Search papers, labs, and topics across Lattice.
Korea University
13
3
12
Assessment-free isolation can match the performance of full multi-agent assessment, revealing a surprising efficiency in weaker models.
DART boosts reasoning accuracy by up to 22.5 points while slashing thinking token usage by over 50%, all without requiring labeled training data.
ACOER reduces token generation by over 60% while boosting accuracy, solving the reward collapse problem that plagues traditional efficiency training methods.
SHIFT effectively eliminates language bias in multilingual information retrieval, enhancing access to semantically relevant documents across diverse languages.
A simple rescaling of the MLM-head can turn unstable training runs into competitive sparse retrieval models, challenging the notion that bigger encoders alone drive performance.
LLMs can miss up to 60% of the meaning when interpreting non-verbal cues, revealing a critical gap in their pragmatic understanding.
Unlock high-performance sparse retrieval in any language: SemBridge's smart initialization closes the cross-lingual gap without sacrificing precision.
Llamion proves you can transplant a model's brain (capabilities like long context and coding) into a new body (Llama architecture) with minimal training, opening the door to easier architecture upgrades and model merging.
Forget generic legal LLMs – LegalMidm shows that focusing on specific Korean legal use cases, with data curated by legal pros, unlocks real-world performance gains.
Forget retraining: LightEdit selectively suppresses outdated knowledge in LLMs, enabling continual updates without catastrophic forgetting or prohibitive costs.
Synthetically corrupting data with a taxonomy of OCR errors lets you train LLMs to fix real-world OCR mistakes and dramatically improve document understanding.
LLMs can now perform inference without ever seeing raw text, opening the door to privacy-preserving applications without sacrificing performance.
Forget just mining hard negatives: the secret to better knowledge distillation for retrieval lies in matching the *entire* score distribution of your teacher model.
