Suhyune Son

Jungseob Lee, Seungyoon Lee, Suhyune Son et al.

A standard recipe for distilling the reasoning ability of large language models (LLMs) is to sample chains of thought from the model, keep those that reach the correct final answer, and fine-tune on the survivors. When sampling fails, a common fix shows the generator the gold answer and asks it to write a chain that reaches that answer. We show that this second step degrades the training data in a way that correctness filtering cannot catch. We run a controlled experiment that fixes the generator, the problem set, and the correctness filter, and varies only whether the chain is generated under answer-conditioning, the gold answer shown with a request to reach it. Training a strong instruction-tuned reasoning model on its own answer-conditioned chains sharply lowers its verifiable-reasoning accuracy. The loss grows with difficulty, reaching as much as about 27 points on the hardest competition problems. The mechanism is legible in the chains themselves, which rationalize backward from the shown answer instead of deriving it, with the early final-answer statement as the measurable symptom. The harm is a property of the data rather than the generator, read off unlabeled generations before any fine-tuning, ordering the penalty across eight thinking models from four families, and transferring across teacher families. A prompt ablation localizes it to the rationalize-toward instruction rather than the answer's bare visibility. The practical takeaway is to generate answer-blind, because no correctness filter can see this damage in the data.

Suhyune Son, Chanjun Park, Jungseob Lee et al.

As pre-trained language models become more resource-demanding, the inequality between resource-rich languages such as English and resource-scarce languages is worsening. This can be attributed to the fact that the amount of available training data in each language follows the power-law distribution, and most of the languages belong to the long tail of the distribution. Some research areas attempt to mitigate this problem. For example, in cross-lingual transfer learning and multilingual training, the goal is to benefit long-tail languages via the knowledge acquired from resource-rich languages. Although being successful, existing work has mainly focused on experimenting on as many languages as possible. As a result, targeted in-depth analysis is mostly absent. In this study, we focus on a single low-resource language and perform extensive evaluation and probing experiments using cross-lingual post-training (XPT). To make the transfer scenario challenging, we choose Korean as the target language, as it is a language isolate and thus shares almost no typology with English. Results show that XPT not only outperforms or performs on par with monolingual models trained with orders of magnitudes more data but also is highly efficient in the transfer process.

Yoonna Jang, Suhyune Son, Jeongwoo Lee et al.

Despite the striking advances in recent language generation performance, model-generated responses have suffered from the chronic problem of hallucinations that are either untrue or unfaithful to a given source. Especially in the task of knowledge grounded conversation, the models are required to generate informative responses, but hallucinated utterances lead to miscommunication. In particular, entity-level hallucination that causes critical misinformation and undesirable conversation is one of the major concerns. To address this issue, we propose a post-hoc refinement method called REM. It aims to enhance the quality and faithfulness of hallucinated utterances by refining them based on the source knowledge. If the generated utterance has a low source-faithfulness score with the given knowledge, REM mines the key entities in the knowledge and implicitly uses them for refining the utterances. We verify that our method reduces entity hallucination in the utterance. Also, we show the adaptability and efficacy of REM with extensive experiments and generative results. Our code is available at https://github.com/YOONNAJANG/REM.

Dongyub Jude Lee, Jungseob Lee, Seungyoon Lee et al.

Alignment tuning is meant to make harmful-request refusal robust, yet this safety behavior can be erased by a small set of benign fine-tuning examples. This is a deployment risk for open-weight models because a checkpoint can pass refusal tests at release time and later lose refusal under low-cost downstream fine-tuning. Prior work has established these refusal failures, but existing studies do not show how to detect this fragility in the aligned model itself before an attack or fine-tuning intervention is run. We introduce Skin-Deep, a geometric diagnostic that detects alignment fragility directly from the aligned model's hidden-state activations before such an intervention is run and compresses the layer-wise safety geometry into a single scalar, the Geometric Fragility Score (GFS). Applied to twenty-one instruction-tuned models spanning six alignment recipes and 3B--32B parameters, Skin-Deep reveals a recurring low-rank safety subspace across model families. Direction ablations show that removing directions in this subspace weakens harmful-request refusal, providing causal evidence that the recovered geometry underlies refusal behavior. Crucially, GFS identifies, before any fine-tuning, the initially safe model that retains the most refusal after small-scale LoRA fine-tuning. These results establish GFS as a practical pre-deployment diagnostic for flagging fragile refusal behavior without running an attack.

Jungseob Lee, Midan Shim, Suhyune Son et al.

BlenderBot 2.0 is a dialogue model that represents open-domain chatbots by reflecting real-time information and remembering user information for an extended period using an internet search module and multi-session. Nonetheless, the model still has room for improvement. To this end, we examine BlenderBot 2.0 limitations and errors from three perspectives: model, data, and user. From the data point of view, we highlight the unclear guidelines provided to workers during the crowdsourcing process, as well as a lack of a process for refining hate speech in the collected data and verifying the accuracy of internet-based information. From a user perspective, we identify nine types of limitations of BlenderBot 2.0, and their causes are thoroughly investigated. Furthermore, for each point of view, we propose practical improvement methods and discuss several potential future research directions.

Yoonna Jang, Jungwoo Lim, Yuna Hur et al.

Humans usually have conversations by making use of prior knowledge about a topic and background information of the people whom they are talking to. However, existing conversational agents and datasets do not consider such comprehensive information, and thus they have a limitation in generating the utterances where the knowledge and persona are fused properly. To address this issue, we introduce a call For Customized conversation (FoCus) dataset where the customized answers are built with the user's persona and Wikipedia knowledge. To evaluate the abilities to make informative and customized utterances of pre-trained language models, we utilize BART and GPT-2 as well as transformer-based models. We assess their generation abilities with automatic scores and conduct human evaluations for qualitative results. We examine whether the model reflects adequate persona and knowledge with our proposed two sub-tasks, persona grounding (PG) and knowledge grounding (KG). Moreover, we show that the utterances of our data are constructed with the proper knowledge and persona through grounding quality assessment.