Making Harmful Behaviors Unlearnable for Large Language Models

Large language models (LLMs) have shown great potential as general-purpose AI assistants in various domains. To meet the requirements of different applications, LLMs are often customized by further fine-tuning. However, the powerful learning ability of LLMs not only enables them to acquire new tasks but also makes them susceptible to learning undesired behaviors. For example, even safety-aligned LLMs can be easily fine-tuned into harmful assistants as the fine-tuning data often contains implicit or explicit harmful content. Can we train LLMs on harmful data without learning harmful behaviors? This paper proposes a controllable training framework that makes harmful behaviors unlearnable during the fine-tuning process. Specifically, we introduce ``security vectors'', a few new parameters that can be separated from the LLM, to ensure LLM's responses are consistent with the harmful behavior. Security vectors are activated during fine-tuning, the consistent behavior makes LLM believe that such behavior has already been learned, there is no need to further optimize for harmful data. During inference, we can deactivate security vectors to restore the LLM's normal behavior. The experimental results show that the security vectors generated by 100 harmful samples are enough to prevent LLM from learning 1000 harmful samples, while preserving the ability to learn other useful information.

Tailoring Personality Traits in Large Language Models via Unsupervisedly-Built Personalized Lexicons

Personality plays a pivotal role in shaping human expression patterns, and empowering and manipulating large language models (LLMs) with personality traits holds significant promise in enhancing the user experience of LLMs. However, prior approaches either rely on fine-tuning LLMs on a corpus enriched with personalized expressions or necessitate the manual crafting of prompts to induce LLMs to produce personalized responses. The former approaches demand substantial time and resources for collecting sufficient training examples while the latter might fail in enabling the precise manipulation of the personality traits at a fine-grained level (e.g., achieving high agreeableness while reducing openness). In this study, we introduce a novel approach for tailoring personality traits within LLMs, allowing for the incorporation of any combination of the Big Five factors (i.e., openness, conscientiousness, extraversion, agreeableness, and neuroticism) in a pluggable manner. This is achieved by employing a set of Unsupervisedly-Built Personalized Lexicons (UBPL) that are utilized to adjust the probability of the next token predicted by the original LLMs during the decoding phase. This adjustment encourages the models to generate words present in the personalized lexicons while preserving the naturalness of the generated texts. Extensive experimentation demonstrates the effectiveness of our approach in finely manipulating LLMs' personality traits. Furthermore, our method can be seamlessly integrated into other LLMs without necessitating updates to their parameters.

DISC-FinLLM: A Chinese Financial Large Language Model based on Multiple Experts Fine-tuning

We propose Multiple Experts Fine-tuning Framework to build a financial large language model (LLM), DISC-FinLLM. Our methodology improves general LLMs by endowing them with multi-turn question answering abilities, domain text processing capabilities, mathematical computation skills, and retrieval-enhanced generation capabilities. We build a financial instruction-tuning dataset named DISC-FIN-SFT, including instruction samples of four categories (consulting, NLP tasks, computing and retrieval-augmented generation). Evaluations conducted on multiple benchmarks demonstrate that our model performs better than baseline models in various financial scenarios. Further resources can be found at https://github.com/FudanDISC/DISC-FinLLM.

Unveiling A Core Linguistic Region in Large Language Models

Brain localization, which describes the association between specific regions of the brain and their corresponding functions, is widely accepted in the field of cognitive science as an objective fact. Today's large language models (LLMs) possess human-level linguistic competence and can execute complex tasks requiring abstract knowledge and reasoning. To deeply understand the inherent mechanisms of intelligence emergence in LLMs, this paper conducts an analogical research using brain localization as a prototype. We have discovered a core region in LLMs that corresponds to linguistic competence, accounting for approximately 1% of the total model parameters. This core region exhibits significant dimension dependency, and perturbations to even a single parameter on specific dimensions can lead to a loss of linguistic competence. Furthermore, we observe that an improvement in linguistic competence does not necessarily accompany an elevation in the model's knowledge level, which might imply the existence of regions of domain knowledge that are dissociated from the linguistic region. Overall, exploring the LLMs' functional regions provides insights into the foundation of their intelligence. In the future, we will continue to investigate knowledge regions within LLMs and the interactions between them.

Orthogonal Subspace Learning for Language Model Continual Learning

Benefiting from massive corpora and advanced hardware, large language models (LLMs) exhibit remarkable capabilities in language understanding and generation. However, their performance degrades in scenarios where multiple tasks are encountered sequentially, also known as catastrophic forgetting. In this paper, we propose orthogonal low-rank adaptation (O-LoRA), a simple and efficient approach for continual learning in language models, effectively mitigating catastrophic forgetting while learning new tasks. Specifically, O-LoRA learns tasks in different (low-rank) vector subspaces that are kept orthogonal to each other in order to minimize interference. Our method induces only marginal additional parameter costs and requires no user data storage for replay. Experimental results on continual learning benchmarks show that our method outperforms state-of-the-art methods. Furthermore, compared to previous approaches, our method excels in preserving the generalization ability of LLMs on unseen tasks.

Are Structural Concepts Universal in Transformer Language Models? Towards Interpretable Cross-Lingual Generalization

Large language models (LLMs) have exhibited considerable cross-lingual generalization abilities, whereby they implicitly transfer knowledge across languages. However, the transfer is not equally successful for all languages, especially for low-resource ones, which poses an ongoing challenge. It is unclear whether we have reached the limits of implicit cross-lingual generalization and if explicit knowledge transfer is viable. In this paper, we investigate the potential for explicitly aligning conceptual correspondence between languages to enhance cross-lingual generalization. Using the syntactic aspect of language as a testbed, our analyses of 43 languages reveal a high degree of alignability among the spaces of structural concepts within each language for both encoder-only and decoder-only LLMs. We then propose a meta-learning-based method to learn to align conceptual spaces of different languages, which facilitates zero-shot and few-shot generalization in concept classification and also offers insights into the cross-lingual in-context learning phenomenon. Experiments on syntactic analysis tasks show that our approach achieves competitive results with state-of-the-art methods and narrows the performance gap between languages, particularly benefiting those with limited resources.

Loose lips sink ships: Mitigating Length Bias in Reinforcement Learning from Human Feedback

Reinforcement learning from human feedback serves as a crucial bridge, aligning large language models with human and societal values. This alignment requires a vast corpus of human feedback to learn a reward model, which is subsequently used to finetune language models. However, we have identified that the reward model often finds shortcuts to bypass its intended objectives, misleadingly assuming that humans prefer longer responses. The emergence of length bias often induces the model to favor longer outputs, yet it doesn't equate to an increase in helpful information within these outputs. In this paper, we propose an innovative solution, applying the Product-of-Experts (PoE) technique to separate reward modeling from the influence of sequence length. In our framework, the main expert concentrates on understanding human intents, while the biased expert targets the identification and capture of length bias. To further enhance the learning of bias, we introduce perturbations into the bias-focused expert, disrupting the flow of semantic information. Experimental results validate the effectiveness of our approach, indicating that language model performance is improved, irrespective of sequence length.

Improving Generalization of Alignment with Human Preferences through Group Invariant Learning

The success of AI assistants based on language models (LLMs) hinges crucially on Reinforcement Learning from Human Feedback (RLHF), which enables the generation of responses more aligned with human preferences. As universal AI assistants, there's a growing expectation for them to perform consistently across various domains. However, previous work shows that Reinforcement Learning (RL) often exploits shortcuts to attain high rewards and overlooks challenging samples. This focus on quick reward gains undermines both the stability in training and the model's ability to generalize to new, unseen data. In this work, we propose a novel approach that can learn a consistent policy via RL across various data groups or domains. Given the challenges associated with acquiring group annotations, our method automatically classifies data into different groups, deliberately maximizing performance variance. Then, we optimize the policy to perform well on challenging groups. Lastly, leveraging the established groups, our approach adaptively adjusts the exploration space, allocating more learning capacity to more challenging data and preventing the model from over-optimizing on simpler data. Experimental results indicate that our approach significantly enhances training stability and model generalization.

RealBehavior: A Framework for Faithfully Characterizing Foundation Models' Human-like Behavior Mechanisms

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.

ReForm-Eval: Evaluating Large Vision Language Models via Unified Re-Formulation of Task-Oriented Benchmarks

Recent years have witnessed remarkable progress in the development of large vision-language models (LVLMs). Benefiting from the strong language backbones and efficient cross-modal alignment strategies, LVLMs exhibit surprising capabilities to perceive visual signals and perform visually grounded reasoning. However, the capabilities of LVLMs have not been comprehensively and quantitatively evaluate. Most existing multi-modal benchmarks require task-oriented input-output formats, posing great challenges to automatically assess the free-form text output of LVLMs. To effectively leverage the annotations available in existing benchmarks and reduce the manual effort required for constructing new benchmarks, we propose to re-formulate existing benchmarks into unified LVLM-compatible formats. Through systematic data collection and reformulation, we present the ReForm-Eval benchmark, offering substantial data for evaluating various capabilities of LVLMs. Based on ReForm-Eval, we conduct extensive experiments, thoroughly analyze the strengths and weaknesses of existing LVLMs, and identify the underlying factors. Our benchmark and evaluation framework will be open-sourced as a cornerstone for advancing the development of LVLMs.