Embedding models play a pivot role in modern NLP applications such as IR and RAG. While the context limit of LLMs has been pushed beyond 1 million tokens, embedding models are still confined to a narrow context window not exceeding 8k tokens, refrained from application scenarios requiring long inputs such as legal contracts. This paper explores context window extension of existing embedding models, pushing the limit to 32k without requiring additional training. First, we examine the performance of current embedding models for long context retrieval on our newly constructed LongEmbed benchmark. LongEmbed comprises two synthetic tasks and four carefully chosen real-world tasks, featuring documents of varying length and dispersed target information. Benchmarking results underscore huge room for improvement in these models. Based on this, comprehensive experiments show that training-free context window extension strategies like position interpolation can effectively extend the context window of existing embedding models by several folds, regardless of their original context being 512 or beyond 4k. Furthermore, for models employing absolute position encoding (APE), we show the possibility of further fine-tuning to harvest notable performance gains while strictly preserving original behavior for short inputs. For models using rotary position embedding (RoPE), significant enhancements are observed when employing RoPE-specific methods, such as NTK and SelfExtend, indicating RoPE's superiority over APE for context window extension. To facilitate future research, we release E5-Base-4k and E5-RoPE-Base, along with the LongEmbed benchmark.
Recent research has shown that large language models (LLMs) can achieve remarkable translation performance through supervised fine-tuning (SFT) using only a small amount of parallel data. However, SFT simply instructs the model to imitate the reference translations at the token level, making it vulnerable to the noise present in the references. Hence, the assistance from SFT often reaches a plateau once the LLMs have achieved a certain level of translation capability, and further increasing the size of parallel data does not provide additional benefits. To overcome this plateau associated with imitation-based SFT, we propose a preference-based approach built upon the Plackett-Luce model. The objective is to steer LLMs towards a more nuanced understanding of translation preferences from a holistic view, while also being more resilient in the absence of gold translations. We further build a dataset named MAPLE to verify the effectiveness of our approach, which includes multiple translations of varying quality for each source sentence. Extensive experiments demonstrate the superiority of our approach in "breaking the plateau" across diverse LLMs and test settings. Our in-depth analysis underscores the pivotal role of diverse translations and accurate preference scores in the success of our approach.
Robust, faithful and harm-free pronoun use for individuals is an important goal for language models as their use increases, but prior work tends to study only one or two of these components at a time. To measure progress towards the combined goal, we introduce the task of pronoun use fidelity: given a context introducing a co-referring entity and pronoun, the task is to reuse the correct pronoun later, independent of potential distractors. We present a carefully-designed dataset of over 5 million instances to evaluate pronoun use fidelity in English, and we use it to evaluate 37 popular large language models across architectures (encoder-only, decoder-only and encoder-decoder) and scales (11M-70B parameters). We find that while models can mostly faithfully reuse previously-specified pronouns in the presence of no distractors, they are significantly worse at processing she/her/her, singular they and neopronouns. Additionally, models are not robustly faithful to pronouns, as they are easily distracted. With even one additional sentence containing a distractor pronoun, accuracy drops on average by 34%. With 5 distractor sentences, accuracy drops by 52% for decoder-only models and 13% for encoder-only models. We show that widely-used large language models are still brittle, with large gaps in reasoning and in processing different pronouns in a setting that is very simple for humans, and we encourage researchers in bias and reasoning to bridge them.
Coherence evaluation aims to assess the organization and structure of a discourse, which remains challenging even in the era of large language models. Due to the scarcity of annotated data, data augmentation is commonly used for training coherence evaluation models. However, previous augmentations for this task primarily rely on heuristic rules, lacking designing criteria as guidance. In this paper, we take inspiration from linguistic theory of discourse structure, and propose a data augmentation framework named CoUDA. CoUDA breaks down discourse coherence into global and local aspects, and designs augmentation strategies for both aspects, respectively. Especially for local coherence, we propose a novel generative strategy for constructing augmentation samples, which involves post-pretraining a generative model and applying two controlling mechanisms to control the difficulty of generated samples. During inference, CoUDA also jointly evaluates both global and local aspects to comprehensively assess the overall coherence of a discourse. Extensive experiments in coherence evaluation show that, with only 233M parameters, CoUDA achieves state-of-the-art performance in both pointwise scoring and pairwise ranking tasks, even surpassing recent GPT-3.5 and GPT-4 based metrics.
In today's fast-paced world, the growing demand to quickly generate comprehensive and accurate Wikipedia documents for emerging events is both crucial and challenging. However, previous efforts in Wikipedia generation have often fallen short of meeting real-world requirements. Some approaches focus solely on generating segments of a complete Wikipedia document, while others overlook the importance of faithfulness in generation or fail to consider the influence of the pre-training corpus. In this paper, we simulate a real-world scenario where structured full-length Wikipedia documents are generated for emergent events using input retrieved from web sources. To ensure that Large Language Models (LLMs) are not trained on corpora related to recently occurred events, we select events that have taken place recently and introduce a new benchmark Wiki-GenBen, which consists of 309 events paired with their corresponding retrieved web pages for generating evidence. Additionally, we design a comprehensive set of systematic evaluation metrics and baseline methods, to evaluate the capability of LLMs in generating factual full-length Wikipedia documents. The data and code are open-sourced at WikiGenBench.
Continual learning (CL) aims to constantly learn new knowledge over time while avoiding catastrophic forgetting on old tasks. We focus on continual text classification under the class-incremental setting. Recent CL studies have identified the severe performance decrease on analogous classes as a key factor for catastrophic forgetting. In this paper, through an in-depth exploration of the representation learning process in CL, we discover that the compression effect of the information bottleneck leads to confusion on analogous classes. To enable the model learn more sufficient representations, we propose a novel replay-based continual text classification method, InfoCL. Our approach utilizes fast-slow and current-past contrastive learning to perform mutual information maximization and better recover the previously learned representations. In addition, InfoCL incorporates an adversarial memory augmentation strategy to alleviate the overfitting problem of replay. Experimental results demonstrate that InfoCL effectively mitigates forgetting and achieves state-of-the-art performance on three text classification tasks. The code is publicly available at https://github.com/Yifan-Song793/InfoCL.
Large language models (LLMs) have demonstrated strong capabilities in various aspects. However, when applying them to the highly specialized, safe-critical legal domain, it is unclear how much legal knowledge they possess and whether they can reliably perform legal-related tasks. To address this gap, we propose a comprehensive evaluation benchmark LawBench. LawBench has been meticulously crafted to have precise assessment of the LLMs' legal capabilities from three cognitive levels: (1) Legal knowledge memorization: whether LLMs can memorize needed legal concepts, articles and facts; (2) Legal knowledge understanding: whether LLMs can comprehend entities, events and relationships within legal text; (3) Legal knowledge applying: whether LLMs can properly utilize their legal knowledge and make necessary reasoning steps to solve realistic legal tasks. LawBench contains 20 diverse tasks covering 5 task types: single-label classification (SLC), multi-label classification (MLC), regression, extraction and generation. We perform extensive evaluations of 51 LLMs on LawBench, including 20 multilingual LLMs, 22 Chinese-oriented LLMs and 9 legal specific LLMs. The results show that GPT-4 remains the best-performing LLM in the legal domain, surpassing the others by a significant margin. While fine-tuning LLMs on legal specific text brings certain improvements, we are still a long way from obtaining usable and reliable LLMs in legal tasks. All data, model predictions and evaluation code are released in https://github.com/open-compass/LawBench/. We hope this benchmark provides in-depth understanding of the LLMs' domain-specified capabilities and speed up the development of LLMs in the legal domain.
In this paper, we introduce Positional Skip-wisE (PoSE) training for efficient adaptation of large language models~(LLMs) to extremely long context windows. PoSE decouples train length from target context window size by simulating long inputs using a fixed context window with manipulated position indices during training. Concretely, we select several short chunks from a long input sequence, and introduce distinct skipping bias terms to modify the position indices of each chunk. These bias terms, along with the length of each chunk, are altered for each training example, allowing the model to adapt to all positions within the target context window without training on full length inputs. Experiments show that, compared with fine-tuning on the full length, PoSE greatly reduces memory and time overhead with minimal impact on performance. Leveraging this advantage, we have successfully extended the LLaMA model to 128k tokens. Furthermore, we empirically confirm that PoSE is compatible with all RoPE-based LLMs and various position interpolation strategies. Notably, by decoupling fine-tuning length from target context window, PoSE can theoretically extend the context window infinitely, constrained only by memory usage for inference. With ongoing advancements for efficient inference, we believe PoSE holds great promise for scaling the context window even further.
Tool-augmented large language models (LLMs) have achieved remarkable progress in tackling a broad range of queries. However, existing work are still in the experimental stage and has limitations in extensibility and robustness, especially facing the real-world applications. In this paper, we consider a more realistic scenario, connecting LLMs with RESTful APIs, which use the commonly adopted REST software architectural style for web service development. To address the practical challenges of planning and API usage, we introduce RestGPT, which leverages LLMs to solve user requests by connecting with RESTful APIs. Specifically, we propose a coarse-to-fine online planning mechanism to enhance the ability of planning and API selection. For the complex scenario of calling RESTful APIs, we also specially designed an API executor to formulate parameters and parse API responses. Experiments show that RestGPT is able to achieve impressive results in complex tasks and has strong robustness, which paves a new way towards AGI.
We uncover a systematic bias in the evaluation paradigm of adopting large language models~(LLMs), e.g., GPT-4, as a referee to score the quality of responses generated by candidate models. We find that the quality ranking of candidate responses can be easily hacked by simply altering their order of appearance in the context. This manipulation allows us to skew the evaluation result, making one model appear considerably superior to the other, e.g., vicuna could beat ChatGPT on 66 over 80 tested queries. To address this issue, we propose two simple yet effective calibration strategies: 1) Multiple Evidence Calibration, which requires the evaluator model to generate multiple detailed pieces of evidence before assigning ratings; 2) Balanced Position Calibration, which aggregates results across various orders to determine the final score. Extensive experiments demonstrate that our approach successfully mitigates evaluation bias, resulting in closer alignment with human judgments. To facilitate future research on more robust large language model comparison, we integrate the techniques in the paper into an easy-to-use toolkit \emph{FairEval}, along with the human annotations.\footnote{\url{https://github.com/i-Eval/FairEval}}