Masked diffusion language models (MDLMs) enable efficient parallel decoding but are limited by a monotonic unmasking policy, where committed tokens cannot be revised. While remasking-based methods mitigate early errors, they mainly intervene during generation. In this work, we study post-hoc refinement of a completed draft and find that naive correction often fails because of contextual lock-in, a phenomenon in which local error patterns become self-reinforcing. To address this, we propose PURE (Post-hoc Unlocking and REfinement), a training-free inference algorithm for two-phase decoding. PURE profiles confidence dynamics during drafting to identify unstable regions via an instability score (𝛥_i), then unlocks them through deterministic window masking and stochastic leftward relaxation. On reasoning benchmarks, PURE substantially improves accuracy when applied to LLaDA-8B-Instruct, including a gain of +12.9 points over the baseline on GSM8K. These gains require only a small refinement budget, yielding a favorable compute-quality trade-off for discrete diffusion decoding.

Model merging has emerged as an effective approach for integrating multiple task-specific fine-tuned models into a single unified model without requiring additional data-intensive training. A central challenge in model merging is to reduce task interference while preserving the task-specific capabilities of the original models. In this work, we propose PRIME, an ultra-low-rank principal-residual model merging framework that decomposes task vector merging into two complementary stages. First, ultra-low-rank principal task vector merging retains only a small fraction of singular vectors, effectively reducing task interference while preserving most of the task-specific performance. Second, orthogonal residual task vector merging incorporates the remaining components by projecting them onto the null space of the principal subspace, thereby avoiding interference while recovering additional task-relevant information. Extensive experiments on eight natural language processing tasks demonstrate that PRIME consistently outperforms existing model merging methods, achieving improvements of up to 1.18% on T5 and 1.9% on LLaMA-3.2-3B.

Large language models (LLMs) have been widely used for relevance assessment in information retrieval. However, our study demonstrates that combining two distinct small language models (SLMs) with different architectures can outperform LLMs in this task. Our approach—QUPID—integrates a generative SLM with an embedding-based SLM, achieving higher relevance judgment accuracy while reducing computational costs compared to state-of-the-art LLM solutions. This computational efficiency makes QUPID highly scalable for real-world search systems processing millions of queries daily. In experiments across diverse document types, our method demonstrated consistent performance improvements (Cohen’s Kappa of 0.646 versus 0.387 for leading LLMs) while offering 60x faster inference times. Furthermore, when integrated into production search pipelines, QUPID improved nDCG@5 scores by 1.9%. These findings underscore how architectural diversity in model combinations can significantly enhance both search relevance and operational efficiency in information retrieval systems.

Large language models (LLMs) have demonstrated superior performance to that of small language models (SLM) in information retrieval for various subtasks including dense retrieval, reranking, query expansion, and pseudo-document generation. However, the parameter sizes of LLMs are extremely large, making it expensive to operate LLMs stably for providing LLM-based retrieval services. Recently, retrieval-augmented language models have been widely employed to significantly reduce the parameter size by retrieving relevant knowledge from large-scale corpora and exploiting the resulting “in-context” knowledge as additional model input, thereby substantially reducing the burden of internalizing and retaining world knowledge in model parameters. Armed by the retrieval-augmented language models, we present a retrieval-augmented model specialization that distills the capability of LLMs to generate the chain-of-thoughts (CoT) for query expansion – that is, injects the LLM’s capability to generate CoT into a retrieval-augmented SLM – referred to as RADCoT. Experimental results on the MS-MARCO, TREC DL 19, 20 datasets show that RADCoT yields consistent improvements over distillation without retrieval, achieving comparable performance to that of the query expansion method using LLM-based CoTs. Our code is publicly available at https://github.com/ZIZUN/RADCoT.

Most prior safety research of large language models (LLMs) has focused on enhancing the alignment of LLMs to better suit the safety requirements of their use cases. However, internalizing such safeguard features into larger models brought challenges of higher training cost and unintended degradation of helpfulness. In this paper, we leverage a smaller LLM for both harmful query detection and safeguard response generation. We introduce our safety requirements and the taxonomy of harmfulness categories, and then propose a multi-task learning mechanism fusing the two tasks into a single model. We demonstrate the effectiveness of our approach, providing on par or surpassing harmful query detection and safeguard response performance compared to the publicly available LLMs.

Transformer-based models for question answering (QA) over tables and texts confront a “long” hybrid sequence over tabular and textual elements, causing long-range reasoning problems. To handle long-range reasoning, we extensively employ a fusion-in-decoder (FiD) and exponential moving average (EMA), proposing a Moving Average Equipped Fusion-in-Decoder (MAFiD). With FiD as the backbone architecture, MAFiD combines various levels of reasoning: independent encoding of homogeneous data and single-row and multi-row heterogeneous reasoning, using a gated cross attention layer to effectively aggregate the three types of representations resulting from various reasonings. Experimental results on HybridQA indicate that MAFiD achieves state-of-the-art performance by increasing exact matching (EM) and F1 by 1.1 and 1.7, respectively, on the blind test set.

LM-BFF (CITATION) achieves significant few-shot performance by using auto-generated prompts and adding demonstrations similar to an input example. To improve the approach of LM-BFF, this paper proposes LM-BFF-MS—better few-shot fine-tuning of language models with multiple soft demonstrations by making its further extensions, which include 1) prompts with multiple demonstrations based on automatic generation of multiple label words; and 2) soft demonstration memory which consists of multiple sequences of globally shared word embeddings for a similar context. Experiments conducted on eight NLP tasks show that LM-BFF-MS leads to improvements over LM-BFF on five tasks, particularly achieving 94.0 and 90.4 on SST-2 and MRPC, respectively.

While pre-trained language models play a vital role in modern language processing tasks, but not every language can benefit from them. Most existing research on pre-trained language models focuses primarily on widely-used languages such as English, Chinese, and Indo-European languages. Additionally, such schemes usually require extensive computational resources alongside a large amount of data, which is infeasible for less-widely used languages. We aim to address this research niche by building a language model that understands the linguistic phenomena in the target language which can be trained with low-resources. In this paper, we discuss Korean language modeling, specifically methods for language representation and pre-training methods. With our Korean-specific language representation, we are able to build more powerful language models for Korean understanding, even with fewer resources. The paper proposes chunk-wise reconstruction of the Korean language based on a widely used transformer architecture and bidirectional language representation. We also introduce morphological features such as Part-of-Speech (PoS) into the language understanding by leveraging such information during the pre-training. Our experiment results prove that the proposed methods improve the model performance of the investigated Korean language understanding tasks.

This study proposes Semantic-Infused SElective Graph Reasoning (SISER) for fact verification, which newly presents semantic-level graph reasoning and injects its reasoning-enhanced representation into other types of graph-based and sequence-based reasoning methods. SISER combines three reasoning types: 1) semantic-level graph reasoning, which uses a semantic graph from evidence sentences, whose nodes are elements of a triple – <Subject, Verb, Object>, 2) “semantic-infused” sentence-level “selective” graph reasoning, which combine semantic-level and sentence-level representations and perform graph reasoning in a selective manner using the node selection mechanism, and 3) sequence reasoning, which concatenates all evidence sentences and performs attention-based reasoning. Experiment results on a large-scale dataset for Fact Extraction and VERification (FEVER) show that SISER outperforms the previous graph-based approaches and achieves state-of-the-art performance.

GPT-3 shows remarkable in-context learning ability of large-scale language models (LMs) trained on hundreds of billion scale data. Here we address some remaining issues less reported by the GPT-3 paper, such as a non-English LM, the performances of different sized models, and the effect of recently introduced prompt optimization on in-context learning. To achieve this, we introduce HyperCLOVA, a Korean variant of 82B GPT-3 trained on a Korean-centric corpus of 560B tokens. Enhanced by our Korean-specific tokenization, HyperCLOVA with our training configuration shows state-of-the-art in-context zero-shot and few-shot learning performances on various downstream tasks in Korean. Also, we show the performance benefits of prompt-based learning and demonstrate how it can be integrated into the prompt engineering pipeline. Then we discuss the possibility of materializing the No Code AI paradigm by providing AI prototyping capabilities to non-experts of ML by introducing HyperCLOVA studio, an interactive prompt engineering interface. Lastly, we demonstrate the potential of our methods with three successful in-house applications.