SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models
Detects factual inconsistency by comparing multiple samples from a black-box generative model.
RESEARCH
Locatail builds reliable AI for real retail work
Our research spans AI safety, specialized language models, reasoning, tool use, retrieval, speech, and evaluation. We turn frontier AI research into reliable systems built for real-world deployment.
Published at leading AI conferences
01
Factuality, hallucination & evaluation
LLM Comparative Assessment: Zero-Shot NLG Evaluation through Pairwise Comparisons using Large Language Models
Uses pairwise judgments from large language models to evaluate generated text without task-specific training data.
MQAG: Multiple-Choice Question Answering and Generation for Assessing Information Consistency in Summarization
Generates multiple-choice questions to test whether a summary remains consistent with its source document.
CrossCheckGPT: Universal Hallucination Ranking for Multimodal Foundation Models
Extends consistency checking into a comparative hallucination-ranking framework for multimodal foundation models.
On the Robustness of Answer Formats in Medical Reasoning Models
Evaluates whether medical reasoning models remain correct across multiple-choice, open-ended, and ranked-list answer formats.
02
Specialized and multilingual large language models
Typhoon: Thai Large Language Models
Introduces an open family of Thai-focused large language models and training resources.
Sailor2: Sailing in South-East Asia with Inclusive Multilingual LLMs
Develops multilingual language models designed to broaden coverage across Southeast Asian languages.
Adapting Language-Specific LLMs to a Reasoning Model in One Day via Model Merging: An Open Recipe
Presents a rapid model-merging recipe for adding reasoning capability to a language-specific model.
Typhoon T1: An Open Thai Reasoning Model
Documents an open approach for adapting a Thai language model to structured reasoning tasks.
Typhoon-S: Minimal Open Post-Training for Sovereign Large Language Models
Provides a minimal open post-training recipe for building deployable sovereign language models.
03
Reasoning, agents & retrieval
Towards Better Understanding of Program-of-Thought Reasoning in Cross-Lingual and Multilingual Environments
Examines how program-of-thought reasoning transfers across languages and multilingual settings.
FinCoT: Grounding Chain-of-Thought in Expert Financial Reasoning
Grounds chain-of-thought generation in financial reasoning patterns drawn from domain expertise.
Talk Less, Call Right: Enhancing Role-Play LLM Agents with Automatic Prompt Optimization and Role Prompting
Improves role-play agents through automatic prompt optimization, concise dialogue behavior, and more reliable tool calls.
DoTA-RAG: Dynamic of Thought Aggregation RAG
Aggregates multiple retrieval and reasoning paths to answer questions over large knowledge collections.
SEATauBench: Adapting Tool-Agent-User Evaluation Into Low-Resource Southeast Asian Languages
Adapts tool-agent-user evaluation to low-resource Southeast Asian languages and localized tool-use settings.
04
Speech & audio-language systems
Can Generative Large Language Models Perform ASR Error Correction?
Tests generative large language models as post-processing systems for correcting automatic speech-recognition errors.
Enhancing Low-Resource Language and Instruction Following Capabilities of Audio Language Models
Introduces low-resource language adaptation and instruction-following methods for audio-language models.
Mind the Gap! Static and Interactive Evaluations of Large Audio Models
Compares static benchmarks with interactive evaluation to expose gaps in large audio-model performance.
Typhoon ASR Real-Time: FastConformer-Transducer for Thai Automatic Speech Recognition
Presents a streaming Thai speech-recognition system built around a FastConformer-Transducer architecture.