{"id":"bf3d5eb9-feda-49ee-a645-bc812b5d8b0f","arxiv_id":"2505.16023","paper_version":3,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"Seven prototypical collaboration behaviors, such as asking for more outputs, asking questions, and adding content, explain most variation in how users follow up with writing assistants in the wild.","lead":"This paper analyzes millions of real writing conversations with two AI assistants to identify seven recurring ways users steer the AI after their first request. The findings link these interaction patterns to specific writing goals and suggest how future AI models could be aligned to real user behavior.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Shared GPT-4o labeling bias may inflate intent-PATH correlations; cross-dataset replication does not rule it out because the same model labels both predictors and outcomes.","rationale":"The reader correctly identified classifier accuracy as the weakest assumption, but the more specific and load-bearing mechanism is shared annotation bias: the same model labels both independent and dependent variables, so errors are not random but systematic, and replication across datasets does little to rule them out. This deepens, rather than replaces, the reader's concern. The paper has real strengths: it uses two large in-the-wild datasets, provides the full prompts, makes the WildChat portion reproducible, and supports its claims with qualitative examples. However, the central quantitative claims about intent-PATH correlations need the proposed human-annotated re-run to separate genuine user behavior from GPT-4o labeling artifacts. Because the reader's verdict is already CONDITIONAL and this test is a reasonable condition for acceptance, the verdict should remain CONDITIONAL (UNCHANGED). If the proposed re-run is not feasible, a weaker alternative would be to report effect sizes and apply a multiple-comparison correction, but that would not fully address the shared-bias mechanism.","tokens_in":29019,"tokens_out":4034,"duration_ms":38522,"concrete_test":"Re-annotate a random sample of 500 sessions per dataset (not label-stratified) with two independent human annotators blind to the paper's hypotheses, using the paper's exact label definitions. Re-run the full pipeline on the human labels: tf-idf of follow-up types, PCA with the same 80–85% variance threshold, and logistic regressions for the six headline intent-PATH pairs. If the reported correlations (e.g., GENERATE CATCHY TEXT→PATH2, GENERATE PROFESSIONAL DOCUMENT→PATH3/PATH4, GENERATE MESSAGE/GENERATE STORY→ADDS CONTENT) remain statistically significant and with comparable coefficient magnitudes in both datasets, the concern is resolved. Additionally, report confusion matrices for GPT-4o versus human labels on this naturalistic sample to expose any systematic bias.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim—specific writing intents correlate with specific PATHs—depends on labels produced by GPT-4o classifiers for both the predictors (writing intents via fWritingI) and the outcomes (follow-up types via fFollowU). The validation in Appendix C.3 is small (110 sessions per classifier), stratified to ensure uniform label presence, and conducted by the paper's own authors. More importantly, it cannot detect systematic label biases that the GPT-4o model applies consistently to both original requests and follow-ups. If the model associates certain lexical cues (e.g., 'give me more', 'another') with REQUESTS MORE OUTPUTS, and similar cues in original requests with GENERATE CATCHY TEXT, the same spurious correlation will appear in BCPWr and WCWr. The paper treats 'significant in both datasets' as strong evidence, but that criterion does not control for shared annotator bias. In addition, the regressions test 7 PATHs × ~19 predictors without multiple-comparison correction; at p≤0.05, one expects 6–7 false positives per dataset, and a shared annotation bias could reproduce the same false positives in both. The qualitative examples illustrate behavior but cannot refute an aggregate artifact. This threatens the load-bearing assertion that the discovered intent-PATH relationships reflect genuine user collaboration rather than GPT-4o's labeling tendencies.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper analyzes two large in-the-wild log corpora of LLM-assisted writing sessions, one from Bing Copilot and one from WildChat. Using GPT-4o classifiers, it labels each user utterance as an original request or one of eleven follow-up types, and labels original requests with fine-grained writing intents. It then applies PCA to the tf-idf representation of follow-up types to define seven Prototypical Human-AI Collaboration Behaviors (PATHs), and runs logistic regressions to correlate writing intents with PATHs. The central claims are that a small set of PATHs accounts for 80-85% of the variance in follow-up behavior in both datasets, and that specific writing intents are significantly associated with specific PATHs, with qualitative examples illustrating behaviors such as brainstorming, staged generation, question asking, and content addition.","tokens_in":29231,"tokens_out":5825,"duration_ms":56951,"significance":"If the empirical findings hold, this is a valuable large-scale characterization of human-LLM collaboration in writing. The two-dataset design, the use of a public dataset for reproducibility, and the mixed-methods grounding in qualitative examples are genuine strengths. The paper also makes concrete, falsifiable descriptive claims about follow-up behavior that go beyond coarse task classification and satisfaction prediction. The main contribution is empirical rather than theoretical, and its value depends heavily on the reliability of the automated labels; the current evidence for that reliability is too thin to support the strength of the downstream claims.","major_comments":[{"comment":"The central regression results depend on labels produced by the same GPT-4o model family for both the predictors (writing intents via fWritingI) and the outcomes (follow-up types via fFollowU). The validation in Appendix C.3 uses only 110 sessions per classifier, with accuracies of 74-89% and kappa of 0.74-0.82. This sample is too small to detect systematic label bias, and a consistent lexical association in GPT-4o labeling--for example, the word \"more\" appearing both in requests labeled GENERATE CATCHY TEXT and in follow-ups labeled REQUESTS MORE OUTPUTS--could manufacture the very correlations reported in §6. Cross-dataset replication does not rule this out because the same labeling model is applied to both datasets. The authors should address this load-bearing concern directly, for example by re-running the full pipeline on a larger human-annotated subset, using an independently trained or different labeler, or providing explicit label-bias diagnostics that show the correlations survive under alternative labels.","section":"§4, §C.3 (Table 8)"},{"comment":"The paper does not correct for multiple comparisons in the regression analysis. With roughly 18 intent predictors and seven PATH targets, running independent logistic regressions at the p≤0.05 level yields an expected number of false positives in the double digits per dataset, even before considering cross-dataset dependence. The statement that an association is meaningful because it is significant in both BCPWr and WCWr does not address this, since a shared labeler can induce correlated errors in both datasets. The authors should report adjusted p-values, false-discovery-rate corrections, or permutation-based null distributions. The thresholding of PCA scores into binary PATH membership, described as \"selected to ensure the accuracy of our logistic regressions,\" is also a modeling choice that can inflate apparent associations; sensitivity of the results to this threshold should be reported.","section":"§C.2 (logistic regressions) and §4"},{"comment":"The abstract and Section 5 state that seven PATHs \"capture 80-85% of the variance,\" but this is true by construction because the number of principal components l was chosen precisely so that the retained components explain 80-85% of the variance. The substantive claim is that a relatively small l suffices, which is reasonable, but the current wording implies that the variance proportion was discovered rather than selected. The authors should present a scree plot or an equivalent explanation of how l was chosen and should report the variance explained as a function of l. The binarization threshold for PATH membership is a second free parameter that should receive the same sensitivity treatment.","section":"§4, §5 (variance claim)"},{"comment":"The logistic regressions treat sessions as independent observations, but the datasets contain many sessions per user; BCPWr has 202k users and 250k sessions, and WCWr has 22k users and 68k sessions. Repeated sessions from the same user are likely correlated, so the reported p-values may be anticonservative. The authors should use cluster-robust standard errors by user or otherwise account for within-user dependence. This is directly relevant to the strength of the statistical significance claims that support RQ2.","section":"§C.2, Table 1"}],"minor_comments":[{"comment":"There is a grammatical error in the phrase \"a important and increasingly prevalent use case\"; it should be \"an important.\"","section":"§1"},{"comment":"The possessive in \"the users original requests\" should be \"the user's original requests.\"","section":"§C.2"},{"comment":"The paper should clarify the unit of analysis for the regression. Table 5 reports a mean of 1.6 original requests per session, but §C.2 describes one-hot intent features for a session. It would be helpful to state explicitly how sessions with multiple original requests are represented in the one-hot encoding and whether the target PATH is defined at the session level.","section":"§4, Table 5"},{"comment":"The heatmap in Figure 3b is difficult to read in the current rendering, especially for distinguishing the loading magnitudes for PATHs 5-7; the accompanying text is clear, but the figure labels should be enlarged or the figure split into two panels.","section":"Figure 3b"},{"comment":"The mapping from prompt labels (GENERATE TITLE, GENERATE CASUAL BIO) to the paper's displayed labels (GENERATE CATCHY TEXT, GENERATE BIO) is shown only in the prompt appendix; the main text should use one consistent naming convention throughout.","section":"Prompt 3, Table 3"},{"comment":"The Limitations section acknowledges English-only data, order-insensitivity, and light-weight qualitative analysis, but it does not mention the shared-labeler concern or the absence of multiple-comparison correction. The sentence \"We ensure the reliability of our analysis through manual validation of all automated components\" overstates the certainty provided by the small validation samples in Appendix C.3.","section":"§8"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid empirical contribution with an interesting, actionable research question, but the validity of the main correlational claims is not yet established given the shared-labeler and multiple-testing issues. I do not think rejection is warranted: the concerns are fixable with additional validation on human-annotated data, alternative labelers, and corrected inference. The paper fits the scope of cs.CL well, and the public WildChat component is a useful reproducibility asset. I would support acceptance after the authors address the load-bearing validity concerns."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The headline is simple: this is the first large-scale, in-the-wild look at how users actually follow up on LLM writing requests, and the core descriptive findings are believable. The PATH construct—seven principal components that account for 80-85% of the variance in follow-up types—is a useful way to organize behavior, and the specific correlations (e.g., 'generate catchy text' → 'request more outputs', 'professional documents' → 'ask questions/add content') make intuitive sense and are backed by example conversations. It does what it claims: goes beyond task classification and satisfaction to characterize multi-turn co-construction.\n\nThe strengths are real. The two datasets are large, the qualitative spot-checks are genuinely informative, and the paper is honest about its limitations—the appendix contains the validation numbers, the label definitions, and the prompts. That level of transparency is exactly what an empirical HCI/NLP paper should look like.\n\nNow the soft spots. The biggest one is also the one the stress-test flags: a single model family (GPT-4o) labels both the writing intents (predictors) and the follow-up types (outcomes). If GPT-4o systematically associates certain phrasings with certain labels—e.g., 'give me more' with 'REQUESTS MORE OUTPUTS' and 'catchy/name/slogan' with 'GENERATE CATCHY TEXT'—then the same correlation appears in both datasets because the labeler is shared, not because user behavior is shared. The validation in Appendix C.3 (roughly 110 sessions per classifier) is small and can't detect that kind of systematic bias. I don't think this is fatal—the qualitative examples make the correlations look real and the paper doesn't overclaim causality—but it is a load-bearing weakness. The paper should report effect sizes, correct for multiple comparisons, test threshold robustness for the PCA binarization, and ideally re-label a sample with a different model or human annotators to check for annotator-driven artifacts.\n\nA smaller issue: the proprietary Bing Copilot data limits full reproducibility, although the WildChat pipeline is public and that part is a step in the right direction. The paper also does not release code for the WildChat analysis, which would help.\n\nOverall verdict: this is a solid, important-for-the-subfield paper that deserves a serious referee. It is not a foundational result yet—the label-bias concern needs to be addressed before I'd treat the intent-PATH correlations as the basis for alignment research—but it is a good contribution that will generate discussion and follow-up work. If I were the editor, I would send it to review with the expectation of a major revision on the statistical rigor.","headline":"A solid first map of how users steer LLM writing sessions in the wild, with a transparent pipeline that deserves refereeing despite a real shared-labeler weakness.","tokens_in":29784,"tokens_out":2433,"would_cite":true,"duration_ms":21751,"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":"A compact set of seven collaboration behaviors explains most of how users steer LLM-assisted writing sessions.","keywords":["human-AI collaboration","LLM-assisted writing","in-the-wild log analysis","follow-up behaviors","prototypical collaboration behaviors","principal component analysis","writing intents","LLM alignment"],"falsifier":"Manually annotate a fresh random sample of roughly 2,000 sessions per dataset with follow-up types and writing intents, retrain or correct the classifiers, rerun the PCA and the intent-PATH regressions, and check whether seven components still explain 80-85% of variance and whether the key correlations, catchy text with requesting more outputs and professional documents with asking questions, reappear; if the structure shifts or dissolves, the discovered PATHs are artifacts of the labeling model.","tokens_in":28782,"feed_emoji":"✍️","tokens_out":7803,"duration_ms":60394,"temperature":0.7,"pith_summary":"The paper tries to establish that the way people steer an LLM after their first request is not idiosyncratic: follow-up messages in writing sessions cluster into a small set of prototypical behaviors, which the authors call PATHs, and seven of these explain 80-85% of the variation across millions of real sessions. The claim matters because most prior log analyses stop at what task users are doing or whether they are satisfied, whereas this work characterizes the interaction itself. Using two assistants with different models, interfaces, and user bases, Bing Copilot and WildChat, the paper finds the same recurring behaviors: revising the request, asking questions, requesting more outputs, changing style, and adding content. It then shows that a user's writing goal predicts which behavior they use, for example brainstorming catchy text leads to requesting more outputs while drafting professional documents leads to asking questions. If correct, this gives a data-driven map of collaboration that points toward aligning LLMs from natural, session-level feedback rather than single-turn explicit preferences.","feed_headline":"Seven behaviors explain most human–AI writing collaborations","feed_subtitle":"Real writing logs show intent drives follow-ups: brainstorming asks for more, drafting asks questions.","key_machinery":"The central object is the PATH, Prototypical Human-AI Collaboration Behavior, a cluster of sessions defined by co-occurring follow-up types. The method works in three steps: a GPT-4o classifier labels each user utterance as an original request or one of eleven follow-up types, such as restates request, elaborates request, requests answers, requests more outputs, changes style, adds content, and removes content; each session becomes a tf-idf vector over these follow-up types, with the tf-idf weighting preventing the most common behaviors from dominating; and PCA on the session-by-type matrix yields principal components, each interpreted as a PATH, retaining the components that jointly explain 80-85% of variance. The same classifier pipeline labels original requests with one of eighteen writing intents, and logistic regressions with intents as predictors and PATH membership as the target turn the components into testable correlations. The PCA is doing the load-bearing work of converting individual utterance labels into a small set of repeatable collaboration styles.","core_discovery":"The central discovery is that follow-up utterances in LLM-assisted writing form a compact behavioral structure rather than noise. After classifying every follow-up into one of eleven high-level types with a GPT-4o labeler and applying PCA to the per-session type frequencies, seven principal components, the PATHs, capture 80-85% of the variance in both datasets, with the same dominant behaviors, restating the request and requesting more outputs, leading in each. Explicit satisfaction signals are rare, appearing in only 1-5% of sessions, so the PATHs are built from the implicit steering users actually do. The paper further claims statistically significant, cross-deployment correlations between writing intents and specific PATHs: generating eye-catching text correlates with requesting more outputs for brainstorming, drafting professional documents, summaries, and technical text correlates with asking follow-up questions to learn domain norms, and adding content appears when generations lack information known only to the user, such as personal experience in resumes or messages. These correlation patterns appear in both Bing Copilot and WildChat logs, which the paper takes as evidence that the behaviors reflect how users collaborate with LLMs rather than how one particular system is built.","pith_inferences":["A testable extension left implicit in the paper: because PATHs come from co-occurrence counts, they ignore turn order, and a sequence-aware model of follow-ups might split the staged long-generation behavior into finer prototypes or reveal that order itself is a collaboration signal.","The intent-PATH correlations suggest a real-time adaptation loop the paper never states: an assistant that recognizes a brainstorming session could proactively supply more alternatives, and one that recognizes professional drafting could volunteer domain norms before being asked.","The same pipeline could be run on non-writing logs, such as coding, image generation, or theorem proving, to see whether the same skeleton behaviors, restate, ask, request-more, add, remove, restyle, reappear, which would make PATHs a general taxonomy of human-AI steering.","Because the classifier accuracies range from 74% to 89%, the cleanest confirmation would be a small lab study where users self-report their intent; if self-reported intents correlate with PATHs more strongly than classifier-derived intents do, the reported effect sizes are lower bounds in the wild."],"forward_implications":["Explicit positive or negative feedback is rare, so aligning LLMs to implicit, natural feedback across multi-turn sessions becomes a tractable and necessary target.","Because intent predicts behavior, alignment can be intent-specific: a model serving a brainstorming request should expect 'give me more' follow-ups, while a professional-drafting session should expect knowledge and norm questions.","Staged generation of long creative texts with follow-ups that grow more specific suggests models should handle under-specified feedback that sharpens over turns.","Content-addition behavior shows professional and communicative drafts commonly lack user-only knowledge, pointing to personalization or proactive question-asking as the fix.","The same PATH structure appearing in two different deployments implies the behaviors are shared styles of human-AI collaboration, not artifacts of one interface."],"supporting_citations":[{"why":"Supplies the public WildChat-1M conversation logs that form one of the two corpora analyzed.","marker":"Zhao et al. (2024a)"},{"why":"Defines the Bing Copilot deployment whose seven-month logs form the other corpus.","marker":"Mehdi (2023)"},{"why":"Provides the PCA formulation used to convert follow-up type frequencies into PATHs.","marker":"Bengio et al. (2013)"},{"why":"Source of the coarse task taxonomy that the writing-session filter fCoarseT builds on.","marker":"Suri et al. (2024)"},{"why":"The conversational follow-up labels that the paper's eleven follow-up types mirror.","marker":"Qu et al. (2018)"},{"why":"The large-scale log task-classification work this analysis extends beyond task labels to collaboration behavior.","marker":"Tamkin et al. (2024)"},{"why":"Prior small-scale qualitative evidence of exploration and question-asking in user-LLM math interactions, which this work scales to writing.","marker":"Collins et al. (2024a)"},{"why":"Satisfaction-estimation approach from logs that this paper contrasts with its behavior-level characterization.","marker":"Lin et al. (2024)"}],"fun_headline_variants":["Seven paths reveal how humans steer AI writing","Users steer, not rate: seven behaviors shape AI writing","Just seven behaviors explain 85% of AI writing collaboration","AI writing: seven behaviors, not ratings, tell the story","Brainstorming asks more, drafting asks questions: AI writing patterns"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything downstream depends on GPT-4o classifiers correctly labeling every session's task, every follow-up type, and every writing intent, but those classifiers were validated on only 110 or 220 manually annotated sessions with accuracies between 74% and 89%, so systematic labeler errors could produce the PATHs and intent correlations rather than real user behavior.","fun_headline_variants_meta":{"raw":{"variants":["Seven paths reveal how humans steer AI writing","Users steer, not rate: seven behaviors shape AI writing","Just seven behaviors explain 85% of AI writing collaboration","AI writing: seven behaviors, not ratings, tell the story","Brainstorming asks more, drafting asks questions: AI writing patterns"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000906,"raw_usage":{"total_tokens":3913,"prompt_tokens":981,"completion_tokens":2932,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":597,"completion_tokens_details":{"reasoning_tokens":2851}},"tokens_in":597,"tokens_out":2932,"duration_ms":19992,"temperature":1.0,"reasoning_tokens":2851,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T15:07:18.592915+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Manually annotate a fresh random sample of roughly 2,000 sessions per dataset with follow-up types and writing intents, retrain or correct the classifiers, rerun the PCA and the intent-PATH regressions, and check whether seven components still explain 80-85% of variance and whether the key correlations, catchy text with requesting more outputs and professional documents with asking questions, reappear; if the structure shifts or dissolves, the discovered PATHs are artifacts of the labeling model.","supporting_citations":[],"review_version":1}