Peilin Jiang

Author directory

2026

Process Reward Models (PRMs) are vulnerable to reward hacking: policies exploit surface cues such as confident phrasing or structured formatting in place of logical correctness, which yields an inverted-U collapse under optimization pressure. We propose ARMOR-PRM, which augments standard PRM training with three regularizers: a counterfactual margin loss, an adversarial invariance loss with low-score anchoring, and a gradient penalty on the reward head. On ProcessBench (three splits; "math" excluded due to training data overlap), ARMOR-PRM improves Detection Rate by 14.9 percentage points and AUROC by 4.1 percentage points over Vanilla PRM, and is the only variant whose Best-of-N reranking accuracy improves monotonically through N=128. A domain-shift diagnostic further shows that objective-level robustness and data coverage are distinct but complementary axes: ARMOR-PRM responds best to iterative hard-negative retraining, suggesting a practical two-stage deployment path.
Multimodal fake news detection requires models to reason over both textual and visual evidence while remaining robust to distribution shifts in social media graphs. Existing approaches often concatenate modality-specific representations or rely on post-hoc fusion, which can leave image and text features in misaligned semantic spaces. Graph-based methods further introduce structural bias: node popularity, degree, or connectivity may become shortcut signals that correlate with labels in the training distribution but do not causally determine veracity. We present CIL, a causal-intervention and LLM-synergized framework for multimodal fake news detection. CIL first uses CLIP to obtain aligned image and text representations and constructs a multi-view semantic graph from cross-modal, intra-modal, and fused similarities. It then uses a large language model to produce high-confidence pseudo labels, which are injected as weak supervisory anchors. Finally, CIL mitigates degree-induced structural confounding with inverse propensity weighting and enforces counterfactual consistency under edge-drop interventions. Experiments on Twitter, PHEME, and Weibo suggest that each component contributes to performance, with the largest gains coming from aligned multimodal representations and causal debiasing.