The performance of Large Language Models (LLMs) relies heavily on the quality of prompts, which are often manually engineered and task-specific, making them costly and non-scalable. We propose a novel approach, Supervisory Prompt Training (SPT). SPT automates the generation of highly effective prompts using a dual LLM system. In this system, one LLM, the generator, performs a task while the other, the corrector, provides feedback and generates improved prompts. In contrast to earlier techniques, both the generator and corrector collaboratively and continuously improve their prompts over time. We also introduce the concept of \textit{impact scores} to measure the sentence-level effectiveness of the prompts. Our method was tested on four benchmarks, testing the level of hallucinations in LLMs. Notably, we were able to increase the accuracy of GPT-4 on GSM8K from 65.8\% to 94.1\% (28.3\% increase). SPT advances LLMs by refining prompts to enhance performance and reduce hallucinations, offering an efficient and scalable alternative to traditional model fine-tuning.
Large language models (LLMs) such as GPT-4 have exhibited remarkable performance in a variety of tasks, but this strong performance often comes with the high expense of using paid API services. In this paper, we are motivated to study building an LLM cascade to save the cost of using LLMs, particularly for performing reasoning (e.g., mathematical, causal) tasks. Our cascade pipeline follows the intuition that simpler questions can be addressed by a weaker but more affordable LLM, whereas only the challenging questions necessitate the stronger and more expensive LLM. To realize this decision-making, we consider the "answer consistency" of the weaker LLM as a signal of the question difficulty and propose several methods for the answer sampling and consistency checking, including one leveraging a mixture of two thought representations (i.e., Chain-of-Thought and Program-of-Thought). Through experiments on six reasoning benchmark datasets, with GPT-3.5-turbo and GPT-4 being the weaker and stronger LLMs, respectively, we demonstrate that our proposed LLM cascades can achieve performance comparable to using solely the stronger LLM but require only 40% of its cost.
Click-through rate (CTR) prediction is a critical task in online advertising and recommendation systems, as accurate predictions are essential for user targeting and personalized recommendations. Most recent cutting-edge methods primarily focus on investigating complex implicit and explicit feature interactions. However, these methods neglect the issue of false correlations caused by confounding factors or selection bias. This problem is further magnified by the complexity and redundancy of these interactions. We propose a CTR prediction framework that removes false correlation in multi-level feature interaction, termed REFORM. The proposed REFORM framework exploits a wide range of multi-level high-order feature representations via a two-stream stacked recurrent structure while eliminating false correlations. The framework has two key components: I. The multi-level stacked recurrent (MSR) structure enables the model to efficiently capture diverse nonlinear interactions from feature spaces of different levels, and the richer representations lead to enhanced CTR prediction accuracy. II. The false correlation elimination (FCE) module further leverages Laplacian kernel mapping and sample reweighting methods to eliminate false correlations concealed within the multi-level features, allowing the model to focus on the true causal effects. Extensive experiments based on four challenging CTR datasets and our production dataset demonstrate that the proposed REFORM model achieves state-of-the-art performance. Codes, models and our dataset will be released at https://github.com/yansuoyuli/REFORM.
As the ever-increasing token limits of large language models (LLMs) have enabled long context as input, prompting with single data samples might no longer an efficient way. A straightforward strategy improving efficiency is to batch data within the token limit (e.g., 8k for gpt-3.5-turbo; 32k for GPT-4), which we call BatchPrompt. We have two initial observations for prompting with batched data. First, we find that prompting with batched data in longer contexts will inevitably lead to worse performance, compared to single-data prompting. Second, the performance of the language model is significantly correlated with the positions and order of the batched data, due to the corresponding change in decoder context. To retain efficiency and overcome performance loss, we propose Batch Permutation and Ensembling (BPE), and a novel Self-reflection-guided EArly Stopping (SEAS) technique. Our comprehensive experimental evaluation demonstrates that BPE can boost the performance of BatchPrompt with a striking margin on a range of popular NLP tasks, including question answering (Boolq), textual entailment (RTE), and duplicate questions identification (QQP). These performances are even competitive with/higher than single-data prompting(SinglePrompt), while BatchPrompt requires much fewer LLM calls and input tokens (For SinglePrompt v.s. BatchPrompt with batch size 32, using just 9%-16% the number of LLM calls, Boolq accuracy 90.6% to 90.9% with 27.4% tokens, QQP accuracy 87.2% to 88.4% with 18.6% tokens, RTE accuracy 91.5% to 91.1% with 30.8% tokens). To the best of our knowledge, this is the first work to technically improve prompting efficiency of large language models. We hope our simple yet effective approach will shed light on the future research of large language models. The code will be released.
Click-Through Rate (CTR) prediction is the most critical task in product and content recommendation, and learning effective feature interaction is the key challenge to exploiting user preferences for products. Some recent research works focus on investigating more sophisticated feature interactions based on soft attention or gate mechanism, while some redundant or contradictory feature combinations are still introduced. According to Global Workspace Theory in conscious processing, human clicks on advertisements ``consciously'': only a specific subset of product features are considered, and the rest are not involved in conscious processing. Therefore, we propose a CTR model that \textbf{D}irectly \textbf{E}nhances the embeddings and \textbf{L}everages \textbf{T}runcated Conscious \textbf{A}ttention during feature interaction, termed DELTA, which contains two key components: (I) conscious truncation module (CTM), which utilizes curriculum learning to apply adaptive truncation on attention weights to select the most critical feature combinations; (II) direct embedding enhancement module (DEM), which directly and independently propagates gradient from the loss layer to the embedding layer to enhance the crucial embeddings via linear feature crossing without introducing any extra cost during inference. Extensive experiments on five challenging CTR datasets demonstrate that DELTA achieves cutting-edge performance among current state-of-the-art CTR methods.
Large Language Models (LLMs) have limited performance when solving arithmetic reasoning tasks and often provide incorrect answers. Unlike natural language understanding, math problems typically have a single correct answer, making the task of generating accurate solutions more challenging for LLMs. To the best of our knowledge, we are not aware of any LLMs that indicate their level of confidence in their responses which fuels a trust deficit in these models impeding their adoption. To address this deficiency, we propose `MathPrompter', a technique that improves performance of LLMs on arithmetic problems along with increased reliance in the predictions. MathPrompter uses the Zero-shot chain-of-thought prompting technique to generate multiple Algebraic expressions or Python functions to solve the same math problem in different ways and thereby raise the confidence level in the output results. This is in contrast to other prompt based CoT methods, where there is no check on the validity of the intermediate steps followed. Our technique improves over state-of-the-art on the MultiArith dataset ($78.7\%\rightarrow92.5\%$) evaluated using 175B parameter GPT-based LLM.
In this work, we propose a novel complementary learning approach to enhance test-time adaptation (TTA), which has been proven to exhibit good performance on testing data with distribution shifts such as corruptions. In test-time adaptation tasks, information from the source domain is typically unavailable and the model has to be optimized without supervision for test-time samples. Hence, usual methods assign labels for unannotated data with the prediction by a well-trained source model in an unsupervised learning framework. Previous studies have employed unsupervised objectives, such as the entropy of model predictions, as optimization targets to effectively learn features for test-time samples. However, the performance of the model is easily compromised by the quality of pseudo-labels, since inaccuracies in pseudo-labels introduce noise to the model. Therefore, we propose to leverage the "less probable categories" to decrease the risk of incorrect pseudo-labeling. The complementary label is introduced to designate these categories. We highlight that the risk function of complementary labels agrees with their Vanilla loss formula under the conventional true label distribution. Experiments show that the proposed learning algorithm achieves state-of-the-art performance on different datasets and experiment settings.
Current lane detection methods are struggling with the invisibility lane issue caused by heavy shadows, severe road mark degradation, and serious vehicle occlusion. As a result, discriminative lane features can be barely learned by the network despite elaborate designs due to the inherent invisibility of lanes in the wild. In this paper, we target at finding an enhanced feature space where the lane features are distinctive while maintaining a similar distribution of lanes in the wild. To achieve this, we propose a novel Repainting and Imitating Learning (RIL) framework containing a pair of teacher and student without any extra data or extra laborious labeling. Specifically, in the repainting step, an enhanced ideal virtual lane dataset is built in which only the lane regions are repainted while non-lane regions are kept unchanged, maintaining the similar distribution of lanes in the wild. The teacher model learns enhanced discriminative representation based on the virtual data and serves as the guidance for a student model to imitate. In the imitating learning step, through the scale-fusing distillation module, the student network is encouraged to generate features that mimic the teacher model both on the same scale and cross scales. Furthermore, the coupled adversarial module builds the bridge to connect not only teacher and student models but also virtual and real data, adjusting the imitating learning process dynamically. Note that our method introduces no extra time cost during inference and can be plug-and-play in various cutting-edge lane detection networks. Experimental results prove the effectiveness of the RIL framework both on CULane and TuSimple for four modern lane detection methods. The code and model will be available soon.