{"id":"0bd34e0c-d2e7-4928-a91b-e6b270792c18","arxiv_id":"2501.05871","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A federated-learning system for Mastodon lets instances share model weights, improving average macro-F1 over local-only moderation on three content moderation tasks.","lead":"This paper proposes FedMod, a system that lets Mastodon servers train content moderation models together by sharing model parameters instead of raw posts. It reports that this collaborative approach improves detection of harmful content, bot posts, and content-warning-worthy posts over each server training alone.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Post-level 80/20 split in §4 leaks the same authors/bot accounts into train and test, inflating macro-F1 and potentially distorting FedMod's reported gains over local models.","rationale":"The reader's label-proxy concern is legitimate and acknowledged in Appendix A.3, but it is not the most load-bearing issue because the paper already conditions its claims on those proxies. The post-level split is a separate, unacknowledged internal-validity problem that can be settled by re-running the same experiments with author-disjoint splits. If the leakage is real, both the absolute macro-F1 values and the FedMod-versus-local deltas are suspect; the central claim that parameter exchange among similar peers yields better proxy-label prediction than local-only training would need to be re-established. The paper otherwise has a clear design, honest discussion of limitations, and useful robustness experiments (perturbed labels, opposing policies, pre-sampling), so the verdict should remain conditional pending this check rather than being rejected outright.","tokens_in":16731,"tokens_out":6547,"duration_ms":69343,"concrete_test":"Re-run the bot content detection and content warning assignment evaluations with an author/account-level split: partition the 44,840 users (or the relevant bot/author sets) into 80/20 groups and assign all posts of a user to one side, stratifying by label distribution where feasible. If FedMod still outperforms local models on all instances with comparable macro-F1 deltas, the leakage concern is resolved; if the gains shrink, disappear, or reverse, the headline claim is an artifact of the split.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4 splits each instance's posts 80/20 'stratified by their respective labels' without grouping by author or account. The dataset makes this consequential: bot detection uses 42,566 posts from only 3,371 bot accounts, and content-warning assignment uses 11,074 posts from 3,905 authors (Section 3.1.3). A random post-level split therefore puts multiple posts from the same account in both training and test. The classifier can memorize account-specific style rather than learn a transferable moderation rule, inflating absolute macro-F1. The bias matters for the central FedMod-versus-local comparison: local models trained on 128 posts can exploit local account overlap directly, while FedMod's averaged parameters may be affected differently by overlapping versus disjoint authors. The paper's strongest claim is that parameter exchange improves proxy-label prediction over local-only training; that claim is not secure until the same result holds on author-disjoint splits.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"FedMod proposes a federated-learning-based content moderation system for Mastodon instances. Using a dataset of 219,577 posts from 50 instances, the paper evaluates three tasks: harmful content detection (with Perspective API labels), bot content detection (self-declared bot flags), and content warning assignment (self-tagged sensitive posts). Each instance trains a local mBERT classifier on 128 labeled posts; FedMod lets each instance average parameters with k similar peers selected by hashtag Jaccard similarity. The paper reports that FedMod improves macro-F1 over local-only training for all three tasks (12.69%, 5.79%, and 9.43%) and shows robustness to label perturbation, peer-count variations, pre-sampling, and divergent moderation policies.","tokens_in":16910,"tokens_out":7263,"duration_ms":69545,"significance":"The problem is timely and important: decentralized moderation is a real bottleneck, and a privacy-preserving parameter-exchange scheme is a sensible proposal. The paper's empirical methodology is thorough in several respects: it uses a multi-instance dataset, reports bootstrapped confidence intervals and Wilcoxon test results, and systematically ablates design choices (number of peers, pre-sampling, label noise, and policy disagreement). These strengths make the headline claims credible if the evaluation-validity concerns are resolved. However, the current evaluation does not fully establish that the reported gains correspond to real moderation effectiveness, because of the author-leakage and proxy-label issues detailed below. The work would be a solid contribution to the measurement and systems community after addressing those points.","major_comments":[{"comment":"The 80:20 split of posts is performed at the post level without grouping by author or bot account. Because the bot detection task contains 42,566 posts from only 3,371 bot accounts and the content-warning task contains 11,074 posts from 3,905 authors (Section 3.1.3), the same account almost certainly appears in both training and test splits of an instance. The fine-tuned mBERT can then memorize account-specific style rather than learn a general moderation rule, which inflates the absolute macro-F1 scores. More importantly for the central claim, the leakage may affect local models and FedMod differently: a local model trained on 128 posts can exploit within-instance account overlap directly, whereas FedMod's parameter averaging with peers may dilute that memorization, making the reported FedMod gains either conservative or spurious depending on the task. I request an author-disjoint (or account-disjoint) split, or a per-account held-out evaluation, for at least the bot and content-warning tasks to confirm that the FedMod-vs-local ranking is not an artifact of this leakage.","section":"§4 \"Training & Testing Sets\" and §3.1.3"},{"comment":"The harmful-content task uses Perspective API scores as ground truth for both training and testing. Consequently, the reported macro-F1 measures how well FedMod agrees with the Perspective API, not how well it moderates according to actual instance administrators. The paper acknowledges this proxy issue (e.g., Section 3.1.3 and Appendix A.3), but the abstract and introduction state improvements in 'automated content moderation' without this qualification. The perturbed-label experiment simulates label noise but does not establish that Perspective labels are a valid surrogate for administrator decisions. I recommend reframing the claims as improvements on surrogate labels, or adding a small human-annotated evaluation set (even a few hundred posts) to show the ranking between FedMod and local models transfers to a more direct measure of moderation quality.","section":"§3.1.3 and §6.1"}],"minor_comments":[{"comment":"The caption of Figure 1 mentions only the distribution of posts per instance, but the text refers to it as showing distributions of harmful posts, bot posts, and content-warning posts. Please align the caption with the actual panels or the references in the text.","section":"Figure 1 and §3.2"},{"comment":"Please clarify whether the online learning sampling of 16 posts at each step is with or without replacement from the training set; this affects whether the reported N=128 corresponds to 128 unique posts.","section":"§4"},{"comment":"There is a typo in 'Perspcetive' (should be 'Perspective') in the Appendix A.5 setup description.","section":"Appendix A.5"},{"comment":"The phrase 'we also re-plot the results of the local content moderation models' is imprecise; these results were not plotted earlier in the same form. Consider saying 'we also plot'.","section":"§6.1"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable fit for the journal's scope and the core idea is worth publishing if the evaluation is tightened. The post-level split causing author leakage is the most serious issue; it is fixable by re-doing the experiments with author-disjoint splits. The Perspective-API surrogate issue is secondary but should be corrected in framing or validated with a small human-labeled set. I would not reject the paper on these grounds."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: FedMod is a genuine incremental contribution — parameter-exchange federated learning for Mastodon moderation with a similarity-based peer-selection mechanism — and the evaluation is honest enough that the conditional verdict is fair. But the author-leakage issue flagged in the stress-test note is real, and it needs fixing before the headline “improves all tasks” claim is secure.\n\nWhat's actually new: most prior work in this line, including the authors' own [7], shares fully trained models. FedMod instead exchanges parameters of partially trained models, and hashsim peer selection based on trending hashtag Jaccard similarity is a simple, plausible idea. The pre-sampling trick for scaling peer selection is a nice touch. The paper also stress-tests itself in useful ways: label perturbation, threshold changes, and opposing moderation policies. The bootstrapped confidence intervals and Wilcoxon tests are appropriate. Credit where due: the paper openly acknowledges that Perspective API labels are a proxy and that human labels would be preferable, and Appendix A.3 gives concrete reasons not to deploy Perspective directly. That's better than many system papers.\n\nSoft spots, in order of severity:\n\n1. The post-level 80/20 split is the main problem. Bot detection uses 42,566 posts from 3,371 accounts; content-warning assignment uses 11,074 posts from 3,905 authors. Splitting posts without grouping by author means the same account appears in both train and test. mBERT can pick up account-specific style, so the absolute macro-F1 numbers are inflated. More importantly for the paper's central comparison, local models can exploit local account overlap directly, while FedMod's averaged parameters may be affected differently. The fix is an author-disjoint split, at least for the bot and content-warning tasks. This should be a required revision, not optional.\n\n2. The harmful-content task trains and tests on Perspective scores. That makes it a surrogate benchmark, not a direct test of what an administrator would moderate. The paper acknowledges this, but the abstract and conclusion still frame it as robust moderation. Reframe the claims or add a small human-labeled validation set.\n\n3. The closest prior work, model sharing from [7], is never used as a baseline. It's fine that FedMod is mechanistically different, but without that comparison the reader can't tell how much of the gain comes from federated learning versus just collaboration itself.\n\nMinor: no code or data release, and the 714MB mBERT footprint per instance per task is mentioned without much discussion of real deployment constraints.\n\nWho this is for: people building moderation tooling for federated systems, and researchers working on decentralized or low-resource FL applications. It deserves a serious referee, but the split issue must be addressed. I'd send it out rather than desk reject, asking for author-disjoint splits and a clearer separation between proxy-label performance and real moderation performance.","headline":"Real evaluation leak in an otherwise solid system paper: author overlap between train and test inflates the headline numbers, and the proxy-label tasks make the moderation claims shakier than the prose admits.","tokens_in":17424,"tokens_out":2603,"would_cite":false,"duration_ms":25799,"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":"Mastodon servers moderate better by sharing model weights","keywords":["Fediverse","Mastodon","content moderation","federated learning","harmful content detection","bot detection","content warnings","peer selection"],"falsifier":"Re-run the FedMod and local-only comparison on the same 50 instances using ground-truth labels from real administrator moderation actions (or a held-out set of human moderation judgments) instead of the three automated and crowd-sourced proxies; if collaborative training does not beat local-only training on that human-labeled test set, the central claim fails.","tokens_in":16531,"feed_emoji":"🛡️","tokens_out":6998,"duration_ms":64929,"temperature":0.7,"pith_summary":"The paper claims that decentralized social networks do not have to choose between weak local moderation and handing moderation to a centralized platform. It introduces FedMod, a system in which Mastodon instances exchange the parameters of their partially trained moderation models with a small number of topically similar peers, so that moderation knowledge spreads through the network rather than the posts themselves. On 219,577 posts from 50 instances, FedMod is reported to improve over local-only training for all three tasks tested, raising average macro-F1 by 12.69% for harmful-content detection, by 5.79% for bot detection, and by 9.43% for content-warning assignment. If the result holds, volunteer administrators with only about 128 labeled posts per task can obtain automated moderation that is closer to what centralized platforms achieve, while keeping user data on their own servers.","feed_headline":"Mastodon servers moderate better by sharing model weights","feed_subtitle":"FedMod lifts harmful-content detection 12.7% over local-only training, without sharing posts.","key_machinery":"The core mechanism is parameter exchange among similar instances: instead of sharing posts or labels, servers share the weights of their partially trained classifiers and average those weights together. Peer similarity is computed as the Jaccard similarity of each instance's trending hashtags (hashsim), which the paper shows predicts whether sharing model updates will help; a pre-sampling stage ranks candidate peers by how many of their users the local instance's users follow, reducing the number of similarity computations by 90%. The parameter averaging is what does the work: it pools linguistic patterns and moderates the effect of individual administrators' labeling mistakes.","core_discovery":"FedMod is a collaborative content-moderation protocol built on federated learning. Each instance first fine-tunes a multilingual transformer classifier on its own small labeled set; it then picks k peers whose weekly trending hashtags are most Jaccard-similar to its own, and the peer group exchanges model parameters and locally averages them with FederatedAveraging to form a shared model. The paper reports that this procedure beats training on local data alone for harmful-content detection, bot-content detection, and content-warning assignment across all 50 instances in the study, reaching average per-instance macro-F1 scores of 0.71, 0.73, and 0.58, respectively. The gains persist when a quarter of the training labels are randomly flipped and when instances use different harmfulness thresholds, and they are largest when instances have very few labeled posts. The optional pre-sampling step cuts peer-search computations by 90% with only a small drop in performance. The paper also shows a limit: if a large fraction of collaborating instances hold completely opposing moderation policies, the shared model degrades below local-only performance.","pith_inferences":["A natural next test is to replace the three proxy label sources (an automated toxicity scorer, self-declared bot flags, and author-set content warnings) with moderation decisions made by real instance administrators; FedMod's advantage should survive that substitution if the mechanism, rather than the label source, is what drives the gains.","Peer selection could be made from signals of actual moderation-policy alignment, such as shared domain blocks or suspension lists, rather than hashtag similarity; this may mitigate the opposing-policy failure mode the paper identifies.","The same parameter-exchange recipe could be applied to other label-scarce decentralized tasks such as spam detection, and the paper's pre-sampling logic suggests it should scale to the full Fediverse."],"forward_implications":["An instance with roughly 128 labeled posts can reach macro-F1 of 0.71 on harmful content, 0.73 on bots, and 0.58 on content warnings by collaborating with two peers, and more peers help when labels are scarcer.","Hashtag-based peer selection (hashsim) beats random selection on every task, so topical affinity is a usable signal for choosing moderation partners.","Because only model parameters change hands, instances gain moderation knowledge without disclosing their labeled posts.","Label noise is partly absorbed by collaboration: with 25% of harmful-content labels flipped, FedMod still beats local models by 14.54%.","Collaborating with instances that enforce opposite moderation policies can hurt; FedMod remains ahead of local models when 25% of peers differ, but not when 50% do."],"supporting_citations":[{"why":"Supplies the proxy labels for harmful-content detection, the basis of the first task's ground truth.","marker":"[1]"},{"why":"Provides the federated-learning implementation used for parameter exchange and aggregation in FedMod.","marker":"[6]"},{"why":"The closest prior system it improves on, which shares fully trained models rather than parameter updates.","marker":"[7]"},{"why":"The multilingual transformer model that each instance fine-tunes for the three classification tasks.","marker":"[13]"},{"why":"Defines federated learning, the conceptual foundation for exchanging parameters instead of data.","marker":"[19]"},{"why":"Defines FederatedAveraging, the aggregation rule that combines peer model parameters in FedMod.","marker":"[21]"}],"fun_headline_variants":["FedMod: Federated learning boosts Mastodon moderation scores","Sharing model weights lifts Fediverse moderation across tasks","Collaborative moderation via weight sharing beats local-only training","Mastodon servers improve moderation by pooling trained models","Federated learning improves content moderation on the Fediverse"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that the proxy labels used for training and testing—automated toxicity scores, self-declared bot flags, and author-set content warnings—are good stand-ins for what real instance administrators would actually moderate; if those proxies misrepresent moderation decisions, the measured improvements may not appear in live Fediverse moderation.","fun_headline_variants_meta":{"raw":{"variants":["FedMod: Federated learning boosts Mastodon moderation scores","Sharing model weights lifts Fediverse moderation across tasks","Collaborative moderation via weight sharing beats local-only training","Mastodon servers improve moderation by pooling trained models","Federated learning improves content moderation on the Fediverse"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000794,"raw_usage":{"total_tokens":3496,"prompt_tokens":944,"completion_tokens":2552,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":560,"completion_tokens_details":{"reasoning_tokens":2474}},"tokens_in":560,"tokens_out":2552,"duration_ms":16786,"temperature":1.0,"reasoning_tokens":2474,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T21:06:02.734882+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the FedMod and local-only comparison on the same 50 instances using ground-truth labels from real administrator moderation actions (or a held-out set of human moderation judgments) instead of the three automated and crowd-sourced proxies; if collaborative training does not beat local-only training on that human-labeled test set, the central claim fails.","supporting_citations":[{"cited_title":"Perspective API","cited_arxiv_id":null,"evidence_quote":"Supplies the proxy labels for harmful-content detection, the basis of the first task's ground truth."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"The closest prior system it improves on, which shares fully trained models rather than parameter updates."}],"review_version":1}