Graph-based social recommendation leverages both the interaction graph and the social graph to model user preferences, especially under sparse feedback. However, users' intricate social behaviors may introduce mismatched social ties that contaminate user representations and harm the models' robustness. The majority of existing methods mitigate this by pruning, rewiring, or assigning edge-wise weights before social aggregation. From users' historical behaviors, we observe that a social neighbor often overlaps with the target user on specific interests but differs in others. Thus, using a single weight for each social connection is insufficient, as it only scales the overall message intensity and fails to selectively suppress the misaligned components within the aggregated message. To fill this gap, we propose Orthogonal Decomposition for Social Recommendation (ODSR), an embedding-space framework that orthogonally decomposes the aggregated social message into an aligned component and an orthogonal deviation, and learns a dimension-wise vector gate to regulate the deviation under ranking supervision. Additionally, we introduce a contrastive regularizer that perturbs representations along deviation directions to enhance robustness against imperfect social signals. Extensive experiments on three datasets show that ODSR consistently outperforms strong baselines, and additional analyses verify the effectiveness of selectively gating the orthogonal deviation.
Rongfeng Guo, Yinxuan Huang, Wei Chen et al.· Proceedings of the 32nd ACM...· 0 citations
Reliable multi-turn tool use requires an agent to preserve an evolving task state and ensure that each action remains consistent with it. However, direct function-calling and ReAct-style policies learn state tracking and action generation within the same autoregressive trajectory. This coupling creates state-action competition: the pressure to produce the next call can overwrite or ignore information accumulated earlier in the interaction. Inspired by Boyd's Observe-Orient-Decide-Act cycle, we introduce OODA-Tool, a typed closed-loop policy designed to mitigate this competition by separating state preservation from action realization. Rather than generating an action directly from the interaction history, OODA-Tool routes each decision through controller-checked intermediate states, ensuring that the final output remains grounded in the current task state. Specifically, Observe reconstructs the task state, Orient determines whether execution is warranted, Decide forms an admissible action structure, and Act realizes the external output. We evaluate OODA-Tool against direct function-calling and ReAct policies using Qwen3 models ranging from 0.6B to 14B across multi-turn, multi-tool, and incomplete-information settings. OODA-Tool consistently improves task success across model sizes, with larger gains on smaller models and on tasks whose actions depend strongly on information accumulated across turns and prior tool results. Controlled variants, stage-level ablations, and transfer evaluations further demonstrate the robustness of these improvements.
Rongfeng Guo, Yinxuan Huang, Yusen Wu et al.· 0 citations