Large language models (LLMs) acquire substantial world knowledge during pre-training, which is further shaped by post-training techniques such as supervised fine-tuning (SFT). However, the impact of SFT on a model’s knowledge remains underexplored, limiting our ability to control knowledge behavior in fine-tuned models. To address this gap, we evaluate closed-book question answering (CBQA) performance across five LLMs from the LLaMA-2 and LLaMA-3 families. Surprisingly, models fine-tuned on 1,920 samples perform up to 14% worse than those fine-tuned on only 240 samples. Furthermore, varying the level of knowledge mastery in the fine-tuning data leads to performance fluctuations of over 12%. To investigate these effects, we analyze model behavior at both the token and parameter levels. Our analysis reveals that up to 90% of parameter updates during SFT do not contribute to knowledge enhancement. Restoring these updates can improve performance on the CBQA task, depending on the characteristics of the fine-tuning data. These insights offer practical guidance for developing fine-tuning strategies that more effectively strengthen model knowledge.

Large language models (LLMs) achieve remarkable advancements by leveraging tools to interact with environments, a critical step toward generalized AI. However, the standard supervised fine-tuning (SFT) approach, which relies on large-scale datasets, often overlooks task-specific characteristics in tool use, leading to performance bottlenecks. To address this issue, we analyze three existing LLMs and uncover key insights: training data can inadvertently impede tool-use behavior, token importance is distributed unevenly, and errors in tool calls fall into a small set of categories. Building on these findings, we propose TL-Training, a task-feature-based framework that mitigates the effects of suboptimal training data, dynamically adjusts token weights to prioritize key tokens during SFT, and incorporates a robust reward mechanism tailored to error categories, optimized through proximal policy optimization. We validate TL-Training by training CodeLLaMA-2-7B and evaluating it on four open-source test sets. Our results demonstrate that the LLM trained by our method matches or surpasses both open- and closed-source LLMs in tool-use performance using only 1,217 training data points. Additionally, our method enhances robustness in noisy environments and improves general task performance, offering a scalable and efficient paradigm for tool-use training in LLMs. Code and data are available at https://github.com/Junjie-Ye/TL-Training.

Mental health issues are worsening in today’s competitive society, such as depression and anxiety. Traditional healings like counseling and chatbots fail to engage effectively, they often provide generic responses lacking emotional depth. Although large language models (LLMs) have the potential to create more human-like interactions, they still struggle to capture subtle emotions. This requires LLMs to be equipped with human-like adaptability and warmth. To fill this gap, we propose the MIND (Multi-agent INner Dialogue), a novel paradigm that provides more immersive psychological healing environments. Considering the strong generative and role-playing ability of LLM agents, we predefine an interactive healing framework and assign LLM agents different roles within the framework to engage in interactive inner dialogues with users, thereby providing an immersive healing experience. We conduct extensive human experiments in various real-world healing dimensions, and find that MIND provides a more user-friendly experience than traditional paradigms. This demonstrates that MIND effectively leverages the significant potential of LLMs in psychological healing.

As general large language models continue to advance, their real-world adaptation through effective fine-tuning remains a significant challenge. We introduce Hierarchical Multilevel Contrastive Learning (HMCL), a new contrastive learning framework that improves task-specific text representation for general models. HMCL integrates 3-level semantic differentiation (positive, weak-positive, and negative) and unifies contrastive learning, pair classification, and ranking objectives into a cohesive optimization strategy. HMCL demonstrates exceptional results across multi-domain and multilingual benchmarks, including text similarity, retrieval, reranking and Retrieval-Augmented Generation (RAG) tasks. It outperforms top unsupervised methods and supervised fine-tuning approaches while maintaining broad compatibility with architectures ranging from BERT to Qwen, 330M to 7B. In real-world merchant consultation scenarios, HMCL shows a 0.70-6.24 point improvement over original fine-tuning methods in large-scale base models. This establishes HMCL as a versatile solution that bridges the gap between general-purpose models and specialized industrial applications.

Referring Expression Comprehension (REC) is a crucial cross-modal task that objectively evaluates the capabilities of language understanding, image comprehension, and language-to-image grounding. Consequently, it serves as an ideal testing ground for Multi-modal Large Language Models (MLLMs). In pursuit of this goal, we have established a new REC dataset characterized by two key features: Firstly, it is designed with controllable varying levels of difficulty, necessitating multi-level fine-grained reasoning across object categories, attributes, and multi-hop relationships. Secondly, it includes negative text and images created through fine-grained editing and generation based on existing data, thereby testing the model’s ability to correctly reject scenarios where the target object is not visible in the image—an essential aspect often overlooked in existing datasets and approaches. Utilizing this high-quality dataset, we conducted comprehensive evaluations of both state-of-the-art specialist models and MLLMs. Our findings indicate that there remains a significant gap in achieving satisfactory grounding performance. We anticipate that our dataset will inspire new approaches to enhance visual reasoning and develop more advanced cross-modal interaction strategies, ultimately unlocking the full potential of MLLMs.

Early exiting can reduce the average latency of pre-trained language models (PLMs) via its adaptive inference mechanism and work with other inference speed-up methods like model pruning, thus drawing much attention from the industry. In this work, we propose a novel framework, BADGE, which consists of two off-the-shelf methods for improving PLMs’ early exiting. We first address the issues of training a multi-exit PLM, the backbone model for early exiting. We propose the novel architecture of block-wise bypasses, which can alleviate the conflicts in jointly training multiple intermediate classifiers and thus improve the overall performances of multi-exit PLM while introducing negligible additional flops to the model. Second, we propose a novel divergence-based early exiting (DGE) mechanism, which obtains early exiting signals by comparing the predicted distributions of two adjacent layers’ exits. Extensive experiments on three proprietary datasets and three GLUE benchmark tasks demonstrate that our method can obtain a better speedup-performance trade-off than the existing baseline methods.\footnote{Code will be made publicly available to the research community upon acceptance.}

Prompt tuning has become a new paradigm for model tuning and it has demonstrated success in natural language pretraining and even vision pretraining. The parameter-efficient prompt tuning methods that optimize soft embeddings while keeping the pretrained model frozen demonstrate advantages in low computation costs and almost lossless performance. In this work, we explore the transfer of prompt tuning to multimodal pretrained models. Specifically, we implement prompt tuning to a unified sequence-to-sequence pretrained model by adding a sequence of learnable embeddings to each layer and finetuning the pretrained model on downstream task with only the learnable embeddings being optimized. Experimental results on a series of multimodal understanding and generation tasks demonstrate that our method OFA-PT can achieve comparable performance with finetuning across a series of multimodal generation and understanding tasks. Additionally, it significantly outperforms the unified multimodal pretrained model with other parameter-efficient tuning methods, e.g., Adapter, BitFit. etc. Besides, in comparison with finetuned models, the prompt-tuned models demonstrate improved robustness against adversarial attacks. We further figure out that experimental factors, including prompt length, prompt depth, and reparameteratization, have great impacts on the model performance, and thus we empirically provide a recommendation for the setups of prompt tuning.

This paper proposes a new method, OFA-OCR, to transfer multimodal pretrained models to text recognition. Specifically, we recast text recognition as image captioning and directly transfer a unified vision-language pretrained model to the end task. Without pretraining on large-scale annotated or synthetic text recognition data, OFA-OCR outperforms the baselines and achieves state-of-the-art performance in the Chinese text recognition benchmark. Additionally, we construct an OCR pipeline with OFA-OCR, and we demonstrate that it can achieve competitive performance with the product-level API.

Recently, a series of works have looked into further improving the adapter-based tuning by manually designing better adapter architectures. Understandably, these manually designed solutions are sub-optimal. In this work, we propose the Learned Adapter framework to automatically learn the optimal adapter architectures for better task adaptation of pre-trained models (PTMs). First, we construct a unified search space for adapter architecture designs. In terms of the optimization method on the search space, we propose a simple-yet-effective method, GDNAS for better architecture optimization. Extensive experiments show that our Learned Adapter framework can outperform the previous parameter-efficient tuning (PETuning) baselines while tuning comparable or fewer parameters. Moreover: (a) the learned adapter architectures are explainable and transferable across tasks. (b) We demonstrate that our architecture search space design is valid.

Natural language spatial video grounding aims to detect the relevant objects in video frames with descriptive sentences as the query. In spite of the great advances, most existing methods rely on dense video frame annotations, which require a tremendous amount of human effort. To achieve effective grounding under a limited annotation budget, we investigate one-shot video grounding and learn to ground natural language in all video frames with solely one frame labeled, in an end-to-end manner. One major challenge of end-to-end one-shot video grounding is the existence of videos frames that are either irrelevant to the language query or the labeled frame. Another challenge relates to the limited supervision, which might result in ineffective representation learning. To address these challenges, we designed an end-to-end model via Information Tree for One-Shot video grounding (IT-OS). Its key module, the information tree, can eliminate the interference of irrelevant frames based on branch search and branch cropping techniques. In addition, several self-supervised tasks are proposed based on the information tree to improve the representation learning under insufficient labeling. Experiments on the benchmark dataset demonstrate the effectiveness of our model.

Existing image captioning systems are dedicated to generating narrative captions for images, which are spatially detached from theimage in presentation. However, texts can also be used as decorations on the image to highlight the key points and increase theattractiveness of images. In this work, we introduce a new taskcalled captioning on image (CapOnImage), which aims to generatedense captions at different locations of the image based on contextual information. To fully exploit the surrounding visual context togenerate the most suitable caption for each location, we propose amulti-modal pre-training model with multi-level pre-training tasksthat progressively learn the correspondence between texts and image locations from easy to difficult. Since the model may generateredundant captions for nearby locations, we further enhance thelocation embedding with neighbor locations as context. For thisnew task, we also introduce a large-scale benchmark called CapOnImage2M, which contains 2.1 million product images, each with anaverage of 4.8 spatially localized captions. Compared with other image captioning model variants, our model achieves the best resultsin both captioning accuracy and diversity aspects.

BERT and other pretrained language models (PLMs) are ubiquitous in modern NLP. Even though PLMs are the state-of-the-art (SOTA) models for almost every NLP task (CITATION), the significant latency during inference prohibits wider industrial usage. In this work, we propose Patient and Confident Early Exiting BERT (PCEE-BERT), an off-the-shelf sample-dependent early exiting method that can work with different PLMs and can also work along with popular model compression methods. With a multi-exit BERT as the backbone model, PCEE-BERT will make the early exiting decision if enough numbers (patience parameter) of consecutive intermediate layers are confident about their predictions. The entropy value measures the confidence level of an intermediate layer’s prediction. Experiments on the GLUE benchmark demonstrate that our method outperforms previous SOTA early exiting methods. Ablation studies show that: (a) our method performs consistently well on other PLMs, such as ALBERT and TinyBERT; (b) PCEE-BERT can achieve different speed-up ratios by adjusting the patience parameter and the confidence threshold. The code for PCEE-BERT can be found at https://github.com/michael-wzhu/PCEE-BERT.

Aspect-based summarization is the task of generating focused summaries based on specific points of interest. Such summaries aid efficient analysis of text, such as quickly understanding reviews or opinions from different angles. However, due to large differences in the type of aspects for different domains (e.g., sentiment, product features), the development of previous models has tended to be domain-specific. In this paper, we propose WikiAsp,1 a large-scale dataset for multi-domain aspect- based summarization that attempts to spur research in the direction of open-domain aspect-based summarization. Specifically, we build the dataset using Wikipedia articles from 20 different domains, using the section titles and boundaries of each article as a proxy for aspect annotation. We propose several straightforward baseline models for this task and conduct experiments on the dataset. Results highlight key challenges that existing summarization models face in this setting, such as proper pronoun handling of quoted sources and consistent explanation of time-sensitive events.

The main purpose of this article is to state the effect of using different methods and models for counterfactual determination and detection of causal knowledge. Nowadays, counterfactual reasoning has been widely used in various fields. In the realm of natural language process(NLP), counterfactual reasoning has huge potential to improve the correctness of a sentence. In the shared Task 5 of detecting counterfactual in SemEval 2020, we pre-process the officially given dataset according to case conversion, extract stem and abbreviation replacement. We use last-5 bidirectional encoder representation from bidirectional encoder representation from transformer (BERT)and term frequency–inverse document frequency (TF-IDF) vectorizer for counterfactual detection. Meanwhile, multi-sample dropout and cross validation are used to improve versatility and prevent problems such as poor generosity caused by overfitting. Finally, our team Ferryman ranked the 8th place in the sub-task 1 of this competition.

Natural language processing (NLP) has been applied to various fields including text classification and sentiment analysis. In the shared task of sentiment analysis of code-mixed tweets, which is a part of the SemEval-2020 competition, we preprocess datasets by replacing emoji and deleting uncommon characters and so on, and then fine-tune the Bidirectional Encoder Representation from Transformers(BERT) to perform the best. After exhausting top3 submissions, Our team MeisterMorxrc achieves an averaged F1 score of 0.730 in this task, and and our codalab username is MeisterMorxrc

Indiscriminately posting offensive remarks on social media may promote the occurrence of negative events such as violence, crime, and hatred. This paper examines different approaches and models for solving offensive tweet classification, which is a part of the OffensEval 2020 competition. The dataset is Offensive Language Identification Dataset (OLID), which draws 14,200 annotated English Tweet comments. The main challenge of data preprocessing is the unbalanced class distribution, abbreviation, and emoji. To overcome these issues, methods such as hashtag segmentation, abbreviation replacement, and emoji replacement have been adopted for data preprocessing approaches. The main task can be divided into three sub-tasks, and are solved by Term Frequency–Inverse Document Frequency(TF-IDF), Bidirectional Encoder Representation from Transformer (BERT), and Multi-dropout respectively. Meanwhile, we applied different learning rates for different languages and tasks based on BERT and non-BERTmodels in order to obtain better results. Our team Ferryman ranked the 18th, 8th, and 21st with F1-score of 0.91152 on the English Sub-task A, Sub-task B, and Sub-task C, respectively. Furthermore, our team also ranked in the top 20 on the Sub-task A of other languages.

Grammatical error diagnosis is an important task in natural language processing. This paper introduces CVTE Character Checking System in the NLP-TEA-4 shared task for CGED 2017, we use Bi-LSTM to generate the probability of every character, then take two kinds of strategies to decide whether a character is correct or not. This system is probably more suitable to deal with the error type of bad word selection, which is one of four types of errors, and the rest are words re-dundancy, words missing and words disorder. Finally the second strategy achieves better F1 score than the first one at all of detection level, identification level, position level.