A significant portion of recent research on Large Language Model (LLM) alignment focuses on developing new policy optimization methods based on Group Relative Policy Optimization (GRPO). Two prominent directions have emerged: (i) a shift toward sequence-level importance sampling weights that better align with the sequence-level rewards used in many tasks, and (ii) alternatives to the PPO-style clipping that aim to avoid the associated loss of training signal and entropy collapse. We introduce Soft Sequence Policy Optimization, an off-policy reinforcement learning objective that incorporates soft gating functions over token-level probability ratios within sequence-level importance weights. We provide theoretical motivation for SSPO and investigate practical modifications to improve optimization behavior. Empirically, we demonstrate that SSPO improves training stability and performance both in mathematical reasoning and coding tasks.

This paper presents the submissions of the iai_MSU team for SemEval-2025 Task 3 – Mu-SHROOM, where we achieved first place in the English language. The task involves detecting hallucinations in model-generated text, which requires systems to verify claims against reliable sources.In this paper, we present our approach to hallucination detection, which employs a three-stage system. The first stage uses a retrieval-based (Lewis et al., 2021) to verify claims against external knowledge sources. The second stage applies the Self-Refine Prompting (Madaan et al., 2023) to improve detection accuracy by analyzing potential errors of the first stage. The third stage combines predictions from the first and second stages into an ensemble.Our system achieves state-of-the-art performance on the competition dataset, demonstrating the effectiveness of combining retrieval-augmented verification with Self-Refine Prompting. The code for the solutions is available on https://github.com/pansershrek/IAI_MSU.