Code secrets are sensitive assets for software developers, and their leakage poses significant cybersecurity risks. While the rapid development of AI code assistants powered by Code Large Language Models (CLLMs), CLLMs are shown to inadvertently leak such secrets due to a notorious memorization phenomenon. This study first reveals that Byte-Pair Encoding (BPE) tokenization leads to unexpected behavior of secret memorization, which we term as gibberish bias. Specifically, we identified that some secrets are among the easiest for CLLMs to memorize. These secrets yield high character-level entropy, but low token-level entropy. Then, this paper supports the biased claim with numerical data. We identified that the roots of the bias are the token distribution shift between the CLLM training data and the secret data. We further discuss how gibberish bias manifests under the “larger vocabulary” trend. To conclude the paper, we discuss potential mitigation strategies and the broader implications on current tokenizer design.

The proliferation of Large Language Models (LLMs) has saturated social media platforms with hyper-realistic posts, rendering traditional detection methods that rely on low-level artifacts or unimodal statistics increasingly ineffective. In this work, we identify a fundamental semantic distinction: humans tend to complement visual content with additional context, while LLMs predominantly describe the visual information. To capture this, UMPIRE employs an orthogonal semantic decomposition mechanism that disentangles textual embeddings into redundant and complementary components. An adaptive gating module dynamically weighs these components to reflect diverse communicative styles. To enforce the desired geometric structure, we introduce a latent contrastive redundancy regularization loss that encourages LLM-generated content to exhibit high semantic redundancy, while human-written content emphasizes complementarity. Experimental results demonstrate that UMPIRE significantly outperforms state-of-the-art detection methods across multiple datasets, achieving up to a 5.38% improvement in accuracy.

Large language models (LLMs) are prone to capturing biases from training corpus, leading to potential negative social impacts. Existing prompt-based debiasing methods exhibit instability due to their sensitivity to prompt changes, while fine-tuning-based techniques incur substantial computational overhead and catastrophic forgetting. In this paper, we propose FairSteer, a novel inference-time debiasing framework without requiring customized prompt design or model retraining. Motivated by the linear representation hypothesis, our preliminary investigation demonstrates that fairness-related features can be encoded into separable directions in the hidden activation space. FairSteer operates in three steps: biased activation detection, debiasing steering vector (DSV) computation, and dynamic activation steering. Specifically, it first trains a lightweight linear classifier to detect bias signatures in activations, and then computes DSVs as intervention directions derived from small contrastive prompt pairs. Subsequently, it performs debiasing by adjusting activations with DSVs in the inference stage. Comprehensive evaluation with six LLMs demonstrates the superiority of FairSteer across question-answering, counterfactual input evaluation and open-ended text generation tasks. Code will be released.

Generating fair and accurate predictions plays a pivotal role in deploying pre-trained language models (PLMs) in the real world. However, existing debiasing methods may inevitably generate incorrect or nonsensical predictions as they are designed and evaluated to achieve parity across different social groups but leave aside individual commonsense facts, resulting in modified knowledge that elicits unreasonable or undesired predictions. This paper introduces a novel debiasing framework that first identifies the encoding locations of biases within language models and then applies the Fairness-Stamp (FAST). FAST focuses on fine-grained, individual bias mitigation and integrates a lightweight network into PLMs, specifically targeting identified biases while preserving essential knowledge and maintaining factual integrity. We also present BiaScope, a new benchmark comprising datasets and metrics designed to evaluate the retention of commonsense knowledge and the generalization across paraphrased social biases. Our extensive experiments across multiple datasets demonstrate that FAST surpasses state-of-the-art baselines with superior debiasing performance while not compromising the overall model capability for knowledge retention and downstream predictions. This highlights the potential of fine-grained debiasing strategies to achieve fairness in PLMs. Code will be publicly available.