High-quality multi-hop instruction data is critical for enhancing the reasoning capabilities of large language models (LLMs) in complex long-context scenarios, e.g., long-form reasoning. Nevertheless, there is currently a notable scarcity of such datasets within the community, and existing data synthesis approaches typically fail to provide explicit modeling of intermediate reasoning steps, resulting in unverifiable and potentially erroneous samples. To mitigate above issue, we design the Concept-Graph based Multi-hop Instructions Synthesis (CGMIS) framework, which constructs long-form reasoning paths via concept graph traversal and automatically generates verifiable multi-hop data. The CGMIS framework not only guarantees the accuracy and verifiability of the synthesized data but also enables the construction of high-quality multi-hop instruction datasets from arbitrary corpora. Experiments show that fine-tuning with CGMIS-generated data achieves state-of-the-art performance across 13 long-context reasoning tasks on various models, using only 10% of the data volume required by existing methods.

Zechen Sun, Yuyang Sun, Zecheng Tang, Juntao Li, Wenpeng Hu, Wenliang Chen, Zhunchen Luo, Guotong Geng, Min Zhang. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Iterative non-autoregressive (NAR) models share a spirit of mixed autoregressive (AR) and fully NAR models, seeking a balance between generation quality and inference efficiency.These models have recently demonstrated impressive performance in varied generation tasks, surpassing the autoregressive Transformer.However, they also face several challenges that impede further development.In this work, we target building more efficient and competitive iterative NAR models.Firstly, we produce two simple metrics to identify the potential problems existing in current refinement processes, and look back on the various iterative NAR models to find the key factors for realizing our purpose.Subsequently, based on the analyses of the limitations of previous inference algorithms, we propose a simple yet effective strategy to conduct efficient refinements without performance declines.Experiments on five widely used datasets show that our final models set the new state-of-the-art performance compared to all previous NAR models, even with fewer decoding steps, and outperform AR Transformer by around one BLEU on average.Our codes and models are available on Github 1 .

Long-context models (LCMs) have shown great potential in processing long input sequences (even more than 100M tokens) conveniently and effectively. With significant progress, recent research has pointed out that LCMs can accurately locate token-level salient information within the context. Yet, the generation performance of these LCMs is far from satisfactory and might result in misaligned responses, such as hallucinations. To enhance the generation capability of LCMs, existing works have investigated the effects of data size and quality for both pre-training and instruction tuning. Though achieving meaningful improvement, previous methods fall short in either effectiveness or efficiency. In this paper, we introduce LOGO (Long cOntext aliGnment via efficient preference Optimization), a training strategy that first introduces preference optimization for long-context alignment. To overcome the GPU memory-bound issue caused by the long sequence, LOGO employs a reference-free preference optimization strategy and adopts a position synthesis method to construct the training data. By training with only 0.3B data on a single 8 x A800 GPU machine for 16 hours, LOGO allows the Llama-3-8B-Instruct-80K model to achieve comparable performance with GPT-4 in real-world long-context tasks while preserving the model's original capabilities on other tasks, e.g., language modeling and MMLU. Moreover, LOGO can extend the model's context window size while enhancing its generation performance.

Recent generative large language models (LLMs) have exhibited incredible instruction-following capabilities while keeping strong task completion ability, even without task-specific fine-tuning. Some works attribute this to the bonus of the new scaling law, in which the continuous improvement of model capacity yields emergent capabilities, e.g., reasoning and universal generalization. However, we point out that recent LLMs still show shortcut learning behavior, where the models tend to exploit spurious correlations between non-robust features and labels for prediction, which might lead to overestimating model capabilities. LLMs memorize more complex spurious correlations (i.e., task \leftrightarrow feature \leftrightarrow label) compared with that learned from previous pre-training and task-specific fine-tuning paradigm (i.e., feature \leftrightarrow label). Based on our findings, we propose FSLI, a framework for encouraging LLMs to Forget Spurious correlations and Learn from In-context information. Experiments on three tasks show that FSFI can effectively mitigate shortcut learning. Besides, we argue not to overestimate the capabilities of LLMs and conduct evaluations in more challenging and complete test scenarios.

Language modeling at scale has proven very effective and brought unprecedented success to natural language models. Many typical representatives, especially decoder-only models, e.g., BLOOM and LLaMA, and encoder-decoder models, e.g., Flan-T5 and AlexaTM, have exhibited incredible instruction-following capabilities while keeping strong task completion ability. These large language models can achieve superior performance in various tasks and even yield emergent capabilities, e.g., reasoning and universal generalization. Though the above two paradigms are mainstream and well explored, the potential of the BERT family, which are encoder-only based models and have ever been one of the most representative pre-trained models, also deserves attention, at least should be discussed. In this work, we adopt XML-R to explore the effectiveness of the BERT family for instruction following and zero-shot learning. We first design a simple yet effective strategy to utilize the encoder-only models for generation tasks and then conduct multi-task instruction tuning. Experimental results demonstrate that our fine-tuned model, Instruct-XMLR, outperforms Bloomz on all evaluation tasks and achieves comparable performance with mT0 on most tasks. Surprisingly, Instruct-XMLR also possesses strong task and language generalization abilities, indicating that Instruct-XMLR can also serve as a good instruction follower and zero-shot learner. Besides, Instruct-XMLR can accelerate decoding due to its non-autoregressive generation manner, achieving around 3 times speedup compared with current autoregressive large language models. Although we also witnessed several limitations through our experiments, such as the performance decline in long-generation tasks and the shortcoming of length prediction, Instruct-XMLR can still become a good member of the family of current large language models.