{"id":"770703e3-0b42-4a05-a058-dddf369b5721","arxiv_id":"2508.15652","paper_version":2,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":5.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":0,"one_line_summary":"ICVs quantify each agent's causal influence on teammates' policy certainty and preference alignment to extract behavioral insights from policies alone, without reward signals.","lead":"This paper proposes Intended Cooperation Values (ICVs), a method based on information-theoretic Shapley values for scoring each agent's effect on teammates' policies in multi-agent reinforcement learning, using only policy distributions rather than rewards. The goal is to explain which agent behaviors support team success in cooperative and competitive settings, which matters for debugging and overseeing trained multi-agent systems.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"ICV causal attribution is not established: the abstract gives no interventional or confounder-controlled definition of 'causal influence,' and the supplied full text does not belong to this paper, so no definitions or experiments can be checked.","rationale":"The reader's weakest-assumption analysis identifies the same fragile premise: policy-distribution changes are attributed causally to the focal agent, but confounders exist in MARL. My read of the abstract agrees and sharpens the concern: no interventional semantics are disclosed, and information-theoretic Shapley values alone do not establish causation. The supplied full text is a different paper, so the body cannot be used to resolve this. Given that the reader's verdict is already UNVERDICTED with low confidence, the appropriate outcome is to leave that verdict unchanged. I am not arguing the method is wrong; I am arguing that the abstract's causal claim is unverified and that a concrete intervention-based check would settle whether the concern lands. No formal verification, code release, or reproducible experiment is available from the supplied material to offset this concern.","tokens_in":32366,"tokens_out":2313,"duration_ms":28572,"concrete_test":"Reconstruct or obtain the ICV estimator and run a controlled experiment where ground-truth causal influence is known by construction: use a two-agent partially observable game in which agent 1's action is the only cause of agent 2's certainty shift, while an unrelated shared environmental feature also shifts agent 2's policy. Compute ICVs from (a) observational rollouts and (b) interventional rollouts that replace agent 1's action with an exogenous sample (a do()-style intervention). If (a) and (b) diverge on the known-causal component—or if (b) is not defined in the paper—the causal interpretation in the abstract is not established. If the paper already contains such an intervention, locate the section and compare its estimator to the observational one.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that ICVs quantify each agent's causal influence on co-players' instrumental empowerment solely from policy distributions. The abstract never states whether the Shapley value is computed over interventional or observational conditionals. If ICVs are computed from observed trajectories, a change in a teammate's decision certainty or preference alignment can be driven by shared environment dynamics, other agents' actions, or the teammate's own exploration; the focal agent's action may merely be correlated with that change. Information-theoretic Shapley values do not by themselves control for confounding—they require a well-defined causal value function and a stated identification strategy. Without an explicit do-style intervention or a causal graphical model with an identifiability argument, the headline 'causal influence' is not supported. The body of the MARL paper is unavailable (the supplied full text is arXiv:2508.15647, a different distributed-systems paper), so the definitions of ICVs, the construction of the value function, and the experiments cannot be inspected. The load-bearing premise—that the ICV estimator identifies causal effects rather than policy covariation—is therefore unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The submission, arXiv:2508.15652, proposes 'Intended Cooperation Values' (ICVs), an information-theoretic Shapley-value method intended to quantify each agent's causal influence on co-players' instrumental empowerment in multi-agent reinforcement learning. The abstract claims that by analyzing only policy distributions—specifically decision (un)certainty and preference alignment—the method can identify which agent behaviors benefit team success in cooperative and competitive tasks, without any reward feedback. However, the full text supplied with the submission is not this paper at all: it is arXiv:2508.15647, 'CausalMesh: A Formally Verified Causally Consistent Distributed Cache with Support for Client Migration,' a distributed-systems paper. The actual MARL paper contains no visible equations, definitions, algorithms, experiments, or evaluation. Thus the submission as provided consists of a title and abstract asserting a specific technical contribution, with no technical content to verify.","tokens_in":32625,"tokens_out":3833,"duration_ms":44475,"significance":"If the claimed results were true, the work could be relevant to MARL explainability and to the broader question of attributing agent influence without reward signals. The stated ambition—extracting meaningful causal insight solely from policy distributions—would be a useful contribution if supported by a rigorous identification argument and empirical validation. However, the significance cannot be assessed from the current submission. The abstract alone provides no derivations, no formal definitions of the quantities being measured, no causal model or identification strategy, and no experimental evidence. The submission therefore does not meet the standard of a verifiable technical claim.","major_comments":[{"comment":"The full text supplied for this submission is an unrelated paper (arXiv:2508.15647, 'CausalMesh'), not the MARL/ICV paper advertised by the title and abstract. No equations, definitions, algorithms, or experiments for ICVs appear anywhere in the submission. The central claim of the paper is entirely unsupported because the actual manuscript is absent. This is a load-bearing defect: the review process cannot evaluate a paper that does not contain the claimed technical content.","section":"Full text (overall submission)"},{"comment":"The abstract states that ICVs 'quantify each agent's causal influence on their co-players' instrumental empowerment' but does not specify whether the underlying Shapley value is computed over interventional conditionals (e.g., do(a_i)) or over observational policy distributions. In the latter case, changes in a teammate's decision certainty or preference alignment could arise from shared environment dynamics, other agents' actions, or the teammate's own exploration, making the focal agent's action merely correlated with—not causally responsible for—the change. Without a stated causal model, a defined intervention, or an identifiability argument, the word 'causal' is not supported. The manuscript needs to provide this information or weaken the claim to 'covariation.'","section":"Abstract, 'causal influence'"},{"comment":"The submission does not define 'decision (un)certainty' or 'preference alignment' mathematically, nor does it state how an information-theoretic Shapley value is constructed from policy distributions. It is impossible to assess whether the proposed ICVs are well-defined, normalized, permutation-invariant, or identifiable from finite samples. The central claim that 'meaningful insights into agent behaviors can be extracted solely by analyzing the policy distribution' is therefore an assertion without formal content.","section":"Abstract, 'solely by analyzing the policy distribution'"},{"comment":"No experiments are reported in the submission. The abstract claims that the method 'identifies which agent behaviors are beneficial to team success,' including fostering deterministic decisions, preserving flexibility, and revealing similar or diverse strategies, but no environments, baselines, metrics, or results are provided. The empirical component of the claim is completely unsubstantiated.","section":"Abstract, 'cooperative and competitive MARL tasks'"}],"minor_comments":[{"comment":"The term 'instrumental empowerment' is used without a formal definition or citation to the empowerment literature. Please define it explicitly (e.g., mutual information between an action and future state entropy) and explain how it relates to the proposed ICV construction.","section":"Abstract"},{"comment":"The acronym 'ICV' is introduced but never expanded in the abstract; the paper should include a notation table and formal definitions of the Shapley-value estimator, including the coalition game and value function.","section":"Abstract"}],"recommendation":"reject","confidential_remarks":"This appears to be a submission-package error: the full text is a completely different paper. The editor may wish to return the submission to the authors to upload the correct MARL manuscript. However, as submitted, the paper contains no technical content matching its title and abstract, so the manuscript in its current form cannot be reviewed as a substantive contribution. The causal-attribution concern is genuine and would remain a major issue even if the correct full text were provided."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: we only see the abstract, and the supplied full text belongs to a different paper (CausalMesh), so this is an evaluation of a promise, not a paper. The promise is interesting. Applying information-theoretic Shapley values to teammates' policy distributions—with decision uncertainty and preference alignment as the features—is a plausible new application of known tools. The underlying problem, making MARL behavior explainable without reward feedback, is real, and the abstract states a clean research question and a concrete output: ICVs with two interpretable channels, determinism and preference alignment. That is enough to be worth a look.\n\nThe soft spot is the word \"causal.\" The abstract says ICVs quantify each agent's causal influence on co-players' empowerment, but it never states the identification strategy. In MARL, a teammate's policy changes because of shared environment dynamics, other agents' actions, and its own exploration, not just the focal agent's action. If the Shapley value is computed over observational conditionals, ICVs measure correlation with team success, not causal effect. The stress-test note is right that this is the load-bearing premise, and it is absent from the abstract. Also absent: any equations, experiments, or evaluation details. I am not faulting the paper for leaving proofs out of an abstract, but the abstract's claims go beyond what the abstract shows.\n\nOne mechanical point: the supplied full text is CausalMesh, a distributed cache paper, not this MARL paper. That is a pipeline issue in what got handed to us, not a judgment of the authors. It just means we cannot check definitions, derivations, or empirical results.\n\nVerdict: unverdictable from this material. But there is enough of a real idea here that I would still want it refereed, with one specific instruction to the referees: the authors need to show either a do-style intervention or a causal graphical model with an identifiability argument, or soften the causal claim to predictive attribution. If the full paper has that argument, this could be a solid contribution to the interpretable RL subfield. If not, it is a Shapley application dressed in causal language. I would send it to peer review and ask that question.","headline":"Only the abstract is in view, so this is a promise, not a paper; the idea is worth checking, but the causal language is doing more work than the abstract can support.","tokens_in":33076,"tokens_out":2127,"would_cite":false,"duration_ms":25789,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that in cooperative and competitive MARL, an agent's causal influence on teammates can be read solely from policy-distribution changes, quantified by Shapley values of instrumental empowerment.","keywords":["multi-agent reinforcement learning","causal attribution","Shapley values","instrumental empowerment","policy distributions","explainability","reward-free analysis","cooperative-competitive MARL"],"falsifier":"In a two-agent gridworld, make agent A's action unobservable to agent B (remove A from B's observation vector) while B still changes its policy because of shared environment dynamics. If ICVs still attribute nonzero causal influence to A on B's decisions, the method measures correlation rather than causation.","tokens_in":32302,"feed_emoji":"🤖","tokens_out":4649,"duration_ms":52315,"temperature":0.7,"pith_summary":"This paper sets out to show that in multi-agent reinforcement learning, an agent's contribution to team success can be identified without any reward or value feedback—simply by watching how its actions shift teammates' policy distributions. It introduces Intended Cooperation Values (ICVs), information-theoretic Shapley values that decompose an agent's action effect into two components: whether it makes teammates more or less decisive, and whether it pulls their preferences toward or away from its own. A sympathetic reader would care because current MARL explainability usually starts from explicit rewards, which are often absent, sparse, or uninformative. If the claim holds, behavior attribution and cooperation diagnostics become available from policies alone.","feed_headline":"Teammates' policy shifts reveal who really helps","feed_subtitle":"A Shapley-based score reads decision certainty and preference alignment to attribute influence without reward feedback.","key_machinery":"Intended Cooperation Values (ICVs): an information-theoretic Shapley-value attribution that measures one agent's causal influence on co-players' instrumental empowerment by scoring changes in decision uncertainty and preference alignment within teammates' policy distributions. This machinery carries the argument by turning 'who helped the team' into a quantity computed solely from policies, with no reward or value function.","core_discovery":"The central claim is that meaningful causal attribution in MARL does not require rewards: the policy distribution itself carries the information needed. ICVs quantify each agent's causal influence on its co-players' instrumental empowerment—the practical room for action a teammate has after observing the agent—via an information-theoretic Shapley value. Concretely, an agent's action effect on a teammate is scored by how much it changes that teammate's decision (un)certainty and how much it aligns or misaligns the teammate's preferences with its own. Across cooperative and competitive tasks, the method identifies behaviors that help the team, distinguishing between actions that foster determi","pith_inferences":["The supplied extract contains only the abstract; the derivation, implementation, and experiments for ICVs are not present here, so the claims rest on the abstract alone.","If policy-only attribution holds, ICVs could serve as a reward-free credit-assignment signal for shaping or diagnosing MARL training where rewards are absent—an extension the paper does not state.","A natural test is to check whether ICV attributions stay stable under shared-environment confounders, for instance when a teammate's policy shift is induced by the environment rather than by the focal agent's action.","The preference-alignment component might generalize to human-agent teams as a signal of whether a robot's actions expand or narrow a human partner's perceived action options."],"forward_implications":["Teams can be audited for individual contribution without any reward or value signal, which matters when rewards are sparse, delayed, or unavailable.","Agent behavior becomes classifiable by its effect: committing teammates to deterministic choices versus preserving their flexibility for future actions.","The degree to which agents adopt similar or diverse strategies becomes measurable from policy data alone.","Cooperative success can be attributed to specific agents in both cooperative and competitive settings, improving explainability of trained MARL systems."],"supporting_citations":[],"fun_headline_variants":["No rewards? Policy alone reveals agent impact","Shapley on policies exposes true team contributors","Action effects read from teammates' uncertainty and preferences","Attributing MARL influence without any reward signal","Policy-based Shapley scores uncover helpful agent behaviors"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The method assumes that when a teammate's policy distribution becomes more or less certain and its preferences shift, that change was caused by the focal agent's action; in reality, shared environment dynamics, other agents' actions, and the teammate's own exploration also move the policy, so the attribution may capture correlation with team success rather than causal influence.","fun_headline_variants_meta":{"raw":{"variants":["No rewards? Policy alone reveals agent impact","Shapley on policies exposes true team contributors","Action effects read from teammates' uncertainty and preferences","Attributing MARL influence without any reward signal","Policy-based Shapley scores uncover helpful agent behaviors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000129,"raw_usage":{"total_tokens":943,"prompt_tokens":714,"completion_tokens":229,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":157}},"tokens_in":458,"tokens_out":229,"duration_ms":3307,"temperature":1.0,"reasoning_tokens":157,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T17:45:00.566761+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"In a two-agent gridworld, make agent A's action unobservable to agent B (remove A from B's observation vector) while B still changes its policy because of shared environment dynamics. If ICVs still attribute nonzero causal influence to A on B's decisions, the method measures correlation rather than causation.","supporting_citations":[],"review_version":1}