acceptodds
Under review as a conference paper at ICLR 2027

Eliciting Interaction-of-Thought Reasoning in LLM Agents for Personalized Recommendation

Abstract

The emergent reasoning capabilities in Large Language Models (LLMs) have given rise to a surge of research interests in developing personalized agentic systems, particularly in the field of information retrieval. However, it remains a challenge to harness the emergent reasoning capabilities of LLMs for personalized recommendation tasks, largely because general natural language corpora inherently lack recommendation-oriented reasoning traces. To bridge this gap, we introduce R2Rec, a reasoning-enhanced recommendation framework that derives recommendation-oriented reasoning supervision directly from user-item interactions. R2Rec samples interaction chains from the user-item graph to expose high-order collaborative signals and transforms them into structured Interaction-of-Thought traces through a progressive, masked agentic pipeline. By grounding intermediate reasoning in progressively revealed interaction evidence while keeping the final conclusion consistent with the user’s actual interaction, these traces provide step-wise supervision for preference reasoning. We then internalize and further refine this capability through post-training: synthesized traces are used for supervised learning of structured reasoning over interaction chains, while recommendation-specific rewards encourage sufficient reasoning decomposition and accurate candidate ranking. Experimental results on three real-world datasets show that R2Rec improves over its backbone LLM by 195.12% on average across three ranking metrics and outperforms state-of-the-art baselines by 9.87%. Furthermore, a human study demonstrates the superior interpretability of R2Rec's reasoning, which achieves a 9.07% improvement over existing methods. Our code is open-source at https://anonymous.4open.science/r/R2Rec-7C5D for reproducibility.

Primary area:
foundation or frontier models, including LLMs
Keywords:
large language model, recommendation system

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_WthKOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes R2Rec, a reasoning-enhanced LLM recommendation framework that leverage collaborative information from user-item interaction graphs. The method first samples interaction chains from the graph and transforms them into structured Interaction-of-Thought reasoning traces. These traces are than used for supervised fine-tuning, followed by reinforcement learning with reasoning-length and ranking-correctness rewards.

Critical strengths
  1. the paper is motivated, and easy to follow.
Critical weaknesses
  1. The proposed framework largely combines established techniques, including user-item graph, CoT prompting, SFT on generated reasoning trace, and GRPO. Similar ideas have been explored in existing graph + reasoning-enhanced recommendation methods. The paper should distinguish its core contribution from prior work.
  2. The latest baselines in the paper are from 2025; the authors should include more recent SOTA methods, as this is a fast-moving domain.
Questions for authors

what fundamentally distinguishes interaction-of-thought from existing approaches? and which LLM is used to generate the Interaction-of-Thought reasoning traces?

7 comments
AuthorsOpenReview ↗

Response to Reviewer WthK - Part 1/2

Thank you for your attention to the core contribution of Interaction-of-Thought, comparisons with more recent methods, and the reasoning-trace generator. Below, we clarify the methodological distinction of R2Rec and provide additional experimental evidence.

W1 & Q1: What fundamentally distinguishes Interaction-of-Thought from existing graph + reasoning approaches?

The core methodological contribution of R2Rec is a collaborative-structure-to-reasoning-supervision process that transforms implicit collaborative relations in the recommendation graph into factual, learnable, step-wise preference reasoning.

This differs from recent methods such as RecLLM-R1 [1] and ReRec [2] in technical focus. RecLLM-R1 mainly studies how reasoning and reinforcement-learning-based post-training can improve LLM recommendation, while ReRec focuses on improving reinforcement fine-tuning for complex recommendation reasoning. R2Rec instead addresses an upstream problem: how to ground the model's reasoning itself in verifiable collaborative evidence from recommendation data before subsequent optimization.

Interaction-of-Thought defines three complementary mechanisms.

First, closed interaction chains define where the reasoning evidence comes from. They extract multi-hop collaborative relations from the user-item graph and turn implicit graph connectivity into explicit, traceable behavioral evidence.

Second, progressive masking keeps intermediate reasoning grounded in evidence. At each step, the model can only use the interaction information revealed up to that point. The resulting reasoning path is therefore built progressively on factual collaborative interactions rather than generated as unconstrained free-form CoT.

Third, outcome-level guidance keeps the final conclusion consistent with observed user behavior. Consequently, Interaction-of-Thought combines intermediate evidence grounding with final outcome consistency.

We therefore define Interaction-of-Thought as:

a mechanism that transforms collaborative interaction structure into evidence-grounded, step-wise reasoning supervision while keeping the final conclusion consistent with observed user behavior.

The key distinction is not simply the textual form of the reasoning trace, but that each reasoning step has explicit factual interaction evidence as its basis.

The resulting IoT traces are subsequently internalized through supervised learning and further refined with recommendation-specific rewards. In other words, SFT and GRPO determine how the constructed reasoning supervision is learned and further optimized, whereas Interaction-of-Thought determines what recommendation evidence the model should reason over and where the supervision comes from.

We further validate these mechanisms through matched controls:

Answer-only supervision removes the synthesized intermediate reasoning traces and learns only the final answer; Shuffled Interaction Chains disrupt the original collaborative structure; Unstructured Interactions preserve the interaction information but remove the structured chain organization.

Method / SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
Answer-only supervision0.355 / 0.480 / 0.4260.528 / 0.606 / 0.5720.402 / 0.536 / 0.478
Shuffled Interaction Chains0.371 / 0.497 / 0.4440.540 / 0.619 / 0.5860.418 / 0.551 / 0.495
Unstructured Interactions0.384 / 0.515 / 0.4600.561 / 0.630 / 0.6000.443 / 0.572 / 0.517
R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545

R2Rec consistently outperforms all three controls. In particular, performance drops when collaborative relations are disrupted or structured evidence is replaced with unstructured interactions even when the interaction content is preserved.

These results support our central positioning: R2Rec improves recommendation by grounding reasoning in factual collaborative interaction evidence and converting user-item structure into effective recommendation reasoning supervision.

AuthorsOpenReview ↗

Response to Reviewer WthK - Part 2/2

W2: Comparison with more recent recommendation-reasoning methods

To align the added baselines with the research question of R2Rec, we prioritize recent methods that directly study LLM reasoning for recommendation.

Specifically, we include VRec(2026.03) [3] and IRMD(2026.07) [4]. VRec studies verification of recommendation reasoning, i.e., how to improve the reliability of the reasoning process, while IRMD studies how to convert explicit reasoning into more efficient implicit reasoning through multi-teacher distillation. They represent recent directions in Reasoning for Recommendation on reasoning verification and reasoning internalization/efficiency, respectively.

R2Rec addresses a complementary question: how reasoning supervision itself can be derived from recommendation interactions and grounded in collaborative evidence.

Accordingly, this comparison does not select baselines according to a particular optimizer, but instead compares different reasoning mechanisms within the same Reasoning-for-Recommendation research context.

Method / SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
VRec0.386 / 0.518 / 0.4620.564 / 0.625 / 0.5960.427 / 0.579 / 0.519
IRMD0.392 / 0.523 / 0.4710.568 / 0.642 / 0.6080.430 / 0.567 / 0.517
R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545

After adding these more recent recommendation-reasoning methods, R2Rec remains stronger across all metrics on all three datasets.

These results show that directly constructing evidence-grounded, step-wise reasoning supervision from high-order collaborative interactions provides a stable and effective recommendation-reasoning signal.

Q2: Which LLM generates the Interaction-of-Thought reasoning traces?

R2Rec uses DeepSeek-R1-Distill-Llama-8B as the downstream recommendation model, while the original Interaction-of-Thought traces are generated by GPT-4o.

We further use the stronger GPT6.1-sol model as the reasoning-trace generator while keeping the downstream recommendation model and subsequent training procedure unchanged.

Trace GeneratorMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
GPT-4o0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545
GPT6.1-sol0.438 / 0.578 / 0.5200.611 / 0.683 / 0.6530.485 / 0.620 / 0.563

Using the stronger generator further improves downstream recommendation performance on all three datasets. This shows that higher-quality synthesized reasoning traces can provide stronger reasoning supervision and directly improve the recommendation-reasoning capability learned by the downstream model.

References

[1] Recllm-r1: A two-stage training paradigm with reinforcement learning and chain-of-thought v1.
[2] ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning. ACL'26.
[3] Verifiable Reasoning for LLM-based Generative Recommendation. SIGIR'26.
[4] Think, But Don't Tell: Implicit Reasoning for LLM-based Sequential Recommendation via Multi-Teacher Distillation. SIGIR'26.

Thank author for the clarification and additional experiments. However, my novelty concerns remains. Graph-based collaborative reasoning, answer-guided reasoning generation are established techniques and progressive marking from A.1 appears to be a prompting strategy. I appreciate the authors' efforts and will increase my score to 2.

Reviewer_5aicOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes R2Rec, which aims to enhance the recommendation reasoning capabilities of large language models by exploiting higher-order collaborative signals in user–item interaction graphs. The method first samples closed interaction chains, then provides an LLM with the chains, prediction questions, and known ground-truth answers to generate natural-language reasoning traces through progressive, masked prompting. The model subsequently learns from these reasoning examples through SFT and is further optimized using GRPO with two reward components: the number of reasoning steps and the rank of the positive item.

Critical strengths
  1. The problem has practical significance. Recommendation data typically lack explicit reasoning supervision, making the construction of reasoning examples from existing interactions a direction worth exploring. Short interaction chains also provide an intuitive way to organize graph structure into inputs for LLMs.

  2. The experiments cover multiple settings. In addition to the default ranking task with 20 candidates, the authors evaluate settings such as ranking 100 candidates. The appendix also includes numerous other experiments to demonstrate the robustness of the method.

Critical weaknesses
  1. Risk of temporal leakage.* Section 4.1.1 states that each user's interactions are sorted chronologically and that the most recent positively rated item is held out. However, it does not clearly specify whether graph construction and sampling of interaction chains across users respect the target prediction time. Even if each user's last interaction is held out, the retained histories of other users may still include interactions that occurred after the current target user's test time.

  2. Insufficient training data. The SFT and RL stages each use only 500 training examples. Since this work primarily uses LoRA, a training approach with low overhead, the amount of training data could be increased further. Although related experiments are provided in the appendix, they remain limited to the order of a thousand examples.

Questions for authors
  1. Which model is used to generate the reasoning traces? Is it the same as the backbone? Could using a stronger model for data synthesis yield greater performance improvements?

  2. Could you add controls under the same SFT and RL configurations using answer-only supervision, as well as shuffled interaction chains or chains replaced with an equal amount of unstructured interaction data, to verify the independent contributions of reasoning traces and graph structure?

  3. The appendix reports performance improvements as the amount of training data increases for the proposed method. Have you conducted comparable experiments that increase the amount of training data for other baselines, such as SASRec? Do traditional baselines achieve larger gains than the proposed method when trained on more data? This is my main concern. I would raise my score if these experiments were added and the training data were scaled to approximately 10,000 examples.

9 comments
AuthorsOpenReview ↗

Response to Reviewer 5aic - Part 1/2

Thank you for raising concrete questions regarding temporal validity, training-data scale, the reasoning-trace generator, and the independent contributions of reasoning traces and graph structure. We conducted additional experiments to directly address these points.

W1: Temporal leakage and strict temporal reconstruction

We conducted a temporal audit of the original graph construction and interaction-chain sampling protocol. The audit confirmed an approximately 14.3% temporal-leakage rate: in these cases, cross-user interactions used to construct the target user's interaction context occurred after the corresponding target prediction time.

To fully eliminate this factor, we reconstructed a strictly time-controlled dataset following the prediction-time constraint. For each target prediction time, we retained only historical interactions that were legally observable before that time. We then rebuilt the user-item graph, resampled the interaction chains, and retrained R2Rec on the reconstructed data.

SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
Original R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545
Strict Temporal Control0.406 / 0.542 / 0.4880.587 / 0.666 / 0.6370.470 / 0.596 / 0.541

After completely removing cross-user interactions occurring after the prediction time and retraining the model, the overall performance remains nearly unchanged across the three datasets.

MovieLens and Amazon Book show only minor changes, while HR@1 on Amazon Electronics slightly increases. The strict temporal reconstruction therefore shows that the main gains of R2Rec do not depend on future cross-user interaction information. When graph construction, chain sampling, and model training are all restricted to temporally valid historical interactions, the method retains performance close to the original results.

W2 & Q3: Scaling to 10,000 training examples and comparison with SASRec

We further scale R2Rec to 10,000 training examples, with 8,000 examples used for supervised reasoning-trace learning and 2,000 examples used for reward-guided refinement.

We also train SASRec with 10,000 training examples under the same evaluation setting.

Method / Training ScaleMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
R2Rec, original scale (500 SFT + 500 RL)0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545
SASRec, 10k0.420 / 0.556 / 0.5030.596 / 0.677 / 0.6460.473 / 0.611 / 0.561
R2Rec, 10k (8k SFT + 2k RL)0.449 / 0.584 / 0.5250.625 / 0.701 / 0.6690.507 / 0.641 / 0.585

As the training scale increases, R2Rec further improves to HR@1 values of 0.449 / 0.625 / 0.507, with corresponding gains in HR@3 and NDCG@3.

At the same 10k scale, R2Rec remains stronger than SASRec across all three datasets and all metrics. These results show that R2Rec is not limited to the small training regime used in the original submission. Increasing the amount of reasoning supervision and reward-guided training data continues to improve the model while preserving its advantage over a traditional sequential recommender at the larger training scale.

Q1: Reasoning-trace generator and stronger synthesis model

The downstream recommendation backbone of R2Rec is DeepSeek-R1-Distill-Llama-8B, while the original Interaction-of-Thought traces are generated by GPT-4o. Thus, the reasoning-trace generator and the downstream recommendation model are separate.

To further evaluate the effect of reasoning-supervision quality, we regenerate Interaction-of-Thought traces using the stronger GPT6.1-sol model while keeping the downstream recommendation model and subsequent training procedure unchanged.

Trace GeneratorMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
GPT-4o0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545
GPT6.1-sol0.438 / 0.578 / 0.5200.611 / 0.683 / 0.6530.485 / 0.620 / 0.563

The stronger trace generator further improves downstream recommendation performance on all three datasets. This shows that the quality of synthesized reasoning supervision affects final recommendation performance and that R2Rec's supervision-construction mechanism can further benefit from a stronger reasoning generator.

AuthorsOpenReview ↗

Response to Reviewer 5aic - Part 2/2

Q2: Answer-only, shuffled-chain, and unstructured-interaction controls

We add the requested matched controls under the same SFT/RL configuration.

Answer-only supervision removes intermediate reasoning traces and learns only the final answer; Shuffled Interaction Chains disrupt the original collaborative structure; Unstructured Interactions preserve the interaction information but no longer organize it into closed interaction chains.

Method / SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
Answer-only supervision0.355 / 0.480 / 0.4260.528 / 0.606 / 0.5720.402 / 0.536 / 0.478
Shuffled Interaction Chains0.371 / 0.497 / 0.4440.540 / 0.619 / 0.5860.418 / 0.551 / 0.495
Unstructured Interactions0.384 / 0.515 / 0.4600.561 / 0.630 / 0.6000.443 / 0.572 / 0.517
R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545

R2Rec outperforms all three controls across all metrics and datasets.

The degradation under Answer-only supervision shows that explicit intermediate reasoning supervision provides useful signals beyond the final answer. The degradation under Shuffled Interaction Chains shows that valid collaborative connectivity independently contributes to reasoning. The Unstructured Interactions result further shows that simply providing the same type of interaction content is insufficient; organizing this information into structured collaborative reasoning evidence provides additional gains.

Interaction-of-Thought therefore does more than generate additional reasoning text: it transforms high-order collaborative evidence in the user-item graph into step-wise reasoning supervision that the model can learn from.

Thanks for the rebuttal. I have also carefully read the comments from the other reviewers and all the rebuttal. I still have two remaining concerns:

  1. Performance comparison with SASRec. It appears that SASRec benefits more from the additional training data than the proposed method. It remains unclear whether SASRec could achieve comparable performance given a larger training dataset. As we know, the performance of methods like SASRec is highly sensitive to the amount of training data. More experiments with varying training data sizes would help clarify this point.

  2. Novelty and necessity of the proposed method. I also noticed the novelty concerns raised by the other reviewers. Existing methods may already be capable of learning reasoning capabilities from scratch, making it unclear what unique advantages the proposed approach brings. More experiments are needed to demonstrate both the superiority and the necessity of the proposed method over existing alternatives.

Reviewer_c5LyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes R2Rec, a model designed to enhance LLM reasoning for personalized recommendation. It constructs interaction chains from user-item graphs to capture high-order collaborative signals and transforms them into structured Interaction-of-Thought traces through progressive masked prompting. These traces are used for SFT, followed by GRPO with recommendation-specific rewards to further improve reasoning and ranking performance. Experiments on three real-world datasets demonstrate promising improvements over existing baselines, supported by ablation studies and transfer evaluations.

Critical strengths
  1. The paper addresses a well-motivated problem. While LLMs have demonstrated strong reasoning capabilities in various domains, effectively leveraging these capabilities for personalized recommendation remains challenging. The motivation is clearly presented, and the paper is generally well organized and easy to follow.

  2. The proposed framework makes reasonable adaptations to the recommendation setting by incorporating high-order collaborative signals from user-item interactions into the reasoning process. In particular, constructing interaction-based reasoning traces provides an interesting way to introduce recommendation-specific supervision.

  3. The experiments cover multiple datasets and include ablation studies and transfer evaluations, with encouraging improvements over the selected baselines.

Critical weaknesses
  1. My main concern is the unclear research gap. There have already been several RL-based recommendation methods, such as RecLLM-R1, ReRec, that aim to improve recommendation through reasoning and reinforcement learning. While the paper includes RecZero as a baseline, it does not sufficiently discuss these recent efforts. It is therefore difficult to see what specific limitation of existing methods R2Rec is trying to address.

  2. The novelty seems somewhat limited. Although interaction chains and masked prompting are designed for the recommendation setting, the overall approach still follows a fairly standard SFT + GRPO pipeline.

  3. It is not clear how much the proposed framework advances beyond existing RL-based recommendation methods. The experimental comparison is also incomplete, as several recent and closely related baselines are missing.

Questions for authors
  1. Could the authors clarify how R2Rec differs from recent RL-based recommendation methods, beyond the construction of interaction-based reasoning traces?

  2. Can the authors include comparisons with more recent RL-based recommendation approaches?

3 comments
AuthorsOpenReview ↗

Response to Reviewer c5Ly - Part 1/2

Thank you for your positive assessment of the motivation, Interaction-of-Thought design, and overall experimental results. Below, we further clarify the research positioning and methodological contribution of R2Rec and provide additional comparisons with recent related methods.

W1 & Q1: How does R2Rec differ from recent methods such as RecLLM-R1 and ReRec?

The core problem addressed by R2Rec is the lack of reasoning supervision in recommendation. User-item interactions reflect user preferences, but they usually do not provide explicit step-wise reasoning annotations explaining how these interactions support preference inference. We therefore focus on how to directly derive recommendation-oriented reasoning supervision from implicit collaborative interactions, with the reasoning process grounded in factual interaction evidence rather than freely generated without such evidence.

R2Rec differs from RecLLM-R1 [1] and ReRec [2] in this technical focus. RecLLM-R1 primarily studies how reasoning and reinforcement-learning-based post-training can improve LLM recommendation, while ReRec focuses on improving reinforcement fine-tuning for complex recommendation reasoning. R2Rec instead addresses an upstream problem before these optimization stages: when recommendation data lack explicit reasoning supervision, what factual evidence should the model reason over, and how can such supervision be constructed?

To this end, we first extract closed interaction chains from the user-item graph, transforming implicit high-order collaborative relations into traceable structured evidence. We then construct Interaction-of-Thought traces through progressive and masked generation, so that each intermediate reasoning step is grounded in progressively revealed collaborative interaction evidence, while the final conclusion remains consistent with observed user behavior.

Accordingly, the core pipeline of R2Rec is:

user-item collaborative structure
→ structured interaction evidence
→ evidence-grounded step-wise reasoning supervision
→ reasoning internalization
→ reward-guided refinement.

SFT is used to internalize the constructed reasoning supervision, while GRPO further refines reasoning and ranking. The main methodological contribution of R2Rec therefore lies in the supervision-construction stage: transforming high-order collaborative structure into factual, learnable, step-wise preference reasoning.

W2: How do we verify that the gains do not simply come from a standard SFT + GRPO pipeline?

To further isolate the contributions of reasoning supervision and collaborative structure, we added three controlled experiments. Answer-only supervision learns only the final answer without synthesized intermediate reasoning traces; Shuffled Interaction Chains disrupt the original collaborative structure; Unstructured Interactions preserve the interaction information but remove the structured chain organization. All other training components are kept aligned with the full method.

Method / SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
Answer-only supervision0.355 / 0.480 / 0.4260.528 / 0.606 / 0.5720.402 / 0.536 / 0.478
Shuffled Interaction Chains0.371 / 0.497 / 0.4440.540 / 0.619 / 0.5860.418 / 0.551 / 0.495
Unstructured Interactions0.384 / 0.515 / 0.4600.561 / 0.630 / 0.6000.443 / 0.572 / 0.517
R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545

R2Rec outperforms all three controls across all metrics on all three datasets.

The degradation under Answer-only supervision shows that intermediate reasoning supervision provides useful learning signals beyond the final answer. The degradation under Shuffled Interaction Chains shows that valid collaborative connectivity is important for constructing effective reasoning evidence. The Unstructured Interactions result further shows that simply providing the same interaction content is insufficient; organizing these facts into relationally structured reasoning evidence provides additional benefit.

These results directly show that the effectiveness of R2Rec cannot be explained by SFT or GRPO alone. The gains come from first learning step-wise reasoning supervision grounded in collaborative interaction evidence, after which post-training internalizes and refines this capability.

Appendix A.13.2 further ablates the reasoning-length reward and ranking-correctness reward, where the combined reward consistently performs best, supporting the effectiveness of the subsequent recommendation-specific refinement.

AuthorsOpenReview ↗

Response to Reviewer c5Ly - Part 2/2

W3 & Q2: Comparison with recent RL-based recommendation methods

Following your suggestion, we additionally compare with RecLLM-R1 [1], ReRec [2], and Rank-GRPO [3] under the same dataset splits and evaluation metrics.

Method / SettingMovieLens HR@1 / HR@3 / NDCG@3Amazon Book HR@1 / HR@3 / NDCG@3Amazon Electronics HR@1 / HR@3 / NDCG@3
RecLLM-R10.379 / 0.516 / 0.4580.550 / 0.637 / 0.5950.412 / 0.557 / 0.490
ReRec0.380 / 0.516 / 0.4590.562 / 0.625 / 0.5930.414 / 0.571 / 0.503
Rank-GRPO0.393 / 0.520 / 0.4660.567 / 0.640 / 0.6070.424 / 0.574 / 0.513
R2Rec0.411 / 0.549 / 0.4930.591 / 0.664 / 0.6320.467 / 0.602 / 0.545

After adding these recent methods, R2Rec remains stronger across all three datasets and all three evaluation metrics.

Together with the controlled experiments above, these results indicate that explicitly constructing evidence-grounded reasoning supervision before post-training is an independent and effective methodological layer. By transforming multi-hop collaborative structure into factual reasoning evidence, R2Rec enables the model to acquire interaction-aware preference reasoning before subsequent recommendation optimization.

References

[1] Recllm-r1: A two-stage training paradigm with reinforcement learning and chain-of-thought v1.
[2] ReRec: Reasoning-Augmented LLM-based Recommendation Assistant via Reinforcement Fine-tuning.
[3] Rank-GRPO: Training LLM-based Conversational Recommender Systems with Reinforcement Learning.

AuthorsOpenReview ↗

Follow-up on Our Response

Dear Reviewer c5Ly,

As the discussion period is approaching its end, we would like to kindly follow up on our response, in case it has not come to your attention.

Addressing your concerns, we clarified that our contribution lies in the supervision-construction stage rather than the post-training pipeline, added three matched controls under an identical SFT/RL configuration showing that the gains cannot be explained by SFT or GRPO alone, and added comparisons with RecLLM-R1, ReRec, and Rank-GRPO, over which R2Rec remains stronger on all three datasets. We have since added two 2026 baselines (TWiCE-Rec, AAAI'26; STAR, Findings of ACL'26).

We would greatly appreciate any feedback on whether these address your concerns, and are glad to run further experiments within the remaining time.

Thank you for your time and effort.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of the Author Response - Part 1/2

We thank the Area Chair and the reviewers for their time during the discussion period. Below we first summarize the points on which the reviewers agreed, and then the concerns raised and the evidence we provided, with a pointer at the end of each item to where the detailed response can be found.

Points of agreement among the reviewers. All three reviewers found the problem well motivated and the paper clear and easy to follow. Reviewers c5Ly and 5aic further noted that recommendation data typically lack explicit reasoning supervision, which makes the construction of reasoning examples from existing interactions a direction worth exploring, and that short interaction chains offer an intuitive way to organize graph structure into inputs for LLMs; Reviewer c5Ly described the construction of interaction-based reasoning traces as "an interesting way to introduce recommendation-specific supervision". Both also acknowledged the coverage of the experiments, spanning multiple datasets, ablation studies, transfer evaluation, an enlarged candidate set, and further robustness analyses in the appendix.

1. Temporal validity (Reviewer 5aic). We audited the graph construction and chain sampling protocol, confirmed a 14.3% temporal-leakage rate, rebuilt a strictly time-controlled dataset retaining only interactions observable before each prediction time, and retrained. Results are essentially unchanged (HR@1 of 0.406 / 0.587 / 0.470 versus 0.411 / 0.591 / 0.467 originally), indicating that the reported gains do not depend on cross-user information from after the prediction time. (Details in our first response to Reviewer 5aic, W1.)

2. Training scale and comparison with classical recommenders (Reviewer 5aic). We first scaled training to 10k examples, where R2Rec reaches an HR@1 of 0.449 / 0.625 / 0.507, above SASRec at the same scale (0.420 / 0.596 / 0.473). Following the further question of whether SASRec would catch up with more data, we extended both methods to 20k, 30k, and 50k examples on Amazon Electronics. The two grow in step, R2Rec leads at every scale, and the HR@1 gap remains between 0.03 and 0.06, reaching 0.643 versus 0.596 at 50k. (Details in our first response to Reviewer 5aic, W2 and Q3, and in our second response, Q1.)

3. Novelty: how R2Rec differs from existing work (Reviewers c5Ly, WthK, 5aic). We positioned our contribution in the construction of reasoning supervision rather than in the SFT + RL pipeline, and set out the differences against representative work published in 2026.

A different role for collaborative information. Existing work uses collaborative information at three levels: as input representations (ThinkRec, WWW'26; ARROW, WSDM'26; NLCF, AAAI'26, which encode CF embeddings or graph structure into the input space of the LLM), as context retrieved at inference (STAR, Findings of ACL'26, which summarizes collaborative neighbors for a teacher to query, with the student imitating the teacher's agent behavior), and as a signal at the reward or outcome level (ReRec, ACL'26, with dual-graph reward shaping; TWiCE-Rec, AAAI'26, which weights the reward by CF confidence). None of these constrains the individual steps of reasoning. R2Rec instead organizes collaborative structure into hop-by-hop traceable behavioral evidence and uses it directly to construct reasoning supervision: every hop of a closed interaction chain is a concrete observed interaction, which specifies both the evidence each intermediate reasoning step may rely on and the endpoint the reasoning must reach.

A different use of answer guidance. TWiCE-Rec (AAAI'26), C-APO (ICLR'26), SelfDR (CIKM'26), and R2END (Findings of ACL'26) all let the teacher model see both the user history and the ground-truth item and generate a free-form explanation, which the first two explicitly call post-hoc rationales. Such explanations ensure that the conclusion agrees with the answer, but their intermediate steps rely mainly on item semantics and are not constrained to be supported by behavioral evidence. We argue that outcome consistency and evidence grounding are two independent conditions, and tested this: a new variant that keeps answer guidance but removes the step-level evidence constraint does not improve over fully free generation and is in fact slightly worse, while full IoT exceeds it by 0.024 / 0.020 / 0.029 in HR@1. Measurements in the appendix support the same distinction: outcome guidance raises the outcome-consistency rate from 43%-57% to 100%, while the step-level evidence-grounded rate of IoT traces reaches 97.75%-99.00%.

AuthorsOpenReview ↗

Summary of the Author Response - Part 2/2

(Section 3, continued.)

The role of progressive masking. We do not present it as a standalone prompting innovation, but as the mechanism that establishes a correspondence between evidence and reasoning steps in a task without step annotations, and that carries this correspondence into the final supervision target. The variant described above measures this mechanism directly.

(Details in our second response to Reviewer WthK, Sections 1-3, which also contains the item-by-item comparison with 2026 work; the basic positioning of Interaction-of-Thought is given in our first response to Reviewer WthK, W1 and Q1, and to Reviewer c5Ly, W1.)

4. Necessity (Reviewers 5aic, WthK). We added three matched controls: answer-only supervision, shuffled interaction chains, and unstructured interactions preserving the same content without the chain organization. All three fall below the full method, indicating that the gains come from organizing collaborative information into traceable evidence and converting it into supervision, rather than merely from exposing the model to that information. On whether pure RL can acquire recommendation reasoning on its own, we scaled pure RL to 5,000 examples, where it saturates between 3,000 and 5,000 (HR@1 of 0.237 / 0.362 / 0.219). Reviewer 5aic recently cited R²ec (NeurIPS'25) as evidence that reasoning can be acquired through RL without reasoning annotations, and we have explained how the two settings differ: R²ec is not a text-only LLM, as it introduces on top of a decoder-only backbone a recommendation head covering the entire item set, the ranking is produced by discriminative scoring rather than generated text, reasoning influences the item scores through a shared hidden state, and its reward includes a continuous similarity term over the whole catalog, whereas our setting produces a textual ranking over 20 candidates. (The matched controls are in our first response to Reviewer 5aic, Q2, and to Reviewer c5Ly, W2; the pure-RL scaling is in our second response to Reviewer 5aic, Q2; the comparison with R²ec is in our third response to Reviewer 5aic.)

5. Updated baselines (Reviewers c5Ly, WthK). We added seven baselines in total: RecLLM-R1, ReRec, Rank-GRPO, VRec, and IRMD in the first round, followed by TWiCE-Rec (AAAI'26) and STAR (Findings of ACL'26). R2Rec outperforms all of them on all three datasets and all three metrics. (Details in our first response to Reviewer c5Ly, W3 and Q2, to Reviewer WthK, W2, and in the section on additional 2026 baselines in our second response to Reviewer WthK.)

6. Reasoning-trace generator (Reviewers 5aic, WthK). The trace generator (GPT-4o) is separate from the downstream recommendation model (DeepSeek-R1-Distill-Llama-8B), and a stronger generator raises downstream performance further, to 0.438 / 0.611 / 0.485. (Details in our first response to Reviewer 5aic, Q1, and to Reviewer WthK, Q2.)

We hope our responses have addressed the concerns of the Area Chair and the reviewers.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.