Tag
This paper presents an empirical study evaluating RAG evaluation metrics from four libraries (Ragas, DeepEval, RAGChecker, Opik) by comparing them to human judgments and standard recall metrics, using a question-answering dataset created from business data.
This paper extends PubHealthBench into a retrieval-augmented setting and evaluates retrieval and generation choices for public health QA, showing hybrid retrieval improves accuracy and that smaller models with retrieval can match larger ones.
This paper introduces RSF-GLLM, a framework that decouples differentiable graph reasoning from LLM generation to address the semantic gap in multi-hop knowledge graph question answering, achieving competitive performance with superior inference efficiency.
Introduces MultAttnAttrib, a training-free method for multimodal attribution in long document QA, along with the MultAttrEval benchmark. It outperforms prompting-based methods and matches frontier models like GPT-5.4.
Introduces MetaFlow, a method that trains large language models to generate zero-shot workflows for tasks by combining supervised fine-tuning and reinforcement learning with execution feedback, achieving strong generalization to untrained tasks and operator sets.
This paper introduces MamaBench and MamaRetrieval, two benchmarks for evaluating medical retrieval-augmented generation in maternal, neonatal, and reproductive health, addressing gaps in existing QA and retrieval datasets.
This paper proposes TriageRA-CCF, a method for adaptive rank budgeting in LoRA for medical question answering. It uses source-side signals (base-model confidence, clinical coverage, counterfactual proxy) to dynamically choose rank budgets, achieving modest accuracy gains on Qwen3-8B and Llama3.1-8B.
AB-RAG is a training-free, backbone-agnostic framework that adaptively retrieves passages for question answering by estimating answer confidence, improving efficiency and accuracy across multiple backbones and datasets.
PASTA is a novel framework for knowledge updating in LLMs that combines data augmentation, question-answering generation, and self-learning DPO to integrate factual information from news articles, achieving accuracy improvement from 0.02 to 0.82 while preserving general capabilities.
Proposes OPI, an ontology-guided framework for multi-hop knowledge graph question answering that leverages a relation-centric ontology graph for bidirectional retrieval and iterative refinement, achieving state-of-the-art results on multiple benchmarks.
This paper tests the assumption that LLMs judge better than they generate in in-context QA, finding generation accuracy exceeds self-evaluation on most benchmarks, with evaluation attending less to context. The findings challenge core assumptions in self-evaluation pipelines.
This paper presents team HSA_CORAL's submission to the FinCausal 2026 shared task, comparing encoder-only, encoder-decoder, and decoder-only LLMs for extractive question answering of cause-effect relations in financial narratives. Fine-tuned GPT-4.1 Mini achieved top scores in the English subtask and third in Spanish.
AgentFinVQA is a multi-agent pipeline for financial chart question answering that decomposes queries into planning, OCR, legend grounding, visual inspection, and verification steps, recording each step in a traceable Model Evaluation Packet. It achieves significant accuracy gains over zero-shot baselines while enabling on-premise deployment and auditability.
This paper models a question-answering forum staffed by expert knowledge workers, studying optimal scheduling to maximize system capacity and stability.
This paper introduces CADE, a framework for time-series question answering that maps each timestep directly into the LLM embedding space and uses a one-directional supervised contrastive loss to align time-series representations with frozen text anchors, outperforming existing baselines on the Time-MQA benchmark.
Introduces IRTS-ToolBench, a benchmark of 1,700 questions for evaluating LLMs and AI agents on irregular time series question answering via tool-grounded reasoning, covering 10 task types across 13 domains.
OmniVideo-100K introduces an automated data engine with entity-anchored scripting and clue-guided QA generation to improve audio-visual reasoning and temporal consistency, achieving significant performance gains across multiple benchmarks.
LakeQA is a new benchmark for exploratory question answering over a million-scale data lake, evaluating multi-hop reasoning and compositionality across text, tables, and knowledge graphs.
Latent Memory introduces a compressed representation approach for external memory in question answering, reducing token consumption and storage requirements while maintaining competitive performance across text-only and multimodal benchmarks.
This paper introduces FullCite, a framework for generating structured inline citations that link each claim to both its source document and specific evidence spans. Evaluated on three QA benchmarks (ASQA, BioASQ, ExpertQA), it finds that while LLMs are good at document-level attribution, they struggle with precise evidence span identification.