Supervised fine-tuning (SFT) is a crucial step for large language models (LLMs), enabling them to align with human instructions and enhance their capabilities in downstream tasks. Substantially increasing instruction data is a direct solution to align the model with a broader range of downstream tasks or notably improve its performance on a specific task. However, we find that large-scale increases in instruction data can damage the world knowledge previously stored in LLMs. To address this challenge, we propose LoRAMoE, a novelty framework that introduces several low-rank adapters (LoRA) and integrates them by using a router network, like a plugin version of Mixture of Experts (MoE). It freezes the backbone model and forces a portion of LoRAs to focus on leveraging world knowledge to solve downstream tasks, to alleviate world knowledge forgetting. Experimental results show that, as the instruction data increases, LoRAMoE can significantly improve the ability to process downstream tasks, while maintaining the world knowledge stored in the LLM. Our code is available at https://github.com/Ablustrund/LoRAMoE.

Tool learning is widely acknowledged as a foundational approach or deploying large language models (LLMs) in real-world scenarios. While current research primarily emphasizes leveraging tools to augment LLMs, it frequently neglects emerging safety considerations tied to their application. To fill this gap, we present ToolSword, a comprehensive framework dedicated to meticulously investigating safety issues linked to LLMs in tool learning. Specifically, ToolSword delineates six safety scenarios for LLMs in tool learning, encompassing malicious queries and jailbreak attacks in the input stage, noisy misdirection and risky cues in the execution stage, and harmful feedback and error conflicts in the output stage. Experiments conducted on 11 open-source and closed-source LLMs reveal enduring safety challenges in tool learning, such as handling harmful queries, employing risky tools, and delivering detrimental feedback, which even GPT-4 is susceptible to. Moreover, we conduct further studies with the aim of fostering research on tool learning safety. The data will be released upon acceptance of the paper.

Large language models are meticulously aligned to be both helpful and harmless. However, recent research points to a potential overkill which means models may refuse to answer benign queries. In this paper, we investigate the factors for overkill by exploring how models handle and determine the safety of queries. Our findings reveal the presence of shortcuts within models, leading to excessive attention to harmful words like ‘kill’ and prompts emphasizing safety will exacerbate overkill. Based on these insights, we introduce Self-Contrastive Decoding (Self-CD), a training-free and model-agnostic strategy, to alleviate this phenomenon. We first extract such excessive attention by amplifying the difference in the model’s output distributions when responding to system prompts that either include or omit an emphasis on safety. Then we determine the final next-token predictions by downplaying the excessive attention via contrastive decoding. Empirical results have indicated that our method has achieved an average reduction of the refusal rate by 20 % while having almost no impact on safety.

Adversarial training is one of the best-performing methods in improving the robustness of deep language models. However, robust models come at the cost of high time consumption, as they require multi-step gradient ascents or word substitutions to obtain adversarial samples. In addition, these generated samples are deficient in grammatical quality and semantic consistency, which impairs the effectiveness of adversarial training. To address these problems, we introduce a novel, effective procedure for instead adversarial training with only clean data. Our procedure, distribution shift risk minimization (DSRM), estimates the adversarial loss by perturbing the input data’s probability distribution rather than their embeddings. This formulation results in a robust model that minimizes the expected global loss under adversarial attacks. Our approach requires zero adversarial samples for training and reduces time consumption by up to 70% compared to current best-performing adversarial training methods. Experiments demonstrate that DSRM considerably improves BERT’s resistance to textual adversarial attacks and achieves state-of-the-art robust accuracy on various benchmarks.

Pretrained language models have achieved remarkable success in various natural language processing tasks. However, pretraining has recently shifted toward larger models and larger data, which has resulted in significant computational and energy costs. In this paper, we propose Influence Subset Selection (ISS) for language model, which explicitly utilizes end-task knowledge to select a tiny subset of the pretraining corpus. Specifically, the ISS selects the samples that will provide the most positive influence on the performance of the end task. Furthermore, we design a gradient matching-based influence estimation method, which can drastically reduce the computation time of influence. With only 0.45% of the data and a three-orders-of-magnitude lower computational cost, ISS outperformed pretrained models (e.g., RoBERTa) on eight datasets covering four domains.

Detecting adversarial samples that are carefully crafted to fool the model is a critical step to socially-secure applications. However, existing adversarial detection methods require access to sufficient training data, which brings noteworthy concerns regarding privacy leakage and generalizability. In this work, we validate that the adversarial sample generated by attack algorithms is strongly related to a specific vector in the high-dimensional inputs. Such vectors, namely UAPs (Universal Adversarial Perturbations), can be calculated without original training data. Based on this discovery, we propose a data-agnostic adversarial detection framework, which induces different responses between normal and adversarial samples to UAPs. Experimental results show that our method achieves competitive detection performance on various text classification tasks, and maintains an equivalent time consumption to normal inference.

Reports of human-like behaviors in foundation models are growing, with psychological theories providing enduring tools to investigate these behaviors. However, current research tends to directly apply these human-oriented tools without verifying the faithfulness of their outcomes. In this paper, we introduce a framework, RealBehavior, which is designed to characterize the humanoid behaviors of models faithfully. Beyond simply measuring behaviors, our framework assesses the faithfulness of results based on reproducibility, internal and external consistency, and generalizability. Our findings suggest that a simple application of psychological tools cannot faithfully characterize all human-like behaviors. Moreover, we discuss the impacts of aligning models with human and social values, arguing for the necessity of diversifying alignment objectives to prevent the creation of models with restricted characteristics.

To enhance the multi-step reasoning capabilities of large language models, researchers have extensively explored prompting methods, notably the Chain-of-Thought (CoT) method which explicitly elicits human-like rationales. However, they have inadvertently overlooked the potential of enhancing model reasoning performance by formulating higher-quality problems. In this work, we start from the problem side and propose Self-Polish (SP), a novel method that facilitates the model’s reasoning by guiding it to progressively refine the given problems to be more comprehensible and solvable. We also explore several automatic prompting varients and propose the Self-Polish prompt bank for the community. SP is orthogonal to all other prompting methods of answer/reasoning side like CoT, allowing for seamless integration with state-of-the-art techniques for further improvement. Thorough experiments show that the proposed method attains notable and consistent effectiveness on five reasoning benchmarks across different models. Furthermore, our method also showcases impressive performance on robustness evaluation. Codes and prompts are available at https://github.com/WooooDyy/Self-Polish.

Dataset bias has attracted increasing attention recently for its detrimental effect on the generalization ability of fine-tuned models. The current mainstream solution is designing an additional shallow model to pre-identify biased instances. However, such two-stage methods scale up the computational complexity of training process and obstruct valid feature information while mitigating bias.To address this issue, we utilize the representation normalization method which aims at disentangling the correlations between features of encoded sentences. We find it also promising in eliminating the bias problem by providing isotropic data distribution. We further propose Kernel-Whitening, a Nystrom kernel approximation method to achieve more thorough debiasing on nonlinear spurious correlations. Our framework is end-to-end with similar time consumption to fine-tuning. Experiments show that Kernel-Whitening significantly improves the performance of BERT on out-of-distribution datasets while maintaining in-distribution accuracy.

Natural language understanding (NLU) models tend to rely on spurious correlations (i.e., dataset bias) to achieve high performance on in-distribution datasets but poor performance on out-of-distribution ones. Most of the existing debiasing methods often identify and weaken these samples with biased features (i.e., superficial surface features that cause such spurious correlations). However, down-weighting these samples obstructs the model in learning from the non-biased parts of these samples. To tackle this challenge, in this paper, we propose to eliminate spurious correlations in a fine-grained manner from a feature space perspective. Specifically, we introduce Random Fourier Features and weighted re-sampling to decorrelate the dependencies between features to mitigate spurious correlations. After obtaining decorrelated features, we further design a mutual-information-based method to purify them, which forces the model to learn features that are more relevant to tasks. Extensive experiments on two well-studied NLU tasks demonstrate that our method is superior to other comparative approaches.