China

When performing reasoning tasks with user-specific requirements, such as strict output formats, large language models (LLMs) often prioritize reasoning over adherence to detailed instructions. Fine-tuning LLMs on supervised datasets to address this is impractical due to high computational costs and limited parameter access. To tackle this, we propose DICE, a lightweight framework that guides small language models (SLMs) to refine LLMs&#39; outputs through chain-of-thought (CoT) correction. DICE decouples the process by first prompting LLMs to generate natural language responses, then using trained SLMs to analyze and refine these outputs to meet structured output specifications. This framework preserves LLMs&#39; broad knowledge and reasoning capabilities while ensuring the outputs conform to user demands. Specifically, DICE first constructs structured CoT adaptation datasets via a two-stage method and subsequently applies a dual-tuning strategy to fine-tune SLMs for generating structured outputs in an analyze-then-answer pattern. Experiments demonstrate that DICE improves the average format accuracy and content correctness of LLM outputs by 35.4\% and 29.4\%, respectively, achieving state-of-the-art (SOTA) performance over other competitive baselines.

EMNLP 2025

DICE: Structured Reasoning in LLMs through SLM-Guided Chain-of-Thought Correction

structured output

model collaboration

chain-of-thought

reasoning

## 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."<br>

*Message from the General Chair, Dirk Hovy*





<br>

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

When performing reasoning tasks with user-specific requirements, such as strict output formats, large language models (LLMs) often prioritize reasoning over adherence to detailed instructions. Fine-tuning LLMs on supervised datasets to address this is impractical due to high computational costs and limited parameter access. To tackle this, we propose DICE, a lightweight framework that guides small language models (SLMs) to refine LLMs' outputs through chain-of-thought (CoT) correction. DICE decouples the process by first prompting LLMs to generate natural language responses, then using trained SLMs to analyze and refine these outputs to meet structured output specifications. This framework preserves LLMs' broad knowledge and reasoning capabilities while ensuring the outputs conform to user demands. Specifically, DICE first constructs structured CoT adaptation datasets via a two-stage method and subsequently applies a dual-tuning strategy to fine-tune SLMs for generating structured outputs in an analyze-then-answer pattern. Experiments demonstrate that DICE improves the average format accuracy and content correctness of LLM outputs by 35.4\% and 29.4\%, respectively, achieving state-of-the-art (SOTA) performance over other competitive baselines.

poster

The wide use of abbreviated column names (derived from English words or Chinese Pinyin) in database tables poses significant challenges for table-centric tasks in natural language processing and database management. Such a column name expansion task, referred to as the NameGuess task, has previously been addressed by fine-tuning Large Language Models (LLMs) on synthetically generated rule-based data. However, the current approaches yield suboptimal performance due to two fundamental limitations: 1) the rule-generated abbreviation data fails to reflect real-world distribution, and 2) the failure of LLMs to follow the rule-sensitive patterns in NameGuess persistently. For the data realism issue, we propose a novel approach that integrates a subsequence abbreviation generator trained on human-annotated data and collects non-subsequence abbreviations to improve the training set. For the rule violation issue, we propose a decoding system constrained on an automaton that represents the rules of abbreviation expansion. We extended the original English NameGuess test set to include non-subsequence and PinYin scenarios. Experimental results show that properly tuned 7/8B moderate-size LLMs with a refined decoding system can surpass the few-shot performance of state-of-the-art LLMs, such as the GPT-4 series. The code and data are presented in the supplementary material.

Realistic Training Data Generation and Rule Enhanced Decoding in LLM for NameGuess

Large language models have significantly enhanced the capacities and efficiency of text generation. On the one hand, they have improved the quality of text-based *steganography*. On the other hand, they have also underscored the importance of *watermarking* as a safeguard against malicious misuse. In this study, we focus on tokenization inconsistency (TI) between Alice and Bob in steganography and watermarking, where TI can undermine robustness. Our investigation reveals that the problematic tokens responsible for TI exhibit two key characteristics: **infrequency** and **temporariness**. Based on these findings, we propose two tailored solutions for TI elimination: *a stepwise verification* method for steganography and *a post-hoc rollback* method for watermarking. Experiments show that (1) compared to traditional disambiguation methods in steganography, directly addressing TI leads to improvements in fluency, imperceptibility, and anti-steganalysis capacity; (2) for watermarking, addressing TI enhances detectability and robustness against attacks.

Addressing Tokenization Inconsistency in Steganography and Watermarking Based on Large Language Models

Large language models (LLMs) have shown remarkable performance across a range of NLP tasks. However, their strong instruction-following capabilities and inability to distinguish instructions from data content make them vulnerable to indirect prompt injection attacks. In such attacks, instructions with malicious purposes are injected into external data sources, such as web documents. When LLMs retrieve this injected data through tools, such as a search engine and execute the injected instructions, they provide misled responses. Recent attack methods have demonstrated potential, but their abrupt instruction injection often undermines their effectiveness. Motivated by the limitations of existing attack methods, we propose **TopicAttack**, which prompts the LLM to generate a fabricated conversational transition prompt that gradually shifts the topic toward the injected instruction, making the injection smoother and enhancing the plausibility and success of the attack. Through comprehensive experiments, TopicAttack achieves state-of-the-art performance, with an attack success rate (ASR) over 90% in most cases, even when various defense methods are applied. We further analyze its effectiveness by examining attention scores. We find that a higher injected-to-original attention ratio leads to a greater success probability, and our method achieves a much higher ratio than the baseline methods.

TopicAttack: An Indirect Prompt Injection Attack via Topic Transition

Recent progress in large language models (LLMs) has given rise to Large Reasoning Models (LRMs) that externalize multi-step, System 2-style reasoning, achieving state-of-the-art results on complex tasks. However, this explicit reasoning introduces notable computational overhead, while traditional LLMs remain efficient but struggle with tasks demanding deep, stepwise thought. In this work, we systematically study the trade-off between efficiency and robustness inherent in System 1 (intuitive, fast) and System 2 (deliberate, explicit) reasoning in modern language models. Through empirical analysis, we show that enforcing concise reasoning on LRMs improves efficiency but can hinder performance, whereas augmenting LLMs with explicit reasoning traces enhances both confidence and accuracy. Motivated by these insights, we propose a curriculum-based distillation framework that incrementally teaches small models to reason, beginning with concise traces and gradually introducing more complex reasoning. Experiments on challenging mathematical benchmarks demonstrate that our approach enables small models to achieve both strong reasoning ability and inference efficiency. Our findings highlight the importance of dynamic, flexible reasoning strategies and staged learning for building practical, adaptable language models.

Teach Small Models to Reason by Curriculum Distillation

Large Language Models (LLMs) often generate repetitive and monotonous outputs, especially in tasks like story generation, due to limited creative diversity when given the same input prompt. To address this challenge, we propose a novel decoding strategy, ***Avoidance Decoding***, that modifies token logits by penalizing similarity to previously generated outputs, thereby encouraging more diverse multi-branch stories. This penalty adaptively balances two similarity measures: (1) Concept-level Similarity Penalty, which is prioritized in early stages to diversify initial story concepts, and (2) Narrative-level Similarity Penalty, which is increasingly emphasized later to ensure natural yet diverse plot development. Notably, our method achieves up to **2.6** times higher output diversity and reduces repetition by an average of 30% compared to strong baselines, while effectively mitigating text degeneration. Furthermore, we reveal that our method activates a broader range of neurons, demonstrating that it leverages the model's intrinsic creative capacity.

Avoidance Decoding for Diverse Multi-Branch Story Generation

Large Language Models (LLMs) exhibit strong reasoning capabilities in complex tasks. However, they still struggle with hallucinations and factual errors in knowledge-intensive scenarios like knowledge graph question answering (KGQA). We attribute this to the semantic gap between structured knowledge graphs (KGs) and unstructured queries, caused by inherent differences in their focuses and structures. Existing methods usually employ resource-intensive, non-scalable workflows reasoning on vanilla KGs, but overlook this gap. To address this challenge, we propose a flexible framework, Enrich-on-Graph (EoG), which leverages LLMs' prior knowledge to enrich KGs, bridge the semantic gap between graphs and queries. EoG enables efficient evidence extraction from KGs for precise and robust reasoning, while ensuring low computational costs, scalability, and adaptability across different methods. Furthermore, we propose three graph quality evaluation metrics to analyze query-graph alignment in KGQA task, supported by theoretical validation of our optimization objectives. Extensive experiments on two KGQA benchmark datasets indicate that EoG can effectively generate high-quality KGs and achieve the state-of-the-art performance.

Enrich-on-Graph: Query-Graph Alignment for Complex Reasoning with LLM Enriching

Radiology reports are critical for clinical decision-making but often lack a standardized format, limiting both human interpretability and machine learning (ML) applications. While large language models (LLMs) have shown strong capabilities in reformatting clinical text, their high computational requirements, lack of transparency, and data privacy concerns hinder practical deployment. To address these challenges, we explore lightweight encoder-decoder models (<300M parameters)—specifically T5 and BERT2BERT—for structuring radiology reports from the MIMIC-CXR and CheXpert Plus datasets. We benchmark these models against eight open-source LLMs (1B–70B parameters), adapted using prefix prompting, in-context learning (ICL), and low-rank adaptation (LoRA) finetuning. Our best-performing lightweight model outperforms all LLMs adapted using prompt-based techniques on a human-annotated test set. While some LoRA-finetuned LLMs achieve modest gains over the lightweight model on the Findings section (BLEU 6.4%, ROUGE-L 4.8%, BERTScore 3.6%, F1-RadGraph 1.1%, GREEN 3.6%, and F1-SRR-BERT 4.3%), these improvements come at the cost of substantially greater computational resources. For example, LLaMA-3-70B incurred more than 400 times the inference time, cost, and carbon emissions compared to the lightweight model. These results underscore the potential of lightweight, task-specific models as sustainable and privacy-preserving solutions for structuring clinical text in resource-constrained healthcare settings.

Structuring Radiology Reports: Challenging LLMs with Lightweight Models

We present PricingLogic, the first benchmark that probes whether Large Language Models (LLMs) can reliably automate tourism-booking prices when multiple, overlapping fare rules apply. Travel agencies are eager to offload this error-prone task to AI systems; however, deploying LLMs without verified reliability could result in significant financial losses and erode customer trust. PricingLogic comprises 300 natural-language booking requests derived from 42 real-world pricing policies, spanning two levels of difficulty: (i) basic customer-type pricing and (ii) bundled-tour calculations involving interacting discounts. Evaluations of a line of LLMs reveal a steep performance drop on the harder tier, exposing systematic failures in rule interpretation and arithmetic reasoning. These results highlight that, despite their general capabilities, today’s LLMs remain unreliable for revenue-critical applications without further safeguards or domain adaptation.

PricingLogic: Evaluating LLMs Reasoning on Complex Tourism Pricing Tasks

The softmax function is crucial in Transformer attention, which normalizes each row of the attention scores with summation to one. **Usually, tokens with larger attention scores are important for the final prediction. However, the softmax function can face a gradient vanishing issue for such important tokens (e.g., probabilities close to one), leading to optimization difficulties for the important tokens so that the performance may not be better.** In this paper, we propose Self-Adjust Softmax (SA-Softmax) to address this issue by modifying softmax(z) to z cdot softmax(z) and its normalized variant frac(z - min(zmin,0))max(0,zmax)-min(zmin,0) cdot softmax(z). We theoretically show that SA-Softmax provides enhanced gradient properties compared to the vanilla softmax function. Moreover, \methodShortName Attention can be seamlessly integrated into existing Transformer models to their attention mechanisms with minor adjustments. We conducted experiments to evaluate the empirical performance of Transformer models using \methodShortName compared to the vanilla softmax function. These experiments, involving models with up to 2.7 billion parameters, are conducted across diverse datasets, language tasks, and positional encoding methods.

Self-Adjust Softmax

We investigate whether internal activations in language models can be used to detect arithmetic errors. Starting with a controlled setting of 3-digit addition, we show that simple probes can accurately decode both the model’s predicted output and the correct answer from hidden states, regardless of whether the model’s output is correct. Building on this, we train lightweight error detectors that predict model correctness with over 90% accuracy. We then extend our analysis to multi-step arithmetic reasoning in the GSM8K dataset and find that probes trained on simple arithmetic generalize well to this more complex setting, maintaining high accuracy and revealing consistent internal representations. Finally, we demonstrate that these probes can guide selective re-prompting of erroneous reasoning steps, improving task accuracy with minimal disruption to correct outputs. Our findings suggest that arithmetic errors can be anticipated from internal activations alone, and that simple probes offer a viable path toward lightweight model self-correction.

Downloads

Next from EMNLP 2025

Realistic Training Data Generation and Rule Enhanced Decoding in LLM for NameGuess

Similar lecture

Fine-Tuned Thoughts: Leveraging Chain-of-Thought Reasoning for Industrial Asset Health Monitoring