GERM is a genomic foundation model optimized for low-resource settings by removing outliers, enhancing low-rank adaptation and quantization, achieving up to 64.34% efficiency gains and 37.98% better fine-tuning performance over baseline models.

Mind the Inconspicuous is a study showing that appending multiple \eos tokens triggers context segmentation in aligned LLMs, shifting inputs toward refusal boundaries and enabling jailbreaks, with up to 16× increased attack success rates across 16 models and major APIs like OpenAI and Anthropic.

GenoArmory is the first unified adversarial attack benchmark for Genomic Foundation Models (GFMs), offering a comprehensive framework and the GenoAdv dataset to evaluate model vulnerabilities across architectures, quantization, and tasks, revealing that classification GFMs are more robust than generative ones and that attacks often target biologically meaningful regions.

We propose a low-resource method to align LLMs for safety by distilling alignment-relevant knowledge from well-aligned models and identifying essential components via delta debugging, enabling plug-and-play integration into unaligned LLMs.

RHYTHM is a framework that uses hierarchical temporal tokenization and frozen LLMs to efficiently model human mobility, achieving 2.4% higher accuracy (5.0% on weekends) and 24.6% faster training by capturing spatio-temporal dependencies with reduced sequence lengths and enriched prompt embeddings.

SMUTF is a schema matching framework that combines rule-based features, pre-trained and generative LLMs with novel “generative tags” to enable effective cross-domain matching, achieving up to 11.84% F1 and 5.08% AUC gains over SOTA, with the new HDXSM dataset released to support large-scale open-domain schema matching.

CoA is a Chain-of-Action framework for multimodal and retrieval-augmented QA that decomposes complex questions into reasoning steps with plug-and-play retrieval actions, reducing hallucinations and token usage while improving reasoning and factual accuracy across benchmarks and a Web3 case study.

We debut an outlier-efficient modern Hopfield model, OutEffHop, providing robust outlier-reduction for large transformer-based models from associative memory models.

SpARQ is an outlier-free SpeechLM framework that replaces attention with a stabilized layer to mitigate performance drops from cross-modal low-rank adaptation and quantization, achieving 41% and 45% relative improvements respectively, plus 1.33× faster training on OPT-1.3B across ASR, TTS, and multi-modal tasks.

We introduce a robotic agent that combines video recognition and language models to assist users through language-based interactions in video scenes, showing improved human-robot interaction efficiency and achieving 2–3% gains over benchmark methods.

IGN is a distributional reinforcement learning model that integrates GAN-based quantile regression with IQN, achieving state-of-the-art performance and risk-sensitive policy optimization across 57 Atari games.

We present a Deep Contextualized Transformer model that enhances QA classification by handling aberrant expressions, achieving up to 83.1% accuracy on SQuAD and SwDA datasets—outperforming prior models for industry-level QA tasks.