A new arXiv paper introduces RECTIFY, an interactive Streamlit workbench that turns RAG evaluation results into auditable, editable repair workflows. It filters non-issues, routes failures into repair families and fine-grained slices, and generates repair cards verifiable via sandbox reruns. For practitioners, it bridges the gap between RAG evaluators that detect failures and the actual decision of what to fix.
Data Science Collective explains diagnosing RAG failures with a groundedness judge
Data Science Collective explains diagnosing RAG failures with a groundedness judge
A new arXiv paper introduces a methodology for refining a primary Retrieval Augmented Generation (RAG) system by integrating an auxiliary feedback RAG system that leverages human feedback. The approach uses a human-in-the-loop implementation to improve accuracy, relevance, and response quality, aiming toward self-improving LLMs.
A new arXiv paper introduces a methodology for refining a primary Retrieval Augmented Generation (RAG) system by integrating an auxiliary feedback RAG system that leverages human feedback. The approach uses a human-in-the-loop implementation to improve accuracy, relevance, and response quality, aiming toward self-improving LLMs.
The paper introduces RAT, a Bayesian evaluation framework for Retrieval-Augmented Generation (RAG) systems. It jointly models retrieval success, abstention behavior, and answer correctness, factorized by pipeline information flow. This provides a unified way to assess component interactions and error propagation in RAG pipelines.
The paper introduces RAT, a Bayesian evaluation framework for Retrieval-Augmented Generation (RAG) systems. It jointly models retrieval success, abstention behavior, and answer correctness, factorized by pipeline information flow. This provides a unified way to assess component interactions and error propagation in RAG pipelines.
A new arXiv paper presents ActReview, a rebuttal-guided post-training framework that teaches LLMs to give reviewers' feedback as concrete, grounded revision plans rather than only identifying weaknesses. It targets Actionable Peer-review Generation, a task relevant as LLMs see growing use in pre-submission self-review.
A new arXiv paper presents ActReview, a rebuttal-guided post-training framework that teaches LLMs to give reviewers' feedback as concrete, grounded revision plans rather than only identifying weaknesses. It targets Actionable Peer-review Generation, a task relevant as LLMs see growing use in pre-submission self-review.
A new arXiv paper introduces a bidirectional reconstruct-and-verify framework that independently verifies whether code patches generated by LLM-based coding agents actually resolve the reported bug. The approach addresses a key gap in autonomous coding, where current methods lack explicit verification signals.
A new arXiv paper introduces a bidirectional reconstruct-and-verify framework that independently verifies whether code patches generated by LLM-based coding agents actually resolve the reported bug. The approach addresses a key gap in autonomous coding, where current methods lack explicit verification signals.