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

TL;DR

UniMod reframes multimodal safety moderation as a dense, multi-attribute reasoning problem, addressing sparse supervision and shortcut learning by decomposing decisions into Evidence, Modality, Risk, Policy, and Answer. It introduces UniTrace for consensus-based trajectory labeling and UniRM, a multi-head scalar reward model with head-wise subspace decoupling and stochastic scheduling to enable stable, multi-dimensional supervision under Single-Sample Single-Label settings, underpinned by a GRPO-based theoretical framework. Empirical results show state-of-the-art multimodal moderation with substantially less training data and enhanced interpretability through explicit grounding across attributes. The work demonstrates both practical gains and theoretical insights into how structured priors and dense feedback mechanisms can outperform mere scaling in complex safety tasks, with clear data–model scaling interactions highlighted.

Abstract

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}.

Figures (9)

• Figure 1: Overview of the UniMod framework. The left panel illustrates the comparison between UniMod and traditional baselines, where UniMod introduces a structured reasoning trajectory comprising Evidence, Modality, Risk, Policy, and Answer. The center panel, UniTrace, demonstrates the consensus mechanism used to select specialized teacher models (e.g., GLM) for labeling each trajectory node. The right panel details the Training stage, where the UniMod is optimized via UniRM. UniRM utilizes a shared VLM backbone with task-specific heads, incorporating head-wise weight subspace decoupling and stochastic head scheduling. A detailed case study illustrating the UniMod reasoning process is provided in App. \ref{['app:case_unimod']}.
• Figure 2: Performance comparison of UniMod against various ablation variants. The first three panels illustrate the training dynamics for Formality, Modality and Risk attributes. The final panel shows the downstream F1 scores for both text and image moderation.
• Figure 3: Ablation study of UniRM and scalability of UniMod. (a-b) Average performance and variance of the UniRM under various ablation settings. (c) Data Scaling: F1 score improvement of UniMod when training data is scaled from $L_1$ to $L_2$. (d) Model Scaling: Comparison of F1 score gains ($\Delta$) across different moderation models when increasing model capacity from 3B to 7B parameters.
• Figure 4: Prompt for generating edge-case samples.
• Figure 5: Prompt used for modality labeling.

Theorems & Definitions (8)

• Lemma 3.1: Search Efficiency
• proof
• Lemma 3.2: Perception Protection
• proof
• Lemma 3.3: Decision Grounding
• proof
• Lemma 3.4
• proof