💬 NLP

Only Ask What You Don't Know: Grounded Delta Planning for Efficient Multi-step RAG

The paper introduces Grounded Delta Planning RAG (GDP-RAG), a framework that improves multi-hop question answering efficiency and accuracy by grounding plans in preliminary retrieval and focusing computation solely on information gaps, thereby achieving state-of-the-art performance with significantly lower costs than existing methods.

Original authors: Wei-Chieh Chou, Xuanjun Chen, Jian-Ren Lin, Claire Lin, Hung-yi Lee, Jyh-Shing Roger Jang

Published 2026-06-23

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Wei-Chieh Chou, Xuanjun Chen, Jian-Ren Lin, Claire Lin, Hung-yi Lee, Jyh-Shing Roger Jang

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to solve a complex mystery, like figuring out who stole the cookie from the jar. You have a massive library of books (the internet) to help you, but you can't read them all at once.

Most current AI systems try to solve this in one of two ways, and both have problems:

The "Step-by-Step" Detective: They ask a question, read a book, ask another question, read another book, and so on. The problem? They often get confused by the sheer amount of paper they've piled up, or they make a small mistake early on that ruins the whole investigation.
The "Over-Planner": They sit down and write a giant, rigid plan before looking at any books. They might plan to check five different books, even though the answer is already in the first one they picked up. This wastes time and money.

The New Solution: "Grounded Delta Planning" (GDP-RAG)

The authors of this paper propose a smarter way to act like a detective. They call it Grounded Delta Planning. Think of it as a "Check-First, Ask-Only-What-You-Need" strategy.

Here is how it works, using a simple analogy:

1. The "Quick Glance" (Preliminary Retrieval)

Before the detective writes a single question, they do a quick, broad scan of the library shelves to grab a few relevant books.

Why? So they don't start from scratch. They know what information is already sitting on the table.

2. The "Gap Finder" (Delta Planning)

Instead of writing a plan to find everything, the detective looks at the books they just grabbed and asks: "What is missing?"

If the books already tell them "The cookie was taken by the dog," they don't need to ask about the dog.
They only write down questions for the Delta (∆)—the specific pieces of information that are still missing.
The Result: They skip the questions they don't need to ask. This is like pruning a tree; you cut off the dead branches (redundant questions) so the tree grows faster and stronger.

3. The "Skeleton Plan" (Skeletal Trajectory)

The detective writes a very short, lean plan. But here is the clever part: next to every question they do ask, they attach a little note saying, "I already know X from my quick glance."

This ensures that as they go deeper into the investigation, they don't forget the facts they found early on. It keeps the whole story connected.

4. The "Double-Check" Loop

Once they start asking their specific questions, they don't just trust the first answer. They use a three-step loop:

Act: Ask the question and get an answer.
Review: Double-check that answer against new evidence to make sure it's true.
Update: If they found something new, they tweak their plan for the next step.

Why is this a big deal?

The paper tested this method on three difficult puzzle datasets (HotpotQA, 2WikiMultiHopQA, and MuSiQue). Here is what they found:

Smarter, Not Harder: The new method got the highest accuracy (60.63%) compared to all other methods. It solved the puzzles better.
Cheaper: Because it asked fewer questions and read fewer unnecessary pages, it cost 22% less to run than the next best method (PAR-RAG) and 68% less than a complex method called KnowTrace.
No Trade-off: Usually, you have to choose between being accurate or being cheap. This method managed to be both the most accurate and the most efficient.

In a nutshell:
Imagine you are cooking a meal.

Old methods either taste the soup every second (too slow) or write a recipe for 50 ingredients before checking what's in your fridge (wasteful).
GDP-RAG looks in the fridge first, sees what you already have, and then writes a recipe only for the ingredients you are missing. You cook faster, spend less money, and the meal tastes better.

The paper concludes that by focusing only on the "missing pieces" (the delta) and grounding the plan in real evidence first, AI can solve complex, multi-step problems much more efficiently without losing accuracy.

Technical Summary: Grounded Delta Planning for Efficient Multi-step RAG

Problem Statement

Multi-hop question answering (QA) remains a significant challenge for Retrieval-Augmented Generation (RAG) systems. Existing approaches generally fall into two categories, both of which face critical limitations:

Step-wise Methods (e.g., ReAct, Self-Ask): These interleave retrieval and reasoning dynamically. While flexible, they suffer from the "lost-in-the-middle" problem due to context overflow and amplify error propagation, where early mistakes cascade through the entire reasoning process.
Plan-based Methods (e.g., ReWOO, PlanRAG): These construct static global plans before execution to isolate sub-steps and reduce cascading errors. However, they often rely solely on parametric knowledge to generate plans, leading to over-planning (generating redundant subqueries for information already available) or under-planning (missing necessary dependencies). Furthermore, strictly following rigid plans creates efficiency issues as the cost of each execution step becomes substantial.

The core challenge is to design a framework that avoids error propagation and context overflow while minimizing computational cost by eliminating redundant reasoning steps and ensuring no critical information is missed.

Methodology: Grounded Delta Planning RAG (GDP-RAG)

The authors propose GDP-RAG, a plan-based framework that shifts the planning paradigm from generating a complete reasoning chain to targeting only the information delta ( $\Delta$ )—the specific knowledge missing after an initial retrieval. The framework operates through three distinct stages:

1. Planning Phase (Grounded Delta Planning)

Instead of decomposing a query using only parametric knowledge, GDP-RAG performs Preliminary Retrieval first to gather a set of relevant documents ( $D_{rel}$ ). The planner then generates a Skeletal Trajectory ( $T_{skel}$ ) based on three design choices:

Grounding: The planner is conditioned on $D_{rel}$ to understand what information is already available.
Gap-aware Query Decomposition: The prompt instructs the planner to identify information present in $D_{rel}$ and generate subqueries only for the remaining gaps ( $\Delta$ ).
Skeletal Trajectory Structure: The output is a sequence of tuples $\langle (\theta_i, q_i) \rangle$ , where $q_i$ is a subquery for missing information, and $\theta_i$ (Thought) captures the evidence from $D_{rel}$ relevant to that step. This preserves evidence across the trajectory, preventing information loss during execution.

This approach yields three structural advantages:

Reasoning-step Compression: Steps that would have been required to retrieve facts already present in $D_{rel}$ are skipped.
Over-planning Pruning: Redundant subqueries based on hallucinated needs are eliminated because the planner verifies assumptions against actual retrieved documents.
Under-planning Refinement: The broader semantics of the initial retrieval help recover dependencies that narrow, targeted subqueries might miss.

2. Trajectory Generation Phase

The skeletal plan is executed via an iterative Act-Review-Update cycle (adapted from PAR-RAG):

Act: Retrieves documents for the specific subquery $q_t$ and generates a provisional answer $\hat{a}_t$ , citing supporting evidence.
Review: Constructs a verification query to perform secondary retrieval, cross-checking $\hat{a}_t$ against new evidence to correct factual inconsistencies.
Update: Commits the verified tuple $(\theta_t, q_t, d_t, a_t)$ to Trajectory Memory and refines the next subquery $q_{t+1}$ based on the accumulated evidence, ensuring runtime adaptability.

3. Answering Phase

The system synthesizes the final answer $A$ by integrating the query $Q$ with the fully grounded and verified trajectory $T$ , ensuring the response is logically consistent and strictly supported by retrieved documents.

Key Contributions

Framework Proposal: GDP-RAG is a training-free, plan-based multi-step RAG framework that retrieves first and plans only for the information delta, yielding concise reasoning trajectories without sacrificing answer quality.
Performance: Extensive experiments demonstrate that GDP-RAG achieves the highest accuracy among compared systems while maintaining significantly lower computational costs.
Ablation Analysis: The authors confirm that each design choice (preliminary retrieval, gap-aware instructions, and skeletal trajectory) contributes uniquely to step efficiency, cost reduction, and answer accuracy.

Experimental Results

The framework was evaluated on three multi-hop QA benchmarks: HotpotQA, 2WikiMultiHopQA, and MuSiQue, using GPT-4.1-mini for generation.

Accuracy: GDP-RAG achieved the highest overall accuracy of 60.63%, surpassing the strongest baseline, PAR-RAG (58.20%), by 2.43 percentage points. It also outperformed PAR-RAG on HotpotQA (64.75% vs. 62.79%) and MuSiQue (44.53% vs. 42.70%).
Efficiency (Cost-of-Pass): The primary metric, cost-of-pass (expected cost to produce one correct answer), was 0.51 for GDP-RAG. This is 22% lower than PAR-RAG (0.65) and 68% lower than KnowTrace (1.57).
Pareto Optimality: GDP-RAG occupies the "high accuracy, low cost" corner of the performance space. No other method achieved both higher accuracy and lower cost. Cheaper methods (e.g., Standard RAG, IRCoT) trailed by 10–23 accuracy points.
Step Reduction: Grounded Delta Planning reduced valid reasoning steps by 61.0% (from 3,561 to 1,388) and pruned 77.0% of redundant steps compared to Direct Planning.

Significance and Claims

The paper claims that GDP-RAG addresses the fundamental inefficiencies of current multi-hop RAG systems by fundamentally altering the planning process. By grounding the planner in preliminary retrieval, the system avoids the "blind" generation of reasoning steps that characterizes both step-wise error propagation and plan-based over-planning.

The authors emphasize that their approach yields concise, reliable reasoning trajectories by focusing computation exclusively on unresolved gaps. The results suggest that "Grounded Delta Planning" allows systems to adapt to task complexity, amortizing the cost of additional hops while maintaining high coverage of necessary information. The framework is presented as a robust, training-free solution that improves both the economic efficiency and the factual reliability of multi-step reasoning in LLMs.