{"id":"890e44d0-e503-41d8-8c82-4305d8132869","arxiv_id":"2502.09869","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A qualitative study of 34 users introduces intentional implicit feedback, actions users knowingly take to steer recommendation algorithms, and shows feedback choices vary with user purposes.","lead":"Researchers interviewed 34 social media users about how they try to shape their recommendation feeds, and propose a new category called intentional implicit feedback: actions like swiping past or deliberately searching that users knowingly perform to influence the algorithm. The paper maps these behaviors to user goals such as increasing content diversity and relevance, which could inform feedback design on platforms like Douyin and Xiaohongshu.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The tripartite taxonomy rests entirely on retrospective self-reports of in-the-moment intention; interview prompts about 'strategies' may manufacture the agency they then measure, so intentional vs unintentional implicit feedback may be a measurement artifact.","rationale":"I agree with the reader's weakest_assumption: the load-bearing point is measurement of intention, not the numerical typos. The numeric swap in Section 5.2 (75.9% vs 88.6%) is real and should be fixed, but it does not threaten the conceptual claim. The purpose-feedback alignment is partly definitional, because feed customization is defined as deliberate action and intentional feedback is defined by intention, so I would not elevate that into the primary attack; it is a framing caveat rather than a falsifying flaw. The authors deserve credit for stating the ground-truth limitation explicitly in Section 5.4. No independent verification exists, but qualitative interview studies routinely rely on self-report; the issue is that the specific construct 'intention at the moment of action' is exactly the kind of thing retrospective interviews can distort. The proposed stimulated-recall check is feasible and would settle whether the concern lands. Because the reader already conditioned the verdict on this weakness, my recommendation is unchanged.","tokens_in":24068,"tokens_out":5162,"duration_ms":51414,"concrete_test":"Run a small validation study with the same population: record 20 users' real browsing sessions on Xiaohongshu/Douyin via screen recording plus touch logs, then conduct stimulated recall within 24 hours by replaying each recorded action and asking, for each action, 'At the moment you did this, did you intend to influence the recommendation feed?' Independently apply the paper's Section 3.3 coding to the resulting momentary-intention labels and to a standard retrospective interview from the same users. Compute Cohen's kappa between the two classifications for the intentional/unintentional implicit boundary. If kappa is below 0.6, the retrospective interview cannot reliably recover in-the-moment intention and the taxonomy's empirical basis is not established; if kappa is above 0.8, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that users consciously employ behaviors previously classified as implicit to shape recommendations, and that this conscious use is missed by the explicit-implicit dichotomy. The only evidence for consciousness is what participants said during semi-structured interviews. In Section 3.2, the protocol explicitly asked about 'strategies for managing content exposure and content preferences and avoiding undesirable content'; this primes participants to narrate their ordinary behavior as strategic. Retrospective verbal reports are reconstructive, and after the fact a user may rationalize a quick swipe as intentional algorithm-shaping even if no such intention was present at the moment. The coding in Section 3.3 therefore sorts a mental state (intention) using data that can manufacture that mental state. The authors honestly flag in Section 5.4 that reported behaviors are 'neither comprehensive nor reflective of the ground truth,' but this is not a peripheral caveat: if retrospective intention does not match in-the-moment intention, the new category collapses into a folk reconstruction, and the main contribution loses its empirical grounding. A secondary internal inconsistency (Section 5.2 swaps the 75.9% and 88.6% percentages from Table 2) should also be corrected, but it is not the core issue.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper reports semi-structured interviews (N=34) with users of Chinese personalized recommendation platforms (e.g., Xiaohongshu, Douyin) to investigate how users provide feedback to shape their recommendation feeds. The authors propose a tripartite taxonomy of feedback: explicit, intentional implicit, and unintentional implicit, arguing that the traditional explicit/implicit dichotomy overlooks cases where users deliberately perform behaviors such as swiping past, ignoring, clicking, dwelling, or searching in order to influence future recommendations. They map these feedback types onto user purposes (content consumption, directed information seeking, content creation/promotion, and feed customization) and present co-occurrence frequencies and extensive illustrative quotes. The paper's contributions are the new concept of intentional implicit feedback, an empirical behavior-purpose mapping, and design implications for supporting purpose-oriented feedback on recommendation platforms.","tokens_in":24293,"tokens_out":8099,"duration_ms":73669,"significance":"If the taxonomy holds, it adds a user-intention dimension to the well-established explicit/implicit feedback distinction and connects it to prior work on folk theories and strategic interaction with algorithms (e.g., [10,43,50]). The study is empirically grounded: 34 interviews, systematic codebook thematic analysis (Section 3.3), detailed behavior tables (Tables 1 and 4), and many direct quotes that let the reader evaluate the interpretation. The authors also explicitly acknowledge in Section 5.4 that reported behaviors are not ground truth, which is a useful caveat. However, the core conceptual contribution turns on the reliability of retrospective self-reports of intention, which is not independently verified.","major_comments":[{"comment":"Section 3.2 (interview procedure) and Section 3.3 (coding of intention): the paper's central distinction between intentional and unintentional implicit feedback rests entirely on participants' retrospective self-reports of their in-the-moment intentions. The protocol explicitly asks about 'strategies for managing content exposure and content preferences and avoiding undesirable content,' which primes participants to narrate ordinary interactions as strategic. Section 5.4 notes that reported behaviors are 'neither comprehensive nor reflective of the ground truth,' but this caveat concerns the completeness of the behavior list, not the possibility that retrospective rationalization creates the intentional/unintentional split. Because this is a load-bearing issue, the paper should either provide additional verification (e.g., diary or log-based confirmation, analysis of unprompted versus prompted mentions of intention, or a follow-up study) or reframe the central claim as being about users' perceptions and folk theories of their own feedback behavior rather than about the actual presence of intention at the moment of action. As written, the evidence establishes that users describe these behaviors as intentional, not that they were intentionally performed.","section":"Sections 3.2-3.3, central claim"},{"comment":"Section 3.3 (data analysis): the coding of intention is a researcher judgment, and no inter-rater reliability statistic (e.g., Cohen's kappa) or equivalent agreement audit is reported. The same observable behavior appears in both intentional and unintentional lists (e.g., 'Initiate a new search' in Table 1 vs. 'Search for information' in Table 4; 'Like a post' in both tables), so the boundary between categories is not self-evident. Without a reliability check, the reader cannot distinguish a stable emergent category from coder interpretation. I recommend adding a quantitative agreement measure or a detailed disagreement audit for the intentional/unintentional coding.","section":"Section 3.3 (data analysis)"}],"minor_comments":[{"comment":"Section 5.2: the percentages from Table 2 are swapped. The text attributes 75.9% to explicit feedback and 88.6% to intentional implicit feedback, but Table 2 shows 88.6% (31/35) for explicit feedback and 75.9% (44/58) for intentional implicit feedback in the feed customization row.","section":"Section 5.2"},{"comment":"Table 1 lists 20 participants for 'Ignore or swipe past a post,' while Section 5.1 states n=21 for this behavior; please reconcile the discrepancy.","section":"Table 1 / Section 5.1"},{"comment":"Section 3.3 uses 'potential correlation patterns' to describe what is actually a code co-occurrence analysis; please replace 'correlation' with 'co-occurrence' to avoid implying a statistical association.","section":"Section 3.3"},{"comment":"Table 5: P26's platform entry reads 'Douyin, Xiaohongshu,' with a trailing comma; this is a minor formatting error.","section":"Table 5"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a solid qualitative HCI study with rich data and transparent reporting. The main risk is the intentionality construct, but this is fixable through careful reframing and (ideally) additional evidence. If the authors address the self-report issue and correct the numeric inconsistencies, I would support publication. Given the current gap between the data and the strength of the claims, major revision seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague, this one is worth reading. The paper's core move is to split implicit feedback into intentional and unintentional implicit feedback, and the interview data do show that users talk about deliberately swiping, ignoring, searching, or dwelling to steer their feeds. That's a real addition to the explicit/implicit dichotomy, and the paper connects it to user purposes (feed customization vs. content consumption) in a way that should be useful to both HCI and recommender-system researchers. The behavior tables and direct quotes give the taxonomy empirical body.\n\nThe method is qualitative, and the central evidence is retrospective self-report of intention. That is the right kind of evidence for a study of perceived agency, but it is not direct evidence of in-the-moment intention. The protocol's explicit questions about 'strategies' could push participants to narrate ordinary behavior as strategic, so the intentional/unintentional line may be somewhat amplified. The authors flag in Section 5.4 that reported behaviors are not ground truth; I'd prefer they also frame the taxonomy as a model of users' perceived feedback, not of their actual mental states.\n\nThere is one genuine error: in Section 5.2, the percentages for explicit and intentional implicit feedback are swapped. Table 2 shows explicit is 88.6% feed customization (31/35) and intentional implicit is 75.9% (44/58). The text says the reverse. That should be fixed. Minor issues: no inter-rater reliability statistic, and the sample skews young, female, and educated—both acknowledged limitations.\n\nThe citation pattern looks reasonable; the paper builds on folk theories and strategic interaction literature without overclaiming novelty. The 'intentional implicit feedback' concept does not appear in the cited prior work.\n\nVerdict: solid paper with a useful conceptual contribution and an honest limitations section. It deserves a serious referee. I'd suggest revision with the percentage fixes and a more careful framing of the self-report basis, but the central taxonomy holds up.","headline":"Introduces a genuinely useful third feedback category, but the self-report basis and a swapped-percentage error mean it needs a careful revision, not a desk reject.","tokens_in":24790,"tokens_out":3398,"would_cite":true,"duration_ms":33412,"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 claims that the standard explicit-implicit feedback split misses a third, intentional kind of feedback that users deliberately perform to shape recommendation feeds.","keywords":["personalized recommendation algorithms","explicit feedback","implicit feedback","intentional implicit feedback","user purpose","semi-structured interview","Xiaohongshu","Douyin"],"falsifier":"Give a panel of users a recommendation feed instrumented with logging and ask them immediately after each swipe, dwell, or search whether they intended to influence the feed; if stated intention does not systematically correspond to the logged action, or if the platform's recommendation changes do not track intentional signals better than unintentional ones, the three-way split would lose its predictive value.","tokens_in":23885,"feed_emoji":"🔄","tokens_out":5771,"duration_ms":52067,"temperature":0.7,"pith_summary":"The paper argues that the standard split of user feedback into explicit and implicit categories misses a central part of how people actually steer recommendation feeds. Based on semi-structured interviews with 34 active users of platforms like Xiaohongshu and Douyin, the authors claim that users often consciously perform behaviors traditionally counted as implicit feedback—swiping past posts, clicking into content, searching topics, even pausing on a video—with the explicit intention of teaching the algorithm. They propose a three-way taxonomy: explicit feedback, intentional implicit feedback, and unintentional implicit feedback. They also report that feedback choices track user purposes: explicit feedback is used mainly for feed customization, unintentional implicit feedback for content consumption, and intentional implicit feedback for increasing diversity and relevance. If right, the finding gives designers a way to recognize and respond to intentional signals that current systems tend to misread or ignore.","feed_headline":"Implicit feedback is not always implicit, user interviews show","feed_subtitle":"Users intentionally swipe, search, and dwell to shape recommendation feeds beyond explicit and implicit input.","key_machinery":"The central object is the three-way feedback taxonomy built from interview coding: explicit feedback (direct preference input), intentional implicit feedback (conscious behavior performed to influence recommendations), and unintentional implicit feedback (natural interaction without feedback intention). The machinery that carries the argument is the researchers' coding procedure: inductive coding of transcripts, deductive mapping onto existing explicit and implicit categories, then splitting implicit feedback by whether participants reported intention, and finally code co-occurrence analysis linking feedback types to four user purposes (content consumption, directed information seeking, content creation and promotion, and feed customization). Each behavior is also characterized by platform feature, polarity (positive or negative), and minimum scope (segment, object, or class), borrowed from prior frameworks, which is how the study shows intentional implicit feedback can carry negative polarity that standard implicit-feedback models assume away.","core_discovery":"The authors claim that the explicit-implicit feedback dichotomy, defined from the system's perspective, cannot fully capture user agency. Users knowingly employ behaviors the platform treats as implicit feedback—ignoring or fast-swiping posts, searching new topics, dwelling on content, even talking aloud for the platform to 'overhear'—to shape future recommendations. The study introduces intentional implicit feedback: behavior performed consciously with the expectation that the platform will interpret it as a preference signal, but without the direct input of explicit feedback. Across 34 interviews the authors found 6 explicit, 9 intentional implicit, and 13 unintentional implicit feedback behaviors, and a systematic alignment between feedback type and purpose: intentional implicit feedback dominates when users want more diverse or more relevant content, while explicit feedback dominates for removing inappropriate content. The conclusion is that the intention dimension belongs inside the feedback taxonomy, and that recognizing intentional implicit feedback would give users a greater sense of control and improve signal accuracy.","pith_inferences":["A natural next step the authors do not take: test whether classifying behavior by inferred intention improves recommendation quality metrics such as dwell time, satisfaction, or perceived control compared with behavior-only models; the taxonomy predicts it should.","The intention dimension connects to strategic classification: if users deliberately shape signals, platforms that ignore intention risk feedback loops where users escalate their strategies, as in the reported 'information cocoon' and platform switching.","Because the evidence is retrospective self-report, the taxonomy could be validated with in-the-moment experience sampling or platform telemetry; without that, it remains a user-perception model rather than a behavioral law.","The participant who talked aloud to be 'overheard' suggests that perceived surveillance itself becomes a feedback channel; a platform that legitimizes voice input could convert an opaque privacy violation into a transparent control."],"forward_implications":["Platforms should treat swiping past, ignoring, searching, and dwelling as potentially intentional signals of preference, not just passive behavioral data.","Implicit feedback is not inherently positive; intentional negative signals such as fast-skipping a previously liked category or refreshing the feed should be captured and validated.","Feedback mechanisms should be surfaced according to purpose: explicit controls for removing unwanted content, subtle intentional signals for diversity and relevance tuning.","Designers can add lightweight confirmations (for example, 'see less of this?') and progress indicators so users know intentional implicit feedback was registered.","Recognizing intentional implicit feedback could reduce the cognitive cost of explicit feedback while preserving user agency."],"supporting_citations":[{"why":"Supplies the baseline definitions of explicit and implicit feedback and the catalogue of implicit behaviors that the study extends.","marker":"[42]"},{"why":"Provides the system-side framing of implicit feedback and the behavior categories used to code user actions.","marker":"[34]"},{"why":"Contributes the comparison of explicit versus implicit properties, including polarity and cognitive effort, which the study modifies.","marker":"[35]"},{"why":"Supplies the minimum-scope concept (segment, object, class) used to characterize each feedback behavior.","marker":"[63]"},{"why":"Documents the positive-polarity assumption for implicit feedback that the findings challenge with intentional negative signals.","marker":"[31]"},{"why":"Frames folk theories, the mechanism by which users decide that certain behaviors will teach the algorithm.","marker":"[15]"},{"why":"Documents strategization, users adapting behavior to shape future recommendations, which the study connects to feedback mechanisms.","marker":"[10]"},{"why":"Lists user strategies for influencing algorithms, grounding the behaviors later classified as intentional implicit feedback.","marker":"[43]"},{"why":"Motivates adding an intention dimension to the explicit-implicit distinction.","marker":"[73]"}],"fun_headline_variants":["Users turn implicit feedback into deliberate signals","Intentional implicit feedback: users steer algorithms","Implicit feedback isn't always accidental, user study finds","Users deliberately exploit implicit feedback to shape feeds","New study: intentional behavior is the hidden implicit feedback"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The taxonomy rests on participants' own retrospective accounts that they performed those behaviors on purpose to influence the algorithm, with no platform log, screen recording, or moment-of-action check confirming the behavior or the intention, a point the authors acknowledge in their limitations.","fun_headline_variants_meta":{"raw":{"variants":["Users turn implicit feedback into deliberate signals","Intentional implicit feedback: users steer algorithms","Implicit feedback isn't always accidental, user study finds","Users deliberately exploit implicit feedback to shape feeds","New study: intentional behavior is the hidden implicit feedback"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000879,"raw_usage":{"total_tokens":3773,"prompt_tokens":888,"completion_tokens":2885,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":504,"completion_tokens_details":{"reasoning_tokens":2815}},"tokens_in":504,"tokens_out":2885,"duration_ms":19293,"temperature":1.0,"reasoning_tokens":2815,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T20:11:47.768096+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Give a panel of users a recommendation feed instrumented with logging and ask them immediately after each swipe, dwell, or search whether they intended to influence the feed; if stated intention does not systematically correspond to the logged action, or if the platform's recommendation changes do not track intentional signals better than unintentional ones, the three-way split would lose its predictive value.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the baseline definitions of explicit and implicit feedback and the catalogue of implicit behaviors that the study extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the system-side framing of implicit feedback and the behavior categories used to code user actions."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the minimum-scope concept (segment, object, class) used to characterize each feedback behavior."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Documents the positive-polarity assumption for implicit feedback that the findings challenge with intentional negative signals."},{"cited_title":"Algorithms ruin everything","cited_arxiv_id":null,"evidence_quote":"Frames folk theories, the mechanism by which users decide that certain behaviors will teach the algorithm."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Lists user strategies for influencing algorithms, grounding the behaviors later classified as intentional implicit feedback."},{"cited_title":"Implicit Interaction","cited_arxiv_id":null,"evidence_quote":"Motivates adding an intention dimension to the explicit-implicit distinction."}],"review_version":1}