Generative process reward models (PRMs) improve step-level verification of mathematical reasoning, yet their effectiveness is limited by the quality of process supervision. The dominant Monte Carlo (MC) estimation derives step labels from final-answer accuracy rather than semantic correctness, producing systematic label noise, while standard pipelines discard the implicit tree structure of rollout data that encodes informative failure patterns. We propose CePRM, which addresses both limitations through Step-Cognitive MC Value Scaling, calibrating raw MC scores with step-level semantic signals via multiplicative adjustment, and pivot-step-directed repair mining, which exploits value cliffs in failed trajectories to construct same-prefix contrastive correction pairs via Sibling Advantage filtering. Experiments on ProcessBench and PRMBench show that CePRM, with a 4B backbone, achieves competitive or superior error localization and process verification against 7B-scale baselines, with notably strong error-side recognition.

Multi-hop question answering (QA) requires integrating evidence across multiple documents and reasoning over interdependent intermediate steps, posing significant challenges for retrieval-augmented generation (RAG) systems. While recent RAG-based approaches interleave retrieval and reasoning to improve factual grounding, they often struggle with complex multi-hop questions due to retrieval–reason mismatch, ignored dependencies among sub-questions, and error propagation during decomposition. In this work, we propose SDRAG, a sub-question dependency-aware RAG framework that explicitly models the intrinsic logic of multi-hop questions as a dependency graph. SDRAG first decomposes the question into fine-grained sub-questions and explicitly models their dependency relations, then refines the graph by pruning unreliable or redundant nodes, and finally traverses the pruned graph to coordinate retrieval and reasoning. To improve graph construction quality, we further introduce a supervised fine-tuning strategy that independently trains task-specific LoRA adapters for decomposition, dependency prediction, and pruning. Experiments on three common multi-hop QA benchmarks show that SDRAG consistently outperforms strong RAG-based baselines, while additional analyses demonstrate the effectiveness of dependency modeling, graph pruning, and fine-tuning.

While Large Language Models (LLMs) have achieved remarkable capabilities, they unintentionally memorize sensitive data, posing critical privacy and security risks.Machine unlearning is pivotal for mitigating these risks, yet existing paradigms face a fundamental dilemma: aggressive unlearning often induces catastrophic forgetting that degrades model utility, whereas conservative strategies risk superficial forgetting, leaving models vulnerable to adversarial recovery. To address this trade-off, we propose AGTAO (Adversarial Gating Training with Adaptive Orthogonality), a unified framework designed to reconcile robust erasure with utility preservation. Specifically, our approach introduces Adaptive Orthogonality (AO) to dynamically mitigate geometric gradient conflicts between forgetting and retention objectives, thereby minimizing unintended knowledge degradation. Concurrently, Adversarial Gating Training (AGT) formulates unlearning as a latent-space min-max game, employing a curriculum-based gating mechanism to simulate and counter internal recovery attempts. Extensive experiments demonstrate that AGTAO achieves a superior trade-off between unlearning efficacy (KUR ≈ 0.01) and model utility (MMLU 58.30).[Code is available at https://anonymous.4open.science/r/AGT-unlearning>.].