Tag
This paper studies hidden states in long horizon language model agents, revealing that memory compression and recall needs are encoded before actions. It proposes the PaMER framework to reduce context consumption while maintaining task performance through state-guided compression and evidence retrieval.
The paper quantifies how improvements in fact-verification scores are partitioned between answer accuracy and evidence quality, using trained DeBERTa checkpoints and LLMs across multiple benchmarks.
The paper introduces PACE, a dataset for evaluating whether AI models can identify hidden conflicts in user requests by retrieving implicit knowledge base facts, and proposes PaceMaker, a multi-agent framework to enhance conflict-aware decision-making.
ECHO is a cognitively inspired, auditable memory architecture for long-horizon agents, evaluated on benchmarks like LoCoMo and LongMemEval-S with high retrieval performance.
FinRCA-Bench is a benchmark designed to evaluate evidence retrieval and reasoning capabilities in financial AI systems, providing a standardized approach for assessment and improvement.
This paper introduces ReMEMBER, a missing-evidence memory framework for streaming dialogue summarization that retrieves and refines evidence from long histories to resolve gaps in current windows under fixed memory budgets, along with a benchmark for evaluation.
ColGraphRAG replaces single-vector bi-encoder similarity with late-interaction MaxSim scoring for ranking graph-linked image candidates in multimodal GraphRAG, improving retrieval and QA accuracy on MultimodalQA.
This paper introduces Tree of Evidence (ToE), a hierarchical and explainable claim verification framework that dynamically retrieves and aggregates multi-source evidence using reinforcement learning. Experiments show 4-24 percentage point improvements over baselines, especially against adversarially poisoned inputs from Generative Engine Optimization.
This paper introduces GranuVistaVQA, a multimodal benchmark with element-level annotations, and GranuRAG, a framework that treats visual elements as first-class retrieval units for verifiable multimodal RAG, achieving up to 29.2% improvement over baselines.