why":"supplies the widely used subjective metrics the paper argues are brittle and insufficient on their own.","marker":"[11]"},{"why":"exemplifies next-action prediction as an objective measure of whether users can follow an agent's policy.","marker":"[16]"},{"why":"exemplifies goal prediction by asking users to identify an agent's strategy from explanations.","marker":"[12]"},{"why":"provides the causal explanation setting in which the proposed counterfactual policy metric would be implemented.","marker":"[9]"},{"why":"measures task completion in a human-agent support setting, grounding the teaming metric in concrete performance.","marker":"[21]"}],"fun_headline_variants":["Objective metrics for XRL: watch what users do","XRL evaluation: measure actions, not opinions","Stop asking users if explanations work—watch them","For explainable RL, objective behavior beats self-reports","Explainable RL: judge by actionability, not self-report"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The crucial assumption is that scoring well on tasks like predicting the agent's next action or completing a team task really does mean the explanation helps people in practice, even though no user study yet verifies that link.","fun_headline_variants_meta":{"raw":{"variants":["Objective metrics for XRL: watch what users do","XRL evaluation: measure actions, not opinions","Stop asking users if explanations work—watch them","For explainable RL, objective behavior beats self-reports","Explainable RL: judge by actionability, not self-report"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000152,"raw_usage":{"total_tokens":1157,"prompt_tokens":852,"completion_tokens":305,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":468,"completion_tokens_details":{"reasoning_tokens":228}},"tokens_in":468,"tokens_out":305,"duration_ms":3784,"temperature":1.0,"reasoning_tokens":228,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T20:48:07.330168+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a between-subjects study in mini-world with two explanation mechanisms, one optimised for next-action prediction and one for task completion. If participants using the better-prediction mechanism still fail to spot a deliberately planted agent fault, or complete debugging tasks no better than a no-explanation control, then next-action prediction does not measure actionability.","supporting_citations":[{"cited_title":"Milani, N","cited_arxiv_id":null,"evidence_quote":"documents the field's overwhelming preference for fidelity and performance metrics and the neglect of human-centred evaluation, framing the gap the paper addresses."},{"cited_title":"Madumal, T","cited_arxiv_id":null,"evidence_quote":"exemplifies next-action prediction as an objective measure of whether users can follow an agent's policy."},{"cited_title":"Huber, M","cited_arxiv_id":null,"evidence_quote":"exemplifies goal prediction by asking users to identify an agent's strategy from explanations."},{"cited_title":"Gyevnar, C","cited_arxiv_id":null,"evidence_quote":"provides the causal explanation setting in which the proposed counterfactual policy metric would be implemented."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"measures task completion in a human-agent support setting, grounding the teaming metric in concrete performance."}],"review_version":1}