Data selection is a critical aspect of Reinforcement Learning with Verifiable Rewards (RLVR) for enhancing the reasoning capabilities of large language models (LLMs). Current data selection methods are largely heuristic-based, lacking theoretical guarantees and generalizability. This work proposes a theoretically-grounded approach using influence functions to estimate the contribution of each data point to the learning objective. To overcome the prohibitive computational cost of policy rollouts required for online influence estimation, we introduce an off-policy influence estimation method that efficiently approximates data influence using pre-collected offline trajectories. Furthermore, to manage the high-dimensional gradients of LLMs, we employ sparse random projection to reduce dimensionality and improve storage and computation efficiency. Leveraging these techniques, we develop Curriculum RL with Off-Policy Influence guidance (CROPI), a multi-stage RL framework that iteratively selects the most influential data for the current policy. Experiments on models up to 7B parameters demonstrate that CROPI significantly accelerates training. On a 1.5B model, it achieves a 2.66x step-level acceleration while using only 10% of the data per stage compared to full-dataset training. Our results highlight the substantial potential of influence-based data selection for efficient RLVR.

Multimodal sarcasm detection has received considerable attention due to its unique role in social networks. Existing methods often rely on feature concatenation to fuse different modalities or model the inconsistencies among modalities. However, sarcasm is often embodied in local and momentary nuances in a subtle way, which causes difficulty for sarcasm detection. To effectively incorporate these nuances, this paper presents Context-Aware Self-Attention Fusion (CAAF) to integrate local and momentary multimodal information into specific words. Furthermore, due to the instantaneous nature of sarcasm, the connotative meanings of words post-multimodal integration generally deviate from their denotative meanings. Therefore, Word Weight Calculation (WWC) is presented to compute the weight of specific words based on CAAF’s fusion nuances, illustrating the inconsistency between connotation and denotation. We evaluate our method on the MUStARD dataset, achieving an accuracy of 76.9 and an F1 score of 76.1, which surpasses the current state-of-the-art IWAN model by 1.7 and 1.6 respectively.

One-hot labels are commonly employed as ground truth in Emotion Recognition in Conversations (ERC). However, this approach may not fully encompass all the emotions conveyed in a single utterance, leading to suboptimal performance. Regrettably, current ERC datasets lack comprehensive emotionally distributed labels. To address this issue, we propose the Emotion Label Refinement (EmoLR) method, which utilizes context- and speaker-sensitive information to infer mixed emotional labels. EmoLR comprises an Emotion Predictor (EP) module and a Label Refinement (LR) module. The EP module recognizes emotions and provides context/speaker states for the LR module. Subsequently, the LR module calculates the similarity between these states and ground-truth labels, generating a refined label distribution (RLD). The RLD captures a more comprehensive range of emotions than the original one-hot labels. These refined labels are then used for model training in place of the one-hot labels. Experimental results on three public conversational datasets demonstrate that our EmoLR achieves state-of-the-art performance.

Recently, fine-tuning the pre-trained language model (PrLM) on labeled sentiment datasets demonstrates impressive performance. However, collecting labeled sentiment dataset is time-consuming, and fine-tuning the whole PrLM brings about much computation cost. To this end, we focus on multi-source unsupervised sentiment adaptation problem with the pre-trained features, which is more practical and challenging. We first design a dynamic feature network to fully exploit the extracted pre-trained features for efficient domain adaptation. Meanwhile, with the difference of the traditional source-target domain alignment methods, we propose a novel asymmetric mutual learning strategy, which can robustly estimate the pseudo-labels of the target domain with the knowledge from all the other source models. Experiments on multiple sentiment benchmarks show that our method outperforms the recent state-of-the-art approaches, and we also conduct extensive ablation studies to verify the effectiveness of each the proposed module.