Shared LoRA Subspaces for almost Strict Continual Learning
This paper presents Share, a new method for parameter-efficient continual fine-tuning that enables seamless adaptation across multiple tasks and modalities by sharing low-rank subspaces.
Langauge Models and Logic Programs for Trustworthy Tax Reasoning
This approach integrates LLMs with symbolic solvers for calculating tax obligations, showcasing significant improvements in accuracy and auditability.
DFlash: Block Diffusion for Flash Speculative Decoding
DFlash introduces a speculative decoding framework using a lightweight block diffusion model, achieving significant acceleration in decoding performance.
Agent2Agent Threats in Safety-Critical LLM Assistants: A Human-Centric Taxonomy
This paper presents a taxonomy exploring vulnerabilities in LLM-based assistants, proposing a framework for rigorous threat modeling.
EvasionBench: A Large-Scale Benchmark for Detecting Managerial Evasion in Earnings Call Q&A
EvasionBench introduces a benchmark for detecting evasive responses in corporate earnings calls, providing valuable insights into managerial communication.
FinCoT: Grounding Chain-of-Thought in Expert Financial Reasoning
FinCoT enhances LLM performance in financial NLP tasks by incorporating domain-specific reasoning blueprints for structured prompting.
Dicta-LM 3.0: Advancing The Frontier of Hebrew Sovereign LLMs
This paper introduces Dicta-LM 3.0, an open-weight LLM trained on Hebrew, highlighting its capability to support NLP applications in low-resource languages.
E-Globe: Scalable ε-Global Verification of Neural Networks via Tight Upper Bounds and Pattern-Aware Branching
E-Globe presents a hybrid verifier that bounds neural network outputs accurately, improving verification scalability and efficiency.
MedErrBench: A Fine-Grained Multilingual Benchmark for Medical Error Detection and Correction
MedErrBench is introduced as the first multilingual benchmark for error detection in clinical texts, offering a scalable framework to evaluate models.
From Code-Centric to Concept-Centric: Teaching NLP with LLM-Assisted “Vibe Coding”
This study presents the Vibe Coding approach, integrating LLMs to enhance students’ understanding of NLP beyond mere code generation.