Tool-Integrated Reasoning (TIR) has emerged as a promising direction by extending Large Language Models' (LLMs) capabilities with external tools during reasoning.Existing TIR methods typically rely on external tool documentation during reasoning.However, this leads to tool mastery difficulty, tool size constraints, and inference inefficiency.To mitigate these issues, we explore Tool-Internalized Reasoning (TInR), aiming at facilitating reasoning with tool knowledge internalized into LLMs.Achieving this goal presents notable requirements, including tool internalization and toolreasoning coordination.To address them, we propose TInR-U, a tool-internalized reasoning framework for unified reasoning and tool usage.TInR-U is trained through a three-phase pipeline: 1) tool internalization with a bidirectional knowledge alignment strategy; 2) supervised fine-tuning warm-up using high-quality reasoning annotations, and 3) reinforcement learning with TInR-specific rewards.We comprehensively evaluate our method across indomain and out-of-domain settings.Experiment results show that TInR-U achieves superior performance in both settings, highlighting its effectiveness and efficiency.Codes are available at https://github.com/travis-xu/TInR.

As Large Language Models (LLMs) increasingly participate in human-AI interactions, evaluating their Theory of Mind (ToM) capabilities - particularly their ability to track dynamic mental states - becomes crucial. While existing benchmarks assess basic ToM abilities, they predominantly focus on static snapshots of mental states, overlooking the temporal evolution that characterizes real-world social interactions. We present DynToM, a novel benchmark specifically designed to evaluate LLMs’ ability to understand and track the temporal progression of mental states across interconnected scenarios. Through a systematic four-step framework, we generate 1,100 social contexts encompassing 5,500 scenarios and 78,100 questions, each validated for realism and quality. Our comprehensive evaluation of ten state-of-the-art LLMs reveals that their average performance underperforms humans by 44.7%, with performance degrading significantly when tracking and reasoning about the shift of mental states. This performance gap highlights fundamental limitations in current LLMs’ ability to model the dynamic nature of human mental states.

The ability to continually learn over time by grasping new knowledge and remembering previously learned experiences is essential for developing an online task-oriented dialogue system (TDS). In this paper, we work on the class incremental learning scenario where the TDS is evaluated without specifying the dialogue domain. We employ contrastive distillation on the intermediate representations of dialogues to learn transferable representations that suffer less from catastrophic forgetting. Besides, we provide a dynamic update mechanism to explicitly preserve the learned experiences by only updating the parameters related to the new task while keeping other parameters fixed. Extensive experiments demonstrate that our method significantly outperforms the strong baselines.

In real-world scenarios, it is crucial to build a lifelong taskoriented dialogue system (TDS) that continually adapts to new knowledge without forgetting previously acquired experiences. Existing approaches mainly focus on mitigating the catastrophic forgetting in lifelong TDS. However, the transfer ability to generalize the accumulated old knowledge to new tasks is underexplored. In this paper, we propose a two-stage lifelong task-oriented dialogue generation method to mitigate catastrophic forgetting and encourage knowledge transfer simultaneously, inspired by the learning process. In the first stage, we learn task-specific masks which adaptively preserve the knowledge of each visited task so as to mitigate catastrophic forgetting. In this stage, we are expected to learn the task-specific knowledge which is tailored for each task. In the second stage, we bring the knowledge from the encountered tasks together and understand thoroughly. To this end, we devise a balanced meta learning strategy for both forward and backward knowledge transfer in the lifelong learning process. In particular, we perform meta-update with a meta-test set sampled from the current training data for forward knowledge transfer. In addition, we employ an uncertainty-based sampling strategy to select and store representative dialogue samples into episodic memory and perform meta-update with a meta-test set sampled from the memory for backward knowledge transfer. With extensive experiments on 29 tasks, we show that MetaLTDS outperforms the strong baselines in terms of both effectiveness and efficiency. For reproducibility, we submit our code at: https: //github.com/travis-xu/MetaLTDS.