Beyond Chunks and Graphs: Retrieval-Augmented Generation through Triplet-Driven Thinking
Shengbo Gong, Xianfeng Tang, Carl Yang, Wei jin
TL;DR
T$^2$RAG introduces a graph-free, triplet-based retrieval-augmented generation framework that improves reasoning performance and reduces costs compared to existing multi-round and graph-based RAG methods.
Contribution
The paper presents T$^2$RAG, a novel triplet-driven RAG framework that simplifies knowledge retrieval and enhances reasoning efficiency without relying on complex graph structures.
Findings
Outperforms state-of-the-art methods with up to 11% accuracy gain
Reduces retrieval costs by up to 45%
Achieves significant improvements across six datasets
Abstract
Retrieval-augmented generation (RAG) is critical for reducing hallucinations and incorporating external knowledge into Large Language Models (LLMs). However, advanced RAG systems face a trade-off between performance and efficiency. Multi-round RAG approaches achieve strong reasoning but incur excessive LLM calls and token costs, while Graph RAG methods suffer from computationally expensive, error-prone graph construction and retrieval redundancy. To address these challenges, we propose TRAG, a novel framework that operates on a simple, graph-free knowledge base of atomic triplets. TRAG leverages an LLM to decompose questions into searchable triplets with placeholders, which it then iteratively resolves by retrieving evidence from the triplet database. Empirical results show that TRAG significantly outperforms state-of-the-art multi-round and Graph RAG methods, achieving an…
Peer Reviews
Decision·ICLR 2026 Conference Withdrawn Submission
- The idea of using purely triplet-based databases for multi-hop reasoning with RAG is interesting. - The proposed online retrieval method is also interesting. With those placeholders, the query triplets can be semantically closer to the relevant triplets in the vector database. - For QA datasets that do not require that many hops, the proposed method may improve accuracy significantly.
- One reason to use graph data is that it can be more friendly when dealing with multi-hop reasoning. Building databases without entity links but with only triplets essentially gives up graph structures during offline building. So, during the online process, the method would, in principle, need to somehow reconstruct the multi-hop links. - This can be especially hard for long-hop reasoning, and this may also be the reason why T2RAG mostly improves datasets with fewer hops, e.g., PopQA, 2Wik
1. The paper proposes T2RAG, an innovative framework that leverages atomic knowledge triplets to go beyond traditional chunk- or graph-based retrieval-augmented generation (RAG) approaches. 2. The authors demonstrate that T2RAG yields strong empirical gains across multiple QA datasets, improving average performance by 11% while reducing retrieval cost by up to 45%, which shows both effectiveness and efficiency.
1. The related-work section could be strengthened. Although GraphRAG and multi-round RAG methods are discussed, the paper lacks comparison with other triplet-based retrieval methods such as SubgraphRAG, which also uses triplets for retrieval. 2. Experimental parameter choices are not always fair. When comparing multi-round methods to direct chunk-retrieval baselines there is an information imbalance: some settings allow more total retrieved content than others. 3. Several implementation detai
1.The work proposes triplet-driven thinking to reduce token footprint and simplify graph dependencies in the RAG pipeline. 2.The split between offline distillation (triplet extraction/embedding) and online iterative resolution is clean and easy to reproduce, with a plausible path to system engineering and deployment. The placeholder-resolution idea encourages explicit evidence grounding rather than opaque, multi-turn prompts; this has the potential to reduce uncontrolled drift. 3.The paper cla
1.The settings of experiments appear to resolve to entity answers or facts naturally expressible as single triplets. This inherently benefits triplet retrieval and makes direct comparisons to multi-round RAG and GraphRAG less fair. The paper should include tasks where answers are compositional, multi-hop, or non-entity (rationales, procedural steps) to validate generality. 2.The placeholder triplet decomposition and iterative resolution are described verbosely, but the novelty claim would be st
* Using triplet with placeholder to represent the required information is new and interesting to me * The experiments are extensive
* The triplet update itself is not complicated, but in Step 2.3, the use of mathematical notation makes the process significantly harder to follow. * T^2RAG is actually a multi rould RAG, but only one multi round baseline is compared, maybe some more multi round baselines should be added.
Videos
No videos yet. Explain this paper in a talk, walkthrough, or lecture? Add one.
Taxonomy
TopicsTopic Modeling · Advanced Graph Neural Networks · Data Quality and Management