China

This paper investigates the enhancement of reasoning capabilities in language models through token-level multi-model collaboration. Our approach selects the optimal tokens from the next token distributions provided by multiple models to perform autoregressive reasoning. Contrary to the assumption that more models yield better results, we introduce a distribution distance-based dynamic selection strategy (DDS) to optimize the multi-model collaboration process. To address the critical challenge of vocabulary misalignment in multi-model collaboration, we propose the concept of minimal complete semantic units (MCSU), which is simple yet enables multiple language models to achieve natural alignment within the linguistic space. Experimental results across various benchmarks demonstrate the superiority of our method. The codes will be released soon.

EMNLP 2025

Dynamic Collaboration of Multi-Language Models based on Minimal Complete Semantic Units

mcsu

multi-model collaboration

llms

### 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.

technical paper

## 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.

Various studies have attempted to remove sensitive or private knowledge from a language model to prevent its unauthorized exposure. However, prior studies have overlooked the inherent complexity and interconnectedness of knowledge, which requires careful examination. To resolve this problem, we first define a new concept called superficial unlearning, which refers to the phenomenon where an unlearning method either fails to erase the interconnected knowledge it should remove or unintentionally erases irrelevant knowledge. Based on the definition, we introduce a novel benchmark, FaithUn, to analyze and evaluate the faithfulness of unlearning in real-world knowledge QA settings. Furthermore, we propose a novel unlearning method, KLUE, which updates only knowledge-related neurons to achieve faithful unlearning. KLUE leverages a regularized explainability method to localize contextual knowledge neurons, updating only these neurons using carefully selected unforgotten samples. Experimental results demonstrate that existing unlearning methods fail to ensure faithful unlearning, while our method shows significant effectiveness in real-world QA unlearning.

FaithUn: Toward Faithful Forgetting in Language Models by Investigating the Interconnectedness of Knowledge

Despite the remarkable capabilities of large language models, current training paradigms inadvertently foster sycophancy—alignment with user-provided information, regardless of factual accuracy. In this paper, we introduce SMART (Sycophancy Mitigation through Adaptive Reasoning Trajectories), reconceptualizing sycophancy as a reasoning optimization problem rather than an output alignment issue. SMART employs a two-stage approach: (1) Uncertainty-Aware Adaptive Monte Carlo Tree Search (UA-MCTS), which dynamically adjusts exploration based on state-level uncertainty; and (2) progress-based reinforcement learning that distills these improved reasoning patterns into model adaptation. Through extensive experiments, we show that SMART significantly outperforms existing baselines in effectively reducing sycophancy while maintaining performance on out-of-distribution inputs. These findings demonstrate the importance of optimizing internal reasoning processes for developing aligned truthful AI assistant.

Sycophancy Mitigation Through Reinforcement Learning with Uncertainty-Aware Adaptive Reasoning Trajectories

Topic modeling is a powerful unsupervised tool for knowledge discovery. However, existing work struggles with generating limited-quality topics that are uninformative and incoherent, which hindering interpretable insights from managing textual data. In this paper, we improve the original variational autoencoder framework by incorporating contextual and graph information to address the above issues. First, the encoder utilizes topic fusion techniques to combine contextual and bag-of-words information well, and meanwhile exploits the constraints of topic alignment and topic sharpening to generate informative topics. Second, we develop a simple word co-occurrence graph information fusion strategy that efficiently increases topic coherence. On three benchmark datasets, our new framework generates more coherent and diverse topics compared to various baselines, and achieves strong performance on both automatic and manual evaluations.

Neural Topic Modeling via Contextual and Graph Information Fusion

Disagreement detection is a crucial task in natural language processing (NLP), particularly in analyzing online discussions and social media content. Large language models (LLMs) have demonstrated significant advancements across various NLP tasks. However, the performance of LLM in disagreement detection is limited by two issues: *conceptual gap* and *reasoning gap*. In this paper, we propose a novel two-stage framework, Concept Alignment and Reasoning Enhancement (CARE), to tackle the issues. The first stage, Concept Alignment, addresses the gap between expert and model by performing **sub-concept taxonomy extraction**, aligning the model's comprehension with human experts. The second stage, Reasoning Enhancement, improves the model's reasoning capabilities by introducing curriculum learning workflow, which includes **rationale to critique** and **counterfactual to detection** for reducing spurious association. Extensive experiments on disagreement detection task demonstrate the effectiveness of our framework, showing superior performance in zero-shot and supervised learning settings, both within and across domains.

CARE: A Disagreement Detection Framework with Concept Alignment and Reasoning Enhancement

Regular updates are essential for maintaining up-to-date knowledge in large language models (LLMs). However, existing training-based model editing methods often struggle to effectively incorporate new knowledge while preserving unrelated general knowledge. To address this challenge, we propose a novel framework called Geometric Knowledge Editing (GeoEdit). GeoEdit utilizes the geometric relationships of parameter updates from fine-tuning to differentiate between neurons associated with new knowledge updates and those related to general knowledge perturbations. By employing a direction-aware knowledge identification method, we avoid updating neurons with directions approximately orthogonal to existing knowledge, thus preserving the model's generalization ability. For the remaining neurons, we integrate both old and new knowledge for aligned directions and apply a ``forget-then-learn'' editing strategy for opposite directions. Additionally, we introduce an importance-guided task vector fusion technique that filters out redundant information and provides adaptive neuron-level weighting, further enhancing model editing performance. Extensive experiments on two publicly available datasets demonstrate the superiority of GeoEdit over existing state-of-the-art methods.

GeoEdit: Geometric Knowledge Editing for Large Language Models

Large language models (LLMs) achieve impressive results on advanced mathematics benchmarks but sometimes fail on basic arithmetic tasks, raising the question of whether they have truly grasped fundamental arithmetic rules or are merely relying on pattern matching. To unravel this issue, we systematically probe LLMs’ understanding of two-integer addition (0 to 2⁶⁴) by testing three crucial properties: commutativity (A+B=B+A), representation invariance via symbolic remapping (e.g., 7mapsto Y), and consistent accuracy scaling with operand length. Our evaluation of 12 leading LLMs reveals a stark disconnect: while models achieve high numeric accuracy (73.8–99.8\%), they systematically fail these diagnostics. Specifically, accuracy plummets to le 7.5\% with symbolic inputs, commutativity is violated in up to 20\% of cases, and accuracy scaling is non-monotonic. Interventions further expose this pattern-matching reliance: explicitly providing rules degrades performance by 29.49\%, while prompting for explanations before answering merely maintains baseline accuracy. These findings demonstrate that current LLMs address elementary addition via pattern matching, not robust rule induction, motivating new diagnostic benchmarks and innovations in model architecture and training to cultivate genuine mathematical reasoning.

Do Large Language Models Truly Grasp Addition? A Rule-Focused Diagnostic Using Two-Integer Arithmetic

Recent studies on hallucination in large language models (LLMs) have been actively progressing in natural language processing. However, the impact of negated text on hallucination with LLMs remains largely unexplored. In this paper, we set three important yet unanswered research questions and aim to address them. To derive the answers, we investigate whether LLMs can recognize contextual shifts caused by negation and still reliably distinguish hallucinations comparable to affirmative cases. We also design the NegHalu dataset by reconstructing existing hallucination detection datasets with negated expressions. Our experiments demonstrate that LLMs struggle to detect hallucinations in negated text effectively, often producing logically inconsistent or unfaithful judgments. Moreover, we trace the internal state of LLMs as they process negated inputs at the token level and reveal the challenges of mitigating their unintended effects.

The Impact of Negated Text on Hallucination with Large Language Models

Large Language Models (LLMs) have demonstrated the capability to refine their generated answers through self-correction, enabling continuous performance improvement over multiple rounds. However, the mechanisms underlying how and why accuracy evolves during this iterative process remain unexplored. To fill this gap, we propose a probabilistic theory to model the dynamics of accuracy change and explain the performance improvements observed in multi-round self-correction. Through mathematical derivation, we establish that the accuracy after the tth round of self-correction is given by: Acct = Upp - αt(Upp - Acc₀), where Acc₀ denotes the initial accuracy, Upp represents the upper bound of accuracy convergence, and α determines the rate of convergence. Based on our theory, these parameters can be calculated and the predicted accuracy curve then can be obtained through only a single round of self-correction. Extensive experiments across diverse models and datasets demonstrate that our theoretical predictions align closely with empirical accuracy curves, validating the effectiveness of the theory. Our work provides a theoretical foundation for understanding LLM self-correction, thus paving the way for further explorations.

A Probabilistic Inference Scaling Theory for LLM Self-Correction

Final-answer-based metrics are commonly used for evaluating large language models (LLMs) on math word problems, often taken as proxies for reasoning ability. However, such metrics conflate two distinct sub-skills: abstract formulation (capturing mathematical relationships using expressions) and arithmetic computation (executing the calculations). Through a disentangled evaluation on GSM8K and SVAMP, we find that the final-answer accuracy of Llama-3 and Qwen2.5 (1B-32B) without CoT is overwhelmingly bottlenecked by the arithmetic computation step and not by the abstract formulation step. Contrary to the common belief, we show that CoT primarily aids in computation, with limited impact on abstract formulation. Mechanistically, we show that these two skills are composed conjunctively even in a single forward pass without any reasoning steps via an abstract-then-compute mechanism: models first capture problem abstractions, then handle computation. Causal patching confirms these abstractions are present, transferable, composable, and precede computation. These behavioural and mechanistic findings highlight the need for disentangled evaluation to accurately assess LLM reasoning and to guide future improvements.

Can LLMs Reason Abstractly Over Math Word Problems Without CoT? Disentangling Abstract Formulation From Arithmetic Computation

Large Language Models (LLMs) have demonstrated strong calibration capabilities, where predicted probabilities align well with correctness, despite prior findings that deep neural networks are often overconfident. Recent studies have linked this behavior to specific components in the final layer, such as entropy neurons and the unembedding matrix’s null space. In this work, we provide a complementary perspective by investigating how calibration evolves throughout the network's depth. Analyzing multiple open-weight models on the MMLU benchmark, we uncover a distinct confidence correction phase in the upper/later layers, where model confidence is actively recalibrated after decision certainty has been reached. Furthermore, we identify a low-dimensional calibration direction in the residual stream whose perturbation significantly improves calibration metrics (ECE and MCE) without harming accuracy. Our findings suggest that calibration is a distributed phenomenon, shaped throughout the network’s forward pass, not just in its final projection, providing new insights into how confidence-regulating mechanisms operate within LLMs.

Downloads

Next from EMNLP 2025

FaithUn: Toward Faithful Forgetting in Language Models by Investigating the Interconnectedness of Knowledge

Similar lecture

Learning to Decode Collaboratively with Multiple Language Models