Austria

ACL 2025

1-2-3 Check: Enhancing Contextual Privacy in LLM via Multi-Agent Reasoning

workshop paper

### Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Message from the General Chair: 
*It is my great pleasure and honor to welcome you to the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), held in beautiful Vienna, Austria, from July 27 to August 1, 2025. ACL2025continues our field’s tradition of excellence in scholarship, innovation, and inclusivity, and I am deeply grateful to the many volunteers who have worked tirelessly to bring this event to life.* 
[Read more](https://drive.google.com/file/d/1GI_hvOpjswAuYdUTromfeDiPpCcqidwg/view?usp=sharing)

To access this event page, you need to log in with the **email address you registered with**. Access credentials will be sent to your email from Underline - subject line "Welcome to ACL 2025". Please be sure to check your spam email folder if you do not see an email confirmation right away.

Please log in

To access this event page, you need to log in with the **email address you registered with**. 

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

Fine-tuning a general-purpose large language model (LLM) for a specific domain or task has become a routine procedure for ordinary users. However, fine-tuning is known to remove the safety alignment features of the model, even when the fine-tuning data does not contain any harmful content. We consider this to be a critical failure mode of LLMs due to the widespread uptake of fine-tuning, combined with the benign nature of the "attack". Most well-intentioned developers are likely unaware that they are deploying an LLM with reduced safety. On the other hand, this known vulnerability can be easily exploited by malicious actors intending to bypass safety guardrails. To make any meaningful progress in mitigating this issue, we first need reliable and reproducible safety evaluations. In this work, we investigate how robust a safety benchmark is to trivial variations in the experimental procedure, and the stochastic nature of LLMs. Our initial experiments expose surprising variance in the results of the safety evaluation, even when seemingly inconsequential changes are made to the fine-tuning setup. Our observations have serious implications for how researchers in this field should report results to enable meaningful comparisons in the future.

Fine-Tuning Lowers Safety and Disrupts Evaluation Consistency

The increasing capability of large language models (LLMs) to generate synthetic content has heightened concerns about their misuse, driving the development of Machine-Generated Text (MGT) detection models. However, these detectors face significant challenges due to the lack of high-quality synthetic datasets for training. To address this issue, we propose SPADE, a structured framework for detecting synthetic dialogues using prompt-based positive and negative samples. Our proposed methods yield 14 new dialogue datasets, which we benchmark against eight MGT detection models. The results demonstrate improved generalization performance when utilizing a mixed dataset produced by proposed augmentation frameworks, offering a practical approach to enhancing LLM application security. Considering that real-world agents lack knowledge of future opponent utterances, we simulate online dialogue detection and examine the relationship between chat history length and detection accuracy. Our open-source datasets, code and prompts can be downloaded from this https URL.

SPADE: Structured Prompting Augmentation for Dialogue Enhancement in Machine-Generated Text Detection

The introduction of advanced reasoning capabilities have improved the problem-solving performance of large language models, particularly on math and coding benchmarks. However, it remains unclear whether these reasoning models are more or less vulnerable to adversarial prompt attacks than their non-reasoning counterparts. In this work, we present a systematic evaluation of weaknesses in advanced reasoning models compared to similar non-reasoning models across a diverse set of prompt-based attack categories. Using experimental data, we find that on average the reasoning-augmented models are slightly more robust than non-reasoning models (42.51% vs 45.53% attack success rate, lower is better). However, this overall trend masks significant category-specific differences: for certain attack types the reasoning models are substantially more vulnerable (e.g., up to 32 percentage points worse on a tree-of-attacks prompt), while for others they are markedly more robust (e.g., 29.8 points better on cross-site scripting injection). Our findings highlight the nuanced security implications of advanced reasoning in language models and emphasize the importance of stress-testing safety across diverse adversarial techniques.

Weakest Link in the Chain: Security Vulnerabilities in Advanced Reasoning Models

Prompt injection remains a major security risk for large language models. However, the efficacy of existing guardrail models in context-aware settings remains underexplored, as they often rely on static attack benchmarks. Additionally, they have over-defense tendencies. We introduce CAPTURE, a novel context-aware benchmark assessing both attack detection and over-defense tendencies with minimal in-domain examples. Our experiments reveal that current prompt injection guardrail models suffer from high false negatives in adversarial cases and excessive false positives in benign scenarios, highlighting critical limitations. To demonstrate our framework's utility, we train CaptureGuard on our generated data. This new model drastically reduces both false negative and false positive rates on our context-aware datasets while also generalizing effectively to external benchmarks, establishing a path toward more robust and practical prompt injection defenses.

CAPTURE: Context-Aware Prompt Injection Testing and Robustness Enhancement

Safeguarding LLMs requires separating harmful prompts from safe ones. However, current safeguard models often rely on superficial keyword patterns, undermining their robustness. However, As safeguard models are typically trained on synthetic data, they may be susceptible to superficial keyword patterns, undermine their robustness. We frame this reliance as a shortcut learning problem and conduct experiments revealing how existing models depend on specific keywords for classification rather than semantic understanding. Performance evaluations across six safety benchmarks show that models perform well when keyword distributions align but degrade on out-of-distribution prompts. Results from our counterfactual analysis demonstrate that current safeguard models are vulnerable to keyword distribution shifts due to shortcut learning. These findings highlight the importance of addressing shortcut learning to enhance the robustness of safeguard models.

Beyond Words: Multilingual and Multimodal Red Teaming of MLLM Safeguards

Multimodal large language models (MLLMs) are increasingly deployed in real-world applications, yet their safety remains underexplored, particularly in multilingual and visual contexts. In this work, we present a systematic red teaming framework to evaluate MLLM safeguards using adversarial prompts translated into seven languages and delivered via four input modalities: plain text, jailbreak prompt + text, text rendered as an image, and jailbreak prompt + text rendered as an image. We find that rendering prompts as images increases attack success rates and reduces refusal rates, with the effect most pronounced in lower-resource languages such as Slovenian, Czech, and Valencian. Our results suggest that vision-based multilingual attacks expose a persistent gap in current alignment strategies, highlighting the need for robust multilingual and multimodal MLLM safety evaluation and mitigation of these risks. We make our code and data available.

Linguistic Diversity in NLP Security

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.

The Mirage of Model Editing: Revisiting Evaluation in the Wild

The rapid expansion of digital information and knowledge across structured and unstructured sources has heightened the importance of Information Retrieval (IR). While dense retrieval methods have substantially improved semantic matching for general queries, they consistently underperform on queries with explicit temporal constraints--often those containing numerical expressions and time specifiers such as "in 2015." Existing approaches to Temporal Information Retrieval (TIR) improve temporal reasoning but often suffer from catastrophic forgetting, leading to reduced performance on non-temporal queries. To address this, we propose Time-Specifier Model Merging (TSM), a novel method that enhances temporal retrieval while preserving accuracy on non-temporal queries. TSM trains specialized retrievers for individual time specifiers and merges them into a unified model, enabling precise handling of temporal constraints without compromising non-temporal retrieval. Extensive experiments on both temporal and non-temporal datasets demonstrate that TSM significantly improves performance on temporally constrained queries while maintaining strong results on non-temporal queries, consistently outperforming other training methods. Our code is available at https://github.com/seungyoonee/TSM.

Temporal Information Retrieval via Time-Specifier Model Merging

A major bottleneck in exam construction involves designing test items (i.e., questions) that accurately reflect key content from domain-aligned curricular materials. For instance, during formative assessments in vocational education and training (VET), exam designers must generate updated test items that assess student learning progress while covering the full breadth of topics in the curriculum. Large language models (LLMs) can partially support this process, but effective use requires careful prompting and task-specific understanding. We propose a new key point extraction method for retrieval-augmented item generation that enhances the process of generating test items with LLMs. We exhaustively evaluated our method using a TREC-RAG approach, finding that prompting LLMs with key content rather than directly using full curricular text passages significantly improves item quality regarding key information coverage by 8%. To demonstrate these findings, we release EdTec-ItemGen, a retrieval-augmented item generation demo tool to support item generation in education.

EdTec-ItemGen: Enhancing Retrieval-Augmented Item Generation Through Key Point Extraction

Large language models (LLMs) often suffer from hallucinations, posing significant challenges for real-world applications. Confidence calibration, as an effective indicator of hallucination, is thus essential to enhance the trustworthiness of LLMs. Prior work mainly focuses on short-form tasks using a single response-level score (macro calibration), which is insufficient for long-form outputs that may contain both accurate and inaccurate claims. In this work, we systematically study atomic calibration, which evaluates factuality calibration at a fine-grained level by decomposing long responses into atomic claims. We further categorize existing confidence elicitation methods into discriminative and generative types, and propose two new confidence fusion strategies to improve calibration. Our experiments demonstrate that LLMs exhibit poorer calibration at the atomic level during long-form generation. More importantly, atomic calibration uncovers insightful patterns regarding the alignment of confidence methods and the changes of confidence throughout generation. This sheds light on future research directions for confidence estimation in long-form generation.

Premium content

Next from ACL 2025

Fine-Tuning Lowers Safety and Disrupts Evaluation Consistency

Stay up to date with the latest Underline news!

PRESENTATIONS

CONFERENCES

COMPANY

RESOURCES