Emotional support is a core capability in human-AI interaction, with applications including psychological counseling, role play, and companionship. However, existing evaluations of large language models (LLMs) often rely on short, static dialogues and fail to capture the dynamic and long-term nature of emotional support. To overcome this limitation, we shift from snapshot-based evaluation to trajectory-based assessment, adopting a user-centered perspective that evaluates models based on their ability to improve and stabilize user emotional states over time. Our framework constructs a large-scale benchmark consisting of 328 emotional contexts and 1,152 disturbance events, simulating realistic emotional shifts under evolving dialogue scenarios. To encourage psychologically grounded responses, we constrain model outputs using validated emotion regulation strategies such as situation selection and cognitive reappraisal. User emotional trajectories are modeled as a first-order Markov process, and we apply causally-adjusted emotion estimation to obtain unbiased emotional state tracking. Based on this framework, we introduce three trajectory-level metrics: Baseline Emotional Level (BEL), Emotional Trajectory Volatility (ETV), and Emotional Centroid Position (ECP). These metrics collectively capture user emotional dynamics over time and support comprehensive evaluation of long-term emotional support performance of LLMs. Extensive evaluations across a diverse set of LLMs reveal significant disparities in emotional support capabilities and provide actionable insights for model development.

Large language models (LLMs) suffer from a lack of decision-making transparency, limiting their deployment in high-stakes domains such as healthcare. We propose a mechanistic interpretability framework that introduces two novel paradigms: Medical Fine-Tuning with Frozen Attention Layers (FTFA) and Posterior Adaptation Transcoders (PAT). FTFA freezes attention layers while fine-tuning only feed-forward network (FFN) parameters, enabling PAT to efficiently adapt pre-trained transcoders on the same data. This approach achieves over 1000× efficiency improvement compared to training transcoders from scratch. We theoretically justify this methodology and demonstrate its cost-effectiveness for cross-domain transfer. Transcoders are sparse autoencoders that replace MLP layers to provide interpretable feature representations. By substituting MLP layers of both base Gemma2-2b and its medical fine-tuned variant with per-layer transcoders, we enable feature-level attribution analysis. Through systematic pruning and node merging of resulting attribution graphs, we construct human-interpretable decision pathways. Our analysis reveals that LLMs employ two parallel mechanisms for medical diagnosis: pattern matching and multi-hop reasoning, with fine-tuned models demonstrating enhanced correct reasoning patterns. This work provides a practical framework for training transcoders on fine-tuned models at minimal cost, enabling broader application of mechanistic interpretability across domains and potentially guiding model training through transcoder-based analysis.

Ruochong Xiong, Qien Li, Wangwang Lian, Yulong Wan, Hanlin Xue, Zhouxing Tan, Han Yang, Fengyu Lu, Junfei Liu. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Existing dialog models are trained with data in an encoder-decoder framework with the same parameters, ignoring the multinomial distribution nature in the dataset. In fact, model improvement and development commonly requires fine-grained modeling on individual data subsets. However, collecting a labeled fine-grained dialogue dataset often requires expert-level domain knowledge and therefore is difficult to scale in the real world. As we focus on better modeling multinomial data for dialog generation, we study an approach that combines the unsupervised clustering and generative model together with a GMM (Gaussian Mixture Model) based encoder-decoder framework. Specifically, our model samples from the prior and recognition distributions over the latent variables by a Gaussian mixture network and the latent layer with the capability to form multiple clusters. We also introduce knowledge distillation to guide and improve the clustering results. Finally, we use a dynamic parameter sharing strategy conditioned on different labels to train different decoders. Experimental results on a widely used dialogue dataset verify the effectiveness of the proposed method.

The purpose of medical dialogue generation is to provide automatic and accurate responses to help doctors provide diagnosis and treatment advice in an efficient w ay. However, due to the subjectivity and open-ended nature of human conversations, the complexity of training dataset varies greatly. Some methods try to solve this problem by using curriculum learning that organizes training data from easy to hard and optimizes the model with an “easy-to-difficult” scheme, but they ignore the intrinsic nature of conversation dataset that has different categories. To solve this problem, in this article, we learn a Dynamic Curriculum with Co-training (DCC) for generative dialogue systems considering both the quality and category of the data. Under our framework, we first t rain two models (query generation model and reply generation model) from dual view by leveraging end-to-end deep Gaussian Mixture Variational Autoencoders (GMVAE) architectures that combine generative and unsupervised clustering tasks together. Then we promote the clustering performance via an ensemble method by combining the clustering distribution of the two networks. Finally, we let the two models determine the training order for each other in each category according to their loss and category confidence dynamically through single-task and multitask curriculum learning. Evaluation results on the widely used medical dialogue generation dataset indicate that our proposed learning approach makes significant improvements compared to strong baselines. varies greatly.

In recent years, the research of the medical dialogue system has attracted much attention. Considering that in the dialogue system, queries with similar meanings tend to have similar replies. In the medical field, this phenomenon is even more prevalent. For queries of the same class, their corresponding replies typically have similar meanings and can be classified into the same category. Having observed that, we propose to improve the neural sequence-to-sequence (Seq2Seq) based medical dialogue system by utilizing this internal relationship of category information between queries and replies. In our model, we first cluster similar queries into the same category according to their query vectors obtained from the encoder. Then we put forward the indirect and direct distillation learning approach to transfer the category information and category center distance from the queries to the replies. In the indirect distillation process, we employ metric learning to learn better representations of replies, in which replies of corresponding queries in the same category are closely grouped together, whereas those with different categories are far apart. In the direct distillation, to transfer the inter-class relationship, we minimize the Kullback-Leibler (KL) divergence between the category center distance distribution of queries and replies. A large number of experimental results on medical datasets have proved that our method is superior to the most advanced one.

Naive neural dialogue generation models tend to produce repetitive and dull utterances. The promising adversarial models train the generator against a well-designed discriminator to push it to improve towards the expected direction. However, assessing dialogues requires consideration of many aspects of linguistics, which are difficult to be fully covered by a single discriminator. To address it, we reframe the dialogue generation task as a multi-objective optimization problem and propose a novel adversarial dialogue generation framework with multiple discriminators that excel in different objectives for multiple linguistic aspects, called AMPGAN, whose feasibility is proved by theoretical derivations. Moreover, we design an adaptively adjusted sampling distribution to balance the discriminators and promote the overall improvement of the generator by continuing to focus on these objectives that the generator is not performing well relatively. Experimental results on two real-world datasets show a significant improvement over the baselines.