As Large Language Models (LLMs) grow increasingly powerful, ensuring their safety and alignment with human values remains a critical challenge.Ideally, LLMs should provide informative responses while avoiding harmful or sensitive disclosures.However, current alignment strategies, reliant on binary refusal (e.g., rejecting prompts or coarse filtering), lack nuance.This leads to overcensorship, failure to detect subtle harm (like withholding public medication information due to misuse concerns), and difficulty with mixed or context-dependent sensitivities, often overcensoring benign content.To overcome these challenges, we introduce HIDDENGUARD, a novel framework for fine-grained, safe generation in LLMs.HIDDENGUARD incorporates PRISM (Representation Router for In-Stream Moderation), which operates alongside the LLM to enable real-time, token-level detection and redaction of harmful content by leveraging intermediate hidden states.This fine-grained approach allows for more nuanced, contextaware moderation, enabling the model to generate informative responses while selectively redacting or replacing sensitive information, rather than outright refusal.We also contribute a comprehensive dataset with token-level finegrained annotations of potentially harmful information across diverse contexts.Our experiments demonstrate that HIDDENGUARD achieves over 90% in F 1 score for detecting and redacting harmful content while preserving the overall utility and informativeness of the model's responses.

Lingrui Mei, Shenghua Liu, Yiwei Wang, Yuyao Ge, Baolong Bi, Jiayu Yao, Jun Wan, Ziling Yin, Jiafeng Guo, Xueqi Cheng. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026.

Baolong Bi, Shenghua Liu, Lingrui Mei, Yiwei Wang, Junfeng Fang, Pengliang Ji, Xueqi Cheng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Yuyao Ge, Shenghua Liu, Baolong Bi, Yiwei Wang, Lingrui Mei, Wenjie Feng, Lizhe Chen, Xueqi Cheng. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

This work explores sequential model editing in large language models (LLMs), a critical task that involves modifying internal knowledge within LLMs continuously through multi-round editing, each incorporating updates or corrections to adjust the model’s outputs without the need for costly retraining. Existing model editing methods, especially those that alter model parameters, typically focus on single-round editing and often face significant challenges in sequential model editing-most notably issues of model forgetting and failure. To address these challenges, we introduce a new model editing method, namely Neuron-level Sequential Editing (NSE), tailored for supporting sequential model editing. Specifically, we optimize the target layer’s hidden states using the model’s original weights to prevent model failure. Furthermore, we iteratively select neurons in multiple layers for editing based on their activation values to mitigate model forgetting. Our empirical experiments demonstrate that NSE significantly outperforms current modifying parameters model editing methods, marking a substantial advancement in the field of sequential model editing. Our code is released on https://anonymous.4open.science/r/NSE-0A8D/.

Large language models (LLMs) exhibit impressive performance across diverse tasks but often struggle to accurately gauge their knowledge boundaries, leading to confident yet incorrect responses.This paper explores leveraging LLMs' internal states to enhance their perception of knowledge boundaries from efficiency and risk perspectives.We investigate whether LLMs can estimate their confidence using internal states before response generation, potentially saving computational resources.Our experiments on datasets like Natural Questions, HotpotQA, and MMLU reveal that LLMs demonstrate significant pre-generation perception, which is further refined post-generation, with perception gaps remaining stable across varying conditions.To mitigate risks in critical domains, we introduce Consistency-based Confidence Calibration (C 3 ), which assesses confidence consistency through question reformulation.C 3 significantly improves LLMs' ability to recognize their knowledge gaps, enhancing the unknown perception rate by 5.6% on NQ and 4.9% on HotpotQA.Our findings suggest that pre-generation confidence estimation can optimize efficiency, while C 3 effectively controls output risks, advancing the reliability of LLMs in practical applications 1 .

“Jailbreak” is a major safety concern of Large Language Models (LLMs), which occurs when malicious prompts lead LLMs to produce harmful outputs, raising issues about the reliability and safety of LLMs. Therefore, an effective evaluation of jailbreaks is very crucial to develop its mitigation strategies. However, our research reveals that many jailbreaks identified by current evaluations may actually be hallucinations—erroneous outputs that are mistaken for genuine safety breaches. This finding suggests that some perceived vulnerabilities might not represent actual threats, indicating a need for more precise red teaming benchmarks. To address this problem, we propose the Benchmark for reliABilitY and jailBreak haLlUcination Evaluation (BabyBLUE). BabyBLUE introduces a specialized validation framework including various evaluators to enhance existing jailbreak benchmarks, ensuring outputs are useful malicious instructions. Additionally, BabyBLUE presents a new dataset as an augmentation to the existing red teaming benchmarks, specifically addressing hallucinations in jailbreaks, aiming to evaluate the true potential of jailbroken LLM outputs to cause harm to human society.

Molecule generation is key to drug discovery and materials science, enabling the design of novel compounds with specific properties.Large language models (LLMs) can learn to perform a wide range of tasks from just a few examples.However, generating valid molecules using representations like SMILES is challenging for LLMs in few-shot settings.In this work, we explore how LLMs can generate 100% valid molecules.We evaluate whether LLMs can use SELFIES, a representation where every string corresponds to a valid molecule, for valid molecule generation but find that LLMs perform worse with SELFIES than with SMILES.We then examine LLMs' ability to correct invalid SMILES and find their capacity limited.Finally, we introduce SmiSelf, a cross-chemical language framework for invalid SMILES correction.SmiSelf converts invalid SMILES to SELFIES using grammatical rules, leveraging SELFIES' mechanisms to correct the invalid SMILES.Experiments show that SmiSelf ensures 100% validity while preserving molecular characteristics and maintaining or even enhancing performance on other metrics.SmiSelf helps expand LLMs' practical applications in biomedicine and is compatible with all SMILES-based generative models.Code is available at https: //github.com/wentao228/SmiSelf.

Retrieval-Augmented Language Models boost task performance, owing to the retriever that provides external knowledge.Although crucial, the retriever primarily focuses on semantics relevance, which may not always be effective for generation.Thus, utility-based retrieval has emerged as a promising topic, prioritizing passages that provide valid benefits for downstream tasks.However, due to insufficient understanding, capturing passage utility accurately remains unexplored.This work proposes SCARLet, a framework for training utility-based retrievers in RALMs, which incorporates two key factors, multi-task generalization and inter-passage interaction.First, SCAR-Let constructs shared context on which training data for various tasks is synthesized.This mitigates semantic bias from context differences, allowing retrievers to focus on learning task-specific utility and generalize across tasks.Next, SCARLet uses a perturbation-based attribution method to estimate passage-level utility for shared context, which reflects interactions between passages and provides more accurate feedback.We evaluate our approach on ten datasets across various tasks, both indomain and out-of-domain, showing that retrievers trained by SCARLet consistently improve the overall performance of RALMs.

Multimodal Retrieval-Augmented Generation (RAG) systems have become essential in knowledge-intensive and open-domain tasks.As retrieval complexity increases, ensuring the robustness of these systems is critical.However, current RAG models are highly sensitive to the order in which evidence is presented, often resulting in unstable performance and biased reasoning, particularly as the number of retrieved items or modality diversity grows.This raises a central question: How does the position of retrieved evidence affect multimodal RAG performance?To answer this, we present the first comprehensive study of position bias in multimodal RAG systems.Through controlled experiments across text-only, imageonly, and mixed-modality tasks, we observe a consistent U-shaped accuracy curve with respect to evidence position.To quantify this bias, we introduce the Position Sensitivity Index (P SI p ) and develop a visualization framework to trace attention allocation patterns across decoder layers.Our results reveal that multimodal interactions intensify position bias compared to unimodal settings, and that this bias increases logarithmically with retrieval range.These findings offer both theoretical and empirical foundations for position-aware analysis in RAG, highlighting the need for evidence reordering or debiasing strategies to build more reliable and equitable generation systems.Our code and experimental resources are available at https://github.com/Theodyy/ Multimodal-Rag-Position-Bias.

As the modern tools of choice for text understanding and generation, large language models (LLMs) are expected to accurately output answers by leveraging the input context. This requires LLMs to possess both context-faithfulness and factual accuracy. While extensive efforts aim to reduce hallucinations through factuality enhancement methods, they also pose risks of hindering context-faithfulness, as factuality enhancement can lead LLMs to become overly confident in their parametric knowledge, causing them to overlook the relevant input context. In this work, we argue that current factuality enhancement methods can significantly undermine the context-faithfulness of LLMs. We first revisit the current factuality enhancement methods and evaluate their effectiveness in enhancing factual accuracy. Next, we evaluate their performance on knowledge editing tasks to assess the potential impact on context-faithfulness. The experimental results reveal that while these methods may yield inconsistent improvements in factual accuracy, they also cause a more severe decline in context-faithfulness, with the largest decrease reaching a striking 69.7\%. To explain these declines, we analyze the hidden states and logit distributions for the tokens representing new knowledge and parametric knowledge respectively, highlighting the limitations of current approaches. Our finding highlights the complex trade-offs inherent in enhancing LLMs. Therefore, we recommend that more research on LLMs' factuality enhancement make efforts to reduce the sacrifice of context-faithfulness.

Large language models (LLMs) acquire information from pre-training corpora, but their stored knowledge can become inaccurate or outdated over time. Model editing addresses this challenge by modifying model parameters without retraining, and prevalent approaches leverage hypernetworks to generate these parameter updates. However, they face significant challenges in lifelong editing due to their incompatibility with LLM parameters that dynamically change during the editing process. To address this, we observed that hypernetwork-based lifelong editing aligns with reinforcement learning modeling and proposed **RLEdit**, an RL-based editing method. By treating editing losses as rewards and optimizing hypernetwork parameters at the full knowledge sequence level, we enable it to precisely capture LLM changes and generate appropriate parameter updates. Our extensive empirical evaluation across several LLMs demonstrates that RLEdit outperforms existing methods in lifelong editing with superior effectiveness and efficiency, achieving a **59.24%** improvement while requiring only **2.11%** of the time compared to most approaches.

Yilong Xu, Jinhua Gao, Xiaoming Yu, Baolong Bi, Huawei Shen, Xueqi Cheng. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

Large Language Models (LLMs) exhibit remarkable capabilities in storing and retrieving vast amounts of factual knowledge. However, they retain outdated or incorrect information from Web corpora. Since full retraining is costly, locate-and-edit model editing methods offer a feasible alternative. Current methods typically follow a two-stage paradigm: (1) identifying critical layers that store knowledge and (2) updating their parameters to store new knowledge. However, both phases have their inherent limitations. Firstly, layer identification is independent of the knowledge being updated, ignoring the differences in knowledge storage patterns. Secondly, parameter updating suffers from high computational overhead due to gradient descent. To solve these, we propose an Explainable and effiCient model Editing method, termed ECE. Specifically, we integrate LLM explainability into the editing process, enabling the adaptive identification of the crucial neurons. Through clustering similar knowledge, we enable batch optimization in a single gradient step, significantly reducing computational time without compromising effectiveness. Extensive experiments demonstrate that ECE can achieve superior performance, showcasing the potential of explainability-driven editing methods for LLMs. Code is available at https://github.com/tianyuzhangterry/ECE.

The dynamic nature of language, particularly evident in the realm of slang and memes on the Internet, poses serious challenges to the adaptability of Large Language Models (LLMs).Traditionally anchored to static datasets, these models often struggle to keep up with the rapid linguistic evolution characteristic of online communities.This research aims to bridge this gap by enhancing LLMs' comprehension of the evolving new concepts on the Internet, without the high cost of continual retraining.In pursuit of this goal, we introduce SLANG, a benchmark designed to autonomously integrate novel data and assess LLMs' ability to comprehend emerging concepts, alongside FOCUS, an approach uses causal inference to enhance LLMs to understand new phrases and their colloquial context.Our benchmark and approach involves understanding real-world instances of linguistic shifts, serving as contextual beacons, to form more precise and contextually relevant connections between newly emerging expressions and their meanings.The empirical analysis shows that our causal inference-based approach outperforms the baseline methods in terms of precision and relevance in the comprehension of Internet slang and memes. 1