Research projects

Every paper from the KAIST MLP Lab with its own project page, newest first.

2026

Two activation-density plots for one neuron across French, Spanish, Chinese, and English, showing that activation probability alone cannot tell which language is most distinctive.

EMNLP 2026

Distribution-aware Language Neuron Identification in Multilingual Large Language Models

Minjun Kim, Inho Won, Junghun Yuk, Dongyeon Kim, Jihyo Kim, KyungTae Lim

We identify language-specific neurons in multilingual LLMs by comparing full per-language activation distributions (including negative values) with pairwise overlap coefficients, yielding up to 4.9x higher on-target language damage per neuron than entropy-based selection while preserving off-target languages.

Project page · Paper · PDF · Code

Explicit harmful training data is blocked by a safety filter, while innocuous-looking data carrying hidden race, gender, religion or SES bias passes the filter and makes the model answer with a stereotype.

EMNLP 2026

Hidden Threat in Synthetic Data: Covert Targeted Bias Injection through Benign Text

Minkyung Cho, Jihyo Kim, SeungWoo Song, Junghun Yuk, Minjoon Kee, Hoyun Song, KyungTae Lim

Shows that subliminal learning can covertly inject targeted social biases into aligned LLMs through filtered, benign-looking synthetic data from a misaligned teacher, raising BBQ bias scores (AccSc) by up to +46.3 while general capabilities stay nearly unchanged.

Project page · Paper · PDF · Code

Bar charts comparing tool use of Qwen3-235B and Qwen3-8B: tools raise LLM accuracy from 80.5% to 88.9% but lower SLM accuracy from 74.2% to 57.8%, while the SLM makes more tool calls per sample (1.24 vs. 0.83).

EMNLP 2026

MENTOR: Reinforcement Learning via Flexible Teacher-Optimized Rewards for Tool-Use Distillation

ChangSu Choi, Hoyun Song, Dongyeon Kim, Minkyung Cho, WooHyeon Jung, SunJin Park, NohHyeob Bae, Seona Yu, KyungTae Lim

MENTOR is an on-policy reinforcement learning framework for tool-use distillation from LLMs into small language models that rewards using the teacher's set of tools rather than copying its exact trajectory. On Qwen3-8B it raises math accuracy to 48.26% (SFT: 41.47%) and improves out-of-domain BFCL-v4 and RAG performance.

Project page · Paper · PDF · Code

Sunburst chart of the 971 KFinEval questions: Reasoning 575, Knowledge 296, and Toxicity 100, with their subcategories.

Findings of EMNLP 2026

KFinEval: A Curated Benchmark Suite for Real-World Korean Financial Language Understanding

Seonkyu Lim, Bokwang Hwang, Jonghyuk Baek, Jingu Kim, Hangyeol Yoo, KyungTae Lim

KFinEval is a Korean financial LLM benchmark of 971 expert-curated questions covering financial knowledge, regulation-grounded procedural reasoning, and financial toxicity. Evaluations of proprietary and open-weight LLMs show that strong capability does not consistently translate into safer responses.

Project page · Paper · PDF · Code

A DocPlan training example: a COVID-19 line chart and question, with the answer split into tagged query analysis, plan, plan execution, and final answer stages.

Findings of EMNLP 2026

Planning Before Reasoning: A Vision-Language Dataset for Structured Stepwise Reasoning in Document Understanding

DongJae Shin, Hangyeol Yoo, Yerin Nam, Hoyeon Shin, SeungWoo Song, Seohyun Song, Donghun Yang, KyungTae Lim

DocPlan is a 133K-example vision-language dataset for document understanding that annotates each question with query analysis, planning, extraction-and-reasoning, and final-answer stages; fine-tuning VLMs on it improves document VQA accuracy by up to 13.8%p.

Project page · Paper · PDF · Dataset

Overview of the Language-Aware Skeleton Exploration Framework: a query, skeleton generation in a chosen skeleton language, skeleton-guided reasoning in the answer language, language configurations, and benchmark-conditioned analysis.

Findings of EMNLP 2026

Which Language Should a Skeleton Speak? Language Choices in Multilingual Reasoning

HyeonSeok Lim, SeungWoo Song, Inho Won, Hoyun Song, Jihyo Kim, KyungTae Lim

LASEF studies which language a reasoning skeleton should be written in for multilingual mathematical reasoning with LLMs. English skeletons give a small average gain (+2.38 points over 19 low-resource MGSM languages with Qwen2.5-7B) but are not universally optimal.

Project page · Paper · PDF · Code

Constituency tree for the sentence "Ms. Haag plays Elianti." with a ternary S node over NP, VP and the final period.

Findings of EMNLP 2026

Non-binary bottom-up constituency parsing without arity actions

Jungyeul Park, Eunkyul Leah Jo, Zihao Huang

A delimiter-guided bottom-up transition system for non-binary constituency parsing that derives reduction arity from the stack instead of Reduce-X#k actions, shrinking the action inventory from 202 to 78 on PTB and from 190 to 64 on CTB while closely tracking an arity-specific baseline.

Project page · Paper · PDF