Tag
Zero-Mem introduces a zero-token memory operation system for LLM agents, eliminating LLM calls and token consumption during memory retrieval by preserving original interaction traces and using deterministic entity-context and temporal structures. It achieves competitive performance on long-memory and long-context QA benchmarks while reducing memory-operation time cost by 57.6%.
This paper introduces SESA, a self-evolving skill-augmented search agent that co-evolves task generation and skill memory via tool-augmented search self-play. It improves accuracy across seven QA benchmarks over baselines while supporting memory-free deployment.
CURV proposes a curriculum learning framework that enhances chart question answering by developing intrinsic visual grounded reasoning in multimodal LLMs, achieving significant improvements on real-world benchmarks.
GROVE is a training-free framework that grows a temporally stratified memory from continuous video streams, supporting both reactive QA and proactive assistance. It achieves state-of-the-art results on benchmarks like MM-lifelong and EgoServe.
This paper introduces Wnuan, a three-stage post-training pipeline for enterprise question answering that combines task-oriented supervision, supervised fine-tuning with general-data replay, and reinforcement learning on residual errors, achieving large gains on the WnuanBench benchmark while measuring general capability costs.
3DZip is a three-stage token compression framework for 3D vision-language models that uses voxelization, diversity-guided anchor selection, and spatial constraint merging to reduce tokens while preserving spatial reasoning. It retains 94.7% of original performance with only 128 tokens and achieves 1.92x faster inference on 3DQA benchmarks.
Proposes Co-E, a training-free system that synchronizes graph and text memory for multi-hop question answering, improving over comparable training-free baselines on six benchmarks.
Opti-Q is a database-inspired optimizer for multi-LLM question answering that plans execution DAGs to optimize answer quality under cost, latency, and energy constraints, achieving significant improvements on benchmarks.
HyCE-RAG is a novel hypergraph-based retrieval-augmented generation framework for multi-hop question answering that constructs explicit evidence chains via confidence-aware heuristic search, outperforming standard RAG and graph-based RAG methods in accuracy, relevance, and faithfulness.
This paper introduces the Relevance-Aware RipGrep Search Agent (RARG), which uses relevance as an execution prior for corpus interaction in agentic search, improving accuracy and efficiency over existing methods.
This paper introduces AILQA, an AI-driven system for legal question answering in the Indian legal context, using embedding and generative models with retrieval-augmented generation (RAG). It evaluates performance on Indian legal texts and the All India Bar Exam, finding that AI-generated responses can sometimes outperform reference answers.
Presents FiT, a diagnostic framework to evaluate small LLMs on cybersecurity QA capabilities before fine-tuning, showing that fine-tuning can degrade vocabulary and parametric knowledge depending on the regime. Provides guidance to avoid unnecessary fine-tuning.
Debate-on-Graph (DoG) is a framework that enhances LLM reasoning by leveraging uncertain knowledge graphs (UKGs) with confidence scores, using a heuristic search and multi-agent debate mechanism to produce reliable answers. It achieves state-of-the-art performance on four QA benchmarks.
This paper introduces FinanceComplexQA, a comprehensive benchmark for evaluating agentic reasoning on industrial-grade financial documents, featuring bilingual support, expert-level questions, and complex layouts across six scenarios and seven tasks.
Introduces LakeQuest, a human-validated benchmark of 9,846 QA pairs across three domains for evaluating end-to-end retrieve-and-synthesize pipelines over heterogeneous data lakes, revealing critical failure modes in modern QA systems.
Introduces CANDI-QA, a dataset to evaluate LLMs on contextual alignment in niche domains, comprising information assistance and applied inference questions. Systematic evaluations across ten LLMs show challenges, and a lightweight neuro-symbolic framework MTSS-Net is proposed as baseline.
ResearchQA is a new benchmark of 6,211 single-paper question-answer pairs from 494 open-access papers across eight domains, designed to evaluate citation-grounded question-answering by requiring verifiable citations and supporting grounded refusal when evidence is insufficient.
CLIR-Bench is a benchmark for multimodal question answering over irregularly sampled clinical time series, constructed from ICU records with 6,600 QA instances across 11 clinical variables. It reveals that existing generalist models struggle with sparse temporal evidence, highlighting the need for stronger irregular time-series reasoning methods.
PolyUQuest is a verifiable, structure-aware web RAG framework that uses heterogeneous graphs to unify hyperlink topology, DOM hierarchy, and entity-relation knowledge, outperforming existing systems on answer correctness, coverage, and faithfulness.
This paper presents the NLPCC 2026 Shared Task 1: Difficulty-Aware Multilingual and Multimodal Medical Instructional Video Understanding Evaluation (DA-MIVQA), which extends previous benchmarks with difficulty-aware annotations and three tracks: temporal answer grounding, video corpus retrieval, and grounding in corpus. The dataset includes medical instructional videos from public channels and aims to evaluate systems under varying reasoning requirements.