China

Recent advances in test-time scaling have enabled Large Language Models (LLMs) to display sophisticated reasoning abilities via extended Chain-of-Thought (CoT) generation. Despite their impressive reasoning abilities, Reasoning LLMs (RLMs) frequently display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations. This raises a deeper research question: *How can we represent the reasoning process of RLMs to map their minds?* To address this, we propose a unified graph-based analytical framework for fine-grained modeling and quantitative analysis of RLM reasoning dynamics. Our method first clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps. Through a comprehensive analysis of derived reasoning graphs, we also reveal that key structural properties, such as exploration density, branching, and convergence ratios, strongly correlate with models&#39; performance. The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and also provides practical insights for prompt engineering and cognitive analysis of LLMs. Code and resources will be released to facilitate future research in this direction.

EMNLP 2025

Mapping the Minds of LLMs: A Graph-Based Analysis of Reasoning LLMs

interpretability

reasoning

### Welcome!
The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. Our Virtual Poster Sessions will take place online Thursday, August 22, 2024.

You are required to register for this event. **Please register [here](https://2024.aclweb.org/registration). **

If you have already registered, please check your inbox for an email from Underline granting you access to ACL 2024 content.

Please register!

ACL 2024

The 62nd Annual Meeting of the Association for Computational Linguistics (ACL 2024) will take place in Bangkok, Thailand from August 11th to 16th, 2024. More information will be announced soon.

Recent advances in test-time scaling have enabled Large Language Models (LLMs) to display sophisticated reasoning abilities via extended Chain-of-Thought (CoT) generation. Despite their impressive reasoning abilities, Reasoning LLMs (RLMs) frequently display unstable behaviors, e.g., hallucinating unsupported premises, overthinking simple tasks, and displaying higher sensitivity to prompt variations. This raises a deeper research question: *How can we represent the reasoning process of RLMs to map their minds?* To address this, we propose a unified graph-based analytical framework for fine-grained modeling and quantitative analysis of RLM reasoning dynamics. Our method first clusters long, verbose CoT outputs into semantically coherent reasoning steps, then constructs directed reasoning graphs to capture contextual and logical dependencies among these steps. Through a comprehensive analysis of derived reasoning graphs, we also reveal that key structural properties, such as exploration density, branching, and convergence ratios, strongly correlate with models' performance. The proposed framework enables quantitative evaluation of internal reasoning structure and quality beyond conventional metrics and also provides practical insights for prompt engineering and cognitive analysis of LLMs. Code and resources will be released to facilitate future research in this direction.

poster

## 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*

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

Reasoning-focused large language models (LLMs) are rapidly evolving across various domains, yet their capabilities in handling complex legal problems remains underexplored. In this paper, we introduce Unilaw-R1, a large language model tailored for legal reasoning. With a lightweight 7-billion parameter scale, Unilaw-R1 significantly reduces deployment cost while effectively tackling three core challenges in the legal domain: insufficient legal knowledge, unreliable reasoning logic, and weak business generalization. To enhance its reasoning capability, we first construct Unilaw-R1-Data, a high-quality dataset containing sim17K distilled and screened chain-of-thought (CoT) samples. Based on this, we adopt a two-stage training strategy combining Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL), which significantly boosts the model’s performance on complex legal reasoning tasks and supports interpretable decision-making in legal AI applications. To assess legal reasoning ability, we also introduce Unilaw-R1-Eval, a dedicated benchmark designed to evaluate models across single- and multi-choice legal tasks. Unilaw-R1 demonstrates strong results on authoritative benchmarks, outperforming all models of similar scale and achieving performance on par with the much larger DeepSeek-R1-Distill-Qwen-32B (54.9). Following domain-specific training, it also showed significant gains on LawBench and LexEval, exceeding Qwen-2.5-7B-Instruct (46.6) by an average margin of 6.6 points.

Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference

Tokenization is a crucial step that bridges human-readable text with model-readable discrete tokens. However, recent studies have revealed that tokenizers can be exploited to elicit unwanted model behaviors. In this work, we investigate incomplete tokens, i.e., undecodable tokens with stray bytes resulting from byte-level byte-pair encoding (BPE) tokenization. We hypothesize that such tokens are heavily reliant on their adjacent tokens and are fragile when paired with unfamiliar tokens. To demonstrate this vulnerability, we introduce improbable bigrams: out-of-distribution combinations of incomplete tokens designed to exploit their dependency. Our experiments show that improbable bigrams are significantly prone to hallucinatory behaviors. Surprisingly, the same phrases have drastically lower rates of hallucination (90% reduction in Llama3.1) when an alternative tokenization is used. We caution against the potential vulnerabilities introduced by byte-level BPE tokenizers, which may introduce blind spots to language models.

Improbable Bigrams Expose Vulnerabilities of Incomplete Tokens in Byte-Level Tokenizers

Concept editing aims to control specific concepts in large language models (LLMs) and is an emerging subfield of model editing. Despite the emergence of various editing methods in recent years, there remains a lack of rigorous theoretical analysis and a unified perspective to systematically understand and compare these methods. To address this gap, we propose a unified paradigm for concept editing methods, in which all forms of conceptual injection are aligned at the neuron level. We study four representative concept editing methods: Neuron Editing (NE), Supervised Fine-tuning (SFT), Sparse Autoencoder (SAE), and Steering Vector (SV). Then we categorize them into two classes based on their mode of conceptual information injection: indirect (NE, SFT) and direct (SAE, SV). We evaluate above methods along four dimensions: editing reliability, output generalization, neuron level consistency, and mathematical formalization. Experiments show that SAE achieves the best editing reliability. In output generalization, SAE captures features closer to human-understood concepts, while NE tends to locate text patterns rather than true semantics. Neuron-level analysis reveals that direct methods share high neuron overlap, as do indirect methods, indicating methodological commonality within each category. Our unified paradigm offers a clear framework and valuable insights for advancing interpretability and controlled generation in LLMs.

Towards a Unified Paradigm of Concept Editing in Large Language Models

The quality of data plays a vital role in the development of Large-scale Generative Models. Understanding how important a data point is for a generative model is essential for explaining its behavior and improving the performance. The influence function provides a framework for quantifying the impact of individual training data on model predictions. However, the high computational cost has hindered their applicability in large-scale applications. In this work, we present RRInf, a novel and principled method for estimating influence function in large-scale generative AI models. We show that influence function estimation can be transformed into a ridge regression problem. Based on this insight, we develop an algorithm that is efficient and scalable to large models. Experiments on noisy data detection and influential data identification tasks demonstrate that RRInf outperforms existing methods in terms of both efficiency and effectiveness for three commonly used large models: RoBERTa-large, Llama-2-13B-chat, and stable-diffusion-v1.5.

RRInf: Efficient Influence Function Estimation via Ridge Regression for Large Language Models and Text-to-Image Diffusion Models

Sparse AutoEncoders (SAEs) have recently been employed as a promising unsupervised approach for understanding the representations of layers of Large Language Models (LLMs). However, with the growth in model size and complexity, training SAEs is computationally intensive, as typically one SAE is trained for each model layer. To address such limitation, we propose \textit{Group-SAE}, a novel strategy to train SAEs. Our method considers the similarity of the residual stream representations between contiguous layers to group similar layers and train a single SAE per group. To balance the trade-off between efficiency and performance, we further introduce \textit{AMAD} (Average Maximum Angular Distance), an empirical metric that guides the selection of an optimal number of groups based on representational similarity across layers. Experiments on models from the Pythia family show that our approach significantly accelerates training with minimal impact on reconstruction quality and comparable downstream task performance and interpretability over baseline SAEs trained layer by layer. This method provides an efficient and scalable strategy for training SAEs in modern LLMs.

Group-SAE: Efficient Training of Sparse Autoencoders for Large Language Models via Layer Groups

We introduce Debate Speech Evaluation as a novel and challenging benchmark for assessing LLM judges. Evaluating debate speeches requires a deep understanding of the speech at multiple levels, including argument strength and relevance, the coherence and organization of the speech, the appropriateness of its style and tone, and so on. This task involves a unique set of cognitive abilities that previously received limited attention in systematic LLM benchmarking. To explore such skills, we leverage a dataset of over 600 meticulously annotated debate speeches and present the first in-depth analysis of how state-of-the-art LLMs compare to human judges on this task. Our findings reveal a nuanced picture: while larger models can approximate individual human judgments in some respects, they differ substantially in their overall judgment behavior. We also investigate the ability of frontier LLMs to generate persuasive, opinionated speech, showing that models may perform at a human level on this task. We make our code and data available at https://anonymous.4open.science/r/BenchmarkingLLMAJ-220C.

Debatable Intelligence: Benchmarking LLM Judges via Debate Speech Evaluation

Large language models (LLMs) often acquire knowledge during pretraining that is undesirable in downstream deployments, e.g., sensitive information or copyrighted content. Existing approaches for removing such knowledge rely on fine-tuning, training low-rank adapters or fact-level editing, but these are either too coarse, too shallow, or ineffective. In this work, we propose PISCES, a novel framework for precisely erasing entire concepts from model parameters by directly editing directions that encode them in parameter space. PISCES uses a disentangler model to decompose MLP vectors into interpretable features, identifies those associated with a target concept using automated interpretability techniques, and removes them from model parameters. Experiments on Gemma 2 and Llama 3.1 over various concepts show that PISCES achieves modest gains in efficacy over leading erasure methods, reducing accuracy on the target concept to as low as 7.7%, while dramatically improving erasure specificity (by up to 31%) and robustness (by up to 41%). Overall, these results demonstrate that feature-based in-parameter editing enables a more precise and reliable approach for removing conceptual knowledge in language models.

Precise In-Parameter Concept Erasure in Large Language Models

We introduce and study artificial impressions--patterns in LLMs’ internal representations of prompts that resemble human impressions and stereotypes. We fit linear probes on generated prompts to predict impressions according to the two-dimensional Stereotype Content Model (SCM). Using these probes, we study the relationship between impressions and downstream model behavior as well as prompt features that inform such impressions. We find that LLMs inconsistently report impressions when prompted, but also that impressions are more consistently linearly decodable from their hidden representations. Additionally, we show that artificial impressions of prompts are predictive of the quality and use of hedging in model responses. We also investigate how particular content, stylistic, and use of dialect in prompts impact LLM impressions.

Artificial Impressions: Evaluating Large Language Model Behavior Through the Lens of Trait Impressions

Stance detection is a task aimed at identifying and analyzing the author's stance from text. Previous studies have primarily focused on the text, which may not fully capture the implicit stance conveyed by the author. To address this limitation, we propose a novel approach that transforms original texts into artificially generated images and uses the visual representation to enhance stance detection. Our approach first employs a text-to-image model to generate candidate images for each text. These images are carefully crafted to adhere to three specific criteria: textual relevance, target consistency, and stance consistency. Next, we introduce a comprehensive evaluation framework to select the optimal image for each text from its generated candidates. Subsequently, we introduce a multimodal stance detection model that leverages both the original textual content and the generated image to identify the author's stance. Experiments demonstrate the effectiveness of our approach and highlight the importance of artificially generated images for stance detection.

Exploring Artificial Image Generation for Stance Detection

Modern large language models (LLMs) employ diverse logical inference mechanisms for reasoning, making the strategic optimization of these approaches critical for advancing their capabilities. This paper systematically investigate the **comparative dynamics** of inductive (System 1) versus abductive/deductive (System 2) inference in LLMs. We utilize a controlled analogical reasoning environment, varying modality (textual, visual, symbolic), difficulty, and task format (MCQ / free-text). Our analysis reveals System 2 pipelines generally excel, particularly in visual/symbolic modalities and harder tasks, while System 1 is competitive for textual and easier problems. Crucially, task format significantly influences their relative advantage, with System 1 sometimes outperforming System 2 in free-text rule-execution. These core findings generalize to broader in-context learning. Furthermore, we demonstrate that advanced System 2 strategies like hypothesis selection and iterative refinement can substantially scale LLM reasoning. This study offers foundational insights and actionable guidelines for strategically deploying logical inference to enhance LLM reasoning.

Downloads

Next from EMNLP 2025

Unilaw-R1: A Large Language Model for Legal Reasoning with Reinforcement Learning and Iterative Inference

Similar lecture

GraphReason: Enhancing Reasoning Capabilities of Large Language Models through A Graph-Based Verification Approach