Austria

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.

ACL 2025

The Mirage of Model Editing: Revisiting Evaluation in the Wild

### 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)

workshop paper

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.

Atomic Calibration of LLMs in Long-Form Generations

Large language models (LLMs) have achieved great success, but their occasional content fabrication, or hallucination, limits their practical application. Hallucination arises because LLMs struggle to admit ignorance due to inadequate training on knowledge boundaries. We call it a limitation of LLMs that they can not accurately express their knowledge boundary, answering questions they know while admitting ignorance to questions they do not know. In this paper, we aim to teach LLMs to recognize and express their knowledge boundary, so they can reduce hallucinations caused by fabricating when they do not know. We propose CoKE, which first probes LLMs' knowledge boundary via internal confidence given a set of questions, and then leverages the probing results to elicit the expression of the knowledge boundary. Extensive experiments show CoKE helps LLMs express knowledge boundaries, answering known questions while declining unknown ones, significantly improving in-domain and out-of-domain performance.

Teaching Large Language Models to Express Knowledge Boundary from Their Own Signals

Despite Large Language Models' advances, document-grounded generation still suffers from factual errors. Current evaluations oversimplify error analysis by applying binary judgements, while costly human-annotated datasets contain under-representative error distributions. To address these challenges, we propose a novel framework named SIS-Fact (Systematic, Interpretable and Scalable Factuality Evaluation), which integrates systematic error typologies, synthetic data generation pipelines, and high-quality interpretable annotations for comprehensive factuality evaluation. Specifically, we first develop ten diverse methods to synthesize six error types in grounded generation, including both intrinsic and extrinsic errors. In this way, we develop SIS-Fact Dataset, a high-quality document-grounded factuality evaluation dataset characterized by challenging errors and interpretable error analysis. Based on SIS-Fact Dataset, we introduce SIS-Fact Evaluator, an advanced factuality evaluation model capable of fine-grained analysis and correction. Our extensive experiments show that SIS-Fact Evaluator achieves SOTA performance in SIS-Fact Dataset while maintaining strong generalization across existing multiple factuality benchmarks.

SIS-Fact: Towards Systematic, Interpretable and Scalable Factuality Evaluation for LLM

Large language models (LLMs) encode vast amounts of knowledge during pre-training (parametric knowledge, or PK) and can further be enhanced by incorporating contextual knowledge (CK). Can LLMs effectively integrate their internal PK with external CK to solve complex problems? In this paper, we investigate the dynamic interaction between PK and CK, categorizing their relationships into four types: Supportive, Complementary, Conflicting, and Irrelevant. To support this investigation, we introduce ECHOQA, a benchmark spanning scientific, factual, and commonsense knowledge. Our results show that LLMs tend to suppress their PK when contextual information is available, even when it is complementary or irrelevant. While tailored instructions can encourage LLMs to rely more on their PK, they still struggle to fully leverage it. These findings reveal a key vulnerability in LLMs, raising concerns about their reliability in knowledge-intensive tasks. We will release our code and dataset to facilitate future research.

Understanding the Interplay between Parametric and Contextual Knowledge for Large Language Models

Evaluating foundation models for crystallographic reasoning requires benchmarks that isolate generalization behavior while enforcing physical constraints. This work introduces, xCrysAlloys, a multiscale multicrystal dataset with two physically grounded evaluation protocols to stress-test multimodal generative models. The Spatial-Exclusion benchmark withholds all supercells of a given radius from a diverse dataset, enabling controlled assessments of spatial interpolation and extrapolation. The Compositional-Exclusion benchmark omits all samples of a specific chemical composition, probing generalization across stoichiometries. Nine vision--language foundation models are prompted with crystallographic images and textual context to generate structural annotations. Responses are evaluated via (i) relative errors in lattice parameters and density, (ii) a physics-consistency index penalizing volumetric violations, and (iii) a hallucination score capturing geometric outliers and invalid space-group predictions. These benchmarks establish a reproducible, physically informed framework for assessing generalization, consistency, and reliability in large-scale multimodal models. Dataset and implementation are available at https://github.com/KurbanIntelligenceLab/StressTestingMMFMinCR.

Stress-Testing Multimodal Foundation Models for Crystallographic Reasoning

We present a novel visual instruction tuning strategy to improve the zero-shot task generalization of multimodal large language models by building a firm text-only knowledge base. Existing work lacks sufficient experimentation on the importance of each modality in the instruction tuning stage, often using a majority of vision-language data while keeping text-only data limited and fixing mixtures of modalities. By incorporating diverse text-only data in the visual instruction tuning stage, we vary vision-language data in various controlled experiments to investigate the importance of modality in visual instruction tuning. Our comprehensive evaluation shows that the text-heavy instruction tuning approach is able to perform on par with traditional vision-heavy mixtures on both modalities across 12 general datasets while using as low as half the total training tokens. We find that simply increasing sufficiently diverse text-only data enables transfer of instruction following ability and domain knowledge across modalities while being more efficient than the vision-language approach.

MLAN: Language-Based Instruction Tuning Preserves and Transfers Knowledge in Multimodal Language Models

Artificial intelligence agents when deployed to solve complex problems, need to first decompose the task into smaller manageable sub-tasks, and further associate tools if one is required to solve the sub-task. If the size of the set of tools to chose from is large, a retrieval system is usually employed to narrow down the tool choices before the LLM can proceed with associating tools to the sub-tasks. This paper focuses on the retrieval problem to identify the set of relevant tools to solve a complex task given a large pool of tools to chose from using retrieval augmented generation (RAG) and we refer to it as ToolReAGT. The proposed approach employs ReAct prompting to perform the retrieval in an iterative fashion to first identify if a tool is required and then associate one or more tools for each sub-task. This deviates from conventional RAG where an n-best list of tools are identified given the complex task directly. Experiments are presented on the UltraTool benchmark corpus with 1000 complex tasks and over 2000 tools to select from. A conventional RAG-system is established as baseline and compared to the ToolReAGt approach, resulting in an 8.9% improved retrieval accuracy score recall@5.

ToolReAGt: Tool Retrieval for LLM-based Complex Task Solution via Retrieval Augmented Generation

Large language models (LLMs) often demonstrate strong performance by leveraging implicit knowledge acquired during pretraining. Analogical reasoning, which solves new problems by referencing similar known examples, offers a structured way to utilize this knowledge, but can also lead to subtle factual errors and hallucinations. In this work, we investigate whether LLMs can recognize the reliability of their own analogical outputs using black-box uncertainty estimation (UE). We evaluate six UE metrics across two reasoning-intensive tasks: mathematical problem solving (GSM8K) and code generation (Codeforces). Our results show that Kernel Language Entropy (KLE) and Lexical Similarity (LexSim) are the most robust indicators of correctness. Moreover, while analogical prompting increases model confidence over direct prompting, most uncertainty arises during the analogy transfer step. These findings highlight the limitations of analogical knowledge transfer in LLMs and demonstrate the potential of UE methods for detecting hallucinated reasoning in black-box settings.

Downloads

Next from ACL 2025

Temporal Information Retrieval via Time-Specifier Model Merging

Similar lecture

The Mirage of Model Editing: Revisiting Evaluation in the Wild

.css-70qvj9{display:-webkit-box;display:-webkit-flex;display:-ms-flexbox;display:flex;-webkit-align-items:center;-webkit-box-align:center;-ms-flex-align:center;align-items:center;}Downloads

Next from ACL 2025

Temporal Information Retrieval via Time-Specifier Model Merging

Similar lecture

The Mirage of Model Editing: Revisiting Evaluation in the Wild

Downloads