CORTEXA
← Browse
arxivcs.AI2026-06-30

Which Tokens Matter? Adaptive Token Selection for RLVR with the Relative Surprisal Index

Outongyi Lv, Yanzhao Zheng, Yuanwei Zhang, Zhenghao Huang, Xingjun Wang, Baohua Dong, Hangcheng Zhu, Yingda Chen

Reinforcement learning (RL) has become a powerful tool for propelling Large Language Models (LLMs) beyond imitation-based training towards more robust reasoning capabilities. Among existing approaches, RL with Verifiable Rewards (RLVR) has emerged as a pivotal paradigm for advancing LLM reasoning. Despite its empirical success, recent studies have offered different insights. One line of inquiry advocates prioritizing high-entropy token positions during training, while another perspective cautions against allowing low-probability tokens to dominate gradient updates. Notably, although high-entropy tokens are usually correlated with low probability, both paradigms empirically yield substantial performance gains. In this work, we argue that evaluating sampled-token probability or entropy in isolation is insufficient to capture the policy optimization dynamics. To resolve this tension, we introduce the Relative Surprisal Index (RSI), a principled, information-theoretic metric that naturally couples the token's entropy with the probability of the selected token. We show that, under mild conditions, RSI is related to the local ratio between the first-order variations of the logit-gradient norm and predictive entropy under a selected-logit perturbation. Building on RSI, we propose RSI Selection (RSI-S), an entropy-adaptive token filtering method that retains tokens within a stable RSI interval. RSI-S successfully reconciles previous contradictory paradigms and filters out both redundant low-surprisal tokens and unstable high-surprisal tail tokens. Empirical evaluations show that RSI-S achieves higher avg@32 accuracy across different model scales (Qwen2.5-1.5B, 3B, and 7B) on AIME and AMC benchmarks: RSI-S improves avg@32 accuracy by 2--3 percentage points over GRPO. Overall, RSI offers a promising perspective for RLVR improvement.

View free PDFSource page

Related papers

arxivcs.AI2026-06-29

Inoculation Adapters: Improved Selective Generalization of Capabilities with Fewer Surprising Backdoors

Maxime Riché, Daniel Tan, Vili Kohonen, Niels Warncke

Inoculation prompting is a selective-generalization technique used against Emergent Misalignment. We introduce inoculation adapters (IA), a family of methods that similarly reduce the optimization pressure to learn undesired traits by strengthening those traits during training. I…

View free PDFSource page
arxivcs.IRcs.AIcs.CLcs.CV2026-07-06

Do All Visual Tokens Matter Equally? Object-Evidence Preserving Token Merging for Vision-Language Retrieval

Suhyeong Park, Junha Jung, Jungwoo Park, Jaewoo Kang

Multi-vector vision-language retrieval preserves fine-grained visual evidence through maximum-similarity late interaction, but dense image-side tokens make storage and scoring expensive. Existing token compression methods reduce this cost, yet they can remove or collapse object-…

View free PDFSource page
arxivcs.AI2026-07-17

MGDT: MLLM-Guided Diffusion Transformer with Relation-Adaptive Mixture-of-Experts for Multimodal Knowledge Graph Completion

Xu Hou, Meiyu Liang, Wei Huang, Yawen Li, Zhe Xue, Wu Liu, et al.

Multimodal Knowledge Graph Completion (MKGC) requires inferring missing entities from structural, textual, and visual cues. Existing diffusion-based MKGC methods usually denoise directly on raw multimodal features. Such a design forces the denoiser to simultaneously perform relat…

View free PDFSource page
arxivcs.LGcs.AI2026-07-20

Selectivity Matters: Source Node Influence Pruning for Unsupervised Graph Domain Adaptation

Ridong Han, Yawen Shen, Zhongnian Li, Tongfeng Sun, Xinzheng Xu, Abdulmotaleb El Saddik

Unsupervised Graph Domain Adaptation (UGDA) aims to facilitate knowledge transfer from a labeled source graph to an unlabeled target graph by mitigating cross-domain distribution shifts. Existing methods primarily focus on node-level feature alignment in latent spaces, relying on…

View free PDFSource page
arxivcs.CVcs.AIcs.ARcs.DCcs.LG2026-07-01

Fusion: A Framework for Unified Sequential Token AdaptatIon in VisiOn TraNsformers

Aravind Pradeep, Samira Nazari, Mahdi Taheri, Christian Herglotz

Vision Transformers achieve strong image classification accuracy but process all image regions with nearly the same computation, even when many regions are redundant or uninformative. Recent adaptive inference methods reduce this cost by selectively compressing tokens or terminat…

View free PDFSource page
arxivcs.LGcs.AI2026-06-30

Geometry-Preserving Orthonormal Initialization for Low-Rank Adaptation in RLVR

Ruijia Zhang, Jiacheng Zhu, Hanqing Zhu, Laixi Shi

Low-rank adaptation (LoRA) and its variants enable parameter-efficient fine-tuning of large language models under the supervised fine-tuning (SFT) paradigm. However, their efficacy and behavior under Reinforcement learning with verifiable rewards (RLVR) are less well understood.…

View free PDFSource page