Tongyao Zhu, Qian Liu, Chang Ma, Jinghan Zhang, Longxu Dou, Junxian He, Shiqi Chen. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Zhengyu Chen, Siqi Wang, Teng Xiao, Yudong Wang, Shiqi Chen, Xunliang Cai, Junxian He, Jingang Wang. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Vision-Language Models (VLMs) combine visual perception with the general capabilities, such as reasoning, of Large Language Models (LLMs). However, the mechanisms by which these two abilities can be combined and contribute remain poorly understood. In this work, we explore to compose perception and reasoning through model merging that connects parameters of different models. Unlike previous works that often focus on merging models of the same kind, we propose merging models across modalities, enabling the incorporation of the reasoning capabilities of LLMs into VLMs. Through extensive experiments, we demonstrate that model merging offers a successful pathway to transfer reasoning abilities from LLMs to VLMs in a training-free manner. Moreover, we utilize the merged models to understand the internal mechanism of perception and reasoning and how merging affects it. We find that perception capabilities are predominantly encoded in the early layers of the model, whereas reasoning is largely facilitated by the middle-to-late layers. After merging, we observe that all layers begin to contribute to reasoning, whereas the distribution of perception abilities across layers remains largely unchanged. These observations shed light on the potential of model merging as a tool for multimodal integration and interpretation.

Large Vision Language Models (VLMs) have long struggled with spatial reasoning tasks. Surprisingly, even simple spatial reasoning tasks, such as recognizing “under” or “behind” relationships between only two objects, pose significant challenges for current VLMs. We believe it is crucial to use the lens of mechanism interpretability, opening up the model and diving into model’s internal states to examine the interactions between image and text tokens during spatial reasoning. Our analysis of attention behaviors reveals significant differences in how VLMs allocate attention to image versus text. By tracing the areas of images that receive the highest attention scores throughout intermediate layers, we observe a notable pattern: errors often coincide with attention being misdirected towards irrelevant objects within the image. Moreover, such attention patterns exhibit substantial differences between familiar (e.g., “on the left side of ”) and unfamiliar (e.g.,“in front of ”) spatial relationships. Motivated by these findings, we propose ADAPTVIS based on inference-time confidence scores to sharpen the attention on highly relevant regions when the model exhibits high confidence, while smoothing and broadening the attention window to consider a wider context when confidence is lower. This training-free decoding method shows significant improvement (e.g., up to a 50 absolute point improvement) on spatial reasoning benchmarks such as WhatsUp and VSR with negligible additional cost.

Recent advances such as OpenAI-o1 and DeepSeek R1 have demonstrated the potential of Reinforcement Learning (RL) to enhance reasoning abilities in Large Language Models (LLMs). While open-source replication efforts have primarily focused on mathematical and coding domains, methods and resources for developing general reasoning capabilities remain underexplored. This gap is partly due to the challenge of collecting diverse and verifiable reasoning data suitable for RL. We hypothesize that logical reasoning is critical for developing general reasoning capabilities, as logic forms a fundamental building block of reasoning. In this work, we present SynLogic, a data synthesis framework and dataset that generates diverse logical reasoning data at scale, encompassing 35 diverse logical reasoning tasks. The SynLogic approach enables controlled synthesis of data with adjustable difficulty and quantity. Importantly, all examples can be verified by simple rules, making them ideally suited for RL with verifiable rewards. In our experiments, we validate the effectiveness of RL training on the SynLogic dataset based on 7B and 32B models. SynLogic leads to state-of-the-art logical reasoning performance among open-source datasets, surpassing DeepSeek-R1-Distill-Qwen-32B by 6 points on BBEH. Furthermore, mixing SynLogic data with mathematical and coding tasks improves the training efficiency of these domains and significantly enhances reasoning generalization. Notably, our mixed training model outperforms DeepSeek-R1-Zero-Qwen-32B across multiple benchmarks. These findings position SynLogic as a valuable resource for advancing the broader reasoning capabilities of LLMs. We will open-source both the data synthesis pipeline and the SynLogic dataset.

Recent advancements in LLM pretraining have featured ever-expanding context windows to process longer sequences. However, our controlled study reveals that models pretrained with shorter context windows consistently outperform their long-context counterparts under a fixed token budget. This finding motivates us to explore an optimal context window scheduling strategy to better balance long-context capability with pretraining efficiency. To this end, we propose SkyLadder, a simple yet effective approach that implements a short-to-long context window transition. SkyLadder preserves strong standard benchmark performance, while matching or exceeding baseline results on long-context tasks. Through extensive experiments, we pretrain 1B-parameter models (up to 32K context) and 3B-parameter models (8K context) on 100B tokens, demonstrating that SkyLadder yields consistent gains of up to 3.7% on common benchmarks, while achieving up to 22% faster training speeds compared to baselines.

While large language models (LLMs) have demonstrated remarkable abilities across various fields, hallucination remains a significant challenge.Recent studies have explored hallucinations through the lens of internal representations, proposing mechanisms to decipher LLMs' adherence to facts.However, these approaches often fail to generalize to out-of-distribution data, leading to concerns about whether internal representation patterns reflect fundamental factual awareness, or only overfit spurious correlations on the specific datasets.In this work, we investigate whether a universal truthfulness hyperplane that distinguishes the model's factually correct and incorrect outputs exists within the model.To this end, we scale up the number of training datasets and conduct an extensive evaluation -we train the truthfulness hyperplane on a diverse collection of over 40 datasets and examine its cross-task, cross-domain, and in-domain generalization.Our results indicate that increasing the diversity of the training datasets significantly enhances the performance in all scenarios, while the volume of data samples plays a less critical role.This finding supports the optimistic hypothesis that a universal truthfulness hyperplane may indeed exist within the model, offering promising directions for future research.Code is publicly available at https://github.com/hkust-nlp/ Universal_Truthfulness_Hyperplane.Tend to overfit

Large language models (LLMs) frequently hallucinate, e.g., making factual errors, yet our understanding of why they make these errors remains limited. In this study, we aim to understand the underlying mechanisms of LLM hallucinations from the perspective of *inner representations*. We discover a pattern associated with hallucinations: correct generations tend to have *sharper* context activations in the hidden states of the in-context tokens, compared to that of the incorrect generations. Leveraging this signal, we propose an entropy-based metric to quantify the *sharpness* among the in-context hidden states and incorporate it into the decoding process, i.e, use the entropy value to adjust the next token prediction distribution to improve the factuality and overall quality of the generated text. Experiments on knowledge-seeking datasets (Natural Questions, HotpotQA, TriviaQA) and hallucination benchmark (TruthfulQA) demonstrate our consistent effectiveness, e.g., up to 8.6 absolute points on TruthfulQA. We believe this study can improve our understanding of hallucinations and serve as a practical solution for hallucination mitigation.

Assessing factuality of text generated by large language models (LLMs) is an emerging yet crucial research area, aimed at alerting users to potential errors and guiding the development of more reliable LLMs. Nonetheless, the evaluators assessing factuality necessitate suitable evaluation themselves to gauge progress and foster advancements. This direction remains under-explored, resulting in substantial impediments to the progress of factuality evaluators. To mitigate this issue, we introduce a benchmark for Factuality Evaluation of large Language Models, referred to as FELM. In this benchmark, we collect responses generated from LLMs and annotate factuality labels in a fine-grained manner. Contrary to previous studies that primarily concentrate on the factuality of world knowledge (e.g. information from Wikipedia), FELM focuses on factuality across diverse domains, spanning from world knowledge to math and reasoning. Our annotation is based on text segments, which can help pinpoint specific factual errors. The factuality annotations are further supplemented by predefined error types and reference links that either support or contradict the statement. In our experiments, we investigate the performance of several LLM-based factuality evaluators on FELM, including both vanilla LLMs and those augmented with retrieval mechanisms and chain-of-thought processes. Our findings reveal that while retrieval aids factuality evaluation, current LLMs are far from satisfactory to faithfully detect factual errors.

As an efficient alternative to conventional full fine-tuning, parameter-efficient fine-tuning (PEFT) is becoming the prevailing method to adapt pretrained language models. In PEFT, a lightweight module is learned on each dataset while the underlying pretrained language model remains unchanged, resulting in multiple compact modules representing diverse skills when applied to various domains and tasks. In this paper, we propose to compose these parameter-efficient modules through linear arithmetic operations in the weight space, thereby integrating different module capabilities. Specifically, we first define an addition and negation operator for the module, and then further compose these two basic operators to perform flexible arithmetic. Our approach requires no additional training and enables highly flexible module composition. We apply different arithmetic operations to compose the parameter-efficient modules for (1) distribution generalization, (2) multi-tasking, (3) detoxifying, and (4) domain transfer. Additionally, we extend our approach to detoxify Alpaca-LoRA, the latest instruction-tuned large language model based on LLaMA. Empirical results demonstrate that our approach produces new and effective parameter-efficient modules that significantly outperform existing ones across all settings.

Meta Reinforcement Learning (MRL) enables an agent to learn from a limited number of past trajectories and extrapolate to a new task. In this paper, we attempt to improve the robustness of MRL. We build upon model-agnostic meta-learning (MAML) and propose a novel method to generate adversarial samples for MRL by using Generative Adversarial Network (GAN). That allows us to enhance the robustness of MRL to adversal attacks by leveraging these attacks during meta training process.