From Sparse Decisions to Dense Reasoning: A Multi-attribute Trajectory Paradigm for Multimodal Moderation

Safety moderation is pivotal for identifying harmful content. Despite the success of textual safety moderation, its multimodal counterparts remain hindered by a dual sparsity of data and supervision. Conventional reliance on binary labels lead to shortcut learning, which obscures the intrinsic classification boundaries necessary for effective multimodal discrimination. Hence, we propose a novel learning paradigm (UniMod) that transitions from sparse decision-making to dense reasoning traces. By constructing structured trajectories encompassing evidence grounding, modality assessment, risk mapping, policy decision, and response generation, we reformulate monolithic decision tasks into a multi-dimensional boundary learning process. This approach forces the model to ground its decision in explicit safety semantics, preventing the model from converging on superficial shortcuts. To facilitate this paradigm, we develop a multi-head scalar reward model (UniRM). UniRM provides multi-dimensional supervision by assigning attribute-level scores to the response generation stage. Furthermore, we introduce specialized optimization strategies to decouple task-specific parameters and rebalance training dynamics, effectively resolving interference between diverse objectives in multi-task learning. Empirical results show UniMod achieves competitive textual moderation performance and sets a new multimodal benchmark using less than 40% of the training data used by leading baselines. Ablations further validate our multi-attribute trajectory reasoning, offering an effective and efficient framework for multimodal moderation. Supplementary materials are available at \href{https://trustworthylab.github.io/UniMod/}{project website}.

💡 Research Summary

The paper tackles the persistent problem of data and supervision sparsity in multimodal safety moderation. While textual moderation has benefited from large‑scale binary labeling, extending the same paradigm to vision‑language models (VLMs) leads to shortcut learning: models exploit superficial statistical cues instead of truly understanding multimodal content. To overcome this, the authors propose UniMod, a paradigm that reframes moderation as a structured reasoning trajectory composed of five sequential anchors: Evidence grounding, Modality assessment, Risk mapping, Policy decision, and Answer generation. By requiring explicit predictions at each anchor, UniMod forces the model to produce dense, interpretable reasoning rather than a single black‑box label.

Dataset construction is handled by UniTrace, a consensus‑based pipeline that leverages three top‑tier VLMs (Seed1.6‑vision‑250815, GLM‑4.5V, Gemini‑2.5‑Pro) as teacher models. For each trajectory node, a majority‑vote (or semantic‑centroid for open‑ended evidence) determines a high‑quality pseudo‑ground truth. A subsequent “expert teacher” phase assigns the most reliable teacher to each node based on performance on a calibration set, ensuring node‑specific expertise and reducing label noise.

Training is driven by UniRM, a multi‑head scalar reward model. UniRM shares a common VLM backbone but attaches separate heads that output scalar scores for each attribute (evidence relevance, modality correctness, risk severity, policy consistency, response quality). Two key optimization tricks are introduced: (1) Head‑wise weight subspace decoupling, which isolates gradient directions of different heads, and (2) Stochastic head scheduling, which randomly activates heads per batch to prevent dominance of any single objective. Rewards are aggregated additively (R_uni = Σ w_k r_k) rather than multiplicatively, preserving a dense reward spectrum and stabilizing the Group Relative Policy Optimization (GRPO) updates.

Theoretical analysis formalizes moderation as a GRPO problem with a tripartite trajectory τ = {τ_p, τ_t, τ_q}. Lemma 3.1 shows that decomposing the task reduces sample complexity because the model first learns to operate in the perception subspace before tackling the decision subspace. Lemma 3.2 (Perception Protection) proves that providing a positive reward for the perception stage prevents negative gradients from penalizing correctly identified evidence when the final decision fails. Lemma 3.3 (Decision Grounding) demonstrates that the final answer stage acts as a semantic regularizer, enforcing cross‑stage consistency. Lemma 3.4 argues that additive aggregation avoids reward degeneracy, keeping the advantage estimator well‑conditioned.

Empirically, UniMod matches or exceeds state‑of‑the‑art (SOTA) textual moderation models while using less than 40 % of the training data. In multimodal benchmarks, it outperforms leading VLM guards such as LlamaGuard‑Vision, GuardReasoner‑VL, and ProGuard, achieving a 7–12 % absolute gain in F1 and markedly reducing errors on complex image‑text threats (e.g., hidden text in images, manipulated captions). Ablation studies confirm that (a) removing the trajectory supervision collapses performance, (b) omitting head‑wise decoupling leads to unstable training, and (c) switching to multiplicative reward aggregation causes the total reward to zero out whenever any stage fails, halting learning.

In summary, UniMod introduces a “sparse decision → dense reasoning” shift that yields interpretable, data‑efficient, and higher‑performing multimodal safety moderation. The work underscores that structural transparency and multi‑attribute supervision can be more impactful than sheer model scaling, offering a promising direction for future safe AI systems across additional modalities and real‑time deployments.