A Theory of Training Profit-Optimal LLMs (arxiv.org)

5h

From Volume to Value: Preference-Aligned Memory Construction for On-Device RAG (arxiv.org)

5h rag

PromptEmbedder: Efficient and Transferable Text Embedding via Dual-LLM Soft Prompting (arxiv.org)

5h

When Do Attention Circuits Form? Developmental Trajectories of Capability and Attention-Sink Emergence Across Three 1B-ClassArchitectures (arxiv.org)

5h

CoRe-MoE: Contrastive Reweighted Mixture of Experts for Multi-Terrain Humanoid Locomotion with Gait Adaptation (arxiv.org)

5h mixture-of-expertsmoe

BenSyc: Benchmarking Conversational Sycophancy and Human Alignment in LLMs for Bengali Contexts (arxiv.org)

5h

OpenRTLSet: A Fully Open-Source Dataset for Large Language Model-based Verilog Module Design (arxiv.org)

5h

MIRAGE: A Polarity-Flipping Encoding Subspace in LLM Agents (arxiv.org)

5h

Early-Token Confidence Predicts Reasoning Quality in Multi-Agent LLM Debate (arxiv.org)

5h

TabClaw: An Interactive and Self-Evolving Agent for Spreadsheet Manipulation and Table Reasoning (arxiv.org)

5h

The Order Matters: Sequential Fine-Tuning of LLaMA for Coherent Automated Essay Scoring (arxiv.org)

5h fine-tuningllama

PADD: Path-Aligned Decompression Distillation for Non-Router Teacher to Guide MoE Student Learning (arxiv.org)

5h moe

Do Vision-Language Models See or Guess? Measuring and Reducing Textual-Prior Reliance with a Phrasing-Controlled Benchmark (arxiv.org)

5h

WebChallenger: A Reliable and Efficient Generalist Web Agent (arxiv.org)

5h

Prefilling-dLLM: Predictive Prefilling for Long-Context Inference in Diffusion Language Models (arxiv.org)

5h

Small Data, Big Noise: Adversarial Training for Robust Parameter-Efficient Fine-Tuning (arxiv.org)

5h fine-tuning

Multilingual Word-Level Forced Alignment with Self-Supervised Representations and Learned Dynamic Programming (arxiv.org)

5h

REAL: A Reasoning-Enhanced Graph Framework for Long-Term Memory Management of LLMs (arxiv.org)

5h

Continual LLM Upcycling: A Predictor-Gated Bank-Wise Sparsity Training Recipe for Dense-to-Sparse LLMs (arxiv.org)

5h

Pushing the Limits of LLM Tool Calling via Experiential Knowledge Integration and Activation (arxiv.org)

5h

Density Field State Space Models: 1-Bit Distillation, Efficient Inference, and Knowledge Organization in Mamba-2 (arxiv.org)

5h

Measuring Human Value Expression in Social Media Texts: Calibrated LLM Annotation and Encoder Transfer (arxiv.org)

5h

Does Reasoning Preserve Alignment? On the Trustworthiness of Large Reasoning Models (arxiv.org)

5h

Attention Amnesia in Hybrid LLMs: When CoT Fine-Tuning Breaks Long-Range Recall, and How to Fix It (arxiv.org)

5h fine-tuning

VISTA: A Versatile Interactive User Simulation Toolkit for Agent Evaluation (arxiv.org)

5h

Multi-Faceted Interactivity Alignment in Full-Duplex Speech Models (arxiv.org)

5h

SpeechJBB: Probing Safety Alignment and Comprehension in Large Audio Language Models under Code-Switched Speech (arxiv.org)

5h

Mechanistic Analysis of Alignment Algorithms in Language Models (arxiv.org)

5h

Streaming Knowledge Compilation: Proactive Materiality-Scored Pinning for Time-Evolving LLM Wikis (arxiv.org)

5h

Benchmarking and Exploring the Capabilities of LLMs for Attack Investigations (arxiv.org)

5h