India

Chain-of-Thought (CoT) ‘reasoning’ promises to enhance the performance and transparency of Large Language Models (LLMs). Models, such as Deepseek R1, are trained via reinforcement learning to automatically generate CoT explanations in their outputs. Their faithfulness, i.e. how well the explanations actually reflect their internal reasoning process, has been called into doubt by recent studies (Chen et al., 2025a; Chua and Evans, 2025). This paper extends previous work by probing Deepseek R1 with 445 logical puzzles under zero- and few-shot settings. We find that whilst the model explicitly acknowledges a strong harmful hint in 94.6% of cases, it reports less than 2% of helpful hints. Further analysis reveals implicit unfaithfulness as the model significantly reduces answer-rechecking behaviour for helpful hints (p&lt;0.01) despite rarely mentioning them in its CoT, demonstrating a discrepancy between its reported and actual decision process. In line with prior reports for GPT, Claude, Gemini and other models, our results for DeepSeek raise concerns about the use of CoT as an explainability technique.

IJCNLP-AACL 2025

Examining the Faithfulness of Deepseek R1’s Chain-of-Thought Reasoning

plausible

deepseek

chain-of-thought

transparency

faithfulness

## Welcome!
"I am excited to welcome you to this year’s edition of the Conference on Empirical Methods in Natural Language Processing! Importantly, it marks the 30th edition of EMNLP. With over 8,000 submissions, more than 3,000 accepted papers, and thousands of attendees, we have come a long way from that first
workshop, which had 14 accepted papers. As the field looks ahead, Suzhou is the fitting location for celebrating this milestone: rooted in a long literary tradition, yet modern and forward-looking, and home to a large share of the NLP community." 

*Message from the General Chair, Dirk Hovy*

EMNLP 2025

Celebrate 30 Years of EMNLP! 
EMNLP 2025 will be held in Suzhou, China from November 5th to November 9th, 2025.

Chain-of-Thought (CoT) ‘reasoning’ promises to enhance the performance and transparency of Large Language Models (LLMs). Models, such as Deepseek R1, are trained via reinforcement learning to automatically generate CoT explanations in their outputs. Their faithfulness, i.e. how well the explanations actually reflect their internal reasoning process, has been called into doubt by recent studies (Chen et al., 2025a; Chua and Evans, 2025). This paper extends previous work by probing Deepseek R1 with 445 logical puzzles under zero- and few-shot settings. We find that whilst the model explicitly acknowledges a strong harmful hint in 94.6% of cases, it reports less than 2% of helpful hints. Further analysis reveals implicit unfaithfulness as the model significantly reduces answer-rechecking behaviour for helpful hints (p<0.01) despite rarely mentioning them in its CoT, demonstrating a discrepancy between its reported and actual decision process. In line with prior reports for GPT, Claude, Gemini and other models, our results for DeepSeek raise concerns about the use of CoT as an explainability technique.

workshop paper

### Welcome to IJCNLP-AACL 2025! 
 It is a great honor to host this joint conference in Mumbai, India, from December 20 to 24, 2025. The joint conferences of IJCNLP and AACL are organized with alternating leadership in the Asia-Pacific region. The event is run by the Asian Federation of Natural Language Processing (AFNLP) in odd years, and by AACL in even years, while it is organized solely by ACL when the annual ACL meeting is held in the region. This year, the conference is primarily organized by AFNLP. 
*Kentaro Inui
MBZUAI, UAE
General Chair, IJCNLP-AACL 2025* 
Read full message and download the Conference Handbook [**here**](https://drive.google.com/file/d/1UTwxkAqSqI-GAoJC3wE1zZt5VP1Y8GX0/view?usp=sharing).

The 14th IJCNLP & 4th AACL will be held in Mumbai, India from December 20th to December 24th, 2025.

Retrieval-Augmented Generation (RAG) enhances language models by grounding outputs in retrieved, relevant documents, effectively addressing key limitations such as hallucinations, updated knowledge, and restricted domain expertise. We propose CREAM-RAG, which stabilizes self-reward signals by applying consistency regularization, reducing hallucinations, and improving factual accuracy. Our framework integrates retrieval, DPO-based self-reward reinforcement learning, and consistency loss into a unified system for more reliable output. Experimental results demonstrated significant performance improvements on the LLaMA 2 7B model. It improved the results on average by 35.04% compared to the base model, suggesting that CREAM-RAG successfully enhanced semantic reasoning and decreased hallucinations.

CREAM-RAG: Enhanced Retrieval Augmented Generation to Limit Hallucination through Consistency-based Self-RAG

LLMs are proving to be adept at machine translation although due to their generative nature they may at times overgenerate in various ways. These overgenerations are different from the neurobabble seen in NMT and range from LLM self-explanations, to risky confabulations, to appropriate explanations, where the LLM is able to act as a human translator would- enabling greater comprehension for the target audience. Detecting and determining the exact nature of the overgenerations is a challenging task. We detail different strategies we have explored for our work in a commercial setting, and present our results.

Fabricator or dynamic translator?

We examine evaluation of faithfulness to input data in the context of hotel highlights – brief LLM-generated summaries that capture unique features of accommodations. Through human evaluation campaigns involving categorical error assessment and span-level annotation, we compare traditional metrics, trainable methods, and LLM-as-a-judge approaches. Our findings reveal that simpler metrics like word overlap correlate surprisingly well with human judgments (r=0.63), often outperforming more complex methods when applied to out-of-domain data. We further demonstrate that while LLMs can generate high-quality highlights, they prove unreliable for evaluation as they tend to severely under- or over-annotate. Our analysis of real-world business impacts shows incorrect and non-checkable information pose the greatest risks. We also highlight challenges in crowdsourced evaluations.

Real-World Summarization: When Evaluation Reaches Its Limits

Hallucinations are one of the most pressing challenges for large language models (LLMs). While numerous methods have been proposed to detect and mitigate them automatically, human evaluation continues to serve as the gold standard. However, these human evaluations of hallucinations show substantial variation in definitions, terminology, and evaluation practices. In this paper, we survey 64 studies involving human evaluation of hallucination published between 2019 and 2024, to investigate how hallucinations are currently defined and assessed. Our analysis reveals a lack of consistency in definitions and exposes several concerning methodological shortcomings. Crucial details, such as evaluation guidelines, user interface design, inter-annotator agreement metrics, and annotator demographics, are frequently under-reported or omitted altogether.

Do My Eyes Deceive Me? A Survey of Human Evaluations of Hallucinations in NLG

Public leaderboards for large language models often rely on aggregate scores that conceal critical information about model behavior. In this paper, we present a methodology for task-aware evaluation that combines (i) correctness metrics aligned with task semantics compliance checks for instruction-following and numeric equivalence for mathematics with (ii) pairwise error-overlap analysis to identify complementary model pairs. We apply this methodology to 17 outputs of recent state of the art and frontier LLMs across multiple-choice QA, instruction-following, and mathematical reasoning tasks. We observe that task-aware metrics can reorder model rankings relative to generic lexical metrics, and that error-overlap patterns vary substantially across model pairs and scenarios. We finally conclude by discussing implications for model selection, routing strategies, and LLM-as-judge calibration, and release our analysis pipeline to support further investigation.

Task-Aware Evaluation and Error-Overlap Analysis for Large Language Models

The detection of hallucinations in multilingual scientific text generated by Large Language Models (LLMs) presents significant challenges for reliable AI systems. This paper describes our submission to the SHROOM-CAP 2025 shared task on scientific hallucination detection across 9 languages. Unlike most approaches that focus primarily on model architecture, we adopted a data-centric strategy that addressed the critical issue of training data scarcity and imbalance. We unify and balance five existing datasets to create a comprehensive training corpus of 124,821 samples (50% correct, 50% hallucinated), representing a 172x increase over the original SHROOM training data. Our approach fine-tuned XLM-RoBERTa-Large with 560 million parameters on this enhanced dataset, achieves competitive performance across all languages, including \textbf{2nd place in Gujarati} (zero-shot language) with Factuality F1 of 0.5107, and rankings between 4th-6th place across the remaining 8 languages. Our results demonstrate that systematic data curation can significantly outperform architectural innovations alone, particularly for low-resource languages in zero-shot settings.

“AGI” team at SHROOM-CAP: Data-Centric Approach to Multilingual Hallucination Detection using XLM-RoBERTa

One of the key challenges of deploying Large Language Models (LLMs) in multilingual scenarios is maintaining output quality across two conditions: factual correctness and linguistic fluency. LLMs are liable to produce text with factual hallucinations, solid-sounding but false information, and fluency errors that take the form of grammatical mistakes, repetition, or unnatural speech patterns. In this paper, we address a two-framework solution for the end-to-end quality evaluation of LLM-generated text in low-resource languages. (1) For hallucination detection, we introduce a retrieval-augmented classification model that utilizes hybrid document retrieval, along with gradient boosting.(2) For fluency detection, we introduce a deep learning model that combines engineered statistical features with pre-trained semantic embeddings using an attention-based mechanism.

Scalar_NITK at SHROOM-CAP: Multilingual Factual Hallucination and Fluency Error Detection in Scientific Publications Using Retrieval-Guided Evidence and Attention-Based Feature Fusion

Large Language Models (LLMs) often generate hallucinations, a critical issue in domains like scientific communication where factual accuracy and fluency are essential. The SHROOM-CAP shared task addresses this challenge by evaluating Factual Mistakes and Fluency Mistakes across diverse languages, extending earlier SHROOM editions to the scientific domain. We present Smurfcat, our system for SHROOM-CAP, which integrates three complementary approaches: uncertainty estimation (white-box and black-box signals), encoder-based classifiers (Multilingual Modern BERT), and decoder-based judges (instruction-tuned LLMs with classification heads). Results show that decoder-based judges achieve the strongest overall performance, while uncertainty methods and encoders provide complementary strengths. Our findings highlight the value of combining uncertainty signals with encoder and decoder architectures for robust, multilingual detection of hallucinations and related errors in scientific publications.

SmurfCat at SHROOM-CAP: Factual but Awkward? Fluent but Wrong? Tackling Both in LLM Scientific QA

Emotions are a fundamental aspect of artistic expression. Due to their abstract nature, there is a broad spectrum of emotion realization in artworks. These are subject to historical change and their analysis requires expertise in art history. In this article, we investigate which aspects of emotional expression can be detected by current (2025) vision language models (VLMs). We present a case study of three VLMs (Llava-Llama and two Qwen models) in which we ask these models four sets of questions of increasing complexity about artworks (general content, emotional content, expression of emotions, and emotion symbols) and carry out a qualitative expert evaluation. We find that the VLMs recognize the content of the images surprisingly well and often also which emotions they depict and how they are expressed. The models perform best for concrete images but fail for highly abstract or highly symbolic images. Reliable recognition of symbols remains fundamentally difficult. Furthermore, the models continue to exhibit the well-known LLM weakness of providing inconsistent answers to related questions.

Artwork Interpretation with Vision Language Models: A Case Study on Emotions and Emotion Symbols

Memes are now a common means of communication on social media. Their humor and short format help messages spread quickly and easily. Propagandistic memes use both words and images to influence opinions and behaviors, often appealing to emotions or ideologies. While propaganda detection has been well-studied in high-resource languages (HRLs), there has been a limited focus on low-resource languages (LRLs), such as Bengali. In this study, we introduce MemeGuard, a new dataset of 3,745 memes for detecting propaganda in Bengali. We tested more than 45 different methods, including both single and combined approaches with fusion. For text, BanglaBERT-1 achieved the best macro F1 score of 80.34%, whereas the CLIP vision transformer scored 78.94% for images. The proposed multimodal model,
which combines BanglaBERT-2 and CLIP via Adaptive Modality Fusion, achieved the highest macro-F1 score of 85.36%. This work establishes a strong baseline and offers valuable insights for future research in Bengali multimodal content analysis.

Downloads

Next from IJCNLP-AACL 2025

CREAM-RAG: Enhanced Retrieval Augmented Generation to Limit Hallucination through Consistency-based Self-RAG

Similar lecture

Analysing Chain of Thought Dynamics: Active Guidance or Unfaithful Post-hoc Rationalisation?