Sarana Nutanong

Patomporn Payoungkhamdee, Tinnakit Udsa, Jian Gang Ngui et al.

Small language models (SLMs) are efficient and scalable, but their multilingual capabilities degrade severely at sub-billion scales, especially for Southeast Asian (SEA) languages. We introduce DuDi, a dual-signal multilingual distillation framework that combines an online sequence-level signal with off-policy and on-policy token-level signals. DuDi further uses a cross-lingual verbalizer to refine teacher feedback and improve teacher-student transferability in multilingual settings. Experiments on SEA-HELM across multiple model families, scales, and teacher-student settings show that DuDi consistently outperforms competitive distillation baselines. Ablations and analyses confirm that sequence-level optimization, token-level supervision, and cross-lingual verbalization provide complementary and transferable learning signals for multilingual SLMs.

Peerawat Chomphooyod, Jian Gang Ngui, Yosephine Susanto et al.

Frontier LLMs perform well in Western contexts, but remain poorly tested on underrepresented cultures such as those in Southeast Asia (SEA). Existing NLI benchmarks are largely Western-centric, translation-derived, or monolingual, limiting their ability to measure culturally grounded reasoning. We introduce SEA-NLI, a native, culturally grounded NLI benchmark covering eight SEA countries in English and native regional languages, verified by native speakers. Across 17 encoder and decoder models, we observe a low performance from all models, especially for knowledge-intensive categories such as Languages and Science and Technology. Our analysis shows that failure cases mainly stem from missing SEA cultural knowledge: SEA-adapted models and culture-aware prompting improve performance, while CoT prompting offers limited gains.

Peerat Limkonchotiwat, Raymond Ng, Sarana Nutanong et al.

Text embeddings are fundamental to many downstream applications, making robustness important for real-world NLP. However, most recent state-of-the-art embedding models are not reproducible because they rely on closed or undisclosed training data, and they remain insufficiently robust for Southeast Asian languages. We present SEA-Embedding, a fully open and reproducible text-embedding pipeline for Southeast Asian languages trained only on publicly available data, and use it to study three core factors of robust embedding design: data composition, training objective, and base encoder initialization. SEA-Embedding achieves state-of-the-art results on SEA-BED while enabling systematic and reproducible analysis of robust text embeddings for the region.

Wannaphong Phatthiyaphaibun, Chompakorn Chaksangchaichot, Peerat Limkonchotiwat et al.

Recently, Automatic Speech Recognition (ASR), a system that converts audio into text, has caught a lot of attention in the machine learning community. Thus, a lot of publicly available models were released in HuggingFace. However, most of these ASR models are available in English; only a minority of the models are available in Thai. Additionally, most of the Thai ASR models are closed-sourced, and the performance of existing open-sourced models lacks robustness. To address this problem, we train a new ASR model on a pre-trained XLSR-Wav2Vec model with the Thai CommonVoice corpus V8 and train a trigram language model to boost the performance of our ASR model. We hope that our models will be beneficial to individuals and the ASR community in Thailand.

Pawitsapak Akarajaradwong, Wuttikrai Lertprasertphakorn, Chompakorn Chaksangchaichot et al.

Free-form legal essay evaluation in NLP treats expert inter-rater stability as a single ceiling number, and treats LLM-judge agreement with that ceiling as evidence of judge stability. We test both assumptions on the Thai bar examination through an identical-inputs protocol: three Bar Council-trained examiners (A, B, C) and a 26-LLM judge panel score the same 15 cross-graded answers from the same four inputs (question, official Bar Council grading regulation, gold answer, candidate answer). The headline finding is asymmetric. On 10 of 15 cells where the rubric prescribes both axes, all 29 raters converge in a tight band: panel agreement is universal. On the remaining 5 cells where the rubric does not prescribe how to grade a correct final answer that omits a decisive statutory citation, the human panel splits between two coherent readings (B/C majority at the upper rubric band, score $6$--$8$; A minority at the lower band, score $1$--$2$). The LLM judge population does not split symmetrically: 22 of 26 LLMs score in or near B/C's contested band, 3 sit in the regulation-silent middle gap, and only 1 (GPT-5.4 Nano) approaches A's band without consistently scoring within it. \emph{Zero LLMs in our 26-judge panel reproduce the minority human reading on the contested cells.} The B/C-direction cluster spans every model size, vendor, and price tier we tested. An instrumented three-LLM anchor sub-panel (Claude 4.6 Opus, Gemini 3.1 Pro, GPT-5.4 Pro) carries determinism probes, input ablations, and bootstrap CIs, and reaches anchor panel $α= 0.77$ on the 15 cells against human-panel $α= 0.36$. The high LLM-panel $α$ reflects systematic convergence on the majority reading rather than balanced reproduction of both readings; a benchmark that selects its LLM judge by maximising agreement with a human reference panel will inherit this asymmetry by construction.

Panuthep Tasawong, Wuttikorn Ponwitayarat, Peerat Limkonchotiwat et al.

Dense retrieval is a basic building block of information retrieval applications. One of the main challenges of dense retrieval in real-world settings is the handling of queries containing misspelled words. A popular approach for handling misspelled queries is minimizing the representations discrepancy between misspelled queries and their pristine ones. Unlike the existing approaches, which only focus on the alignment between misspelled and pristine queries, our method also improves the contrast between each misspelled query and its surrounding queries. To assess the effectiveness of our proposed method, we compare it against the existing competitors using two benchmark datasets and two base encoders. Our method outperforms the competitors in all cases with misspelled queries. Our code and models are available at https://github. com/panuthept/DST-DenseRetrieval.

Peerat Limkonchotiwat, Wuttikorn Ponwitayarat, Lalita Lowphansirikul et al.

Self-supervised sentence representation learning is the task of constructing an embedding space for sentences without relying on human annotation efforts. One straightforward approach is to finetune a pretrained language model (PLM) with a representation learning method such as contrastive learning. While this approach achieves impressive performance on larger PLMs, the performance rapidly degrades as the number of parameters decreases. In this paper, we propose a framework called Self-supervised Cross-View Training (SCT) to narrow the performance gap between large and small PLMs. To evaluate the effectiveness of SCT, we compare it to 5 baseline and state-of-the-art competitors on seven Semantic Textual Similarity (STS) benchmarks using 5 PLMs with the number of parameters ranging from 4M to 340M. The experimental results show that STC outperforms the competitors for PLMs with less than 100M parameters in 18 of 21 cases.

Jitkapat Sawatphol, Can Udomcharoenchaikit, Sarana Nutanong

Authorship verification (AV) aims to identify whether a pair of texts has the same author. We address the challenge of evaluating AV models' robustness against topic shifts. The conventional evaluation assumes minimal topic overlap between training and test data. However, we argue that there can still be topic leakage in test data, causing misleading model performance and unstable rankings. To address this, we propose an evaluation method called Heterogeneity-Informed Topic Sampling (HITS), which creates a smaller dataset with a heterogeneously distributed topic set. Our experimental results demonstrate that HITS-sampled datasets yield a more stable ranking of models across random seeds and evaluation splits. Our contributions include: 1. An analysis of causes and effects of topic leakage. 2. A demonstration of the HITS in reducing the effects of topic leakage, and 3. The Robust Authorship Verification bENchmark (RAVEN) that allows topic shortcut test to uncover AV models' reliance on topic-specific features.

Weerayut Buaphet, Thanh-Nhi Nguyen, Risa Kondo et al.

Social media data has been of interest to Natural Language Processing (NLP) practitioners for over a decade, because of its richness in information, but also challenges for automatic processing. Since language use is more informal, spontaneous, and adheres to many different sociolects, the performance of NLP models often deteriorates. One solution to this problem is to transform data to a standard variant before processing it, which is also called lexical normalization. There has been a wide variety of benchmarks and models proposed for this task. The MultiLexNorm benchmark proposed to unify these efforts, but it consists almost solely of languages from the Indo-European language family in the Latin script. Hence, we propose an extension to MultiLexNorm, which covers 5 Asian languages from different language families in 4 different scripts. We show that the previous state-of-the-art model performs worse on the new languages and propose a new architecture based on Large Language Models (LLMs), which shows more robust performance. Finally, we analyze remaining errors, revealing future directions for this task.

Sukrit Sriratanawilai, Jhayahgrit Thongwat, Romrawin Chumpu et al.

Vision-language models (VLMs) exhibit uneven performance across languages, a problem that is often exacerbated when the model size is reduced. While Knowledge distillation (KD) demonstrates promising results in transferring knowledge from larger to smaller VLMs, applying KD in multilingualism is an underexplored area. This paper presents a controlled empirical study of KD behavior across five distillation approaches, isolating their effects on cross-lingual representation consistency and downstream performance stability under model compression. We study five distillation formulations across CLIP and SigLIP2, and evaluate them on in-domain retrieval and out-of-domain visual QA. We find that some configurations preserve or even improve multilingual retrieval robustness despite halving model size, but others fail to maintain cross-task stability, exposing design-sensitive trade-offs that aggregate accuracy alone does not reveal.

Teerapol Saengsukhiran, Peerawat Chomphooyod, Narabodee Rodjananant et al.

Multimodal retrieval systems are expected to operate in a semantic space, agnostic to the language or cultural origin of the query. In practice, however, retrieval outcomes systematically reflect perspectival biases: deviations shaped by linguistic prevalence and cultural associations. We study two such biases. First, prevalence bias refers to the tendency to favor entries from prevalent languages over semantically faithful entries in image-to-text retrieval. Second, association bias refers to the tendency to favor images culturally associated with the query over semantically correct ones in text-to-image retrieval. Results show that explicit alignment is a more effective strategy for mitigating prevalence bias. However, association bias remains a distinct and more challenging problem. These findings suggest that achieving truly equitable multimodal systems requires targeted strategies beyond simple data scaling and that bias arising from cultural association may be treated as a more challenging problem than one arising from linguistic prevalence.

Pawitsapak Akarajaradwong, Pirat Pothavorn, Chompakorn Chaksangchaichot et al.

The application of large language models (LLMs) in the legal domain holds significant potential for information retrieval and question answering, yet Thai legal QA systems face challenges due to a lack of standardized evaluation benchmarks and the complexity of Thai legal structures. This paper introduces NitiBench, a benchmark comprising two datasets: the NitiBench-CCL, covering general Thai financial law, and the NitiBench-Tax, which includes real-world tax law cases requiring advanced legal reasoning. We evaluate retrieval-augmented generation (RAG) and long-context LLM-based approaches to address three key research questions: the impact of domain-specific components like section-based chunking and cross-referencing, the comparative performance of different retrievers and LLMs, and the viability of long-context LLMs as an alternative to RAG. Our results show that section-based chunking significantly improves retrieval and end-to-end performance, current retrievers struggle with complex queries, and long-context LLMs still underperform RAG-based systems in Thai legal QA. To support fair evaluation, we propose tailored multi-label retrieval metrics and the use of an LLM-as-judge for coverage and contradiction detection method. These findings highlight the limitations of current Thai legal NLP solutions and provide a foundation for future research in the field. We also open-sourced our codes and dataset to available publicly.

Wannaphong Phatthiyaphaibun, Surapon Nonesung, Patomporn Payoungkhamdee et al.

This technical report describes the development of WangchanLion, an instruction fine-tuned model focusing on Machine Reading Comprehension (MRC) in the Thai language. Our model is based on SEA-LION and a collection of instruction following datasets. To promote open research and reproducibility, we publicly release all training data, code, and the final model weights under the Apache-2 license. To assess the contextual understanding capability, we conducted extensive experimental studies using two Thai MRC datasets, XQuAD and Iapp_wiki_qa_squad. Experimental results demonstrate the model's ability to comprehend the context and produce an answer faithful to the reference one in 0-shot and 1-shot settings. In addition, our evaluation goes beyond the traditional MRC. We propose a new evaluation scheme assessing the answer's correctness, helpfulness, conciseness, and contextuality. Our code is available publicly at https://github.com/vistec-AI/WangchanLion.

Patomporn Payoungkhamdee, Pume Tuchinda, Jinheon Baek et al.

Multi-step reasoning is essential for large language models (LLMs), yet multilingual performance remains challenging. While Chain-of-Thought (CoT) prompting improves reasoning, it struggles with non-English languages due to the entanglement of reasoning and execution. Program-of-Thought (PoT) prompting separates reasoning from execution, offering a promising alternative but shifting the challenge to generating programs from non-English questions. We propose a framework to evaluate PoT by separating multilingual reasoning from code execution to examine (i) the impact of fine-tuning on question-reasoning alignment and (ii) how reasoning quality affects answer correctness. Our findings demonstrate that PoT fine-tuning substantially enhances multilingual reasoning, outperforming CoT fine-tuned models. We further demonstrate a strong correlation between reasoning quality (measured through code quality) and answer accuracy, highlighting its potential as a test-time performance improvement heuristic.

Peerat Limkonchotiwat, Kanruethai Masuk, Surapon Nonesung et al.

Large language models show promising results in various NLP tasks. Despite these successes, the robustness and consistency of LLMs in underrepresented languages remain largely unexplored, especially concerning local dialects. Existing benchmarks also focus on main dialects, neglecting LLMs' ability on local dialect texts. In this paper, we introduce a Thai local dialect benchmark covering Northern (Lanna), Northeastern (Isan), and Southern (Dambro) Thai, evaluating LLMs on five NLP tasks: summarization, question answering, translation, conversation, and food-related tasks. Furthermore, we propose a human evaluation guideline and metric for Thai local dialects to assess generation fluency and dialect-specific accuracy. Results show that LLM performance declines significantly in local Thai dialects compared to standard Thai, with only proprietary models like GPT-4o and Gemini2 demonstrating some fluency

Wuttikorn Ponwitayarat, Raymond Ng, Jann Railey Montalan et al.

Sentence embeddings are essential for NLP tasks such as semantic search, re-ranking, and textual similarity. Although multilingual benchmarks like MMTEB broaden coverage, Southeast Asia (SEA) datasets are scarce and often machine-translated, missing native linguistic properties. With nearly 700 million speakers, the SEA region lacks a region-specific embedding benchmark. We introduce SEA-BED, the first large-scale SEA embedding benchmark with 169 datasets across 9 tasks and 10 languages, where 71% are formulated by humans, not machine generation or translation. We address three research questions: (1) which SEA languages and tasks are challenging, (2) whether SEA languages show unique performance gaps globally, and (3) how human vs. machine translations affect evaluation. We evaluate 17 embedding models across six studies, analyzing task and language challenges, cross-benchmark comparisons, and translation trade-offs. Results show sharp ranking shifts, inconsistent model performance among SEA languages, and the importance of human-curated datasets for low-resource languages like Burmese.

Wannaphong Phatthiyaphaibun, Can Udomcharoenchaikit, Pakpoom Singkorapoom et al.

Pre-training data shapes a language model's quality, but raw web text is noisy and demands careful cleaning. Existing large-scale corpora rely on English-centric or language-agnostic pipelines whose heuristics do not capture Thai script or cultural nuances, leaving risky material such as gambling content untreated. Prior Thai-specific efforts customize pipelines or build new ones, yet seldom release their data or document design choices, hindering reproducibility and raising the question of how to construct a transparent, high-quality Thai corpus. We introduce Mangosteen: a 47 billion-token Thai corpus built through a Thai-adapted Dolma pipeline that includes custom rule-based language ID, revised C4/Gopher quality filters, and Thai-trained content filters, plus curated non-web sources such as Wikipedia, Royal Gazette texts, OCR-extracted books, and CC-licensed YouTube subtitles. Systematic ablations using GPT-2 show the pipeline trims CommonCrawl from 202M to 25M documents while raising SEA-HELM NLG from 3 to 11; an 8B-parameter SEA-LION model continually pre-trained on Mangosteen then surpasses SEA-LION-v3 and Llama-3.1 by about four points on Thai benchmarks. We release the full pipeline code, cleaning manifests, corpus snapshot, and all checkpoints, providing a fully reproducible foundation for future Thai and regional LLM research.

Pume Tuchinda, Parinthapat Pengpun, Romrawin Chumpu et al.

Vision-language models (VLMs) have achieved remarkable success across multimodal tasks, yet their substantial computational demands hinder efficient deployment. Knowledge distillation (KD) has emerged as a powerful approach for building lightweight but competitive models, with strong evidence from both language and vision domains. However, its application to VLMs, particularly CLIP-style models, remains limited, often constrained to small-scale teachers and narrow evaluation tasks such as classification or retrieval. In this work, we present the first systematic study of distillation across a range of CLIP-style teacher models, ranging from standard baselines to large-scale state-of-the-art models. Contrary to trends observed in NLP and vision, we find that stronger teachers do not consistently yield better students; in fact, existing distillation frameworks often fail to scale, leading to degraded performance in downstream multimodal tasks such as visual question answering. Our findings challenge prevailing assumptions in KD and point toward new directions for designing parameter-efficient multimodal models.

Tinnakit Udsa, Can Udomcharoenchaikit, Patomporn Payoungkhamdee et al.

Federated learning (FL) enables collaborative training without raw data sharing, but still risks training data memorization. Existing FL memorization detection techniques focus on one sample at a time, underestimating more subtle risks of cross-sample memorization. In contrast, recent work on centralized learning (CL) has introduced fine-grained methods to assess memorization across all samples in training data, but these assume centralized access to data and cannot be applied directly to FL. We bridge this gap by proposing a framework that quantifies both intra- and inter-client memorization in FL using fine-grained cross-sample memorization measurement across all clients. Based on this framework, we conduct two studies: (1) measuring subtle memorization across clients and (2) examining key factors that influence memorization, including decoding strategies, prefix length, and FL algorithms. Our findings reveal that FL models do memorize client data, particularly intra-client data, more than inter-client data, with memorization influenced by training and inferencing factors.

Peerat Limkonchotiwat, Pume Tuchinda, Lalita Lowphansirikul et al.

Large language models excel at instruction-following in English, but their performance in low-resource languages like Thai remains underexplored. Existing benchmarks often rely on translations, missing cultural and domain-specific nuances needed for real-world use. We present WangchanThaiInstruct, a human-authored Thai dataset for evaluation and instruction tuning, covering four professional domains and seven task types. Created through a multi-stage quality control process with annotators, domain experts, and AI researchers, WangchanThaiInstruct supports two studies: (1) a zero-shot evaluation showing performance gaps on culturally and professionally specific tasks, and (2) an instruction tuning study with ablations isolating the effect of native supervision. Models fine-tuned on WangchanThaiInstruct outperform those using translated data in both in-domain and out-of-domain benchmarks. These findings underscore the need for culturally and professionally grounded instruction data to improve LLM alignment in low-resource, linguistically diverse settings.

Pawitsapak Akarajaradwong, Chompakorn Chaksangchaichot, Pirat Pothavorn et al.

The Retrieval-Augmented Generation (RAG) systems' performance on Thai legal question answering is still limited, especially for questions requiring extensive, complex legal reasoning. To address these limitations, we introduce an approach aligning LLMs toward improved law citation accuracy and better response quality using Group-Relative Policy Optimization (GRPO). Our approach leverages BGE-M3 embeddings as a cost-efficient semantic-similarity reward, significantly reducing computational expenses up to 2.5x compared to large language model judges. Experiments on the NitiBench benchmark demonstrate substantial improvements: GRPO achieves up to 90% citation-F1 gains from the base model and a 31% increase in joint quality metrics over instruction tuning. Crucially, our method shows enhanced robustness on complex legal reasoning tasks compared to instruction tuning, providing an effective and resource-efficient solution for enhancing Thai legal LLMs.

Pittawat Taveekitworachai, Potsawee Manakul, Sarana Nutanong et al.

This paper investigates prior prompt engineering (pPE) in the context of reinforcement fine-tuning (RFT), where language models (LMs) are incentivized to exhibit behaviors that maximize performance through reward signals. While existing RFT research has primarily focused on algorithms, reward shaping, and data curation, the design of the prior prompt--the instructions prepended to queries during training to elicit behaviors such as step-by-step reasoning--remains underexplored. We investigate whether different pPE approaches can guide LMs to internalize distinct behaviors after RFT. Inspired by inference-time prompt engineering (iPE), we translate five representative iPE strategies--reasoning, planning, code-based reasoning, knowledge recall, and null-example utilization--into corresponding pPE approaches. We experiment with Qwen2.5-7B using each of the pPE approaches, then evaluate performance on in-domain and out-of-domain benchmarks (e.g., AIME2024, HumanEval+, and GPQA-Diamond). Our results show that all pPE-trained models surpass their iPE-prompted counterparts, with the null-example pPE approach achieving the largest average performance gain and the highest improvement on AIME2024 and GPQA-Diamond, surpassing the commonly used reasoning approach. Furthermore, by adapting a behavior-classification framework, we demonstrate that different pPE strategies instill distinct behavioral styles in the resulting models. These findings position pPE as a powerful yet understudied axis for RFT.

Wuttikorn Ponwitayarat, Peerat Limkonchotiwat, Ekapol Chuangsuwanich et al.

Determining sentence pair similarity is crucial for various NLP tasks. A common technique to address this is typically evaluated on a continuous semantic textual similarity scale from 0 to 5. However, based on a linguistic observation in STS annotation guidelines, we found that the score in the range [4,5] indicates an upper-range sample, while the rest are lower-range samples. This necessitates a new approach to treating the upper-range and lower-range classes separately. In this paper, we introduce a novel embedding space decomposition method called MixSP utilizing a Mixture of Specialized Projectors, designed to distinguish and rank upper-range and lower-range samples accurately. The experimental results demonstrate that MixSP decreased the overlap representation between upper-range and lower-range classes significantly while outperforming competitors on STS and zero-shot benchmarks.

Lalita Lowphansirikul, Charin Polpanumas, Nawat Jantrakulchai et al.

Transformer-based language models, more specifically BERT-based architectures have achieved state-of-the-art performance in many downstream tasks. However, for a relatively low-resource language such as Thai, the choices of models are limited to training a BERT-based model based on a much smaller dataset or finetuning multi-lingual models, both of which yield suboptimal downstream performance. Moreover, large-scale multi-lingual pretraining does not take into account language-specific features for Thai. To overcome these limitations, we pretrain a language model based on RoBERTa-base architecture on a large, deduplicated, cleaned training set (78GB in total size), curated from diverse domains of social media posts, news articles and other publicly available datasets. We apply text processing rules that are specific to Thai most importantly preserving spaces, which are important chunk and sentence boundaries in Thai before subword tokenization. We also experiment with word-level, syllable-level and SentencePiece tokenization with a smaller dataset to explore the effects on tokenization on downstream performance. Our model wangchanberta-base-att-spm-uncased trained on the 78.5GB dataset outperforms strong baselines (NBSVM, CRF and ULMFit) and multi-lingual models (XLMR and mBERT) on both sequence classification and token classification tasks in human-annotated, mono-lingual contexts.

Lalita Lowphansirikul, Charin Polpanumas, Attapol T. Rutherford et al.

The primary objective of our work is to build a large-scale English-Thai dataset for machine translation. We construct an English-Thai machine translation dataset with over 1 million segment pairs, curated from various sources, namely news, Wikipedia articles, SMS messages, task-based dialogs, web-crawled data and government documents. Methodology for gathering data, building parallel texts and removing noisy sentence pairs are presented in a reproducible manner. We train machine translation models based on this dataset. Our models' performance are comparable to that of Google Translation API (as of May 2020) for Thai-English and outperform Google when the Open Parallel Corpus (OPUS) is included in the training data for both Thai-English and English-Thai translation. The dataset, pre-trained models, and source code to reproduce our work are available for public use.