Semantic Textual Similarity (STS) constitutes a critical research direction in computational linguistics and serves as a key indicator of the encoding capabilities of embedding models. Driven by advances in pre-trained language models and contrastive learning, leading sentence representation methods have reached an average Spearman’s correlation score of approximately 86 across seven STS benchmarks in SentEval. However, further progress has become increasingly marginal, with no existing method attaining an average score higher than 86.5 on these tasks. This paper conducts an in-depth analysis of this phenomenon and concludes that the upper limit for Spearman’s correlation scores under contrastive learning is 87.5. To transcend this ceiling, we propose an innovative approach termed Pcc-tuning, which employs Pearson’s correlation coefficient as a loss function to refine model performance beyond contrastive learning. Experimental results demonstrate that Pcc-tuning can markedly surpass previous state-of-the-art strategies with only a minimal amount of fine-grained annotated samples.
论文检索
输入标题、作者或关键词，从 2,271 篇学术成果中精准定位
Tools have become a mainstay of LLMs, allowing them to retrieve knowledge not in their weights, to perform tasks on the web, and even to control robots. However, most ontologies and surveys of tool-use have assumed the core challenge for LLMs is choosing the tool. Instead, we introduce a framework for tools more broadly which guides us to explore a model’s ability to detect “silent” tool errors, and reflect on how to plan. This more directly aligns with the increasingly popular use of models as tools. We provide an initial approach to failure recovery with promising results both on a controlled calculator setting and embodied agent planning.
Explicit multi-step reasoning, such as chain-of-thought, is widely adopted in the community to explore the better performance of language models (LMs). We report on the systematic strategy that LMs use in this process.Our controlled experiments reveal that LMs rely more heavily on heuristics, such as lexical overlap, in the earlier stages of reasoning when more steps are required to reach an answer. Conversely, their reliance on heuristics decreases as LMs progress closer to the final answer. This suggests that LMs track only a limited number of future steps and dynamically combine heuristic strategies with rational ones in solving tasks involving multi-step reasoning.
MolTRES: Improving Chemical Language Representation Learning for Molecular Property Prediction
PDF ↗Chemical representation learning has gained increasing interest due to the limited availability of supervised data in fields such as drug and materials design. This interest particularly extends to chemical language representation learning, which involves pre-training Transformers on SMILES sequences - textual descriptors of molecules. Despite its success in molecular property prediction, current practices often lead to overfitting and limited scalability due to early convergence. In this paper, we introduce a novel chemical language representation learning framework, called MolTRES, to address these issues. MolTRES incorporates generator-discriminator training, allowing the model to learn from more challenging examples that require structural understanding. In addition, we enrich molecular representations by transferring knowledge from scientific literature by integrating external materials embedding. Experimental results show that our model outperforms existing state-of-the-art models on popular molecular property prediction tasks.
Large language models (LLMs) are typically fine-tuned on diverse and extensive datasets sourced from various origins to develop a comprehensive range of skills, such as writing, reasoning, chatting, coding, and more. Each skill has unique characteristics, and these datasets are often heterogeneous and imbalanced, making the fine-tuning process highly challenging. Balancing the development of each skill while ensuring the model maintains its overall performance requires sophisticated techniques and careful dataset curation. In this work, we propose a general, model-agnostic, reinforcement learning framework, Mixture-of-Skills (MoS), that learns to optimize data usage automatically during the fine-tuning process. This framework ensures the optimal comprehensive skill development of LLMs by dynamically adjusting the focus on different datasets based on their current learning state. To validate the effectiveness of MoS, we conduct extensive experiments using three diverse LLM backbones on two widely used benchmarks and demonstrate that MoS substantially enhances model performance. Building on the success of MoS, we propose MoSpec, an adaptation for task-specific fine-tuning, which harnesses the utilities of various datasets for a specific purpose. Our work underlines the significance of dataset rebalancing and present MoS as a powerful, general solution for optimizing data usage in the fine-tuning of LLMs for various purposes.
A SMART Mnemonic Sounds like “Glue Tonic”: Mixing LLMs with Student Feedback to Make Mnemonic Learning Stick
PDF ↗Keyword mnemonics are memorable explanations that link new terms to simpler keywords.Prior work generates mnemonics for students, but they do not train models using mnemonics students prefer and aid learning.We build SMART, a mnemonic generator trained on feedback from real students learning new terms.To train SMART, we first fine-tune LLaMA-2 on a curated set of user-written mnemonics.We then use LLM alignment to enhance SMART: we deploy mnemonics generated by SMART in a flashcard app to find preferences on mnemonics students favor.We gather 2684 preferences from 45 students across two types: **expressed** (inferred from ratings) and **observed** (inferred from student learning), yielding three key findings.First, expressed and observed preferences disagree; what students *think* is helpful does not always capture what is *truly* helpful.Second, Bayesian models can synthesize complementary data from multiple preference types into a single effectiveness signal.SMART is tuned via Direct Preference Optimization on this signal, which resolves ties and missing labels in the typical method of pairwise comparisons, augmenting data for LLM output quality gains. Third, mnemonic experts assess SMART as matching GPT-4 at much lower deployment costs, showing the utility of capturing diverse student feedback to align LLMs in education.
The proliferation of Large Language Models (LLMs) has driven considerable interest in fine-tuning them with domain-specific data to create specialized language models. Nevertheless, such domain-specific fine-tuning data often contains contextually sensitive personally identifiable information (PII). Direct fine-tuning LLMs on this data without privacy protection poses a risk of data leakage of sensitive PII during inference time. To address this challenge, we introduce Contextual Privacy Protection Language Models (CPPLM), a novel paradigm for fine-tuning LLMs that effectively injects domain-specific knowledge while safeguarding inference-time data privacy. Our work offers a theoretical analysis for model design and delves into various techniques such as corpus curation, penalty-based unlikelihood in training loss, and instruction-based tuning, etc. Extensive experiments across diverse datasets and scenarios demonstrate the effectiveness of our approaches. In particular, instruction tuning with both positive and negative examples, stands out as a promising method, effectively protecting private data while enhancing the model’s knowledge. Our work underscores the potential for Large Language Models as robust contextual privacy protection learners. The complete code and data for the work can be found at https://github.com/Yijia-Xiao/PPLM.
Flashcard schedulers rely on 1) *student models* to predict the flashcards a student knows; and 2) *teaching policies* to pick which cards to show next via these predictions.Prior student models, however, just use study data like the student’s past responses, ignoring the text on cards. We propose **content-aware scheduling**, the first schedulers exploiting flashcard content.To give the first evidence that such schedulers enhance student learning, we build KARL, a simple but effective content-aware student model employing deep knowledge tracing (DKT), retrieval, and BERT to predict student recall.We train KARL by collecting a new dataset of 123,143 study logs on diverse trivia questions.KARL bests existing student models in AUC and calibration error.To ensure our improved predictions lead to better student learning, we create a novel delta-based teaching policy to deploy KARL online.Based on 32 study paths from 27 users, KARL improves learning efficiency over SOTA, showing KARL’s strength and encouraging researchers to look beyond historical study data to fully capture student abilities.
As the utilization of Large Language Models (LLMs) becomes more widespread, there is a growing demand for their ability to handle more complex and longer external knowledge across various use cases. Most existing evaluations of the open-ended question answering (ODQA) task, which necessitates the use of external knowledge, focus solely on whether the model provides the correct answer. However, even when LLMs answer correctly, they often fail to provide an obvious source for their responses. Therefore, it is necessary to jointly evaluate and verify the correctness of the answers and the appropriateness of grounded evidence in complex external contexts. To address this issue, we examine the phenomenon of discrepancies in abilities across two distinct tasks—QA and evidence selection—when performed simultaneously, from the perspective of task alignment. To verify LLMs’ task alignment, we introduce a verification framework and resources considering both semantic relevancy and structural diversity of the given long context knowledge. Through extensive experiments and detailed analysis, we provide insights into the task misalignment between QA and evidence selection. Our code and resources will be available upon acceptance.
Enhancing Training Data Attribution for Large Language Models with Fitting Error Consideration
PDF ↗The black-box nature of large language models (LLMs) poses challenges in interpreting results, impacting issues such as data intellectual property protection and hallucination tracing. Training data attribution (TDA) methods are considered effective solutions to address these challenges.Most recent TDA methods rely on influence functions, assuming the model achieves minimized empirical risk. However, achieving this criterion is difficult, and sourcing accuracy can be compromised by fitting errors during model training. In this paper, we introduce a novel TDA method called Debias and Denoise Attribution (DDA), which enhances influence functions by addressing fitting errors. Specifically, the debias strategy seeks to improve the performance of influence functions by eliminating the knowledge bias present in the base model before fine-tuning, while the denoise strategy aims to reduce discrepancies in influence scores arising from varying degrees of fitting during the training process through smoothing techniques.Experimental results demonstrate that our method significantly outperforms existing approaches, achieving an averaged AUC of 91.64%. Moreover, DDA exhibits strong generality and scalability across various sources and different-scale models like LLaMA2, QWEN2, and Mistral.
Large language models (LLMs) can solve complex multi-step problems, but little is known about how these computations are implemented internally. Motivated by this, we study how LLMs answer multi-hop queries such as “The spouse of the performer of Imagine is”. These queries require two information extraction steps: a latent one for resolving the first hop (“the performer of Imagine”) into the bridge entity (John Lennon), and another for resolving the second hop (“the spouse of John Lennon”) into the target entity (Yoko Ono). Understanding how the latent step is computed internally is key to understanding the overall computation. By carefully analyzing the internal computations of transformer-based LLMs, we discover that the bridge entity is resolved in the early layers of the model. Then, only after this resolution, the two-hop query is solved in the later layers. Because the second hop commences in later layers, there could be cases where these layers no longer encode the necessary knowledge for correctly predicting the answer. Motivated by this, we propose a novel “back-patching” analysis method whereby a hidden representation from a later layer is patched back to an earlier layer. We find that in up to 66% of previously incorrect cases there exists a back-patch that results in the correct generation of the answer, showing that the later layers indeed sometimes lack the needed functionality. Overall our methods and findings open further opportunities for understanding and improving latent reasoning in transformer-based LLMs.
The deployment of language models brings challenges in generating reliable text, especially when these models are fine-tuned with human preferences. To extract the encoded knowledge in these models without (potentially) biased human labels, unsupervised probing techniques like Contrast-Consistent Search (CCS) have been developed (Burns et al., 2022). However, salient but unrelated features in activation space can mislead these probes (Farquhar et al., 2023). Addressing this, we propose a cluster-normalization method to minimize the impact of such features by clustering and normalizing activations of contrast pairs before applying unsupervised probing techniques. While this approach does not address the issue of distinguishing between latent knowledge and that portrayed by a simulated agent—a major issue in the literature of eliciting latent knowledge (Paul Christiano and Xu, 2021)—it still significantly improves the accuracy of probes in identifying the intended knowledge amidst distractions.
In-context learning (ICL) has demonstrated excellent performance across various downstream NLP tasks, especially when synergized with powerful large language models (LLMs). Existing studies evaluate ICL methods primarily based on downstream task performance. This evaluation protocol overlooks the significant cost associated with the demonstration configuration process, i.e., tuning the demonstration as the ICL prompt. However, in this work, we point out that the evaluation protocol leads to unfair comparisons and potentially biased evaluation, because we surprisingly find the correlation between the configuration costs and task performance. Then we call for a two-dimensional evaluation paradigm that considers both of these aspects, facilitating a fairer comparison.Finally, based on our empirical finding that the optimized demonstration on one language model generalizes across language models of different sizes, we introduce a simple yet efficient strategy that can be applied to any ICL method as a plugin, yielding a better trade-off between the two dimensions according to the proposed evaluation paradigm.
Recent information retrieval (IR) models are pre-trained and instruction-tuned on massive datasets and tasks, enabling them to perform well on a wide range of tasks and potentially generalize to unseen tasks with instructions. However, existing IR benchmarks focus on a limited scope of tasks, making them insufficient for evaluating the latest IR models. In this paper, we propose MAIR (Massive Instructed Retrieval Benchmark), a heterogeneous IR benchmark that includes 126 distinct IR tasks across 6 domains, collected from existing datasets. We benchmark state-of-the-art instruction-tuned text embedding models and re-ranking models. Our experiments reveal that instruction-tuned models generally achieve superior performance compared to non-instruction-tuned models on MAIR Additionally, our results suggest that current instruction-tuned text embedding models and re-ranking models still lack effectiveness in specific long-tail tasks. MAIR is publicly available at https://github.com/sunnweiwei/Mair.
Recently, there has been a growing interest in studying how to construct better code instruction tuning data. However, we observe Code models trained with these datasets exhibit high performance on HumanEval but perform worse on other benchmarks such as LiveCodeBench. Upon further investigation, we find that many datasets suffer from severe data leakage. After cleaning up most of the leaked data, some well-known high-quality datasets perform poorly. This discovery reveals a new challenge: identifying which dataset genuinely qualify as high-quality code instruction data. To address this, we propose an efficient code data pruning strategy for selecting good samples. Our approach is based on three dimensions: instruction complexity, response quality, and instruction diversity. Based on our selected data, we present XCoder, a family of models finetuned from LLaMA3. Our experiments show Xcoder achieves new state-of-the-art performance using fewer training data, which verify the effectiveness of our data strategy. Moreover, we perform a comprehensive analysis on the data composition and find existing code datasets have different characteristics according to their construction methods, which provide new insights for future code LLMs.
Modeling lexical semantics is a challenging task, often suffering from interpretability pitfalls. In this paper, we delve into the generation of dictionary-like sense definitions and explore their utility for modeling word meaning. We fine-tuned two Llama models and include an existing T5-based model in our evaluation. Firstly, we evaluate the quality of the generated definitions on existing English benchmarks, setting new state-of-the-art results for the Definition Generation task. Next, we explore the use of definitions generated by our models as intermediate representations subsequently encoded as sentence embeddings. We evaluate this approach on lexical semantics tasks such as the Word-in-Context, Word Sense Induction, and Lexical Semantic Change, setting new state-of-the-art results in all three tasks when compared to unsupervised baselines.
The massive parameters and computational demands hinder the widespread application of Large Language Models (LLMs). Network pruning provides a practical solution to this problem. However, existing pruning works for LLMs mainly focus on unstructured pruning or necessitate post-pruning fine-tuning. The former relies on special hardware to accelerate computation, while the latter may need substantial computational resources. In this paper, we introduce a retraining-free structured pruning method called SoBP (Structured Optimal Brain Pruning). It leverages global first-order information to select pruning structures, then refines them with a local greedy approach, and finally adopts module-wise reconstruction to mitigate information loss. We assess the effectiveness of SoBP across 14 models from 3 LLM families on 8 distinct datasets. Experimental results demonstrate that SoBP outperforms current state-of-the-art methods.
Current approaches for detecting text reuse do not focus on recontextualization, i.e., how the new context(s) of a reused text differs from its original context(s). In this paper, we propose a novel framework called TRoTR that relies on the notion of topic relatedness for evaluating the diachronic change of context in which text is reused. TRoTR includes two NLP tasks: TRiC and TRaC. TRiC is designed to evaluate the topic relatedness between a pair of recontextualizations. TRaC is designed to evaluate the overall topic variation within a set of recontextualizations. We also provide a curated TRoTR benchmark of biblical text reuse, human-annotated with topic relatedness. The benchmark exhibits an inter-annotator agreement of .811. We evaluate multiple, established SBERT models on the TRoTR tasks and find that they exhibit greater sensitivity to textual similarity than topic relatedness. Our experiments show that fine-tuning these models can mitigate such a kind of sensitivity.
Contribution of Linguistic Typology to Universal Dependency Parsing: An Empirical Investigation
PDF ↗Universal Dependencies (UD) is a global initiative to create a standard annotation for the dependency syntax of human languages. Addressing its deviation from typological principles, this study presents an empirical investigation of a typologically motivated transformation of UD proposed by William Croft. Our findings underscore the significance of the transformations across diverse languages and highlight their advantages and limitations.
Knowledge Graphs (KGs) often suffer from incomplete knowledge, which restricts their utility. Recently, Contrastive Learning (CL) has been introduced to Knowledge Graph Completion (KGC), significantly improving the discriminative capabilities of KGC models and setting new benchmarks in performance. However, existing contrastive methods primarily focus on individual triples, overlooking the broader structural connectivities and topologies of KGs. This narrow focus limits a comprehensive understanding of the graph’s structural knowledge. To address this gap, we propose StructKGC, a novel contrastive learning framework designed to flexibly accommodate the diverse topologies inherent in KGs. We introduce four contrastive tasks specifically tailored to KG data: Vertex-level CL, Neighbor-level CL, Path-level CL, and Relation composition level CL. These tasks are trained synergistically during the fine-tuning of pre-trained language models (PLMs), allowing for a more nuanced capture of subgraph semantics. To validate the effectiveness of our method, we perform a comprehensive set of experiments on several real-world datasets. The experimental results demonstrate that our approach achieves SOTA performance under standard supervised and low-resource settings. Furthermore, we observe that the various structure-aware tasks introduced can mutually reinforce each other, resulting in consistent performance enhancements.