China

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.

EMNLP 2025

A Probabilistic Inference Scaling Theory for LLM Self-Correction

inference scaling

self-correction

theory

### 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**. 

ACL 2025

Welcome to The 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025)

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.

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.

Calibration Across Layers: Understanding Calibration Evolution in LLMs

Code-mixed text—where multiple languages are used within the same utterance—is increasingly common in both spoken and written communication. However, it presents significant challenges for machine learning models due to the interplay of distinct grammatical structures, effectively forming a hybrid language. While fine-tuning large language models (LLMs) such as GPT-3, or Llama-3 on code-mixed data has led to performance improvements, these models still lag behind their monolingual counterparts and incur high computational costs due to the large number of trainable parameters. In this paper, we focus on the task of sentiment detection in code-mixed text and propose a Hybrid Language Model (HLM) that combines a multilingual encoder (e.g., mBERT) with a lightweight decoder (e.g., Sarvam-1) (< 3B parameters). Despite having significantly fewer trainable parameters, HLM achieves sentiment classification performance comparable to that of fine-tuned Large Language Models (LLMs) (>7B) parameters). Furthermore, our results demonstrate that HLM significantly outperforms models trained individually, underscoring its effectiveness for low-resource, code-mixed sentiment analysis.

Power doesn't reside in size: A Low Parameter Hybrid Language Model (HLM) for Sentiment Analysis in Code-mixed data

Large Language Models (LLMs), despite their remarkable capabilities, are hampered by hallucinations. A particularly challenging variant, knowledge overshadowing, occurs when one piece of activated knowledge inadvertently masks another relevant piece, leading to erroneous outputs even with high-quality training data. Current understanding of overshadowing is largely confined to inference-time observations, lacking deep insights into its origins and internal mechanisms during model training. Therefore, we introduce **PhantomCircuit, a novel framework designed to comprehensively analyze and detect knowledge overshadowing.** By innovatively employing knowledge circuit analysis, PhantomCircuit dissects the internal workings of attention heads, tracing how competing knowledge pathways contribute to the overshadowing phenomenon and its evolution throughout the training process. Extensive experiments demonstrate PhantomCircuit’s effectiveness in identifying such instances, offering novel insights into this elusive hallucination and providing the research community with a new methodological lens for its potential mitigation.

Pierce the Mists, Greet the Sky: Decipher Knowledge Overshadowing via Knowledge Circuit Analysis

Methods for learning taxonomies from data have been widely studied. We study a specific version of this task, called commonality identification, where only the set of entities is given and we need to find meaningful ways to group those entities. While LLMs should intuitively excel at this task, it is difficult to directly use such models in large domains. In this paper, we instead use LLMs to describe the different properties that are satisfied by each of the entities individually. We then use pre-trained embeddings to cluster these properties, and finally group entities that have properties which belong to the same cluster. To achieve good results, it is paramount that the properties predicted by the LLM are sufficiently diverse. We find that this diversity can be improved by prompting the LLM to structure the predicted properties into different facets of knowledge.

Grouping Entities with Shared Properties using Multi-Facet Prompting and Property Embeddings

While Knowledge Editing has been extensively studied in monolingual settings, it remains underexplored in multilingual contexts. This survey systematizes recent research on Multilingual Knowledge Editing (MKE), a growing subdomain of model editing focused on ensuring factual edits generalize reliably across languages. We present a comprehensive taxonomy of MKE methods, covering parameter-based, memory-based, fine-tuning, and hypernetwork approaches. We survey available benchmarks, summarize key findings on method effectiveness and transfer patterns, identify challenges in cross-lingual propagation, and highlight open problems related to language anisotropy, evaluation coverage, and edit scalability. Our analysis consolidates a rapidly evolving area and lays the groundwork for future progress in editable language-aware LLMs.

Editing Across Languages: A Survey of Multilingual Knowledge Editing

We investigate how Large Language Models (LLMs) distinguish between memorization and generalization at the neuron level. Through carefully designed tasks, we identify distinct neuron subsets responsible for each behavior. Experiments on both a GPT-2 model trained from scratch and a pretrained LLaMA-3.2 model fine-tuned with LoRA show consistent neuron-level specialization. We further demonstrate that inference-time interventions on these neurons can steer the model's behavior toward memorization or generalization. To assess robustness, we evaluate intra-task and inter-task consistency, confirming that these neuron-behavior associations reflect generalizable patterns rather than dataset-specific artifacts. Our findings reveal modular structure in LLMs and enable controlling memorization and generalization behaviors at inference time.

Neuron-Level Differentiation of Memorization and Generalization in Large Language Models

Despite the impressive chain-of-thought(CoT) reasoning ability of large language models (LLMs), its underlying mechanisms remains unclear. In this paper, we explore the inner workings of LLM's CoT ability via the lens of neurons in the feed-forward layers. We propose an efficient method to identify reasoning-critical neurons by analyzing their activation patterns under reasoning chains of varying quality. Based on it, we devise a rather simple intervention method that directly stimulates these reasoning-critical neurons, to guide the generation of high-quality reasoning chains. Extended experiments validate the effectiveness of our method and demonstrate the critical role these identified neurons play in CoT reasoning. Our code and data will be publicly available.

Enhancing Chain-of-Thought Reasoning via Neuron Activation Differential Analysis

Large language models (LLMs) struggle with compositional generalisation, limiting their ability to systematically combine learned components to interpret novel inputs. While architectural modifications, fine-tuning, and data augmentation improve compositionality, they often have limited adaptability, face scalability constraints, or yield diminishing returns on real data. To address this, we propose CARMA, an intervention that enhances the stability and robustness of compositional reasoning in LLMs while preserving fine-tuned performance. CARMA employs mutual information regularisation and layer-wise stability constraints to mitigate feature fragmentation, ensuring structured representations persist across and within layers. We evaluate CARMA on inverse dictionary modelling and sentiment classification, measuring its impact on semantic consistency, performance stability, and robustness to lexical perturbations. Results show that CARMA reduces the variability introduced by fine-tuning, stabilises token representations, and improves compositional reasoning. While its effectiveness varies across architectures, CARMA's key strength lies in reinforcing learned structures rather than introducing new capabilities, making it a scalable auxiliary method. These findings suggest that integrating CARMA with fine-tuning can improve compositional generalisation while maintaining task-specific performance in LLMs.

CARMA: Enhanced Compositionality in LLMs via Advanced Regularisation and Mutual Information Alignment

Large Language Models (LLMs) with safe-alignment training are powerful instruments with robust language comprehension capability. Typically LLMs undergo careful alignment training involving human feedback to ensure the acceptance of safe inputs while rejection of harmful or unsafe ones. However, these humongous models are still vulnerable to jailbreak attacks, in which malicious users attempt to generate harmful outputs that safety-aligned LLMs are trained to avoid. In this study, we find that the safety mechanisms in LLMs are predominantly prevalent in the middle-to-late layers. Based on this observation, we introduce a novel white-box jailbreak method SABER (Safety Alignment Bypass via Extra Residuals) that connects two intermediate layer s and e such that s<e with a residual connection, achieving an improvement of 51% over the best performing baseline GCG on HarmBench test set. Moreover, model demonstrates only a marginal shift in perplexity when evaluated on the validation set of HarmBench.

Downloads

Next from EMNLP 2025

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

Similar lecture

Confidence v.s. Critique: A Decomposition of Self-Correction Capability for LLMs