Retrieval-augmented generation (RAG) has been extensively employed to mitigate hallucinations in large language models (LLMs). However, existing methods for multi-hop reasoning tasks often lack global planning, increasing the risk of falling into local reasoning impasses. Insufficient exploitation of retrieved content and the neglect of latent clues fail to ensure the accuracy of reasoning outcomes. To overcome these limitations, we propose Recursive Evaluation and Adaptive Planning (REAP), whose core idea is to explicitly maintain structured sub-tasks and facts related to the current task through the Sub-task Planner (SP) and Fact Extractor (FE) modules. SP maintains a global perspective, guiding the overall reasoning direction and evaluating the task state based on the outcomes of FE, enabling dynamic optimization of the task-solving trajectory. FE performs fine-grained analysis over retrieved content to extract reliable answers and clues. These two modules incrementally enrich a logically coherent representation of global knowledge, enhancing the reliability and the traceability of the reasoning process. Furthermore, we propose a unified task paradigm design that enables effective multi-task fine-tuning, significantly enhancing SP's performance on complex, data-scarce tasks. We conduct extensive experiments on multiple public multi-hop datasets, and the results demonstrate that our method significantly outperforms existing RAG methods in both in-domain and out-of-domain settings, validating its effectiveness in complex multi-hop reasoning tasks.
Figures & tables
Figure 1: Overall framework of REAP. After the Decomposer breaks down the question, the Sub-task Planner and Fact Extractor iterate to guide the reasoning trajectory and gather the facts, and the Synthesizer produces the final answer.
Models
HotpotQA
2Wiki
MuSiQue ‡
Bamboogle ‡
CEM
F1
ACC †
CEM
F1
ACC †
CEM
F1
ACC †
CEM
F1
ACC †
Naive
24.6
31.8
34.6
31.6
35.1
32.8
6.2
10.6
10.0
13.6
17.7
21.6
Standard RAG
41.6
48.6
51.2
36.2
38.5
38.2
9.8
13.5
12.6
20.8
30.9
32.0
FT-Standard RAG
53.2
61.2
65.4
53.6
57.1
56.2
18.2
26.4
22.6
42.4
54.4
51.2
SuRe
29.4
33.8
40.4
24.4
24.6
28.2
5.8
11.6
11.4
10.4
19.1
20.0
IRCoT
42.2
51.4
54.8
41.2
36.5
41.8
18.2
18.6
20.4
35.2
30.1
44.8
Table 1: Performance comparisons between REAP and the baselines on four MHQA benchmarks. Bold numbers indicate the best result, while underlined number indicates the second-best result. ‡ represents out-of-domain datasets.
Models
HQA
WQA
MQA ‡
BQA ‡
F1
F1
F1
F1
w/o replan
64.9
78.6
34.2
61.6
w/o verify
65.1
78.0
34.8
60.8
w/o clue
64.6
76.5
35.2
62.7
REAP
68.0
79.6
38.3
65.2
Table 2: The ablation study on HotpotQA (HQA), 2WikimultihopQA (WQA), MuSiQue (MQA) and Bamboogle (BQA).
Figure 2: Ablation study on the effectiveness of Multi-task Fine-tuning.
Models
HotpotQA
2Wiki
MuSiQue
Bamboogle
CEM
F1
ACC †
CEM
F1
ACC †
CEM
F1
ACC †
CEM
F1
ACC †
Naive 70B
32.4
39.7
40.4
36.8
39.3
38.8
11.6
15.1
17.4
37.6
37.0
44.0
StandardRAG 70B
51.0
56.3
61.2
46.8
47.2
46.6
16.8
21.1
24.8
35.2
46.1
46.4
SuRe 70B
35.8
43.1
46.0
27.6
34.6
33.2
12.2
20.4
17.4
10.4
19.1
20.0
IRCoT 70B
55.8
49.5
60.4
57.0
33.1
35.8
21.0
19.4
24.3
32.0
38.8
40.8
Iter-RetGen 70B
53.8
47.6
65.0
51.4
41.5
53.6
17.2
21.6
24.6
41.6
47.5
52.8
Table 3: Performance comparison of REAP and non-fine-tuned baselines on four MHQA benchmarks under the Llama-3.1-70B-Instruct setting. Bold numbers indicate the best result, while underlined number indicates the second-best result.
Figure 3: Module substitution experiment validating the efficiency gains of asymmetric configuration.
Multi-hop question answering studies how systems combine evidence across documents, facts, or modalities to answer questions requiring linked reasoning.
Large language models (LLMs) often generate inaccurate answers due to their reliance on static internal knowledge. Retrieval-augmented generation (RAG) addresses this limitation by integrating external knowledge and excelling at single-hop queries. However, it struggles with multi-hop questions that require cross-document reasoning. Existing methods, such as graph structured RAG or question decomposition, often lack dynamic decomposition and effective filtering, which leads to lower efficiency and accuracy. To overcome these limitations, we propose Dynamic Decomposition and Filtering for Multi-Hop Reasoning-Augmented Generation (D2F-ReAG), a novel paradigm that adaptively controls reasoning depth by judging the reliability of the root-level reasoning. If the root reasoning is reliable, the model directly generates the answer. Otherwise, the question is logically decomposed into sub-questions, and the verified reasoning derived from these sub-questions is used to refine the root reasoning. Experiments on three multi-hop benchmarks demonstrate the effectiveness of our method in handling complex multi-hop questions.
Jiaoyang Li, Junhao Ruan, Shengwei Tang +4
Northeastern University, Shenyang, China · Kunming University of Science and Technology, Kunming, China · NiuTrans Research, Shenyang, China
Multi-hop question answering remains challenging for Retrieval-Augmented Generation (RAG) because existing approaches either propagate errors across iterative retrieval rounds or over-generate reasoning steps, increasing cost without improving accuracy. We propose Grounded Delta Planning RAG (GDP-RAG), a plan-based framework that targets only the information delta based on three simple design choices: (1) preliminary retrieval to ground planning before execution, (2) a gap-conditioned planning prompt that asks only for missing information, and (3) a skeletal trajectory that pairs each subquery with a Thought capturing evidence from preliminary retrieval and carrying it through to the final answer. GDP-RAG focuses computation on unresolved gaps, yielding concise, reliable reasoning trajectories. Extensive experiments on HotpotQA, 2WikiMultiHopQA, and MuSiQue show that GDP-RAG achieves the highest accuracy (60.63%) among all compared systems while maintaining a cost-of-pass of 0.51, 22% lower than PAR-RAG (0.65) and 68% lower than KnowTrace (1.57), with no method achieving both higher accuracy and lower cost.
Wei-Chieh Chou, Xuanjun Chen, Jian-Ren Lin +3
Dept. of Computer Science and Information Engineering, National Taiwan University · Graduate Institute of Communication Engineering, National Taiwan University · Department of Economics, National Taiwan University +2
Retrieval-augmented generation (RAG) connects large language models (LLMs) to external knowledge, but single-round retrieval is often insufficient for complex multi-hop questions. To enhance search capabilities for complex tasks, most existing works integrate multi-round iterative retrieval with reasoning processes via end-to-end training. While these approaches improve problem-solving performance, they still face challenges in task reasoning and model training, especially ambiguous retrieval execution paths and sparse rewards in end-to-end reinforcement learning (RL), which can lead to inaccurate retrieval results and lower performance. We attribute these failures to hierarchical credit entanglement: a single final reward updates planning and execution together, so the model cannot clearly separate plan errors from retrieval errors. We propose APEX-Searcher, which uses a Refining Credit Assignment paradigm: planning is optimized by RL with a plan-level reward, while execution is learned by SFT. Extensive experiments show consistent gains in both multi-hop RAG and task planning across benchmarks.
Kun Chen, Qingchao Kong, Zhao Feifei +1
University of Chinese Academy of Sciences · MAIS, Institute of Automation, Chinese Academy of Sciences · Wenge Technology Co., Ltd