LLMs, correcting errors or outdated responses while ensuring no negative impacts are created.

In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks without fine-tuning by leveraging contextual information provided within a prompt. However, ICL relies not only on contextual clues but also on the global knowledge acquired during pretraining for the next token prediction. Analyzing this process has been challenging due to the complex computational circuitry of LLMs. This paper investigates the balance between in-context information and pretrained bigram knowledge in token prediction, focusing on the induction head mechanism, a key component in ICL. Leveraging the fact that a two-layer transformer can implement the induction head mechanism with associative memories, we theoretically analyze the logits when a two-layer transformer is given prompts generated by a bigram model. In the experiments, we design specific prompts to evaluate whether the outputs of a two-layer transformer align with the theoretical results.

As large language models (LLMs) are applied across diverse domains, the ability to selectively unlearn specific information has become increasingly essential. For instance, LLMs are expected to provide confidential information to authorized internal users, such as employees or trusted partners, while withholding it from external users, including the general public and unauthorized entities. In response to this challenge, we propose a novel method termed ``in-context knowledge unlearning'', which enables the model to selectively forget information in test-time based on the context of the query. Our method fine-tunes pre-trained LLMs to enable prompt unlearning of target knowledge within the context, while preserving other knowledge. Experiments on the TOFU and AGE datasets using Llama2-7B/13B and Mistral-7B models show our method achieves up to 95% forgetting accuracy while retaining 80% of unrelated knowledge, significantly outperforming baselines in both in-domain and out-of-domain scenarios. Further investigation into the model's internal behavior revealed that while fine-tuned LLMs generate correct predictions in the middle layers and maintain them up to the final layer, they make the decision to forget at the last layer, i.e., ``LLMs pretend to forget''. Our findings offer valuable insights into enhancing the robustness of unlearning mechanisms in LLMs, setting a foundation for future research in the field.

In the current landscape of large language models (LLMs), the process of instruction tuning serves as an essential step. Considering the high computing power overhead, data-efficient instruction tuning was proposed to reduce the training data size in this process, aiming at selecting high-quality instructional data. Nevertheless, we argue that most current data-efficient instruction-tuning methods are highly dependent on the quality of the original instruction-tuning dataset. When it comes to datasets synthesized by LLMs, a common scenario in this field, dirty samples will even be selected with a higher probability than other samples. To address these challenges, we utilized external knowledge (relevant examples or paragraphs) to evaluate those samples synthesized by LLMs with an in-context-based relative predictive entropy. Based on the new metric, we proposed a framework, dubbed as \textbf{RECOST}, which integrates external-knowledge-base re-ranking and diversity-consistent sampling into a single pipeline. Through extensive experiments on several synthetic datasets (Alpaca and Alpaca-gpt4), we demonstrate the effectiveness of our method and achieve even better results with only \textbf{1\%} of the full dataset.