AI Research Paper Feed

TLDR: Agentic Proposing is a framework that improves reasoning in large language models by dynamically composing modular skills, achieving high accuracy with fewer data points.

TLDR: CSR-Bench is a benchmark for evaluating the cross-modal safety and reliability of multimodal large language models, revealing systematic alignment gaps and trade-offs between over-rejection and safety.

TLDR: GFlowPO is a new framework for optimizing language model prompts using a probabilistic approach and dynamic memory updates, leading to better performance in various language tasks.

TLDR: The paper introduces structured bipolar argumentation frameworks (SBAFs) that allow agents to reject arguments based on doubt and to focus on individual sentences rather than whole arguments, providing new semantics that do not require accepting all defendable arguments.

TLDR: The study introduces CADS, a method for generating high-quality synthetic multimodal data to improve multimodal large language models, resulting in the creation of MMSynthetic-20K and the high-performing R1-SyntheticVL model.

TLDR: The paper introduces a novel framework using Toxicity Association Graphs to detect covert toxicity in multimodal data, offering a new metric for measuring hidden toxicity and outperforming existing methods in interpretability and detection accuracy.

TLDR: This paper provides a convergence proof for Proximal Policy Optimization (PPO) by interpreting its update scheme as approximate policy gradient ascent and addresses an issue in Generalized Advantage Estimation (GAE).

TLDR: The study introduces AC-MIL, a novel framework for assessing risks in live streaming by analyzing user behaviors and coordination patterns, significantly improving detection accuracy and interpretability over existing methods.

TLDR: The paper introduces a method called positive-negative pairing for prompt selection in reinforcement learning with verifiable rewards, leading to improved performance on deterministic reasoning tasks by amplifying rare event signals.

TLDR: This paper presents a method to integrate formal domain knowledge into large language models by compiling ontological specifications into tools that enforce semantic constraints during knowledge generation.

TLDR: Memora is a memory representation system that balances abstraction and specificity to improve retrieval relevance and reasoning effectiveness in agent memory systems.

TLDR: EGSPO improves large language model training by using token-level gradient modulation to enhance performance on mathematical reasoning tasks with minimal computational overhead.

TLDR: MemCast introduces a memory-driven framework for time series forecasting that leverages past experiences to improve prediction accuracy and adaptability over time.

TLDR: MeetBench-XL introduces a comprehensive evaluation framework and a dual-policy AI agent for enhancing real-time meeting assistance in enterprise environments.

TLDR: TodyComm is a dynamic communication algorithm for multi-agent systems that adapts to changing conditions, improving task effectiveness and efficiency.

TLDR: The study develops a Turkish-specific Retrieval-Augmented Generation (RAG) dataset and benchmarks various methods, finding that complex methods like HyDE significantly improve accuracy over simpler baselines.

TLDR: HySparse is a new hybrid sparse attention model that improves performance and reduces memory usage by using full attention layers as oracles for token selection and sharing KV caches with sparse layers.

TLDR: ELIQ is a label-free framework designed to assess the quality of AI-generated images, adapting to evolving generative models without the need for human annotations.

TLDR: DeepDFA is a framework that integrates temporal logic into neural networks, outperforming traditional models in sequential tasks by bridging subsymbolic learning and symbolic reasoning.

TLDR: Accordion-Thinking enables LLMs to self-regulate reasoning step granularity, achieving efficient and readable reasoning with reduced computational overhead.