acceptodds
Under review as a conference paper at ICLR 2027

Break the Chain, Gain a Margin: A Theory of Multi-Step Reasoning in LLMs

Abstract

Despite advances in retrieval-augmented reasoning through reinforcement learning (RL), it remains unclear when and why language models struggle with multi-step reasoning. We study this issue using a graph-based mathematical framework, in which a language model (LM) identifies valid reasoning steps by using statements as nodes and correct links as edges. We present a key theoretical result suggesting that such LMs are capable of internally identifying long reasoning chains. However, without multi-step reasoning, their confidence (margin) in distinguishing true from false links decreases as chains grow longer, making long-range dependencies increasingly error-prone. Motivated by this analysis, we develop theoretical principles for multi-step reasoning that break long chains into shorter segments with larger training margins. We instantiate these principles in PHyRR (Plan-Conditioned Hypergraph Retrieval-Reasoning), a reinforcement learning framework for question answering (QA). PHyRR first decomposes queries into smaller sub-goals during planning, then performs hypergraph retrieval and reasoning adaptively based on the resulting plan. Its RL rewards further encourage reasoning steps that are theoretically predicted to be more effective. Experiments confirm the predicted margin behavior and show that PHyRR outperforms baselines across four QA benchmarks, with the largest improvements on multi-hop reasoning tasks.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models, Retrieval-augmented Reasoning, Multi-hop question answering

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qKwKOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2(+1)
Soundness
2
Read review
Paper summary

This paper proposes PHyRR, a RL framework for QA, which combines planning and retrieval and is optimized through RL. Experiments with Qwen2.5-1.5B and 3B show improvements over baselines on both single-hop and multi-hop QA tasks, while ablation studies examine the contributions of planning and the reward design.

Critical strengths
  1. The overall design of PHyRR is reasonable, which first decomposes complex questions into sub-goals and then uses the plan to guide retrieval and reasoning.
  2. The experimental results show promising improvements over the evaluated baselines across four datasets and two model sizes.
Critical weaknesses
  1. The presentation of this paper is not good enough. Related work is placed in the appendix. There is no model diagram to help readers understand the framework. Figures 1–3 are overly compressed and difficult to read.
  2. Some important baselines are missing. For example, GraphRAG-R1 is closely related to this work and is cited, but it is not included in the experiments.
  3. The proposed mechanism needs more evidence. Ablation studies does not examine planning without the extra rewards, making it hard to separate their effects.
Questions for authors
  1. How does your work compare to GraphRAG-R1?
  2. Could you add an ablation that keeps sub-goal decomposition but disables additional rewards to directly support the performance gains of the proposed mechanism?
1 comment

We thank the reviewer for the constructive feedback. We address each point below.

Question1:

We thank the reviewer for pointing this out. GraphRAG-R1 trains a GraphRAG agent with GRPO, but differs from PHyRR in three ways: it has no explicit plan, while PHyRR generates a plan before reasoning that also guides retrieval (Eqs. 6–8); it requires an SFT stage on data generated by another LLM to address cold start, while PHyRR is trained with RL only and the plan itself mitigates cold start, and this removes both the cost of generating SFT data with an external LLM and the cost of a separate training stage. Finally, its rewards control the number of retrieval calls, while PHyRR’s rewards target grounding and plan consistency. We did not include GraphRAG-R1 as a baseline because it uses a different setting from ours: it retrieves pairwise triplets combined with text passages rather than hypergraphs, and it requires an additional SFT stage with externally generated data. A fair comparison would therefore require re-implementing both its retrieval and its training pipeline in our setting. Instead, we selected one representative method per setting (Table 3). We chose Graph-R1 for RL-based graph retrieval because, like PHyRR, it retrieves from a knowledge hypergraph and is trained only with RL, making it the most directly comparable method. Due to the limited rebuttal period, we could not complete a fair comparison with GraphRAG-R1 under our setting.

Question2:

Thank you for this suggestion. We added the requested ablation (HG + PL, no MR) in Table 2. Planning improves average F1 over the modified hypergraph compared with the hypergraph-only model.

About Presentation: We agree and have revised the paper accordingly: (i) a related work section is now in the main text (Section 2); (ii) we added a framework figure (Figure 1 in Section 4) showing the full pipeline and its components; (iii) the former Figures 1–3 (now Figures 2–4) are enlarged so that labels and legends are readable.

Reviewer_i5t4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies multi-step reasoning through a graph-embedding formulation in which passages are nodes and valid support relations are directed edges forming a partially ordered set. Its central theoretical claim is that long reasoning chains remain representable in a dimension determined by graph width, but the valid connections between reasoning steps may become harder to learn reliably as the chains grow longer. Motivated by this analysis, the authors propose PHyRR, which combines initial planning, plan-conditioned hypergraph retrieval, adaptive retrieval–reasoning, and a multi-component reward optimized using GRPO. The paper evaluates the framework on HotpotQA, 2WikiMultiHopQA, PopQA, and TriviaQA with two Qwen2.5 backbones, alongside synthetic embedding experiments, component ablations, and an analysis of early-stage learning behavior.

Critical strengths

S1: Consistent empirical improvements. PHyRR outperforms the evaluated baselines across four QA benchmarks and both model sizes. In particular, it achieves substantial improvements on multi-hop QA tasks, demonstrating the practical effectiveness of the proposed framework.

S2: Evidence from component-level evaluations. The ablation studies and early-stage training analysis show that planning and multi-component rewards contribute to performance improvements, particularly for smaller models. These findings provide useful empirical support for the overall framework.

Critical weaknesses

W1: Insufficient validation of the theoretical claims. Although the paper provides mathematical arguments linking reasoning depth, representation margins, and learning difficulty, the experiments do not directly establish whether these relationships hold in actual LLM reasoning. The theoretical evaluations are largely limited to synthetic graph embeddings, and the paper lacks analysis demonstrating that PHyRR improves the proposed margin or reduces reasoning errors through the claimed mechanism. Furthermore, the proof of Theorem 2 for general partially ordered sets (posets) appears incomplete, as it treats each chain independently without showing how valid reasoning relationships between different chains are preserved.

W2: Limited novelty relative to existing planning-based QA methods. Question decomposition, planning, and adaptive retrieval have already been extensively explored in multi-hop QA and agentic reasoning frameworks, including CoopRAG [1] and multi-agent approaches [2, 3]. While PHyRR integrates these techniques with hypergraph retrieval and RL, its methodological distinction from existing work [3, 4, 5] remains unclear. The absence of comparisons with planning-aware baselines further limits the evidence for its novelty.

[1] Cooperative Retrieval-Augmented Generation for Question Answering: Mutual Information Exchange and Ranking by Contrasting Layers (NeurIPS, 2025)

[2] MindSearch: Mimicking Human Minds Elicits Deep AI Searcher (ICLR, 2025)

[3] Improving Retrieval-Augmented Generation through Multi-Agent Reinforcement Learning (NeurIPS, 2025)

[4] Graph-R1: Towards Agentic GraphRAG Framework via End-to-end Reinforcement Learning (ICML, 2026)

[5] DECOR: Learning to Decompose and Collaborate in Deep Search via Multi-Agent Reinforcement Learning (ICML, 2026)

Questions for authors

Q1: Theoretical validity. In Appendix C, the proof of Theorem 2 appears to overlook valid relations between nodes belonging to different chains. For example, consider a width-2 poset with (a→c) and (b→c), where (a) and (b) are incomparable. Under the chain partition ({a,c}) and ({b}), Eq. (20) assigns zero inner product to (b) and (c), despite their valid relation. How does the proposed construction account for such cross-chain relations?

Q2: Empirical verification of the theory. Can the authors provide direct evidence that planning increases representation margins or reduces reasoning errors in actual LLMs? An analysis across different reasoning depths, rather than only synthetic embedding experiments, would help establish the connection between the theoretical claims and the improvements of PHyRR.

Q3: Novelty and baseline comparisons. Planning, question decomposition, and plan-guided retrieval have already been extensively explored in multi-hop QA and agentic reasoning [1–4]. Given these existing approaches, what is the fundamental methodological novelty of PHyRR beyond integrating planning, hypergraph retrieval, and RL? Could the authors compare against stronger planning-aware baselines or evaluate Graph-R1 with an equivalent planning component under matched computational budgets to better isolate the contributions of PHyRR?

[1] D²Plan: Dual-Agent Dynamic Global Planning for Complex Retrieval-Augmented Reasoning (ACL, 2026).

[2] Retrieval as Generation: A Unified Framework with Self-Triggered Information Planning (ACL, 2026).

[3] THOR: A Theta-Gamma Hierarchical Oscillatory Reasoning Framework for Multi-hop QA (ACL, 2026).

[4] DecoupleSearch: Decouple Planning and Search via Hierarchical Reward Modeling (EMNLP, 2025).

3 comments

Q1: Theoretical validity

We thank the reviewer for this observation, but there seems to be a misunderstanding. As a clarification, the partitions are over edges, not the nodes. We explain the process in the mentioned example. In this case, there are indeed two partitions (chains) C1=(a,c)C_1=\\{(a,c)\\} and C2=(b,c)C_2=\\{(b,c)\\}; both are directed graphs on the same set of nodes a,b,c\\{a,b,c\\}, and hence we embed all nodes (nodes with no connection will receive zero vectors for both keys and queries, which is consistent with our definition of embedding). This gives us two sets of embeddings:

&q^{(1)}(a)=\begin{pmatrix}1\\\\-1\end{pmatrix},\\ q^{(1)}(b)=\begin{pmatrix}0\\\\0\end{pmatrix},\\ q^{(1)}(c)=\begin{pmatrix}1\\\\-2\end{pmatrix},\\\\ &k^{(1)}(a)=\begin{pmatrix}1\\\\1\end{pmatrix},\\ k^{(1)}(b)=\begin{pmatrix}0\\\\0\end{pmatrix},\\ k^{(1)}(c)=\begin{pmatrix}2\\\\1\end{pmatrix} \end{aligned}$$ and $$\begin{aligned} &q^{(2)}(a)=\begin{pmatrix}0\\\\0\end{pmatrix},\\ q^{(2)}(b)=\begin{pmatrix}1\\\\-1\end{pmatrix},\\ q^{(2)}(c)=\begin{pmatrix}1\\\\-2\end{pmatrix},\\\\ &k^{(2)}(a)=\begin{pmatrix}0\\\\0\end{pmatrix},\\ k^{(2)}(b)=\begin{pmatrix}1\\\\1\end{pmatrix},\\ k^{(2)}(c)=\begin{pmatrix}2\\\\1\end{pmatrix} \end{aligned}$$ Now, we build our final embedding by taking $$e_1=\begin{pmatrix}1\\\\0\end{pmatrix},\\ e_2=\begin{pmatrix}0\\\\1\end{pmatrix}$$ and following Eq. (20) in the appendix. For example, we have $$q(c)=q^{(1)}(c)\otimes e_1+q^{(2)}(c)\otimes e_2=\begin{pmatrix}1\\\\-2\\\\0\\\\0\end{pmatrix}+\begin{pmatrix}0\\\\0\\\\1\\\\-2\end{pmatrix}=\begin{pmatrix}1\\\\-2\\\\1\\\\-2\end{pmatrix}$$ The full list becomes: $$\begin{aligned} &q(a)=\begin{pmatrix}1\\\\-1\\\\0\\\\0\end{pmatrix},\\ q(b)=\begin{pmatrix}0\\\\0\\\\1\\\\-1\end{pmatrix},\\ q(c)=\begin{pmatrix}1\\\\-2\\\\1\\\\-2\end{pmatrix},\\\\ &k(a)=\begin{pmatrix}1\\\\1\\\\0\\\\0\end{pmatrix},\\ k(b)=\begin{pmatrix}0\\\\0\\\\1\\\\1\end{pmatrix},\\ k(c)=\begin{pmatrix}2\\\\1\\\\2\\\\1\end{pmatrix} \end{aligned}$$ Should there be further questions, we will be happy to discuss them.

Q2: Empirical verification of the theory

We thank the reviewer for this suggestion. We added an analysis of the actual Qwen2.5 models (Appendix C), comparing PHyRR with its variant without planning on 2WikiMultiHopQA. Rather than measuring margins inside the LLMs directly, which we leave to future work, we test two observable predictions of our theory.

Our theory predicts that the margin shrinks with the height of the inference chain (its longest sequential path), whereas parallel branches only add width, which does not affect the margin. If planning helps by breaking long chains into shorter ones, its benefit should therefore appear on questions with taller chains. To test this, we group the questions of 2WikiMultiHopQA by height: Height 1: comparison questions, which consist of independent one-step lookups. Height 2: compositional, inference, and bridge-comparison questions, which require sequential two-step chains, where the second step depends on the answer to the first.

The table below reports, separately for each model size, the effect of planning, computed as the score of PHyRR minus the score of its variant without planning, in percentage points (pp). Share is the fraction of test questions in each group. The first number in each cell is the change within that group. The number in parentheses is the same change weighted by the group's share, i.e., its contribution to the overall change; the weighted values of the two heights therefore add up to the All row.

ModelHeightShareΔEMΔF1
Qwen2.5-3B125%−3.1 (−0.8)−3.1 (−0.8)
275%+5.2 (+3.9)+2.4 (+1.8)
All100%+3.1+1.1
Qwen2.5-1.5B125%−12.5 (−3.1)−13.8 (−3.4)
275%+8.3 (+6.3)+5.2 (+3.9)
All100%+3.1+0.5

Consistent with the theory, planning improves EM and F1 on height-2 questions for both model sizes, which account for the overall gain. On height-1 questions, there is no chain to decompose, so the plan only adds steps, each an additional opportunity for error. Since 2WikiMultiHopQA contains only chains of height 1 and 2, these results are consistent trends rather than statistically conclusive evidence, and the trend at larger heights remains to be verified.

Q3: Novelty and baseline comparisons

We thank the reviewer for these references. We would first like to clarify that the framework alone is not the only novelty of our work. One part of our contribution is the theoretical analysis of how chain length affects the learnability of reasoning links, and PHyRR is developed as an instantiation of this analysis. Independent of this theoretical contribution, the framework itself also introduces new elements: PHyRR generates a single plan before reasoning, uses it to anchor both the policy and the retrieval over a refined knowledge hypergraph, and trains the whole process with RL only. So, beyond using planning, hypergraph retrieval, and RL, PHyRR introduces new mechanisms in each of them:

  1. Plan-conditioned retrieval scoring: In the cited methods, the plan guides which queries the model generates. In PHyRR, the plan also changes how retrieval ranks results: each query is matched to its closest sub-goal, and entity and hyperedge scores are reweighted by their similarity to that sub-goal (Eqs. 6–8). Retrieval itself thus becomes plan-aware.

  2. A fixed plan as an anchor for RL training: MindSearch, MMOA-RAG, DECOR, and D²Plan distribute planning, searching, or filtering across multiple agents. PHyRR uses a single policy that generates one plan before reasoning and keeps it fixed as an anchor for the whole trajectory. D²Plan requires an SFT stage before RL, and CoopRAG, GRIP, and THOR rely on prompting or supervised training; DecoupleSearch requires value models and tree search at inference. PHyRR uses a single policy and a single plan, generated once and kept fixed for the whole trajectory, and is trained with RL only. This fixed plan mitigates the cold-start problem of RL training.

  3. Multi-component reward for process quality: Most RL-based methods reward only the final outcome and the output format (e.g., Search-R1, Graph-R1). PHyRR adds two process rewards that evaluate the trajectory itself (Eqs. 10–13): a reasoning reward and a retrieval reward.

  4. Refined knowledge hypergraph: PHyRR merges synonymous entities and redundant hyperedges using contextual embeddings and verification (Appendix D), reducing the ambiguity of hypergraph retrieval.

We selected one representative method per setting (Table 3) and chose Graph-R1 for RL with hypergraph-based retrieval because it is the closest to PHyRR. Due to the limited rebuttal period, we could not complete the suggested run of Graph-R1 with an equivalent planning component or other plan baselines. But we agree that this is a valuable suggestion, and we plan to include a comparison with a planning-based baseline in future work.

Thanks for your detailed responses regarding theoratical clarification, empirical verification, and comparisons with baselines. Although the authors clarified the theoretical validity of Theorem 2 via a running example, it seems that the remaining concerns have been partially (or insufficiently addressed. Therefore, I maintain my score.

Reviewer_nzooOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper asks why LLMs struggle to learn long multi-step reasoning chains even when they have the capacity to represent them. The authors model statements and valid support relations as a poset inference graph, in which an edge is predicted by the sign of a query–key inner product. The main theoretical claim separates width from height: the required embedding dimension is bounded by the poset's width, while for a fixed dimension the representation margin has an upper bound that decreases with chain height. Together with an SVD perturbation analysis under noisy edge sampling, this motivates splitting long internal chains into shorter segments through external actions. Based on this view, the authors propose PHyRR. It generates an initial sub-goal plan, uses the plan to guide hypergraph retrieval, and trains the policy with GRPO on answer, format, reasoning, and retrieval rewards. Experiments with Qwen2.5-1.5B/3B on HotpotQA, 2WikiMultiHopQA, PopQA, and TriviaQA show consistent gains over Search-R1 and Graph-R1, supported by ablations and synthetic poset experiments.

Critical strengths
  1. Interesting separation of representation and learnability. The distinction between representing a long chain and learning it reliably is useful. The width/height separation gives this question a concrete mathematical form. Despite my concerns about the proof, a correct result of this kind could help explain why decomposition helps.
  2. Substantial empirical gains. In Table 1, PHyRR improves average F1 over Graph-R1 by 4.27 points (3B) and 9.03 points (1.5B), with gains on all four datasets. Table 2 also suggests that planning helps when the additional rewards are enabled.
Critical weaknesses
  1. Theorem 2 and its proof (Appendix C) contain substantive errors.

    • Inconsistent normalization: Definition 2 implies ∥A∥F2=n2N\|A\|_F^2 = n^2N, while Appendix B/C uses NN, so the two margins differ by a factor of nn. Under Definition 2, a total order admits a rank-two construction (scores proportional to j−i+1j-i+1) whose margin approaches 3\sqrt{3} as the chain grows. This contradicts the vanishing bound in Theorem 2.
    • QQ in Eq. (17) is not orthogonal: For n=3n=3, columns 2 and 3 are (−2,1,1)⊤/6(-2,1,1)^\top/\sqrt{6} and (−1,−1,2)⊤/6(-1,-1,2)^\top/\sqrt{6}, and their inner product is 1/21/2. The variational form of the nuclear norm therefore does not apply, and Eq. (19) is not justified.
  2. The link between theory and method is weak, and the experiments are hard to attribute.

    • A relaxed upper bound on the margin does not show that the trained model attains a larger margin, and no margin measurement is reported for the Qwen models. The rewards (length, keyword, and overlap checks) are not connected to the theoretical quantities.
    • Incomplete ablation: Table 2 lacks the w/o MR (PL+HG) row, so the effect of planning alone is unknown. The analysis also does not separate the effect of putting the plan in the prompt from that of plan-guided retrieval reweighting (ψV,ψH\psi_V,\psi_H, Eqs. 6–8).
Questions for authors
  1. How do you reconcile the factor-of-nn normalization gap between Definition 2 and Appendix B? How can Part 2 of Theorem 2 be repaired, given that QQ in Eq. (17) is not orthogonal? How should Eqs. (20)–(21) be modified to handle cross-chain relations u→vu\to v (u∈Cku\in C_k, v∈Ck′v\in C_{k'}) in a general width-ww poset?
  2. Please reconcile the numbers in Secs. 4.2–4.3 with Table 2, and report (a) PHyRR w/o MR (planning kept) and (b) the plan kept in the prompt with the retrieval weights ψV,ψH\psi_V,\psi_H removed.
  3. Which LLM builds and cleans GH\mathcal{G}_H and generates the initial plan? Does Graph-R1 use the same refined hypergraph and the same training and inference compute budget as PHyRR? Please also report the missing reward hyperparameters.
1 comment

Question 1:

About the inconsistent scale in Definition 2, the reviewer is correct. The scaling factor 1/n1/n in the two equations in Eq. (1) must be removed; the expressions and the result are then consistent with the appendix. Note that our intention is to have an embedding whose corresponding map $ A_{ij} = \langle q(u_i), k(u_j) \rangle $ is comparable to a common adjacency matrix, that is, $ \|A\|_F^2 = N . $

This is achieved by Eq. (1) without the 1/n1/n factors, since then ∣A∣F2=Tr(D)=N\\|A\\|_F^2 = \mathrm{Tr}(D) = N, and this is what we followed in the proof. By this convention, the margin of the rank-two embedding of a chain mentioned by the reviewer is of order O(1/n)O(1/n), which is consistent with the vanishing bound of Theorem 2.

We checked the construction of QQ in Eq. (17) and confirm that it is not orthonormal. It seems that we had only partially verified the orthogonality of the columns of QQ. Note that it is not essential for QQ to be orthonormal, as long as it has a unit operator norm, since Eq. (19) holds for any such matrix: $ \|A\|_* = \max_{\|P\|_{\mathrm{op}} \le 1} \mathrm{Tr}(P^\top A). $

However, we checked that the operator norm of this QQ also grows (it equals n/2\sqrt{n/2}), so the matrix QQ is not applicable.

Fortunately, there are multiple constructions that can resolve the issue. Here is an explicit construction of QQ. Take B=(bij)B = (b_{ij}) with

\begin{cases} 1 & j \ge i,\\\\ 0 & j < i, \end{cases}$$ and take \$ Q = B\\,(B^\\top B)^{-1/2}. \$ This matrix is clearly orthonormal, as it is the orthogonal polar factor of $B$, and we can prove that it has the correct required sign pattern ($q_{ij} > 0$ for $j \ge i$ and $q_{ij} < 0$ for $j < i$), with \$ \\sum_{j \\ge i} q_{ij} = \\Theta(n \\log n). \$ Replacing it in the proof leads to the desired result. Since the sign patterns of $Q$ and of the weight matrix $A$ are identical, Eqs. (18)–(19) give $$c\\, n \log n \cdot \epsilon \\;\le\\; \mathrm{Tr}(Q^\top A) \\;\le\\; \\|A\\|_* \\;\le\\; \sqrt{m}\\,\\|A\\|_F \\;=\\; \sqrt{mN} \\;\le\\; n\sqrt{m},$$ and hence $\epsilon \le c\sqrt{m}/\log n$ for a chain, and $\epsilon \le c\sqrt{m}/\log h$ for a general poset. However, the upper bound on the speed of vanishing of the margin should therefore be changed from $O(\sqrt{m/h})$ to \$ \\epsilon \\;\\le\\; c\\,\\frac{\\sqrt{m}}{\\log h}. \$ This is indeed much slower than our original proposal, but it is sufficient to support our general discussion, since the margin still vanishes as the chain grows. We still believe that the speed should be closer to polynomial, but this needs a completely different proof strategy and can be a great follow-up to this initial study. **Question2**: We thank the reviewer for this comment. We have clarified the description in the revised text as follows: 1. In Section 4.3, the values 3.12% and 6.09% are the relative drops in average EM, not F1, when removing the plan: 1.63/52.23 = 3.12% for Qwen2.5-3B and 2.76/45.31 = 6.09% for Qwen2.5-1.5B. The corresponding F1 drops are 0.92 points (1.57%) and 1.61 points (3.10%). 2. In Section 4.2, the 2Wiki example (41.41 → 28.90) compares the full model with the variant without both planning and the multi-component reward, not without planning alone (38.27). The revised text now states this explicitly. 3. We also corrected two rounding errors in the Δ column of Table 2 (+1.64 → +1.63 and +1.62 → +1.61). Following the reviewer's suggestion, we added (a) PHyRR w/o MR (planning kept, i.e., PL+HG) to Table 2. Due to the limited rebuttal period, we could not yet complete (b), keeping the plan in the prompt with the retrieval weights removed. Regarding the reviewer's concern about how the multi-component reward relates to our theoretical analysis, we clarify that not all reward terms come from the theory. Some follow it directly, such as the upper length bound, which limits the number of internal reasoning steps, keeping internal chains short. The answer-in-retrieval term rewards obtaining the needed facts through retrieval, so that fewer links must be handled internally. The plan-consistency term keeps the trajectory a single structured chain. Some of them are practical; for example, the answer and format terms are standard in RL-based retrieval, and the keyword term encourages explicit reasoning. **Question3**: We address the reviewer's questions below: - The knowledge hypergraph was built with GPT-4o-mini. This is done once, offline, before training. The initial plan is generated by the same Qwen2.5-Instruct model used as the backbone (3B or 1.5B, respectively). - Graph-R1 in Table 1 uses its original hypergraph construction. Also, all RL-based baseline methods and PHyRR are trained with the same backbones, training data, and number of training steps. At inference, PHyRR requires one additional generation for the plan.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.