7179 papers
The paper introduces ActiveUltraFeedback, an efficient active learning pipeline that leverages uncertainty estimates and novel selection strategies like Double Reverse Thompson Sampling to generate high-quality preference data, enabling Large Language Models to achieve superior alignment performance with as little as one-sixth of the annotated data required by static baselines.
Mousse is a novel optimizer that improves upon the Muon algorithm by integrating Shampoo's Kronecker-factored preconditioning to adaptively handle the heavy-tailed curvature of deep neural networks, thereby achieving faster training convergence with negligible computational overhead.
The paper introduces OOD-MMSafe, a benchmark revealing significant causal blindness in current Multimodal Large Language Models regarding hidden consequences, and proposes the Consequence-Aware Safety Policy Optimization (CASPO) framework to effectively mitigate these risks by shifting safety alignment from intent detection to consequence projection.
This paper introduces MUGEN, a comprehensive benchmark revealing that Large Audio-Language Models struggle with multi-audio understanding as input scaling increases, and demonstrates that combining training-free strategies like Audio-Permutational Self-Consistency with Chain-of-Thought can significantly improve performance.
This paper introduces CVS, a training-free data selection method that identifies high-quality vision-language samples by measuring how much a question alters a frozen model's assessment of answer validity, thereby improving multimodal reasoning performance while reducing computational costs.
AutoAgent is a self-evolving multi-agent framework that integrates evolving cognition, on-the-fly contextual decision-making, and elastic memory orchestration to enable autonomous agents to adaptively learn from experience and make reliable, context-aware decisions in dynamic environments without external retraining.
This paper introduces RbtAct, a framework that leverages peer review rebuttals as implicit supervision to train large language models to generate more actionable and specific review feedback through a novel perspective-conditioned task and a new dataset called RMR-75K.
This paper introduces EXPLORE-Bench, a benchmark derived from real first-person videos to evaluate the ability of multimodal large language models to perform long-horizon egocentric scene prediction, revealing significant performance gaps compared to humans and demonstrating that stepwise reasoning offers partial improvements at a computational cost.
The paper proposes "Ego," an efficient personalization method for vision-language models that extracts visual tokens representing target concepts via internal attention mechanisms to serve as memory, enabling strong performance across single-concept, multi-concept, and video personalization tasks without requiring additional training stages or external modules.
World2Mind is a training-free toolkit that enhances foundation models' allocentric spatial reasoning by constructing structured cognitive maps and an Allocentric-Spatial Tree, enabling significant performance gains and even allowing text-only models to achieve complex 3D spatial reasoning comparable to advanced multimodal systems.
This paper demonstrates that simulation-based inference (SBI) is a viable and potentially superior alternative to traditional empirical tuning for determining neutrino interaction model parameters, as it successfully reproduces and slightly improves upon the MicroBooNE collaboration's tuned GENIE configuration while also approximating the NuWro simulation.
This paper introduces the concept of "opaque serial depth" to formally quantify the limits of internal reasoning in language models without externalized chain-of-thought, providing computational bounds for various architectures and an open-source tool to analyze the potential for non-externalized reasoning.
This paper proposes and validates a hybrid quantum-classical framework that integrates a Long Short-Term Memory (LSTM) network with a Quantum Circuit Born Machine (QCBM) to significantly improve financial volatility forecasting accuracy on high-frequency stock market data compared to traditional classical models.
This paper proposes ACP-SL, an adaptive channel pruning scheme for Split Learning that utilizes a label-aware channel importance scoring module to compress smashed data, thereby significantly reducing communication overhead while improving test accuracy and training efficiency.
The paper introduces MITRA, an on-premise Retrieval-Augmented Generation (RAG) system designed to enhance knowledge retrieval in large-scale physics collaborations like CMS by employing a novel automated pipeline for document processing and a two-tiered vector database architecture to accurately answer context-aware questions while ensuring data privacy.
This paper introduces the smoothing pseudo-projector, a lightweight, multigrid-inspired module that corrects hidden representations in transformer-based models to suppress noise from label-irrelevant inputs, thereby improving training dynamics and robustness without altering the core architecture.
This paper introduces MA-EgoQA, a novel benchmark and dataset featuring 1,700 questions across five categories designed to evaluate the ability of AI models to understand and reason over multiple long-horizon egocentric videos from embodied agents, alongside a proposed baseline model named EgoMAS that highlights current limitations in system-level multi-agent understanding.
This paper introduces SCENEBench, a comprehensive benchmark suite designed to evaluate Large Audio Language Models on critical non-speech and cross-lingual audio understanding tasks relevant to assistive and industrial applications, revealing significant performance gaps and latency challenges in current state-of-the-art models.
This paper introduces HR-GAT, a hierarchical resolution graph attention network that leverages geospatial data to predict spectrum demand with 21% higher accuracy than baseline models, effectively addressing spatial autocorrelation challenges to enable more efficient spectrum sharing and policy-making.
This paper introduces the Dynamics-Aware Policy Learning (DAPL) framework, which leverages explicit world modeling to learn contact-induced dynamics, enabling robots to achieve robust extrinsic dexterity in cluttered environments without hand-crafted heuristics and significantly outperforming existing manipulation methods in both simulation and real-world deployments.