Despite near-perfect results reported in the literature, the effectiveness of model editing in real-world applications remains unclear. To bridge this gap, we introduce QAEdit, a new benchmark aligned with widely used question answering (QA) datasets, and WILD, a task-agnostic evaluation framework designed to better reflect real-world usage of model editing. Our single editing experiments show that current editing methods perform substantially worse than previously reported (38.5% vs. 96.8%). We demonstrate that it stems from issues in the synthetic evaluation practices of prior work. Among them, the most severe is the use of teacher forcing during testing, which leaks both content and length of the ground truth, leading to overestimated performance. Furthermore, we simulate practical deployment by sequential editing, revealing that current approaches fail drastically with only 1000 edits. This work calls for a shift in model editing research toward rigorous evaluation and the development of robust, scalable methods that can reliably update knowledge in LLMs for real-world use.

Text watermarking, which modify tokens to embed watermark, has proven effective in detecting machine-generated texts. Yet its application to low-entropy texts like code and mathematics presents significant challenges. A fair number of tokens in these texts are hardly modifiable without changing the intended meaning, causing statistical measures to falsely indicate the absence of a watermark. Existing research addresses this issue by rely mainly on a limited number of high-entropy tokens, which are considered flexible for modification, and accurately reflecting watermarks. However, their detection accuracy remains suboptimal, as they neglect strong watermark evidences embedded in low entropy tokens modified through watermarking. To overcome this limitation, we introduce Bayes’ Rule derived Watermark Detector (BRWD), which exploit watermark information from every token, by leveraging the posterior probability of watermark’s presence. We theoretically prove the optimality of our method in terms of detection accuracy, and demonstrate its superiority across various datasets, models, and watermark injection strategies. Notably, our method achieves up to 50% and 70% relative improvements in detection accuracy over the best baselines in code generation and math problem-solving tasks, respectively. Our code is available at https://github.com/cczslp/BRWD.

As large language models (LLMs) are applied across diverse domains, the ability to selectively unlearn specific information is becoming increasingly essential. For instance, LLMs are expected to selectively 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.Therefore, we propose a novel method termed ìn-context knowledge unlearning”, which enables the model to selectively forget information in test-time based on the query context.Our method fine-tunes pre-trained LLMs to enable prompt unlearning of target knowledge within the context, while preserving unrelated information. Experiments on TOFU, AGE and RWKU datasets using Llama2-7B/13B and Mistral-7B models demonstrate that our method achieves up to 95% forget accuracy while retaining 80% of unrelated knowledge, significantly outperforming baselines in both in-domain and out-of-domain scenarios.Further investigation of the model’s internal behavior revealed that while fine-tuned LLMs generate correct predictions in the middle layers and preserve them up to the final layer. However, the decision to forget is made only at the last layer, i.e. LLMs pretend to forget”.Our findings offer valuable insight into the improvement of the robustness of the unlearning mechanisms in LLMs, laying a foundation for future research in the field.

While auxiliary information has become a key to enhancing Large Language Models (LLMs), relatively little is known about how LLMs merge these contexts, specifically contexts generated by LLMs and those retrieved from external sources.To investigate this, we formulate a systematic framework to identify whether LLMs’ responses are attributed to either generated or retrieved contexts.To easily trace the origin of the response, we construct datasets with conflicting contexts, i.e., each question is paired with both generated and retrieved contexts, yet only one of them contains the correct answer.Our experiments reveal a significant bias in several LLMs (GPT-4/3.5 and Llama2) to favor generated contexts, even when they provide incorrect information.We further identify two key factors contributing to this bias: i) contexts generated by LLMs typically show greater similarity to the questions, increasing their likelihood of being selected; ii) the segmentation process used in retrieved contexts disrupts their completeness, thereby hindering their full utilization in LLMs.Our analysis enhances the understanding of how LLMs merge diverse contexts, offers valuable insights for advancing current LLM augmentation methods, and highlights the risk of generated misinformation for retrieval-augmented LLMs.

Recent large language models (LLMs) have demonstrated remarkable generalization abilities in mathematics and logical reasoning tasks.Prior research indicates that LLMs pre-trained with programming language data exhibit high mathematical and reasoning abilities; however, this causal relationship has not been rigorously tested. Our research aims to verify which programming languages and features during pre-training affect logical inference performance. Specifically, we pre-trained decoder-based language models from scratch using datasets from ten programming languages (e.g., Python, C, Java) and three natural language datasets (Wikipedia, Fineweb, C4) under identical conditions. Thereafter, we evaluated the trained models in a few-shot in-context learning setting on logical reasoning tasks: FLD and bAbi, which do not require commonsense or world knowledge. The results demonstrate that nearly all models trained with programming languages consistently outperform those trained with natural languages, indicating that programming languages contain factors that elicit logic inference performance. In addition, we found that models trained with programming languages exhibit a better ability to follow instructions compared to those trained with natural languages. Further analysis reveals that the depth of Abstract Syntax Trees representing parsed results of programs also affects logical reasoning performance. These findings will offer insights into the essential elements of pre-training for acquiring the foundational abilities of LLMs.

Despite the success of large language models (LLMs) in natural language generation, much evidence shows that LLMs may produce incorrect or nonsensical text. This limitation highlights the importance of discerning when to trust LLMs, especially in safety-critical domains. Existing methods often express reliability by confidence level, however, their effectiveness is limited by the lack of objective guidance. To address this, we propose CONfidence-Quality-ORDer-preserving alignment approach (CONQORD), which leverages reinforcement learning guided by a tailored dual-component reward function. This function integrates quality reward and order-preserving alignment reward functions. Specifically, the order-preserving reward incentivizes the model to verbalize greater confidence for responses of higher quality to align the order of confidence and quality. Experiments demonstrate that CONQORD significantly improves the alignment performance between confidence and response accuracy, without causing over-cautious. Furthermore, the aligned confidence provided by CONQORD informs when to trust LLMs, and acts as a determinant for initiating the retrieval process of external knowledge. Aligning confidence with response quality ensures more transparent and reliable responses, providing better trustworthiness.

In textual backdoor attacks, attackers insert poisoned samples with triggered inputs and target labels into training datasets to manipulate model behavior, threatening the model’s security and reliability. Current defense methods can generally be categorized into inference-time and training-time ones. The former often requires a part of clean samples to set detection thresholds, which may be hard to obtain in practical application scenarios, while the latter usually requires an additional retraining or unlearning process to get a clean model, significantly increasing training costs. To avoid these drawbacks, we focus on developing a practical defense method before model training without using any clean samples. Our analysis reveals that with the help of a pre-trained language model (PLM), poisoned samples, different from clean ones, exhibit mismatched relationship and shared characteristics. Based on these observations, we further propose a two-stage poison detection strategy solely leveraging insights from PLM before model training. Extensive experiments confirm our approach’s effectiveness, achieving better performance than current leading methods more swiftly. Our code is available at https://github.com/Ascian/PKAD.

While Large Language Models (LLMs) have achieved remarkable performance in many tasks, much about their inner workings remains unclear. In this study, we present novel experimental insights into the resilience of LLMs, particularly GPT-4, when subjected to extensive character-level permutations. To investigate this, we first propose the Scrambled Bench, a suite designed to measure the capacity of LLMs to handle scrambled input, in terms of both recovering scrambled sentences and answering questions given scrambled context. The experimental results indicate that multiple advanced LLMs demonstrate the capability akin to typoglycemia, a phenomenon where humans can understand the meaning of words even when the letters within those words are scrambled, as long as the first and last letters remain in place. More surprisingly, we found that only GPT-4 nearly flawlessly processes inputs with unnatural errors, a task that poses significant challenges for other LLMs and often even for humans. Specifically, GPT-4 can almost perfectly reconstruct the original sentences from scrambled ones, decreasing the edit distance by 95%, even when all letters within each word are entirely scrambled. It is counter-intuitive that LLMs can exhibit such resilience despite severe disruption to input tokenization caused by scrambled text.