{"id":"482164b0-42cf-4da8-8d0e-fabd999701b7","arxiv_id":"2505.05211","paper_version":1,"verdict":"ACCEPT","confidence":"MODERATE","novelty_score":1.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A taxonomy and literature review of strategic classification and performative prediction, organized into robustness, fairness, and improvement/causality perspectives.","lead":"This paper surveys the field of incentive-aware machine learning, where people can change their data to get better outcomes from algorithms. It groups the research into three viewpoints: making algorithms robust to gaming, ensuring fairness, and recognizing when strategic changes are genuine improvement.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's own concessions (Sections 4.3 and 6) admit that performative prediction and other topics do not fit the utility-based Stackelberg framework, so the claimed unified framework is less comprehensive than the abstract suggests.","rationale":"The reader's weakest assumption was that papers may not be cleanly assigned to one primary perspective, potentially undermining the three-way partition. My concern is related but distinct: even when the partition is accepted, the 'unified framework' language is not fully supported by the paper's own content. Section 4.3 and Section 6 contain explicit admissions that significant strands of the literature do not share the utility-based Stackelberg microfoundation or the three-perspective taxonomy. This is a concrete, textual basis for the concern, not a speculative worry about labeler subjectivity. \n\nI do not think this changes the verdict. The paper is a clearly written survey whose organizational value survives the overstatement: the three perspectives remain a helpful lens, the notation is useful, and the author is candid about exceptions. An ACCEPT verdict with the caveat that the abstract's 'unified framework' be softened is appropriate. Since the reader already noted the taxonomy's limitations, my additional point is a refinement rather than a new rejection. Hence UNCHANGED, with partial agreement: the reader identified a related weakness, but the more precise load-bearing issue is the mismatch between the claimed unified framework and the admit-ted exceptions like performative prediction.","tokens_in":17686,"tokens_out":7832,"duration_ms":78102,"concrete_test":"Construct a classification table of every paper cited in Sections 3-5. For each paper, mark three binary properties: (a) the model is an instance of Equations (1)-(3) with a utility-maximizing agent; (b) the model explicitly allows causal features that change the ground-truth label; (c) the model's primary contribution is a fairness or downstream-impact analysis. Then add a fourth row for papers in Section 4.3 and Section 6 (performative prediction, alternating leaders, self-selection). If a substantial fraction of entries do not satisfy (a), or if any perspective lacks at least one paper satisfying the corresponding formal objective in the Section 2 notation, the 'unified framework' claim fails and the abstract should be revised to say the Stackelberg model unifies the robustness core and is extended informally to the other two perspectives.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim has two parts: that the field can be organized into three perspectives, and that these perspectives share a common modeling framework based on the Stackelberg interaction in Section 2 (Equations (1)-(3)). The second part is the more load-bearing because it underwrites the abstract's promise of a 'unified framework encapsulating models for these perspectives.' The paper itself provides direct textual evidence that this framework is narrower than claimed. In Section 4.3, performative prediction (Perdomo et al. 2020) is described as 'another framework' whose 'main distinction' is that it uses smoothness assumptions on how x leads to x̂ 'instead of focusing on the agents' utility functions.' That is a different microfoundation from Equation (3), yet performative prediction is one of the most active lines in the area. Similarly, Section 6 lists 'a handful of topics related to incentive-aware ML settings that we did not touch upon, as they did not directly fit under one of our three outlined perspectives,' including alternating-leader games, econometrics for strategic agents, and strategic self-selection. These are not isolated edge cases; they are research programs with their own models. \n\nMoreover, the formal framework in Section 2 is essentially a robustness-perspective model. Equation (1) minimizes loss against a fixed ground-truth label h*(x), which presupposes the 'gaming only' assumption stated at the start of the Stackelberg protocol. The causality subsection relaxes this assumption by allowing some features to change y, but it introduces no analogous optimization problem. The fairness perspective is never given a formal objective in the Section 2 notation; it is described in prose as the study of downstream societal impacts.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper is a survey of incentive-aware machine learning, organizing the literature into three perspectives: robustness, fairness, and improvement/causality. It introduces a common modeling framework in Section 2 based on a Stackelberg interaction between a principal and strategically responding agents, formalized through ground-truth functions, agent utility functions, and offline and online objectives (Equations (1)-(3)). The survey then reviews representative work in each perspective, including causal models, performative prediction, fairness, and heterogeneous agents, and concludes with open directions and a candid list of omitted topics. The paper makes no original technical claims; its contribution is a taxonomy and a consistent notation for comparing existing results.","tokens_in":17932,"tokens_out":7499,"duration_ms":72333,"significance":"As a synthesis, the paper is valuable: it offers a readable entry point into a fast-growing area, proposes a useful three-way categorization, and carefully acknowledges its limitations (Sections 4.3 and 6). It gathers a broad set of references and consistently translates them into a common notation. The value of the survey, however, depends on the accuracy of its literature summaries and on how well the stated 'unified framework' matches the actual scope of the field. Those two points are the main sources of the concerns raised below. If the authors tighten the claims and correct the identified factual error, the survey will be a reliable and helpful resource.","major_comments":[{"comment":"The abstract and introduction claim a 'unified framework encapsulating models for these perspectives,' but the paper's own scope limitations contradict this. Section 4.3 describes performative prediction as 'another framework' that relies on smoothness assumptions on x to x-hat 'instead of focusing on the agents' utility functions,' which is a different microfoundation from Equation (3). Section 6 then lists several active research programs (alternating-leader games, econometrics for strategic agents, strategic self-selection) that 'did not directly fit under one of our three outlined perspectives.' The Section 2 formalism therefore unifies only the utility-based, best-response models within the three perspectives, not the entire field. Please revise the abstract and introduction to scope the claim explicitly (e.g., 'a unified framework for the utility-based Stackelberg models studied in the three perspectives') and position performative prediction and the Section 6 topics as adjacent frameworks rather than parts of the unified framework.","section":"Abstract, Section 1, Section 2"},{"comment":"The text attributes the 'strategic Littlestone dimension' to Ahmadi et al. [2021] (The Strategic Perceptron), but the reference list shows that Ahmadi et al. [2024] is the paper titled 'Strategic Littlestone Dimension: Improved bounds on online strategic classification.' The same sentence states that Ahmadi et al. [2021] investigates whether learnability of a concept class implies strategic learnability for general classes; the Strategic Perceptron paper instead proposes a margin-based algorithm for linear classifiers and does not address general concept-class learnability. Please correct the citation and the accompanying characterization, and adjust the sentence about Cohen et al. [2024a] and Ahmadi et al. [2021] accordingly.","section":"Section 3.2, paragraph on strategic learnability"},{"comment":"The abstract promises that the unified framework covers 'causal settings,' but Section 2 gives no formal model for the causal case. It only states that some features are causal and that several cited papers use structural causal graphs; no analogue of Equations (1)-(3) is written down for how ground truth changes when agents modify causal features. To support the claim of a unified framework, please either include a formal statement of the causal extension (e.g., how h*(x) is updated after causal feature changes) and reference the specific formal models in the cited papers, or soften the claim to say that causal settings are reviewed in Section 4.2 rather than being part of the unified framework.","section":"Section 2, Causality subsection; abstract"}],"minor_comments":[{"comment":"The text contains typographical and typesetting artifacts (e.g., 'c an', 'i nto', 'differentiating', and corrupted symbols like '/BD {sign}' in the loss definitions). These should be cleaned up for the final version.","section":"Abstract and throughout"},{"comment":"The summary of Levanon and Rosenfeld [2021] says 'the paper does not provide theoretical guarantees,' but that paper includes both theoretical and experimental content; consider clarifying what aspect lacks guarantees.","section":"Section 3.1"},{"comment":"The statement that 'most (if not all) of the papers discussed so far' focus on homogeneous populations is slightly incomplete, since some improvement-perspective papers (e.g., Alon et al. [2020], Haghtalab et al. [2020]) already model heterogeneous agents; a brief qualifier would improve accuracy.","section":"Section 5, first paragraph"},{"comment":"The subsection on performative prediction is very short and the transition to it is abrupt; given that the paper's own footnote in Section 1 lists performative prediction as part of the field, a few more sentences explaining why it does not fit the Section 2 framework would help the reader.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The paper is a literature review and does not claim new technical results, so the main risks are accuracy and scoping. The misattribution in Section 3.2 is a concrete factual error that must be fixed; the overstatement of the unified framework in the abstract is also worth addressing because it will shape how readers use the survey. The listed omitted topics and the performative prediction discussion suggest the authors are aware of the boundaries, so I expect the revisions to be straightforward. I recommend major_revision rather than reject because the core taxonomy is useful and the issues are correctable without changing the paper's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick take on arXiv:2505.05211. It is a literature review published in SIGEcom Exchanges, not a new-results paper. The two things to know: it is currently the cleanest single entry point to strategic classification and adjacent work, and its three-way split (robustness vs fairness vs improvement/causality) is a sensible organizer. Treat the taxonomy as a teaching device, not a theorem.\n\nWhat it does well: consistent notation across offline and online settings, clear explanation of the Stackelberg protocol, manipulation graphs, partial information, and heterogeneity. It gives accurate one-paragraph summaries of the main papers in the area — Hardt et al., Dong et al., Chen et al., Miller et al., Shavit et al., Bechavod et al., Hu/Milli, and others. I spot-checked the ones I know well and the characterizations are fair. It is also honest about scope: Section 6 lists alternating-leader games, econometrics for strategic agents, and strategic self-selection as topics that did not fit the three perspectives. That candor is real.\n\nSoft spots, in proportion. The abstract promises a \"unified framework encapsulating models for these perspectives,\" but Section 2 does not deliver that. The formal model is a robustness-perspective Stackelberg game with utility-maximizing agents (Equation 3). Fairness gets no formal objective in that notation; it is discussed in prose. Improvement and causality are motivated well but have no analogous optimization problem. And Section 4.3 explicitly calls performative prediction \"another framework\" based on smoothness assumptions rather than utility functions. So the unified framework is really a unified robustness/Stackelberg language, with fairness and causality as perspectives attached to it. Still useful, but the abstract oversells it.\n\nOne minor citation error: Section 3.2 attributes the \"strategic Littlestone dimension\" to Ahmadi et al. [2021] (the Strategic Perceptron paper), but the 2024 Ahmadi et al. paper is the one that introduces that dimension. Worth fixing.\n\nSelf-citation is present, but the papers cited are real and the summaries match the literature; I don't see it as a red flag.\n\nBottom line: this is a good survey for newcomers and for course reading. It advances no new science, and the unified-framework claim should be softened. For SIGEcom Exchanges, it deserves peer review and likely acceptance with minor revisions. I would cite it as an entry point.","headline":"A genuinely useful survey of incentive-aware ML whose three-perspective taxonomy is the main contribution; the 'unified framework' language oversells Section 2, but the paper deserves a serious referee.","tokens_in":18507,"tokens_out":3798,"would_cite":true,"duration_ms":35056,"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":"This survey organizes incentive-aware machine learning into three perspectives—robustness, fairness, and improvement/causality—unified by a common Stackelberg model of a principal committing to a rule and agents best-responding.","keywords":["incentive-aware machine learning","strategic classification","Stackelberg games","performative prediction","algorithmic fairness","causal inference","online learning","agent heterogeneity"],"falsifier":"Code the papers cited in the survey's own three perspective sections, recording for each whether its objective is robustness only, fairness only, improvement only, or a combination; if a substantial share (say, more than one third) of the papers are naturally classified as mixed, the three-way partition is weakened as a description of the field.","tokens_in":17463,"feed_emoji":"🎯","tokens_out":6746,"duration_ms":60577,"temperature":0.7,"pith_summary":"The paper argues that incentive-aware machine learning—where people alter their data strategically to get better outcomes from an algorithm—is best understood as three perspectives on one underlying interaction. The robustness perspective asks how a decision-maker can keep a rule accurate when agents try to game it. The fairness perspective asks which groups bear the costs of that strategic behavior. The improvement and causality perspective asks how to distinguish genuine self-improvement from gaming, and how to incentivize the former. A sympathetic reader comes away with a common notation and a map of the field that makes results from different papers directly comparable.","feed_headline":"Three perspectives, one Stackelberg model for strategic ML","feed_subtitle":"A survey maps robustness, fairness, and improvement/causality research onto one shared game-theoretic model.","key_machinery":"The load-bearing object is the Stackelberg interaction formalized in Equations (1)–(3): nature draws a feature vector $x$, the principal commits to and announces a rule $f$, agents report $\\hat{x}(f)$ maximizing a utility of the form $u(x,\\hat{x};f)=\\mathrm{val}(\\hat{x};f)-\\mathrm{cost}(x,\\hat{x})$, and the principal's loss $\\ell$ is evaluated at the reported point. In the offline case the objective is $\\min_f \\mathbb{E}_{x\\sim D}[\\ell(h^\\star(x), f(\\hat{x}(f)))]$, and in the online case it is Stackelberg regret against the best fixed rule. The survey uses this skeleton to place every variant: causal features enter by letting changes to some coordinates change the true label, heterogeneity enters through different distributions or cost functions, and partial information enters through randomized or opaque rules. This shared notation is what lets the robustness, fairness, and improvement/causality results be stated and compared in one vocabulary.","core_discovery":"The central claim is that the field has a shared mathematical skeleton: a Stackelberg game in which a principal publicly commits to a decision rule, agents observe the rule and respond with a best-response report, and the principal's loss or regret is evaluated on the reported points. The paper writes this skeleton explicitly for offline settings, where the principal minimizes expected loss at the Stackelberg equilibrium rule, and for online settings, where the principal minimizes Stackelberg regret against the best fixed rule. It then shows that the three perspectives differ in what the principal optimizes and what the agents' actions mean: gaming in the robustness view, a cost or benefit to subpopulations in the fairness view, and a change to genuinely causal features in the improvement/causality view. If the survey is right, every model variant in the literature—continuous adaptation or manipulation graph, full or partial information, rational or biased best response, homogeneous or heterogeneous agents—is a configuration of this single interaction.","pith_inferences":["If the taxonomy holds, papers that do not fit the three perspectives, such as alternating-move games or strategic label manipulation, are not counterexamples but a neighboring problem class defined by dropping one of the framework's assumptions.","A reader could test the taxonomy's fit by coding the papers cited in the survey for whether their stated objective is purely robustness, purely fairness, purely improvement, or a mix; a high mixed share would suggest the field is better described by a two-dimensional grid than by three disjoint boxes.","The survey's own distinction between causal and proxy features implies that interventions meant to reduce gaming, such as making rules less transparent, may also suppress genuine improvement; this trade-off is a natural next theoretical target."],"forward_implications":["A newcomer to incentive-aware ML can use the paper's taxonomy to classify any paper as robustness, fairness, or improvement/causality and translate its model into the shared Stackelberg notation.","Online strategic learning results become comparable through the Stackelberg regret benchmark, so different algorithms can be ranked by the same quantity.","The causal-feature split turns the design problem into a choice: if strategic changes affect only proxy features, the goal is suppression; if they affect causal features, the goal is incentivization.","The open questions the survey names—gaming versus improvement, agent heterogeneity, partial information about the rule—are all expressible as concrete modeling choices inside the same framework."],"supporting_citations":[{"why":"Supplies the original Stackelberg formulation of strategic classification that the survey's offline model and robustness perspective are built on.","marker":"Hardt et al. [2016]"},{"why":"Introduces the online strategic classification setting and the convexity approach that underlies the survey's online Stackelberg regret framework.","marker":"Dong et al. [2018]"},{"why":"Provides the nearly tight online linear classification algorithm and regret benchmark used to illustrate the robustness perspective.","marker":"Chen et al. [2020]"},{"why":"Formalizes the claim that strategic classification is causal modeling, the foundation of the improvement/causality perspective.","marker":"Miller et al. [2020]"},{"why":"Establishes the causal strategic linear regression model with separate objectives, grounding the causal section.","marker":"Shavit et al. [2020]"},{"why":"Defines the fairness perspective by showing disparate effects of strategic manipulation on different subpopulations.","marker":"Hu et al. [2019]"},{"why":"Defines social burden and the accuracy-versus-agent-utility trade-off that anchors the fairness discussion.","marker":"Milli et al. [2019]"},{"why":"Supplies the performative prediction framework that the survey contrasts with utility-based strategic models.","marker":"Perdomo et al. [2020]"}],"fun_headline_variants":["One Stackelberg game to explain gaming, fairness, and causality","Strategic ML's three faces share one game-theoretic model","Incentive-aware ML: Robustness, fairness, and causality unite","The common core of strategic ML: a Stackelberg roadmap"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy assumes that every paper can be assigned to one primary perspective (robustness, fairness, or improvement/causality), so if a large part of the literature is genuinely mixed, the survey's organizing partition would misrepresent the field.","fun_headline_variants_meta":{"raw":{"variants":["One Stackelberg game to explain gaming, fairness, and causality","Strategic ML's three faces share one game-theoretic model","Incentive-aware ML: Robustness, fairness, and causality unite","The common core of strategic ML: a Stackelberg roadmap"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00029,"raw_usage":{"total_tokens":1656,"prompt_tokens":862,"completion_tokens":794,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":478,"completion_tokens_details":{"reasoning_tokens":721}},"tokens_in":478,"tokens_out":794,"duration_ms":7917,"temperature":1.0,"reasoning_tokens":721,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:08:55.407878+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Code the papers cited in the survey's own three perspective sections, recording for each whether its objective is robustness only, fairness only, improvement only, or a combination; if a substantial share (say, more than one third) of the papers are naturally classified as mixed, the three-way partition is weakened as a description of the field.","supporting_citations":[{"cited_title":"Strategic classification from revealed preferences","cited_arxiv_id":null,"evidence_quote":"Introduces the online strategic classification setting and the convexity approach that underlies the survey's online Stackelberg regret framework."}],"review_version":1}