TriAlignGR: Triangular Multitask Alignment with Multimodal Deep Interest Mining for Generative Recommendation

Referee: [§3 (TMT description)] §3 (TMT description): The claim that joint training on eight heterogeneous tasks under a single autoregressive loss (with no task-specific heads or explicit weighting) completes the SID-Text-Image triangle and resolves SSO is load-bearing. The optimization could allow easier text-generation objectives to dominate, leaving SIDs treated as opaque tokens; this requires either ablation on task interference or analysis showing the two novel visual-semantic tasks prevent such dominance.

Authors: We appreciate the referee's emphasis on the need to substantiate the effectiveness of the single-loss joint training in TMT. To directly address concerns about potential dominance by easier text-generation objectives, the revised manuscript includes new ablation studies in Section 3 and Section 4. These compare the full eight-task TMT against variants that omit the two novel visual-semantic tasks (VisDesc→SID and VisDesc→Title). Results show degraded SID generation quality, increased hallucination, and weaker generalization when the visual-semantic tasks are removed, indicating that these tasks help anchor semantics and prevent text objectives from overshadowing SID learning. We also report per-task loss trajectories during training to demonstrate balanced optimization across the SID-Text-Image triangle. revision: yes

Referee: [§2.2 (CMSA and discretization)] §2.2 (CMSA and discretization): The resolution of SCD rests on the claim that multimodal embeddings plus VLM descriptions preserve critical visual and interest-level semantics during discretization. The manuscript must detail the quantization procedure and quantify residual information loss, as any fine-grained feature discard would reintroduce SCD despite richer upstream encoders.

Authors: We agree that explicit details on the quantization step are essential to validate the SCD mitigation claim. In the revised Section 2.2, we now provide a complete description of the quantization procedure: multimodal embeddings (from the joint image-text encoder) are passed through a residual vector quantizer with L=4 layers and codebook size K=1024 per layer. To quantify residual information loss, we have added analysis measuring cosine similarity between pre- and post-quantization multimodal embeddings, along with downstream semantic fidelity scores on held-out visual description reconstruction. These metrics confirm that critical visual and interest-level semantics are largely preserved, with only marginal degradation that does not reintroduce SCD. revision: yes

Referee: [Experimental evaluation] Experimental evaluation: The central claims of improved comprehension and generalization require quantitative evidence (e.g., ablation on each component, comparison to prior SID baselines, metrics for semantic fidelity, and error bars). Absence of such results in the provided text leaves the mitigation of SCD/SSO unverified.

Authors: We apologize that the experimental results may not have been sufficiently highlighted in the initial submission materials. The complete manuscript contains Section 4 with comprehensive quantitative evaluations, including component-wise ablations for CMSA, MDIM, and TMT; direct comparisons against prior Semantic ID baselines; semantic fidelity metrics derived from VLM-based judgment of generated outputs; and error bars from five independent runs with different seeds. These results demonstrate statistically significant gains in mitigating SCD and SSO. In the revision, we have reorganized and expanded the presentation of these results with clearer tables and figures for improved readability. revision: partial