Balancing Enhancement, Harmlessness, and General Capabilities: Enhancing Conversational LLMs with Direct RLHF

In recent advancements in Conversational Large Language Models (LLMs), a concerning trend has emerged, showing that many new base LLMs experience a knowledge reduction in their foundational capabilities following Supervised Fine-Tuning (SFT). This process often leads to issues such as forgetting or a decrease in the base model's abilities. Moreover, fine-tuned models struggle to align with user preferences, inadvertently increasing the generation of toxic outputs when specifically prompted. To overcome these challenges, we adopted an innovative approach by completely bypassing SFT and directly implementing Harmless Reinforcement Learning from Human Feedback (RLHF). Our method not only preserves the base model's general capabilities but also significantly enhances its conversational abilities, while notably reducing the generation of toxic outputs. Our approach holds significant implications for fields that demand a nuanced understanding and generation of responses, such as customer service. We applied this methodology to Mistral, the most popular base model, thereby creating Mistral-Plus. Our validation across 11 general tasks demonstrates that Mistral-Plus outperforms similarly sized open-source base models and their corresponding instruct versions. Importantly, the conversational abilities of Mistral-Plus were significantly improved, indicating a substantial advancement over traditional SFT models in both safety and user preference alignment.

ICE-GRT: Instruction Context Enhancement by Generative Reinforcement based Transformers

The emergence of Large Language Models (LLMs) such as ChatGPT and LLaMA encounter limitations in domain-specific tasks, with these models often lacking depth and accuracy in specialized areas, and exhibiting a decrease in general capabilities when fine-tuned, particularly analysis ability in small sized models. To address these gaps, we introduce ICE-GRT, utilizing Reinforcement Learning from Human Feedback (RLHF) grounded in Proximal Policy Optimization (PPO), demonstrating remarkable ability in in-domain scenarios without compromising general task performance. Our exploration of ICE-GRT highlights its understanding and reasoning ability to not only generate robust answers but also to provide detailed analyses of the reasons behind the answer. This capability marks a significant progression beyond the scope of Supervised Fine-Tuning models. The success of ICE-GRT is dependent on several crucial factors, including Appropriate Data, Reward Size Scaling, KL-Control, Advantage Normalization, etc. The ICE-GRT model exhibits state-of-the-art performance in domain-specific tasks and across 12 general Language tasks against equivalent size and even larger size LLMs, highlighting the effectiveness of our approach. We provide a comprehensive analysis of the ICE-GRT, underscoring the significant advancements it brings to the field of LLM.

A Self-enhancement Approach for Domain-specific Chatbot Training via Knowledge Mining and Digest

Large Language Models (LLMs), despite their great power in language generation, often encounter challenges when dealing with intricate and knowledge-demanding queries in specific domains. This paper introduces a novel approach to enhance LLMs by effectively extracting the relevant knowledge from domain-specific textual sources, and the adaptive training of a chatbot with domain-specific inquiries. Our two-step approach starts from training a knowledge miner, namely LLMiner, which autonomously extracts Question-Answer pairs from relevant documents through a chain-of-thought reasoning process. Subsequently, we blend the mined QA pairs with a conversational dataset to fine-tune the LLM as a chatbot, thereby enriching its domain-specific expertise and conversational capabilities. We also developed a new evaluation benchmark which comprises four domain-specific text corpora and associated human-crafted QA pairs for testing. Our model shows remarkable performance improvement over generally aligned LLM and surpasses domain-adapted models directly fine-tuned on domain corpus. In particular, LLMiner achieves this with minimal human intervention, requiring only 600 seed instances, thereby providing a pathway towards self-improvement of LLMs through model-synthesized training data.

Balancing Specialized and General Skills in LLMs: The Impact of Modern Tuning and Data Strategy

This paper introduces a multifaceted methodology for fine-tuning and evaluating large language models (LLMs) for specialized monetization tasks. The goal is to balance general language proficiency with domain-specific skills. The methodology has three main components: 1) Carefully blending in-domain and general-purpose data during fine-tuning to achieve an optimal balance between general and specialized capabilities; 2) Designing a comprehensive evaluation framework with 45 questions tailored to assess performance on functionally relevant dimensions like reliability, consistency, and business impact; 3) Analyzing how model size and continual training influence metrics to guide efficient resource allocation during fine-tuning. The paper details the design, data collection, analytical techniques, and results validating the proposed frameworks. It aims to provide businesses and researchers with actionable insights on effectively adapting LLMs for specialized contexts. We also intend to make public the comprehensive evaluation framework, which includes the 45 tailored questions and their respective scoring guidelines, to foster transparency and collaboration in adapting LLMs for specialized tasks.

Improving the Generalization Ability in Essay Coherence Evaluation through Monotonic Constraints

Coherence is a crucial aspect of evaluating text readability and can be assessed through two primary factors when evaluating an essay in a scoring scenario. The first factor is logical coherence, characterized by the appropriate use of discourse connectives and the establishment of logical relationships between sentences. The second factor is the appropriateness of punctuation, as inappropriate punctuation can lead to confused sentence structure. To address these concerns, we propose a coherence scoring model consisting of a regression model with two feature extractors: a local coherence discriminative model and a punctuation correction model. We employ gradient-boosting regression trees as the regression model and impose monotonicity constraints on the input features. The results show that our proposed model better generalizes unseen data. The model achieved third place in track 1 of NLPCC 2023 shared task 7. Additionally, we briefly introduce our solution for the remaining tracks, which achieves second place for track 2 and first place for both track 3 and track 4.

GLUECons: A Generic Benchmark for Learning Under Constraints

Recent research has shown that integrating domain knowledge into deep learning architectures is effective -- it helps reduce the amount of required data, improves the accuracy of the models' decisions, and improves the interpretability of models. However, the research community is missing a convened benchmark for systematically evaluating knowledge integration methods. In this work, we create a benchmark that is a collection of nine tasks in the domains of natural language processing and computer vision. In all cases, we model external knowledge as constraints, specify the sources of the constraints for each task, and implement various models that use these constraints. We report the results of these models using a new set of extended evaluation criteria in addition to the task performances for a more in-depth analysis. This effort provides a framework for a more comprehensive and systematic comparison of constraint integration techniques and for identifying related research challenges. It will facilitate further research for alleviating some problems of state-of-the-art neural models.

Cross Aggregation Transformer for Image Restoration

Recently, Transformer architecture has been introduced into image restoration to replace convolution neural network (CNN) with surprising results. Considering the high computational complexity of Transformer with global attention, some methods use the local square window to limit the scope of self-attention. However, these methods lack direct interaction among different windows, which limits the establishment of long-range dependencies. To address the above issue, we propose a new image restoration model, Cross Aggregation Transformer (CAT). The core of our CAT is the Rectangle-Window Self-Attention (Rwin-SA), which utilizes horizontal and vertical rectangle window attention in different heads parallelly to expand the attention area and aggregate the features cross different windows. We also introduce the Axial-Shift operation for different window interactions. Furthermore, we propose the Locality Complementary Module to complement the self-attention mechanism, which incorporates the inductive bias of CNN (e.g., translation invariance and locality) into Transformer, enabling global-local coupling. Extensive experiments demonstrate that our CAT outperforms recent state-of-the-art methods on several image restoration applications. The code and models are available at https://github.com/zhengchen1999/CAT.

Point-Voxel Adaptive Feature Abstraction for Robust Point Cloud Classification

Great progress has been made in point cloud classification with learning-based methods. However, complex scene and sensor inaccuracy in real-world application make point cloud data suffer from corruptions, such as occlusion, noise and outliers. In this work, we propose Point-Voxel based Adaptive (PV-Ada) feature abstraction for robust point cloud classification under various corruptions. Specifically, the proposed framework iteratively voxelize the point cloud and extract point-voxel feature with shared local encoding and Transformer. Then, adaptive max-pooling is proposed to robustly aggregate the point cloud feature for classification. Experiments on ModelNet-C dataset demonstrate that PV-Ada outperforms the state-of-the-art methods. In particular, we rank the $2^{nd}$ place in ModelNet-C classification track of PointCloud-C Challenge 2022, with Overall Accuracy (OA) being 0.865. Code will be available at https://github.com/zhulf0804/PV-Ada.

Dynamic Relevance Graph Network for Knowledge-Aware Question Answering

This work investigates the challenge of learning and reasoning for Commonsense Question Answering given an external source of knowledge in the form of a knowledge graph (KG). We propose a novel graph neural network architecture, called Dynamic Relevance Graph Network (DRGN). DRGN operates on a given KG subgraph based on the question and answers entities and uses the relevance scores between the nodes to establish new edges dynamically for learning node representations in the graph network. This explicit usage of relevance as graph edges has the following advantages, a) the model can exploit the existing relationships, re-scale the node weights, and influence the way the neighborhood nodes' representations are aggregated in the KG subgraph, b) It potentially recovers the missing edges in KG that are needed for reasoning. Moreover, as a byproduct, our model improves handling the negative questions due to considering the relevance between the question node and the graph entities. Our proposed approach shows competitive performance on two QA benchmarks, CommonsenseQA and OpenbookQA, compared to the state-of-the-art published results.

Long-term Causal Effects Estimation via Latent Surrogates Representation Learning

Estimating long-term causal effects based on short-term surrogates is a significant but challenging problem in many real-world applications, e.g., marketing and medicine. Despite its success in certain domains, most existing methods estimate causal effects in an idealistic and simplistic way - ignoring the causal structure among short-term outcomes and treating all of them as surrogates. However, such methods cannot be well applied to real-world scenarios, in which the partially observed surrogates are mixed with their proxies among short-term outcomes. To this end, we develop our flexible method, Laser, to estimate long-term causal effects in the more realistic situation that the surrogates are observed or have observed proxies.Given the indistinguishability between the surrogates and proxies, we utilize identifiable variational auto-encoder (iVAE) to recover the whole valid surrogates on all the surrogates candidates without the need of distinguishing the observed surrogates or the proxies of latent surrogates. With the help of the recovered surrogates, we further devise an unbiased estimation of long-term causal effects. Extensive experimental results on the real-world and semi-synthetic datasets demonstrate the effectiveness of our proposed method.