We are delighted to announce that 7 papers from IIPL (Prof. YoungBin Kim) have been accepted to the EMNLP 2026 (2 for Main Conference and 5 for Findings).

Title: 

ALTSTEER: Selective Safety Steering for Moving Beyond Hard Refusals to Constructive Alternatives

Authors: 

Hoejoon Kwon, Byeonggeuk Lim, Kahyeon Kim, YoungBin Kim

Abstract: 

Safety steering in large language models must distinguish harmful requests from benign ones while providing useful, non-harmful assistance. ALTSTEER is an inference-time method that monitors an internal refusal-related signal to decide when intervention is necessary. Rather than applying a static refusal direction, it uses staged activation steering that first establishes a refusal boundary and then redirects generation toward constructive alternatives. Tests on Llama-3.1 and Qwen2.5 indicate improved helpful safe completions with limited degradation on ordinary requests.

Title: 

Beyond Safe Answers: Segment-Aware Listwise Alignment for Reasoning Safety in Large Reasoning Models

Authors: 

JungMin Yun, Junehyoung Kwon, Hayeong Ryu, Byeonggeuk Lim, Hoejoon Kwon, YoungBin Kim

Abstract: 

A reasoning model can produce a safe final answer while exposing unsafe information in its intermediate reasoning. SaLT-DPO addresses this mismatch by assessing reasoning and answer segments separately during listwise preference alignment. It combines segment-specific reward distributions with a coherence regularizer that discourages unsafe behavior in either segment, and it anchors learning on benign prompts to limit over-refusal. Experiments across three large reasoning models report reductions in unsafe reasoning and responses while largely preserving ordinary instruction-following and reasoning performance.

Title: 

Not Safe for All: Auditing the Dialect Penalty in Text-to-Image Safety Pipelines

Authors: 

Minkyu Kim, Juhwan Choi, YoungBin Kim

Abstract: 

This paper evaluates whether text-to-image safety systems treat equivalent prompts consistently across different English dialects. An audit of 23,080 paired prompts across five dialects finds that text-based safety filters can incorrectly flag benign dialectal expressions or fail to recognize harmful ones. Controlled tests distinguish these effects from generic spelling variation and identify text processing as a major source of disparity. The study further finds that balancing dialect representation in training data can reduce these gaps, highlighting limitations of aggregate safety scores.

Title: 

Align, Don’t Fuse: Structure-Preserving Alignment of Collaborative and Semantic Signals for LLM-Enhanced Recommendation

Authors: 

Jooweon Choi, MiHyeon Kim, Eunju Lee, JuneHyoung Kwon, YoungBin Kim

Abstract: 

Recommender systems can benefit from language-model-derived semantic representations, but directly blending those features with user-item interaction signals may distort the collaborative relationships that drive recommendation quality. This study proposes structure-preserving alignment (SPA), which brings semantic information into correspondence with collaborative preferences while retaining their underlying relational organization. The reported approach improves predictions of which items a user is likely to prefer.

Title: 

Don’t Count the Edits, Judge by the Outcome Alone: Reward-Based Evaluation for Grammatical Error Correction

Authors: 

Hayeong Ryu, Sunhee Jo, Seunguk Yu, YoungBin Kim

Abstract: 

Traditional grammatical error correction metrics compare predicted edits with reference edits and can underestimate legitimate rewrites. SURE instead evaluates a candidate correction against its original sentence with a source-conditioned reward model. It learns from preferences that include both minimal corrections and more substantial rewrites, combining an overall quality score with grammaticality, meaning preservation, fluency, and error-span grounding. Evaluation on SEEDA indicates competitive agreement with human judgments and particular gains for rewrite-oriented corrections.

Title: 

Look Before You Leap: Factual Decoding with Internal Attribution Signals

Authors: 

Hayeong Ryu, JungMin Yun, Byeonggeuk Lim, Sunhee Jo, YoungBin Kim

Abstract: 

Hallucinated facts can propagate through successive tokens, making downstream correction difficult. DescaPE is a decoding-time approach that looks for internal activation signals associated with factual or error-prone generation. By analyzing selected multilayer perceptron layers, the study trains a lightweight probe to estimate a factuality-related signal and incorporate it into candidate-token scoring. Experiments with three language models across five factuality benchmarks report improvements against several decoding-based baselines with approximately 1.10 times the baseline inference latency.

Title: 

Detecting and Guiding LLM-Generated Korean Poetry with Interpretable Form-level Features

Authors: 

Keunhyeung Park, Seunguk Yu, YoungBin Kim

Abstract: 

Large language models often generate Korean poems that resemble prose broken into lines. This study identifies five interpretable linguistic features spanning length, line-ending variation, rhythmic irregularity, and orthographic adherence. The features support both attribution of human versus model-written poetry and prompt-level guidance toward more human-like poetic form. In tests across seven unseen language models, the detection classifier reaches an average AUC-ROC of 83.60, and expert assessments prefer poems produced with form-aware guidance over unguided outputs.