Ercong Nie

Hengyuan Zhang, Zhihao Zhang, Mingyang Wang et al.

Mechanistic Interpretability (MI) has emerged as a vital approach to demystify the opaque decision-making of Large Language Models (LLMs). However, existing reviews primarily treat MI as an observational science, summarizing analytical insights while lacking a systematic framework for actionable intervention. To bridge this gap, we present a practical survey structured around the pipeline: "Locate, Steer, and Improve." We formally categorize Localizing (diagnosis) and Steering (intervention) methods based on specific Interpretable Objects to establish a rigorous intervention protocol. Furthermore, we demonstrate how this framework enables tangible improvements in Alignment, Capability, and Efficiency, effectively operationalizing MI as an actionable methodology for model optimization. The curated paper list of this work is available at https://github.com/rattlesnakey/Awesome-Actionable-MI-Survey.

Ercong Nie, Sheng Liang, Helmut Schmid et al.

Multilingual Pretrained Language Models (MPLMs) have shown their strong multilinguality in recent empirical cross-lingual transfer studies. In this paper, we propose the Prompts Augmented by Retrieval Crosslingually (PARC) pipeline to improve the zero-shot performance on low-resource languages (LRLs) by augmenting the context with semantically similar sentences retrieved from a high-resource language (HRL) as prompts. PARC improves the zero-shot performance on three downstream tasks (binary sentiment classification, topic categorization and natural language inference) with multilingual parallel test sets across 10 LRLs covering 6 language families in both unlabeled settings (+5.1%) and labeled settings (+16.3%). PARC-labeled also outperforms the finetuning baseline by 3.7%. We find a significant positive correlation between cross-lingual transfer performance on one side, and the similarity between the high- and low-resource languages as well as the amount of low-resource pretraining data on the other side. A robustness analysis suggests that PARC has the potential to achieve even stronger performance with more powerful MPLMs.

Bolei Ma, Ercong Nie, Helmut Schmid et al.

Multilingual pretrained language models (MPLMs) have demonstrated substantial performance improvements in zero-shot cross-lingual transfer across various natural language understanding tasks by finetuning MPLMs on task-specific labelled data of a source language (e.g. English) and evaluating on a wide range of target languages. Recent studies show that prompt-based finetuning surpasses regular finetuning in few-shot scenarios. However, the exploration of prompt-based learning in multilingual tasks remains limited. In this study, we propose the ProFiT pipeline to investigate the cross-lingual capabilities of Prompt-based Finetuning. We conduct comprehensive experiments on diverse cross-lingual language understanding tasks (sentiment classification, paraphrase identification, and natural language inference) and empirically analyze the variation trends of prompt-based finetuning performance in cross-lingual transfer across different few-shot and full-data settings. Our results reveal the effectiveness and versatility of prompt-based finetuning in cross-lingual language understanding. Our findings indicate that prompt-based finetuning outperforms vanilla finetuning in full-data scenarios and exhibits greater advantages in few-shot scenarios, with different performance patterns dependent on task types. Additionally, we analyze underlying factors such as language similarity and pretraining data size that impact the cross-lingual performance of prompt-based finetuning. Overall, our work provides valuable insights into the cross-lingual prowess of prompt-based finetuning.

Zijing Wang, Yongkang Liu, Mingyang Wang et al.

Multimodal Large Language Models (MLLMs) rely on strong linguistic reasoning inherited from their base language models. However, multimodal instruction fine-tuning paradoxically degrades this text's reasoning capability, undermining multimodal performance. To address this issue, we propose a training-free framework to mitigate this degradation. Through layer-wise vision token masking, we reveal a common three-stage pattern in multimodal large language models: early-modal separation, mid-modal alignment, and late-modal degradation. By analyzing the behavior of MLLMs at different stages, we propose a plateau-guided model merging method that selectively injects base language model parameters into MLLMs. Experimental results based on five MLLMs on nine benchmarks demonstrate the effectiveness of our method. Attention-based analysis further reveals that merging shifts attention from diffuse, scattered patterns to focused localization on task-relevant visual regions. Our repository is on https://github.com/wzj1718/PlaM.

Yongkang Liu, Zijing Wang, Mengjie Zhao et al.

This work presents \textsc{ChunkFT}, a memory-efficient fine-tuning framework that reformulates full-parameter fine-tuning around a dynamically activated working set. \textsc{ChunkFT} enables gradient computation for arbitrary sub-tensors without modifying the network architecture, providing an algorithmic foundation for optimizing arbitrary sub-networks while avoiding standard dense gradient computation. We provide a theoretical convergence analysis of \textsc{ChunkFT} in the deterministic setting. Empirically, we apply \textsc{ChunkFT} to fine-tune Llama 3-8B and Llama 3-70B using a single RTX 4090-24GB GPU and 2$\times$ H800-80GB GPUs, respectively. Full-parameter fine-tuning of a 7B model with a 1K input length requires only 13.72GB of GPU memory. The results demonstrate the effectiveness of \textsc{ChunkFT} in memory usage, running time, and optimization quality. Moreover, downstream evaluations on language understanding, mathematical reasoning, and MT-Bench show that \textsc{ChunkFT} consistently outperforms existing memory-efficient baselines. Notably, \textsc{ChunkFT} achieves performance comparable to, and in some cases exceeding, full-parameter fine-tuning. Our repository is on https://github.com/misonsky/chunk.

Ercong Nie, Helmut Schmid, Hinrich Schütze

Pretrained multilingual encoder models can directly perform zero-shot multilingual tasks or linguistic probing by reformulating the input examples into cloze-style prompts. This is accomplished by predicting the probabilities of the label words at the masked token position, without requiring any updates to the model parameters. However, the performance of this method is limited by the model's bias toward predicting label words which frequently occurred during the pretraining. These words typically receive high probabilities. To address this issue, we combine the models with calibration techniques which modify the probabilities of label words predicted by the models. We first validate the effectiveness of a proposed simple calibration method together with other existing techniques on monolingual encoders in both zero- and few-shot scenarios. We subsequently employ these calibration techniques on multilingual encoders, resulting in substantial performance improvements across a wide range of tasks.

Ercong Nie, Helmut Schmid, Hinrich Schütze

Constituency parsing plays a fundamental role in advancing natural language processing (NLP) tasks. However, training an automatic syntactic analysis system for ancient languages solely relying on annotated parse data is a formidable task due to the inherent challenges in building treebanks for such languages. It demands extensive linguistic expertise, leading to a scarcity of available resources. To overcome this hurdle, cross-lingual transfer techniques which require minimal or even no annotated data for low-resource target languages offer a promising solution. In this study, we focus on building a constituency parser for $\mathbf{M}$iddle $\mathbf{H}$igh $\mathbf{G}$erman ($\mathbf{MHG}$) under realistic conditions, where no annotated MHG treebank is available for training. In our approach, we leverage the linguistic continuity and structural similarity between MHG and $\mathbf{M}$odern $\mathbf{G}$erman ($\mathbf{MG}$), along with the abundance of MG treebank resources. Specifically, by employing the $\mathit{delexicalization}$ method, we train a constituency parser on MG parse datasets and perform cross-lingual transfer to MHG parsing. Our delexicalized constituency parser demonstrates remarkable performance on the MHG test set, achieving an F1-score of 67.3%. It outperforms the best zero-shot cross-lingual baseline by a margin of 28.6% points. These encouraging results underscore the practicality and potential for automatic syntactic analysis in other ancient languages that face similar challenges as MHG.

Zijing Wang, Mingyang Wang, Ercong Nie et al.

Towards more general and human-like intelligence, large language models should seamlessly integrate both multilingual and multimodal capabilities; however, extending an existing multimodal model to many languages typically requires expensive multilingual multimodal data construction and repeated end-to-end retraining. We study a training-free alternative: injecting multilingual capability into an existing multimodal model by composing residual updates in the shared language model backbone. The key challenge is that multilingual and multimodal updates are heterogeneous, reflecting different functional roles in the shared model. To address this, we propose Direction- and Magnitude-aware Multilingual Multimodal merging (DiM3), which selectively composes the two updates at each parameter dimension while preserving the original vision encoder and multimodal projector. Experiments on multilingual benchmarks in both text-only and vision-language settings, covering 57 languages across LLaVA- and Qwen-based backbones, show that DiM3 consistently outperforms existing merging baselines, substantially improves multilingual performance over the original multimodal model, and remains competitive with dedicated multilingual multimodal fine-tuning while largely retaining general multimodal ability. We further show that DiM3 can be directly applied to already trained multilingual multimodal models and still yield additional gains. Further interpretability analysis shows that DiM3 primarily reshapes intermediate-layer semantic representations, strengthening cross-lingual alignment under both text-only and multimodal inputs while preserving higher-layer task-sensitive structure. Our repository is on https://github.com/wzj1718/DiM3.

Minghan Li, Xinxuan Lv, Junjie Zou et al.

Modern information retrieval must reconcile short, ambiguous queries with increasingly diverse and dynamic corpora. Query expansion (QE) remains a core technique for mitigating vocabulary mismatch, but its design space has been reshaped by pre-trained and large language models (PLMs/LLMs). This survey reviews QE methods in the PLM/LLM era and provides a unified view of the emerging landscape. We first summarize how different model families enable new expansion behaviors, including stronger contextualization, more controllable generation, and instruction-following. We then organize recent techniques along four complementary design dimensions: where expansion is injected in the pipeline, how it is grounded and interacts with corpus evidence, how it is learned or aligned, and how structured knowledge such as knowledge graphs is incorporated. Beyond taxonomy, we synthesize application patterns and deployment considerations across representative retrieval settings, highlighting practical trade-offs among effectiveness, controllability, grounding quality, and operating cost. Finally, we outline open challenges and future directions toward more reliable, safe, efficient, and continually adaptive QE under real-world constraints.

Xiaoqian Li, Ercong Nie, Sheng Liang

The remarkable ability of Large Language Models (LLMs) to understand and follow instructions has sometimes been limited by their in-context learning (ICL) performance in low-resource languages. To address this, we introduce a novel approach that leverages cross-lingual retrieval-augmented in-context learning (CREA-ICL). By extracting semantically similar prompts from high-resource languages, we aim to improve the zero-shot performance of multilingual pre-trained language models (MPLMs) across diverse tasks. Though our approach yields steady improvements in classification tasks, it faces challenges in generation tasks. Our evaluation offers insights into the performance dynamics of retrieval-augmented in-context learning across both classification and generation domains.

Yongkang Liu, Jiayang Yu, Mingyang Wang et al.

Argumentation generation has attracted substantial research interest due to its central role in human reasoning and decision-making. However, most existing argumentative corpora focus on non-interactive, single-turn settings, either generating arguments from a given topic or refuting an existing argument. In practice, however, argumentation is often realized as multi-turn dialogue, where speakers defend their stances and employ diverse argumentative strategies to strengthen persuasiveness. To support deeper modeling of argumentation dialogue, we present the first large-scale \textbf{S}trategic \textbf{A}rgumentative \textbf{D}ialogue dataset, SAD, consisting of 392,822 examples. Grounded in argumentation theories, we annotate each utterance with five strategy types, allowing multiple strategies per utterance. Unlike prior datasets, SAD requires models to generate contextually appropriate arguments conditioned on the dialogue history, a specified stance on the topic, and targeted argumentation strategies. We further benchmark a range of pretrained generative models on SAD and present in-depth analysis of strategy usage patterns in argumentation.

Xiaoqian Li, Ercong Nie, Sheng Liang

The promise of Large Language Models (LLMs) in Natural Language Processing has often been overshadowed by their limited performance in low-resource languages such as Bangla. To address this, our paper presents a pioneering approach that utilizes cross-lingual retrieval augmented in-context learning. By strategically sourcing semantically similar prompts from high-resource language, we enable multilingual pretrained language models (MPLMs), especially the generative model BLOOMZ, to successfully boost performance on Bangla tasks. Our extensive evaluation highlights that the cross-lingual retrieval augmented prompts bring steady improvements to MPLMs over the zero-shot performance.

Zheyu Zhang, Han Yang, Bolei Ma et al.

Large Language Models (LLMs) demonstrate remarkable performance on a variety of natural language understanding (NLU) tasks, primarily due to their in-context learning ability. This ability could be applied to building babylike models, i.e. models at small scales, improving training efficiency. In this paper, we propose a "CoThought" pipeline, which efficiently trains smaller "baby" language models (BabyLMs) by leveraging the Chain of Thought prompting of LLMs. Our pipeline restructures a dataset of less than 100M in size using GPT-3.5-turbo, transforming it into task-oriented, human-readable texts that are comparable to the school texts for language learners. The BabyLM is then pretrained on this restructured dataset in a RoBERTa fashion. In evaluations across 4 benchmarks, our BabyLM outperforms the vanilla RoBERTa in 10 linguistic, NLU, and question-answering tasks by more than 3 points, showing a superior ability to extract contextual information. These results suggest that compact LMs pretrained on small, LLM-restructured data can better understand tasks and achieve improved performance.

Sikuan Yan, Ahmed Bahloul, Ercong Nie et al.

Memory-augmented LLM agents enable interactions that extend beyond finite context windows by storing, updating, and reusing information across sessions. However, training such agents with reinforcement learning in multi-session environments is challenging because memory turns the agent's past actions into part of its future environment. Once different rollouts write, update, or delete different memories, they no longer share the same intermediate memory state, making trajectory-level comparisons fundamentally unfair. This violates a key assumption behind group-relative methods such as GRPO, where rollouts are compared as if they were sampled from the same effective environment. Consequently, trajectory-level rewards provide noisy or biased credit signals for long-horizon memory operations. To address this challenge, we introduce Memory-R2, a training framework for long-horizon memory-augmented LLM agents. Its core algorithm, LoGo-GRPO, combines local and global group-relative optimization. The global objective preserves end-to-end learning from long-horizon trajectory-level rewards, while local rerollouts compare different memory-operation outcomes from the same intermediate memory state, yielding fairer group comparisons and more precise supervision for memory construction. Beyond credit assignment, Memory-R2 jointly optimizes memory formation and memory evolution with a shared-parameter co-learning design, where a fact extractor and a memory manager are instantiated from the same LLM backbone through role-specific prompts. To stabilize multi-step RL over long memory horizons, we adopt a progressive curriculum that increases the training horizon from 8 to 16 to 32 sessions. Together, these components provide an effective training paradigm for memory-augmented LLM agents in long-horizon multi-session settings.

Sikuan Yan, Sicheng Dong, Haotong Wang et al.

Memory has become an increasingly important component of agentic systems, as these systems are expected to reason over long-term experience. However, prior work has largely focused on unimodal memory, leaving multimodal memory relatively underexplored despite its central role in real-world applications. Compared with unimodal settings, multimodal memory introduces additional challenges, including heterogeneous input integration, person-centric information alignment, and evidence aggregation across different granularities. We present PyraVid, a hierarchical multimodal memory framework inspired by Event Segmentation Theory from cognitive science. PyraVid organizes long videos into a coarse-to-fine pyramid structure, enabling structured memory access and effective evidence aggregation. It further supports structure-guided memory expansion with pruning, allowing the retrieval of related events with strong causal connectivity but low semantic similarity while reducing noise. Experiments on multiple long-video understanding benchmarks show that PyraVid consistently improves performance across datasets, model scales, and question types, highlighting the effectiveness of hierarchical multimodal memory for long-horizon reasoning.

Ercong Nie, Helmut Schmid, Hinrich Schütze

Language confusion -- where large language models (LLMs) generate unintended languages against the user's need -- remains a critical challenge, especially for English-centric models. We present the first mechanistic interpretability (MI) study of language confusion, combining behavioral benchmarking with neuron-level analysis. Using the Language Confusion Benchmark (LCB), we show that confusion points (CPs) -- specific positions where language switches occur -- are central to this phenomenon. Through layer-wise analysis with TunedLens and targeted neuron attribution, we reveal that transition failures in the final layers drive confusion. We further demonstrate that editing a small set of critical neurons, identified via comparative analysis with a multilingual-tuned counterpart, substantially mitigates confusion while largely preserving general competence and fluency. Our approach matches multilingual alignment in confusion reduction for many languages and yields cleaner, higher-quality outputs. These findings provide new insights into the internal dynamics of LLMs and highlight neuron-level interventions as a promising direction for robust, interpretable multilingual language modeling. Code and data are available at: https://github.com/ercong21/lang_confusion.

Yihong Liu, Mingyang Wang, Amir Hossein Kargaran et al.

Large Language Models (LLMs) are capable of recalling multilingual factual knowledge present in their pretraining data. However, most studies evaluate only the final model, leaving the development of factual recall and crosslingual consistency throughout pretraining largely unexplored. In this work, we trace how factual recall and crosslingual consistency evolve during pretraining, focusing on OLMo-7B as a case study. We find that both accuracy and consistency improve over time for most languages. We show that this improvement is primarily driven by the fact frequency in the pretraining corpus: more frequent facts are more likely to be recalled correctly, regardless of language. Yet, some low-frequency facts in non-English languages can still be correctly recalled. Our analysis reveals that these instances largely benefit from crosslingual transfer of their English counterparts -- an effect that emerges predominantly in the early stages of pretraining. We pinpoint two distinct pathways through which multilingual factual knowledge acquisition occurs: (1) frequency-driven learning, which is dominant and language-agnostic, and (2) crosslingual transfer, which is limited in scale and typically constrained to relation types involving named entities. We release our code and data to facilitate further research at https://github.com/cisnlp/multilingual-fact-tracing.

Shuzhou Yuan, William LaCroix, Hardik Ghoshal et al.

Large Language Models (LLMs) are increasingly employed as AI tutors due to their scalability and potential for personalized instruction. However, off-the-shelf LLMs often underperform in educational settings: they frequently reveal answers too readily, fail to adapt their responses to student uncertainty, and remain vulnerable to emotionally manipulative prompts. To address these challenges, we introduce CoDAE, a framework that adapts LLMs for educational use through Chain-of-Thought (CoT) data augmentation. We collect real-world dialogues between students and a ChatGPT-based tutor and enrich them using CoT prompting to promote step-by-step reasoning and pedagogically aligned guidance. Furthermore, we design targeted dialogue cases to explicitly mitigate three key limitations: over-compliance, low response adaptivity, and threat vulnerability. We fine-tune four open-source LLMs on different variants of the augmented datasets and evaluate them in simulated educational scenarios using both automatic metrics and LLM-as-a-judge assessments. Our results show that models fine-tuned with CoDAE deliver more pedagogically appropriate guidance, better support reasoning processes, and effectively resist premature answer disclosure.

Shuzhou Yuan, Ercong Nie, Mario Tawfelis et al.

Hate speech detection is a socially sensitive and inherently subjective task, with judgments often varying based on personal traits. While prior work has examined how socio-demographic factors influence annotation, the impact of personality traits on Large Language Models (LLMs) remains largely unexplored. In this paper, we present the first comprehensive study on the role of persona prompts in hate speech classification, focusing on MBTI-based traits. A human annotation survey confirms that MBTI dimensions significantly affect labeling behavior. Extending this to LLMs, we prompt four open-source models with MBTI personas and evaluate their outputs across three hate speech datasets. Our analysis uncovers substantial persona-driven variation, including inconsistencies with ground truth, inter-persona disagreement, and logit-level biases. These findings highlight the need to carefully define persona prompts in LLM-based annotation workflows, with implications for fairness and alignment with human values.

Yongkang Liu, Xingle Xu, Ercong Nie et al.

Parameter-Efficient Fine-Tuning (PEFT) methods achieve performance comparable to Full Fine-Tuning (FFT) while requiring significantly fewer computing resources, making it the go-to choice for researchers. We find that although PEFT can achieve competitive results on some benchmarks, its performance falls short of FFT in complex tasks, such as reasoning and instruction-based fine-tuning. In this paper, we compare the characteristics of PEFT and FFT in terms of representational capacity and robustness based on optimization theory. We theoretically demonstrate that PEFT is a strict subset of FFT. By providing theoretical upper bounds for PEFT, we show that the limited parameter space constrains the model's representational ability, making it more susceptible to perturbations. Experiments on 15 datasets encompassing classification, generation, reasoning, instruction fine-tuning tasks and 11 adversarial test sets validate our theories. We hope that these results spark further research beyond the realms of well established PEFT. The source code is in the anonymous Github repository\footnote{https://github.com/misonsky/PEFTEval}.

Ercong Nie, Bo Shao, Zifeng Ding et al.

This paper introduces BMIKE-53, a comprehensive benchmark for cross-lingual in-context knowledge editing (IKE) across 53 languages, unifying three knowledge editing (KE) datasets: zsRE, CounterFact, and WikiFactDiff. Cross-lingual KE, which requires knowledge edited in one language to generalize across others while preserving unrelated knowledge, remains underexplored. To address this gap, we systematically evaluate IKE under zero-shot, one-shot, and few-shot setups, incorporating tailored metric-specific demonstrations. Our findings reveal that model scale and demonstration alignment critically govern cross-lingual IKE efficacy, with larger models and tailored demonstrations significantly improving performance. Linguistic properties, particularly script type, strongly influence performance variation across languages, with non-Latin languages underperforming due to issues like language confusion. Code and data are publicly available at: https://github.com/ercong21/MultiKnow/.

Mingyang Wang, Heike Adel, Lukas Lange et al.

Multilingual language models (MLMs) store factual knowledge across languages but often struggle to provide consistent responses to semantically equivalent prompts in different languages. While previous studies point out this cross-lingual inconsistency issue, the underlying causes remain unexplored. In this work, we use mechanistic interpretability methods to investigate cross-lingual inconsistencies in MLMs. We find that MLMs encode knowledge in a language-independent concept space through most layers, and only transition to language-specific spaces in the final layers. Failures during the language transition often result in incorrect predictions in the target language, even when the answers are correct in other languages. To mitigate this inconsistency issue, we propose a linear shortcut method that bypasses computations in the final layers, enhancing both prediction accuracy and cross-lingual consistency. Our findings shed light on the internal mechanisms of MLMs and provide a lightweight, effective strategy for producing more consistent factual outputs.

Minghan Li, Ercong Nie, Siqi Zhao et al.

Query expansion with large language models is promising but often relies on hand-crafted prompts, manually chosen exemplars, or a single LLM, making it non-scalable and sensitive to domain shift. We present an automated, domain-adaptive QE framework that builds in-domain exemplar pools by harvesting pseudo-relevant passages using a BM25-MonoT5 pipeline. A training-free cluster-based strategy selects diverse demonstrations, yielding strong and stable in-context QE without supervision. To further exploit model complementarity, we introduce a two-LLM ensemble in which two heterogeneous LLMs independently generate expansions and a refinement LLM consolidates them into one coherent expansion. Across TREC DL20, DBPedia, and SciFact, the refined ensemble delivers consistent and statistically significant gains over BM25, Rocchio, zero-shot, and fixed few-shot baselines. The framework offers a reproducible testbed for exemplar selection and multi-LLM generation, and a practical, label-free solution for real-world QE.

Sikuan Yan, Xiufeng Yang, Zuchao Huang et al.

Large Language Models (LLMs) have demonstrated impressive capabilities across a wide range of NLP tasks, but they remain fundamentally stateless, constrained by limited context windows that hinder long-horizon reasoning. Recent efforts to address this limitation often augment LLMs with an external memory bank, yet most existing pipelines are static and heuristic-driven, lacking a learned mechanism for deciding what to store, update, or retrieve. We present Memory-R1, a reinforcement learning (RL) framework that equips LLMs with the ability to actively manage and utilize external memory through two specialized agents: a Memory Manager that learns structured operations, including ADD, UPDATE, DELETE, and NOOP; and an Answer Agent that pre-selects and reasons over relevant entries. Both agents are fine-tuned with outcome-driven RL (PPO and GRPO), enabling adaptive memory management with minimal supervision. With only 152 training QA pairs, Memory-R1 outperforms strong baselines and generalizes across diverse question types, three benchmarks (LoCoMo, MSC, LongMemEval), and multiple model scales (3B-14B).

Linyang He, Peili Chen, Ercong Nie et al.

Inspired by cognitive neuroscience studies, we introduce a novel `decoding probing' method that uses minimal pairs benchmark (BLiMP) to probe internal linguistic characteristics in neural language models layer by layer. By treating the language model as the `brain' and its representations as `neural activations', we decode grammaticality labels of minimal pairs from the intermediate layers' representations. This approach reveals: 1) Self-supervised language models capture abstract linguistic structures in intermediate layers that GloVe and RNN language models cannot learn. 2) Information about syntactic grammaticality is robustly captured through the first third layers of GPT-2 and also distributed in later layers. As sentence complexity increases, more layers are required for learning grammatical capabilities. 3) Morphological and semantics/syntax interface-related features are harder to capture than syntax. 4) For Transformer-based models, both embeddings and attentions capture grammatical features but show distinct patterns. Different attention heads exhibit similar tendencies toward various linguistic phenomena, but with varied contributions.

Bolei Ma, Ercong Nie, Shuzhou Yuan et al.

Prompt-based methods have been successfully applied to multilingual pretrained language models for zero-shot cross-lingual understanding. However, most previous studies primarily focused on sentence-level classification tasks, and only a few considered token-level labeling tasks such as Named Entity Recognition (NER) and Part-of-Speech (POS) tagging. In this paper, we propose Token-Level Prompt Decomposition (ToPro), which facilitates the prompt-based method for token-level sequence labeling tasks. The ToPro method decomposes an input sentence into single tokens and applies one prompt template to each token. Our experiments on multilingual NER and POS tagging datasets demonstrate that ToPro-based fine-tuning outperforms Vanilla fine-tuning and Prompt-Tuning in zero-shot cross-lingual transfer, especially for languages that are typologically different from the source language English. Our method also attains state-of-the-art performance when employed with the mT5 model. Besides, our exploratory study in multilingual large language models shows that ToPro performs much better than the current in-context learning method. Overall, the performance improvements show that ToPro could potentially serve as a novel and simple benchmarking method for sequence labeling tasks.

Shuzhou Yuan, Ercong Nie, Bolei Ma et al.

Large Language Models (LLMs) possess outstanding capabilities in addressing various natural language processing (NLP) tasks. However, the sheer size of these models poses challenges in terms of storage, training and inference due to the inclusion of billions of parameters through layer stacking. While traditional approaches such as model pruning or distillation offer ways for reducing model size, they often come at the expense of performance retention. In our investigation, we systematically explore the approach of reducing the number of layers in LLMs. Surprisingly, we observe that even with fewer layers, LLMs maintain similar or better performance levels, particularly in prompt-based fine-tuning for text classification tasks. Remarkably, in certain cases, models with a single layer outperform their fully layered counterparts. These findings offer valuable insights for future work aimed at mitigating the size constraints of LLMs while preserving their performance, thereby opening avenues for significantly more efficient use of LLMs.

Linyang He, Ercong Nie, Sukru Samet Dindar et al.

We introduce XCOMPS in this work, a multilingual conceptual minimal pair dataset covering 17 languages. Using this dataset, we evaluate LLMs' multilingual conceptual understanding through metalinguistic prompting, direct probability measurement, and neurolinguistic probing. By comparing base, instruction-tuned, and knowledge-distilled models, we find that: 1) LLMs exhibit weaker conceptual understanding for low-resource languages, and accuracy varies across languages despite being tested on the same concept sets. 2) LLMs excel at distinguishing concept-property pairs that are visibly different but exhibit a marked performance drop when negative pairs share subtle semantic similarities. 3) Instruction tuning improves performance in concept understanding but does not enhance internal competence; knowledge distillation can enhance internal competence in conceptual understanding for low-resource languages with limited gains in explicit task performance. 4) More morphologically complex languages yield lower concept understanding scores and require deeper layers for conceptual reasoning.

Shuzhou Yuan, Ercong Nie, Michael Färber et al.

Large Language Models (LLMs) exhibit strong In-Context Learning (ICL) capabilities when prompts with demonstrations are used. However, fine-tuning still remains crucial to further enhance their adaptability. Prompt-based fine-tuning proves to be an effective fine-tuning method in low-data scenarios, but high demands on computing resources limit its practicality. We address this issue by introducing a prompt-based parameter-efficient fine-tuning (PEFT) approach. GNNavi leverages insights into ICL's information flow dynamics, which indicates that label words act in prompts as anchors for information propagation. GNNavi employs a Graph Neural Network (GNN) layer to precisely guide the aggregation and distribution of information flow during the processing of prompts by hardwiring the desired information flow into the GNN. Our experiments on text classification tasks with GPT-2 and Llama2 show GNNavi surpasses standard prompt-based fine-tuning methods in few-shot settings by updating just 0.2% to 0.5% of parameters. We compare GNNavi with prevalent PEFT approaches, such as prefix tuning, LoRA and Adapter in terms of performance and efficiency. Our analysis reveals that GNNavi enhances information flow and ensures a clear aggregation process.

Shuzhou Yuan, Ercong Nie, Lukas Kouba et al.

Detoxification, the task of rewriting harmful language into non-toxic text, has become increasingly important amid the growing prevalence of toxic content online. However, high-quality parallel datasets for detoxification, especially for hate speech, remain scarce due to the cost and sensitivity of human annotation. In this paper, we propose a novel LLM-in-the-loop pipeline leveraging GPT-4o-mini for automated detoxification. We first replicate the ParaDetox pipeline by replacing human annotators with an LLM and show that the LLM performs comparably to human annotation. Building on this, we construct ParaDeHate, a large-scale parallel dataset specifically for hatespeech detoxification. We release ParaDeHate as a benchmark of over 8K hate/non-hate text pairs and evaluate a wide range of baseline methods. Experimental results show that models such as BART, fine-tuned on ParaDeHate, achieve better performance in style accuracy, content preservation, and fluency, demonstrating the effectiveness of LLM-generated detoxification text as a scalable alternative to human annotation.

Lovisa Hagström, Ercong Nie, Ruben Halifa et al.

Language model (LM) re-rankers are used to refine retrieval results for retrieval-augmented generation (RAG). They are more expensive than lexical matching methods like BM25 but assumed to better process semantic information and the relations between the query and the retrieved answers. To understand whether LM re-rankers always live up to this assumption, we evaluate 6 different LM re-rankers on the NQ, LitQA2 and DRUID datasets. Our results show that LM re-rankers struggle to outperform a simple BM25 baseline on DRUID. Leveraging a novel separation metric based on BM25 scores, we explain and identify re-ranker errors stemming from lexical dissimilarities. We also investigate different methods to improve LM re-ranker performance and find these methods mainly useful for NQ. Taken together, our work identifies and explains weaknesses of LM re-rankers and points to the need for more adversarial and realistic datasets for their evaluation.

Shuzhou Yuan, Ercong Nie, Yinuo Sun et al.

Large language models (LLMs) frequently produce false refusals, declining benign requests that contain terms resembling unsafe queries. We address this challenge by introducing two comprehensive benchmarks: the Exaggerated Safety Benchmark (XSB) for single-turn prompts, annotated with "Focus" keywords that identify refusal-inducing triggers, and the Multi-turn Scenario-based Exaggerated Safety Benchmark (MS-XSB), which systematically evaluates refusal calibration in realistic, context-rich dialog settings. Our benchmarks reveal that exaggerated refusals persist across diverse recent LLMs and are especially pronounced in complex, multi-turn scenarios. To mitigate these failures, we leverage post-hoc explanation methods to identify refusal triggers and deploy three lightweight, model-agnostic approaches, ignore-word instructions, prompt rephrasing, and attention steering, at inference time, all without retraining or parameter access. Experiments on four instruction-tuned Llama models demonstrate that these strategies substantially improve compliance on safe prompts while maintaining robust safety protections. Our findings establish a reproducible framework for diagnosing and mitigating exaggerated refusals, highlighting practical pathways to safer and more helpful LLM deployments.

Chunkit Chan, Yauwai Yim, Hongchuan Zeng et al.

Theory of Mind (ToM), the ability to infer mental states in others, is pivotal for human social cognition. Existing evaluations of ToM in LLMs are largely limited to English, neglecting the linguistic diversity that shapes human cognition. This limitation raises a critical question: can LLMs exhibit Multilingual Theory of Mind, which is the capacity to reason about mental states across diverse linguistic contexts? To address this gap, we present XToM, a rigorously validated multilingual benchmark that evaluates ToM across five languages and incorporates diverse, contextually rich task scenarios. Using XToM, we systematically evaluate LLMs (e.g., DeepSeek R1), revealing a pronounced dissonance: while models excel in multilingual language understanding, their ToM performance varies across languages. Our findings expose limitations in LLMs' ability to replicate human-like mentalizing across linguistic contexts.

Linyang He, Ercong Nie, Helmut Schmid et al.

This study investigates the linguistic understanding of Large Language Models (LLMs) regarding signifier (form) and signified (meaning) by distinguishing two LLM assessment paradigms: psycholinguistic and neurolinguistic. Traditional psycholinguistic evaluations often reflect statistical rules that may not accurately represent LLMs' true linguistic competence. We introduce a neurolinguistic approach, utilizing a novel method that combines minimal pair and diagnostic probing to analyze activation patterns across model layers. This method allows for a detailed examination of how LLMs represent form and meaning, and whether these representations are consistent across languages. We found: (1) Psycholinguistic and neurolinguistic methods reveal that language performance and competence are distinct; (2) Direct probability measurement may not accurately assess linguistic competence; (3) Instruction tuning won't change much competence but improve performance; (4) LLMs exhibit higher competence and performance in form compared to meaning. Additionally, we introduce new conceptual minimal pair datasets for Chinese (COMPS-ZH) and German (COMPS-DE), complementing existing English datasets.

Yongkang Liu, Ercong Nie, Shi Feng et al.

Current state-of-the-art dialogue systems heavily rely on extensive training datasets. However, challenges arise in domains where domain-specific training datasets are insufficient or entirely absent. To tackle this challenge, we propose a novel data \textbf{A}ugmentation framework for \textbf{M}ulti-\textbf{D}omain \textbf{D}ialogue \textbf{G}eneration, referred to as \textbf{AMD$^2$G}. The AMD$^2$G framework consists of a data augmentation process and a two-stage training approach: domain-agnostic training and domain adaptation training. We posit that domain corpora are a blend of domain-agnostic and domain-specific features, with certain representation patterns shared among diverse domains. Domain-agnostic training aims to enable models to learn these common expressive patterns. To construct domain-agnostic dialogue corpora, we employ a \textit{\textbf{de-domaining}} data processing technique used to remove domain-specific features. By mitigating the effects of domain-specific features, the model trained on the de-domained corpora can effectively learn common expression patterns in different domains. Subsequently, we adapt the learned domain-agnostic features to the target domain through domain adaptation training. We conduct experiments on Chinese dialogue datasets from five different domains and show that AMD$^2$G achieves superior performance compared to both direct training on the target domain corpus and collective training on all five domain corpora. Our work underscores AMD$^2$G as a viable alternative solution for low-resource multi-domain dialogue generation. Code and data associated with our work are available on GitHub repository$^{\text 1}$.

Ercong Nie, Shuzhou Yuan, Bolei Ma et al.

Probing the multilingual knowledge of linguistic structure in LLMs, often characterized as sequence labeling, faces challenges with maintaining output templates in current text-to-text prompting strategies. To solve this, we introduce a decomposed prompting approach for sequence labeling tasks. Diverging from the single text-to-text prompt, our prompt method generates for each token of the input sentence an individual prompt which asks for its linguistic label. We test our method on the Universal Dependencies part-of-speech tagging dataset for 38 languages, using both English-centric and multilingual LLMs. Our findings show that decomposed prompting surpasses the iterative prompting baseline in efficacy and efficiency under zero- and few-shot settings. Moreover, our analysis of multilingual performance of English-centric LLMs yields insights into the transferability of linguistic knowledge via multilingual prompting.