Supervised Fine-Tuning (SFT) accelerates task-specific large language models (LLMs) development, but the resulting proliferation of fine-tuned models incurs substantial memory overhead. Delta compression addresses this by retaining a single pre-trained LLM with multiple compressed delta weights. However, existing methods fail on models fine-tuned with large-scale datasets. We find that larger SFT data scale amplifies delta parameter magnitude, singular values, and entropy, exacerbating compression errors. To tackle this, we propose D-QRELO ( Delta Compression via Quantization and Rsidual Low-Rank), a novel training- and data-free delta compression method. It combines coarse-grained one-bit quantization to capture the dominant structure of the delta, followed by compensated residual low-rank approximation to recover fine-grained details from the smaller residual error. Experiments on various LLMs spanning dense and MoE architectures across multiple domains under this challenging setting demonstrate that D-QRELO outperforms existing methods. Moreover, we establish key design principles for delta compression through extensive empirical analysis, demonstrating how task difficulty, architecture, and layer positioning create predictable patterns that can guide optimal compression strategies in production systems.

Human scanpaths offer rich and reliable clues about the cognitive mechanisms underlying language comprehension. Decoder-only language models, typically large language models (LLMs), have proven to exhibit striking parallels with human cognitive processes. In this study, we investigate to what extent language models can be endowed with human-like gaze shifts. Besides, by probing scanpath through eye model, analogous to probing language through language models, we ask whether such modeling can yield novel knowledge of the cognitive machinery of sense making.This study presents a novel plug-and-play module, EyeLM, to transform an autoregressive language model into an autoregressive eye model, thus facilitating a probabilistic spatial modeling of human explicit attention. Our EyeLM module, powered by LLMs, achieves competitive performance with novel cognitive probing capabilities. By probing EyeLM, we can reach the predictability and uncertainty of the scanpath. Exhibiting aligned patterns with prior knowledge about human reading comprehension, these probabilistic measures of scanpath act as promising predictors of human comprehension skills.

Large Vision-Language Models (LVLMs) have achieved remarkable progress in multimodal tasks, yet their reliability is persistently undermined by hallucinations—generating text that contradicts visual input. Recent studies often attribute these errors to inadequate visual attention. In this work, we analyze the attention mechanisms via the logit lens, uncovering a distinct anomaly we term **Vocabulary Hijacking**. We discover that specific visual tokens, defined as **Inert Tokens**, disproportionately attract attention. Crucially, when their intermediate hidden states are projected into the vocabulary space, they consistently decode to a fixed set of unrelated words (termed **Hijacking Anchors**) across layers, revealing a rigid semantic collapse. Leveraging this semantic rigidity, we propose **Hijacking Anchor-Based Identification (HABI)**, a robust strategy to accurately localize these Inert Tokens. To quantify the impact of this phenomenon, we introduce the **Non-Hijacked Visual Attention Ratio (NHAR)**, a novel metric designed to identify attention heads that remain resilient to hijacking and are critical for factual accuracy. Building on these insights, we propose **Hijacking-Aware Visual Attention Enhancement (HAVAE)**, a training-free intervention that selectively strengthens the focus of these identified heads on salient visual content. Extensive experiments across multiple benchmarks demonstrate that HAVAE significantly mitigates hallucinations with **no additional computational overhead**, while preserving the model’s general capabilities.

Fine-tuning Large Language Models (LLMs) with multimodal encoders on modality-specific data expands the modalities that LLMs can handle, leading to the formation of Multimodal LLMs (MLLMs). However, this paradigm heavily relies on resource-intensive and inflexible fine-tuning from scratch with new multimodal data. In this paper, we propose MMER (Multi-modality Expansion and Retention), a training-free approach that integrates existing MLLMs for effective multimodal expansion while retaining their original performance. Specifically, MMER reuses MLLMs’ multimodal encoders while merging their LLM parameters. By comparing original and merged LLM parameters, MMER generates binary masks to approximately separate LLM parameters for each modality. These decoupled parameters can independently process modality-specific inputs, reducing parameter conflicts and preserving original MLLMs’ fidelity. MMER can also mitigate catastrophic forgetting by applying a similar process to MLLMs fine-tuned on new tasks. Extensive experiments show significant improvements over baselines, proving that MMER effectively expands LLMs’ multimodal capabilities while retaining 99% of the original performance, and also markedly mitigates catastrophic forgetting.

Foundation models and their checkpoints have significantly advanced deep learning, boosting performance across various applications. However, fine-tuned models often struggle outside their specific domains and exhibit considerable redundancy. Recent studies suggest that combining a pruned fine-tuned model with the original pre-trained model can mitigate forgetting, reduce interference when merging model parameters across tasks, and improve compression efficiency. In this context, developing an effective pruning strategy for fine-tuned models is crucial. Leveraging the advantages of the task vector mechanism, we preprocess fine-tuned models by calculating the differences between them and the original model. Recognizing that different task vector subspaces contribute variably to model performance, we introduce a novel method called **N**eural **P**arameter **S**earch (**NPS**) for slimming down fine-tuned models. This method enhances pruning efficiency by searching through neural parameters of task vectors within low-rank subspaces. Our method has three key applications: enhancing knowledge transfer through pairwise model interpolation, facilitating effective knowledge fusion via model merging, and enabling the deployment of compressed models that retain near-original performance while significantly reducing storage costs. Extensive experiments across vision, NLP, and multi-modal benchmarks demonstrate the effectiveness and robustness of our approach, resulting in substantial performance gains.

Grice’s Quantity Maxims dictate that human speakers aim for the optimal quantity of information during conversation. To empower LLMs to self-repair their responses toward optimal quantity and improve their attentive listening skills, we propose Q-Tuning and Q-Traveling, which draw on heuristic path-finding to enable decoder-only LLMs to travel among multiple “Q-alternatives” (Quantity Alternatives) and search for the optimal quantity in coordination with a conversation goal. Automatic and human evaluations demonstrate the effectiveness of Q-Tuning and Q-Traveling in constructing human-like, user-centered conversation agents.

Human empathy builds on the shared pragmatic common ground among different languages. However, existing human empathy data is limited to English. Inspired by multilingual coactivation as the neurocognitive underpinning of human bilingual proficiency, which predicts empathy, we integrate language-independent diffusion processes to facilitate the cross-lingual transfer of empathy. Taking Chinese language varieties as the target domain, automatic and human evaluations demonstrate successful transfers of source empathy into target contexts without compromising linguistic naturalness. The results of this work offer empirical clues on the importance of pragmatic transferability of empathy and its cross-lingual effects in conversation.

For a conversation to help and support, speakers should maintain an “effect-effort” tradeoff. As outlined in the gist of “Cognitive Relevance Principle”, helpful speakers should optimize the “cognitive relevance” through maximizing the “cognitive effects” and minimizing the “processing effort” imposed on listeners. Although preference learning methods have given rise a boon of studies in pursuit of“effect-optimization”, none have delved into the critical “effort-optimiazation” to fully cultivate the awareness of “optimal relevance” into thecognition of conversation agents. To address this gap, we integrate the “Cognitive Relevance Principle” into emotional support agents in the environment of multi-turn conversation. The results demonstrate a significant and robust improvement against the baseline systems with respect to response quality, human-likedness and supportivenss. This study offers compelling evidence for the effectiveness of the “Relevance Principle” in generating human-like, helpful, and harmless emotional support conversations. The source code will be available at https://github.com/CN-Eyetk/VLESA-ORL.git

Please find the attached PDF file for the extended abstract of our study.

To provide effective support, it is essential for a skilled supporter to emotionally resonate with the help-seeker’s current emotional state. In conversational interactions, this emotional alignment is further influenced by the comforting strategies employed by the supporter. Different strategies guide the interlocutors to align their emotions in nuanced patterns. However, the incorporation of strategy into emotional alignment in the context of emotional support agents remains underexplored. To address this limitation, we propose an improved emotional support agent called Emstremo. Emstremo aims to achieve strategic control of emotional alignment by perceiving and responding to the user’s emotions. Our system’s state-of-the-art performance emphasizes the importance of integrating emotions and strategies in modeling conversations that provide emotional support.

Eye-tracking data in Chinese languages present unique challenges due to the non-alphabetic and unspaced nature of the Chinese writing systems. This paper introduces the first deeply-annotated joint Mandarin-Cantonese eye-tracking dataset, from which we achieve a unified eye-tracking prediction system for both language varieties. In addition to the commonly studied first fixation duration and the total fixation duration, this dataset also includes the second fixation duration, expressing fixation patterns that are more relevant to higher-level, structural processing. A basic comparison of the features and measurements in our dataset revealed variation between Mandarin and Cantonese on fixation patterns related to word class and word position. The test of feature usefulness suggested that traditional features are less powerful in predicting the second-pass fixation, to which the linear distance to root makes a leading contribution in Mandarin. In contrast, Cantonese eye-movement behavior relies more on word position and part of speech.