Reinforcement learning with verifiable rewards (e.g., GRPO) significantly advances LLM reasoning. However, in complex tasks, GRPO frequently encounters the "zero-advantage problem": when all rollouts for a query fail, the relative advantage collapses to zero, wasting computational budget. While resampling with more budget is a common remedy, the static sampling policy inherently constrains reasoning exploration, limiting the success rate. In this paper, we propose Lorem Perturbation for Exploration (LoPE), a simple training framework to break this exploration bottleneck. Motivated by the intuition that task-irrelevant prompt perturbations can shift the output distribution, LoPE prepends sequences stochastically assembled from the Lorem Ipsum vocabulary to the prompts before resampling, forcing the model to explore broader reasoning pathways. Experiments across 1.7B, 4B, and 7B models show LoPE significantly outperforms resampling with original prompts. Our findings, including the effectiveness of other low-perplexity Latin-based random sequences, establish LoPE as a strong baseline for broadening LLM exploration.

Large Language Models (LLMs) for complex reasoning is often hindered by high computational costs and latency, while resource-efficient Small Language Models (SLMs) typically lack the necessary reasoning capacity. Existing collaborative approaches, such as cascading or routing, operate at a coarse granularity by offloading entire queries to LLMs, resulting in significant computational waste when the SLM is capable of handling the majority of reasoning steps. To address this, we propose RelayLLM, a novel framework for efficient reasoning via token-level collaborative decoding. Unlike routers, RelayLLM empowers the SLM to act as an active controller that dynamically invokes the LLM only for critical tokens via a special command, effectively "relaying" the generation process. We introduce a two-stage training framework, including warm-up and Group Relative Policy Optimization (GRPO) to teach the model to balance independence with strategic help-seeking. Empirical results across six benchmarks demonstrate that RelayLLM achieves an average accuracy of 49.52%, effectively bridging the performance gap between the two models. Notably, this is achieved by invoking the LLM for only 1.07% of the total generated tokens, offering a 98.2% cost reduction compared to performance-matched random routers.

Self-evolving has emerged as a key paradigm for improving foundational models such as Large Language Models (LLMs) and Vision Language Models (VLMs) with minimal human intervention. While recent approaches have demonstrated that LLM agents can self-evolve from scratch with little to no data, VLMs introduce an additional visual modality that typically requires at least some seed data, such as images, to bootstrap the self-evolution process. In this work, we present Multi-model Multi-modal Zero (MM-Zero), the first RL-based framework to achieve zero-data self-evolution for VLM reasoning. Moving beyond prior dual-role (Proposer and Solver) setups, MM-Zero introduces a multi-role self-evolving training framework comprising three specialized roles: a Proposer that generates structured visual concepts and formulates questions; a Coder that translates these concepts into executable code (e.g., Python, SVG) to render visual images; and a Solver that performs multimodal reasoning over the generated visual content. All three roles are initialized from the same base model and trained using Group Relative Policy Optimization (GRPO), with carefully designed reward mechanisms that integrate execution feedback, visual verification, and difficulty balancing. Our experiments show that MM-Zero improves VLM reasoning performance across a wide range of multimodal benchmarks. MM-Zero establishes a scalable path toward self-evolving multi-model systems for multimodal models, extending the frontier of self-improvement beyond the conventional two-model paradigm.

Test-time scaling improves the reasoning performance of large language models but incurs substantial cost in both total computation and latency. Existing adaptive sampling methods partially mitigate this issue by dynamically deciding when to stop sampling, yet they typically rely on heuristic rules or rely on distribution assumptions. In this work, we formulate adaptive sampling as a Markov decision process (MDP). We train a lightweight sampling controller with reinforcement learning (RL) to jointly balance answer correctness, latency, and computation cost. At each round, the controller decides to stop sampling or to acquire additional samples. Our method is lightweight which only relies on statistics of final answers, and can be trained and deployed on CPU. We further show that the resulting framework admits an interpretation as the Lagrangian relaxation of a constrained optimization problem with explicit budget constraints. Experiments against strong baselines such as ASC and ESC show that our method achieves improved trade-offs among answer correctness, sampling rounds, and total samples required.

In-Context Learning (ICL) emerges as a key feature for Large Language Models (LLMs), allowing them to adapt to new tasks by leverageing task-specific examples without updating model parameters. However, ICL faces challenges with increasing numbers of examples due to performance degradation and quadratic computational costs. In this paper, we propose Logit Arithmetic Reweighting Approach (LARA), a novel framework that enhances ICL by using logit-based ensembling of multiple demonstrations. Our approach divides long input demonstrations into parallelizable shorter inputs to significantly reduce memory requirements, and then effectively aggregate the information by reweighting logits of each group via a non-gradient optimization approach. We further introduce Binary LARA (B-LARA), a variant that constrains weights to binary values to simplify the search space and reduces memory usage by filtering out less informative demonstration groups. Experiments on BBH and MMLU demonstrate that LARA and B-LARA outperform all baseline methods in both accuracy and memory efficiency. We also conduct extensive analysis to show that LARA generalizes well to scenarios of varying numbers of examples from limited to many-shot demonstrations.

Large pre-trained language models (PLMs) have achieved remarkable success, making them highly valuable intellectual property due to their expensive training costs. Consequently, model watermarking, a method developed to protect the intellectual property of neural models, has emerged as a crucial yet underexplored technique. The problem of watermarking PLMs has remained unsolved since the parameters of PLMs will be updated when fine-tuned on downstream datasets, and then embedded watermarks could be removed easily due to the catastrophic forgetting phenomenon. This study investigates the feasibility of watermarking PLMs by embedding backdoors that can be triggered by specific inputs. We employ contrastive learning during the watermarking phase, allowing the representations of specific inputs to be isolated from others and mapped to a particular label after fine-tuning. Moreover, we demonstrate that by combining weight perturbation with the proposed method, watermarks can be embedded in a flatter region of the loss landscape, thereby increasing their robustness to watermark removal. Extensive experiments on multiple datasets demonstrate that the embedded watermarks can be robustly extracted without any knowledge about downstream tasks, and with a high success rate.

Tables are widely used in research and business, which are suitable for human consumption, but not easily machine-processable, particularly when tables are present in images. One of the main challenges to extracting data from images of tables is accurately recognizing table structures, especially for complex tables with cross rows and columns. In this study, we propose a novel multi-modal pre-training model for table structure recognition, named TableVLM.With a two-stream multi-modal transformer-based encoder-decoder architecture, TableVLM learns to capture rich table structure-related features by multiple carefully-designed unsupervised objectives inspired by the notion of masked visual-language modeling. To pre-train this model, we also created a dataset, called ComplexTable, which consists of 1,000K samples to be released publicly. Experiment results show that the model built on pre-trained TableVLM can improve the performance up to 1.97% in tree-editing-distance-score on ComplexTable.

Distantly supervision automatically generates plenty of training samples for relation extraction. However, it also incurs two major problems: noisy labels and imbalanced training data. Previous works focus more on reducing wrongly labeled relations (false positives) while few explore the missing relations that are caused by incompleteness of knowledge base (false negatives). Furthermore, the quantity of negative labels overwhelmingly surpasses the positive ones in previous problem formulations. In this paper, we first provide a thorough analysis of the above challenges caused by negative data. Next, we formulate the problem of relation extraction into as a positive unlabeled learning task to alleviate false negative problem. Thirdly, we propose a pipeline approach, dubbed ReRe, that first performs sentence classification with relational labels and then extracts the subjects/objects. Experimental results show that the proposed method consistently outperforms existing approaches and remains excellent performance even learned with a large quantity of false positive samples. Source code is available online at https://github.com/redreamality/RERE-relation-extraction.