Hybrid-thinking language models expose both cheap non-thinking and expensive thinking modes, creating a deployment problem: deciding when additional reasoning is worth its token cost and override risk. We formulate this problem as budgeted opportunity allocation, where a controller first obtains a cheap seed answer, estimates the marginal value and risk of paid thinking, and preserves the seed unless an adoption gate accepts the thinking result. We propose Bora-Budget, a budget-aware inference framework for frozen LLMs, and instantiate it as Bora-Switch-Calibrated. The controller uses seed-state features to predict helpfulness, harmfulness, and thinking cost, then allocates thinking through either frontier operating points or an AutoThreshold policy calibrated to a total token budget. Across LLM scales on math dataset, the method substantially improves over random allocation and approaches uniform-thinking accuracy under lower token budgets. Transfer results on multiple-choice, code, and hard-QA tasks show promising but uneven generalization: trace length remains a strong baseline, harmful overrides are reduced but not eliminated, and learning how much to think remains open. These results position Bora-Budget as a seed-preserving control layer for budgeted inference rather than a replacement for stronger always-thinking baselines.

In this paper, we survey recent advances in meta-reasoning for Large Reasoning Models (LRMs). Although LRMs have made remarkable progress, they face critical limitations including overthinking on simple queries, under-thinking on complex problems, and ineffective exploration lacking strategic direction. We examine how meta-reasoning addresses these challenges by providing the necessary mechanisms to dynamically regulate and optimize the reasoning process, and propose a framework to categorize existing studies into three core components: the monitor which provides self-perception of the reasoning state, the controller which dictates the meta-policy for regulating actions, and the mechanism which executes the algorithmic mapping between them. By systematically reviewing these developments and foundational challenges, we aim to map the current landscape and identify promising directions for creating more adaptive, efficient, and self-regulating reasoning systems.

Recent researches show that pre-trained models (PTMs) are beneficial to Chinese Word Segmentation (CWS). However, PTMs used in previous works usually adopt language modeling as pre-training tasks, lacking task-specific prior segmentation knowledge and ignoring the discrepancy between pre-training tasks and downstream CWS tasks. In this paper, we propose a CWS-specific pre-trained model MetaSeg, which employs a unified architecture and incorporates meta learning algorithm into a multi-criteria pre-training task. Empirical results show that MetaSeg could utilize common prior segmentation knowledge from different existing criteria and alleviate the discrepancy between pre-trained models and downstream CWS tasks. Besides, MetaSeg can achieve new state-of-the-art performance on twelve widely-used CWS datasets and significantly improve model performance in low-resource settings.

Incorporating lexicons into character-level Chinese NER by lattices is proven effective to exploitrich word boundary information. Previous work has extended RNNs to consume lattice inputsand achieved great success. However, due to the DAG structure and the inherently unidirectionalsequential nature, this method precludes batched computation and sufficient semantic interaction. In this paper, we propose PLTE, an extension of transformer encoder that is tailored for ChineseNER, which models all the characters and matched lexical words in parallel with batch process-ing. PLTE augments self-attention with positional relation representations to incorporate latticestructure. It also introduces a porous mechanism to augment localness modeling and maintainthe strength of capturing the rich long-term dependencies. Experimental results show that PLTEperforms up to 11.4 times faster than state-of-the-art methods while realizing better performance. We also demonstrate that using BERT representations further substantially boosts the performanceand brings out the best in PLTE.