Modern coding agents are increasingly deployed for realistic software engineering, yet existing benchmarks primarily evaluate static, fixed-target tasks, failing to capture the long-horizon, multi-turn nature of real-world development. We introduce MACBench, a novel benchmark evaluating agents on sustained frontend web application development where text-based user requirements continuously evolve within a persistent workspace. Comprising 28 repository instances across 8 domains, MACBench features 699 dialogue turns and 956 evaluation items. It explicitly models realistic interaction patterns, challenging agents to maintain cross-turn coherence and continuous executability. Our dual-branch evaluation framework assesses both process robustness and final artifact quality, uniquely integrating intermediate viability checks with bounded runtime-feedback repair. Benchmarking 4 contemporary models reveals a severe reliability gap: the best-performing agent achieves only a 70.6% checklist success rate and a 10.7% strict instance success rate, struggling to satisfy requirements accumulated over long sessions. Furthermore, ablation studies reveal diminishing and sometimes negative returns from runtime-feedback repair: while it improves checklist success by an average of 10.4% and aids immediate failure recovery, excessive repair attempts frequently degrade final functional outcomes, suggesting they may introduce regressions or side effects in the evolving codebase. Finally, failure analysis identifies complex cross-turn application logic as the dominant bottleneck, accounting for 33–50% of unresolved errors.

Long-Context Question Answering (LCQA), a challenging task, aims to reason over long-context documents to yield accurate answers to questions. Existing long-context Large Language Models (LLMs) for LCQA often struggle with the “lost in the middle” issue. Retrieval-Augmented Generation (RAG) mitigates this issue by providing external factual evidence. However, its chunking strategy disrupts the global long-context information, and its low-quality retrieval in long contexts hinders LLMs from identifying effective factual details due to substantial noise. To this end, we propose LongRAG, a general, dual-perspective, and robust LLM-based RAG system paradigm for LCQA to enhance RAG’s understanding of complex long-context knowledge (i.e., global information and factual details). We design LongRAG as a plug-and-play paradigm, facilitating adaptation to various domains and LLMs. Extensive experiments on three multi-hop datasets demonstrate that LongRAG significantly outperforms long-context LLMs (up by 6.94%), advanced RAG (up by 6.16%), and Vanilla RAG (up by 17.25%). Furthermore, we conduct quantitative ablation studies and multi-dimensional analyses, highlighting the effectiveness of the system’s components and fine-tuning strategies.Data and code are available at https://github.com/QingFei1/LongRAG.

In this paper, we propose a novel end-to-end framework called KBRD, which stands for Knowledge-Based Recommender Dialog System. It integrates the recommender system and the dialog generation system. The dialog generation system can enhance the performance of the recommendation system by introducing information about users’ preferences, and the recommender system can improve that of the dialog generation system by providing recommendation-aware vocabulary bias. Experimental results demonstrate that our proposed model has significant advantages over the baselines in both the evaluation of dialog generation and recommendation. A series of analyses show that the two systems can bring mutual benefits to each other, and the introduced knowledge contributes to both their performances.