W1. "Drift" is not defined, and the proposed mechanisms do not demonstrably address it<
This is the central concept of the paper, so its vagueness undermines both the method design and the interpretation of the results.«
• No formal definition. Style drift and decision drift are described only informally. Their relation to standard notions (covariate shift in P(X), concept shift in P(Y X), label shift in P(Y)) is never discussed. "Style" is operationally defined as whatever the hand-written rule G(;) removes, which is circular. No experiment verifies that the two kinds of drift exist or are actually separated; the E-scores in Figure 3 are statistics of the prediction distribution, not measurements of drift.
Label shift is ignored even though the bot ratio differs substantially across datasets (about 35% / 50% / 56%).
• Style alignment never involves the target domain. L style is computed only on labeled sources (positives come from other source domains), so it cannot reduce source-target style drift; the only target-side handling of style is the fixed preprocessing G(;). Moreover, Z style is never fed into any classifier head (the heads use zaG and Zah, Eqs. 3-4), so the claimed benefit of "preserving class-discriminative style information" is not realized in the decision.« • The transferability estimator measures confidence, not correctness. Sep, Var and Mar (Eq.6) depend only on the shape of the prediction distribution. A confidently wrong, or even labelflipped, classifier would receive a high score. This contradicts the claim that the method differs from confidence-based approaches. In particular, Mar = 2|mean(p) - 1/2| is maximized when all target samples are predicted as the same class, so it rewards collapse and is confounded with the target class prior. In addition, the normalization of these statistics to [0, 1] is undefined: each channel yields a single scalar per target. If min-max normalization is performed across the three targets, information from other target domains is used, which would not be available in deployment.-
• Rank-based fusion implicitly assumes a balanced target. The direct-transfer path averages rank-transformed predictions and thresholds the result at 0.5, which forces roughly half of the target samples to be predicted as bots. Fox8, the only balanced target, is also where direct transfer works best. The rank transform is also transductive, which conflicts with the per-sample/per-batch inference in Algorithm 1. Finally, Eq. 9 mixes percentile ranks, shifted-andclipped probabilities and sigmoid probabilities, which are not on a common scale.-
W2. The empirical evidence does not support the main claims«
• The ablation contradicts the routing mechanism. In Table 2, removing routing (using direct transfer only) improves Macro-F1 from 0.819 to 0.850 on Fox8 and from 0.571 to 0.587 on TwiBot20. Removing semantic adaptation also gives 0.850 on Fox8, removing L style givesm 0.844 on Fox8, and removing the semantic channel gives 0.574 on TwiBot20. The router assigns about 55% weight to semantic adaptation on Fox8 and about 98% to behavioral transfer on TwiBot20, yet both choices are worse than plain direct transfer. The advantage of DDDA is therefore essentially limited to BotSim24, where about 96% of the weight goes to semantic adaptation, which follows Song et al. (2025), i.e., the OOD baseline that is already the strongest baseline on BotSim24. It is unclear how much of the gain comes from DDDA's own design. With only three targets and a hand-crafted routing formula (a mid, products of E-scores), the router cannot be meaningfully validated.~
• Missing comparisons with mainstream methods. BotSSCL and BotBuster are cited but not
compared. Established bot detectors (e.g., BotRGCN, RGT, SimpleHGN, BotMoE, LMBot) are absent, and the graph and metadata information in TwiBot20 is unused. Classic UDA baselines (DANN, CDAN, MDD, and M3SDA for the multi-source setting) are also missing. Instead, the comparison includes PoCoMT (object detection), UASA (OOD detection) and SSA (open-set adaptation for vision-language models), which are not designed for this task. Several baselines also appear degenerate: MGDIL obtains 0.338 Macro-F1 on TwiBot20, and on BotSim24 its precision/recall are 0.357/0.499, i.e., near-constant predictions. This raises doubts about how the baselines were reproduced and tuned.«
• Evaluation protocol. The sensitivity analysis (Appendix C.2) reports target-domain metrics and
discusses per-dataset optimal values of ß and a_dir, which suggests that target labels were used for model selection in a UDA setting. The text states that a mid = 0.5 is optimal on all three targets, whereas the implementation uses 0.55. Three seeds are run but no standard deviations are reported. DDDA's target accuracy on BotSim24 (0.788) is higher than its source accuracy (0.746), which is unusual and deserves explanation. The claim of good performance across backbones is contradicted by DDDA (Qwen0.5B) obtaining 0.573 Macro-F1 on BotSim24.~
• The semantic logit l_sem = l_sh + l_spe requires the source index k, which does not exist for target samples. How target inference uses the source-specific heads is not specified.« • The main loss I fuse trains a fused output _fuse, but inference never uses _fuse; the fusion weights n are unconstrained.«
• The estimation of tada and the TTA procedure (updated parameters, loss, steps, learning rate) are not described. If TTA updates the encoder , the behavioral features (which depend on Ф) also change.«
• The behavioral features are inconsistent between the main text and Appendix A (3 + 4 vs. 6 + 10 features). T is not specified; with an average of 18 posts (std 43) on BotSim24, temporal statistics are unstable or undefined for many users. The emotion lexicon and topic model are not specified, and features are not normalized across domains despite very different post counts.- • With batch size 8 and K = 2 sources, many anchors in the multi-positive contrastive loss will have empty positive sets."
• "Routing" produces global target-level weights shared by all samples; the name overstates the mechanism.