Chain-of-thought traces drive strong LLM accuracy but dominate inference cost: output length sets both monetary cost and decoding latency. We ask which tokens in a reasoning trace actually carry the answer. A token-level analysis of the model’s own next-token distribution identifies three redundancies - discourse-element removal, tokenizer-aware abbreviation, and substitution of prose causal connectives with an arrow "X -> Y". All three leave the predictive distribution close to the original under Jensen-Shannon divergence and preserve the greedy next token in the vast majority of prefixes. We instantiate the analysis as CONCISON, a single training-free system prompt that operationalises the three rules and protects numeric, equation, and code tokens verbatim. Across 18 models from six families (2B-120B) and five reasoning benchmarks, CONCISON cuts output tokens by 24-62% on strong models at -1.1pp average accuracy, outperforms token-budget prompting in a single decoding call, and matches trained routers on the accuracy-cost frontier without any training, model pool, or extra infrastructure.

Large Language Models (LLMs) demonstrate strong multilingual capabilities but are costly to deploy due to their size and computational demands. To mitigate this, compression techniques such as pruning and quantization are widely used. However, these methods face two key limitations: (1) they assume access to high-quality instruction or calibration data, which is often unavailable for low-resource languages; and (2) they aim to preserve multilingual generality, making them inefficient for language-specific applications. We introduce LangCompress, a language-aware compression framework that enhances existing compression methods for targeted deployment. LangCompress is method-agnostic and improves state-of-the-art pruning and quantization approaches. It features two core components: an iterative self-supervised pipeline for generating instruction data in the target language, and a vocabulary simplification strategy that reduces the LM head to focus on key tokens. Experiments on perplexity, translation, and summarization tasks show that LangCompress improves performance in the target language. The code and data are publicly available.

Vietnamese is the native language of over 98 million people in the world. However, existing Vietnamese Question Answering (QA) datasets do not explore the model’s ability to perform advanced reasoning and provide evidence to explain the answer. We introduce VIMQA, a new Vietnamese dataset with over 10,000 Wikipedia-based multi-hop question-answer pairs. The dataset is human-generated and has four main features: (1) The questions require advanced reasoning over multiple paragraphs. (2) Sentence-level supporting facts are provided, enabling the QA model to reason and explain the answer. (3) The dataset offers various types of reasoning to test the model’s ability to reason and extract relevant proof. (4) The dataset is in Vietnamese, a low-resource language. We also conduct experiments on our dataset using state-of-the-art Multilingual single-hop and multi-hop QA methods. The results suggest that our dataset is challenging for existing methods, and there is room for improvement in Vietnamese QA systems. In addition, we propose a general process for data creation and publish a framework for creating multilingual multi-hop QA datasets. The dataset and framework are publicly available to encourage further research in Vietnamese QA systems.