Pith. sign in

REVIEW 3 major objections 5 minor 2 cited by

Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read Interaction-of-thought reasoning lifts LLM recommendation hit rate by 10.48 percent on average.

desk verdict Plausible reasoning-enhanced recommender, but the evaluation may leak the target item into the prompt; the central numbers need verification before they can be believed. read the letter →

arxiv 2506.05069 v2 pith:KR43MMWC submitted 2025-06-05 cs.IR cs.AI

classification cs.IRcs.AI
keywords interaction-of-thoughtLLMrecommendationreinforcementlearningsupervisedfine-tuninginteractionchainsinterpretabilityHitRatioGRPO
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper proposes R2Rec, a reasoning-enhanced recommendation framework that gives large language models a structured way to reason over user-item interaction data. It samples short interaction chains from the user-item graph, converts them into step-by-step interaction-of-thoughts through a progressive, masked prompting strategy, and then internalizes this reasoning with supervised fine-tuning followed by reinforcement learning. The central claim is that this approach makes LLMs better recommenders: R2Rec outperforms classical and LLM-based baselines on MovieLens, Amazon Book, and Amazon Electronics, with an average 10.48% improvement in HitRatio@1 over the second-best baseline and a 131.81% gain over the original LLM. The paper also claims the generated reasoning chains make recommendations interpretable by revealing the decision process.

What carries the argument

The central object is the interaction-of-thought, a structured reasoning trace built from a closed interaction chain u0 → i0 → u1 → i1 → u0, where u0 is the target user. The progressive, masked prompting strategy injects the correct answer into the prompt but instructs the LLM to reason as if the answer were unknown, generating stepwise inferences that draw only on information before each hop. This converts implicit behavioral patterns into explicit, inspectable reasoning. The two-stage training pipeline carries the optimization: supervised fine-tuning with LoRA imitates high-quality traces, and GRPO reinforcement learning refines the policy using a reward that combines a reasoning-step count reward with a ranking-correctness reward weighted 1:2.

What would settle it

Re-run the MovieLens and Amazon experiments after explicitly removing the ground-truth item from the five liked items and from every sampled interaction chain, then recompute HitRatio@1; if the advantage over baselines disappears or reverses, the reported gains stem from context leakage rather than from interaction-of-thought reasoning.

Watch

Extended reading notes

Core claim

R2Rec claims that a recommendation LLM trained to reason over interaction chains substantially outperforms both classical recommenders and existing LLM-based recommenders. Interaction chains of the form u0 → i0 → u1 → i1 → u0 are sampled from the user-item graph with a closure constraint, then transformed by a progressive, masked prompting strategy into interaction-of-thoughts: natural-language reasoning traces in which each step relies only on prior steps and deliberately ignores the injected correct answer. A two-stage pipeline, LoRA-based supervised fine-tuning followed by GRPO reinforcement learning with reasoning-step and ranking-correctness rewards, teaches the LLM to generate such reasoning autonomously. In experiments, R2Rec reaches HitRatio@1 scores of 0.404 on MovieLens, 0.586 on Amazon Book, and 0.460 on Amazon Electronics, beating all baselines, and its ablations attribute most of the gain to the interaction chain, the reasoning chain, and the SFT warm start.

Load-bearing premise

The evaluation assumes that the ground-truth item is never among the five liked items shown in the prompt and never appears inside the sampled interaction chains; if it does, the model can identify the target by direct match and the reported HitRatio gains are inflated.

Editorial extensions

If this is right

  • If the central claim holds, LLM recommendation systems can be improved by teaching the model to reason over local graph structure rather than only over item metadata and text.
  • Reinforcement learning can partially substitute for scarce explicit reasoning annotations in domains where user feedback is implicit, extending the RL-reasoning recipe beyond math and code.
  • The interaction-of-thought format makes recommendation decisions traceable: a user could inspect why a particular item was ranked first.
  • The cross-domain experiments suggest that reasoning over interaction chains transfers across domains, so a model trained on one dataset can improve recommendations in another without retraining on that domain.
  • The ablations indicate that both the interaction chain and the reasoning chain contribute distinct signals, which points to further gains from richer chain construction and more detailed reasoning supervision.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A testable question the paper leaves open is whether the reported gains come from genuine reasoning or partly from the ground-truth item appearing in the prompt's liked-item list or interaction chains; excluding that item from all prompt context and recomputing HitRatio@1 would settle it.
  • The same interaction-of-thought recipe could be carried over to other implicit-feedback tasks such as search ranking or content moderation, where behavioral chains are abundant but natural-language reasoning labels are not.
  • The ranking-correctness reward only rewards the position of a single most-recent positive item, so the method as evaluated rewards one correct slot rather than the quality of the whole ranked list; a list-level reward might change which behaviors the RL stage reinforces.
  • The 131.81% gain over the original LLM is best read together with the comparison against tuned baselines, since the base model may be a weak zero-shot recommender; the headline relative gain should not be compared to gains measured against a stronger starting point.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes R2Rec, a recommendation framework that augments an LLM (DeepSeek-R1-Distill-Llama-8B) with structured interaction-of-thought reasoning. Interaction chains are sampled from the user–item graph under a closure constraint, converted into stepwise reasoning traces via a progressive, masked prompting strategy, and internalized through a two-stage pipeline of supervised fine-tuning and GRPO-based reinforcement learning. Experiments on MovieLens-1M, Amazon Book, and Amazon Electronics report consistent HitRatio@1 improvements over classical and LLM-based baselines, with an average 10.48% gain over the second-best baseline, plus ablations, transferability experiments, and a case study intended to demonstrate interpretability.

Significance. If the reported gains are reproducible under a leakage-free evaluation protocol, R2Rec is a useful contribution: it shows how local graph structure can be converted into token-level reasoning traces for LLM recommenders, and it provides an SFT+RL recipe that is more scalable than pure distillation. Strengths include the closure-constrained chain construction, the clear two-stage training design, ablations that isolate the contributions of interaction chains, reasoning chains, SFT, and RL, a cross-domain transfer study, and a released code link. The interpretability claim is plausible but currently undercut by the answer-conditioned trace generation, and the quantitative claim is vulnerable to a prompt-leakage concern in the evaluation protocol.

major comments (3)
  1. [§4.1.1–§4.1.2, §3.2.1] The evaluation does not state that the ground-truth item (the most recent positively rated item) is excluded from the two prompt components built from the user's own interaction history. In §3.2.1, the interaction-chain items i0 and i1 are selected from Ishared ⊆ Iu0, so either item can be the ground-truth item; the five 'recently liked items' of §4.1.2 can likewise include it. If this overlap occurs, the prompt reveals the target item with its rating before the model ranks the candidate set, while none of the baselines receive this signal, which would inflate the reported HitRatio@1 values and the average 10.48% gain. Please (a) state explicitly whether such overlap is prevented, (b) report the number and fraction of evaluation users for whom the ground-truth item appears in the liked-items list or in the sampled interaction chains under the current construction, and (c) rerun the comparison with all overlap excluded.
  2. [Table 1, §4.2.2] The 'second-best baseline' comparison is misreported for Amazon Book: SASRec's HitRatio@1 of 0.531 is higher than LLaRA's 0.528, yet the text reports a 10.99% improvement over the second-best baseline, which is computed against LLaRA. Using the actual second-best value gives 0.586/0.531 − 1 ≈ 10.36%, and the average improvement over the three datasets should be recomputed accordingly. In addition, the table reports means of three runs but no standard deviations, confidence intervals, or significance tests; given the small margin on MovieLens (0.404 vs. 0.380), the comparison is currently not statistically supported.
  3. [§3.2.2, §3.3.1, §4.5] The paper claims that explicit reasoning chains 'reveal the decision process,' but the SFT traces are generated by injecting the correct answer A into the prompt and instructing the model to ignore it (Eq. (2), Appendix A.1). Such traces are conditioned on the gold label and cannot by themselves demonstrate that the model's ranking decisions are caused by the stated reasoning; the case study in §4.5 is a single anecdote. The performance claim does not depend on this point, but the interpretability claim should be softened or supported by an analysis that separates reasoning produced without answer injection from reasoning that is merely post-hoc rationalization.
minor comments (5)
  1. [§4.2 heading] The heading contains a typo: 'Comparsion' should be 'Comparison'.
  2. [§3.2.1, Eq. (1)] The set definition Uik = u|(u,ik) ∈ E is missing set-builder notation; it should be Uik = {u | (u,ik) ∈ E} for clarity.
  3. [§4.1.1] The dataset description says users with at least six ratings above 3 are retained, but the prompt contains five liked items; please clarify whether the liked items are drawn from interactions strictly before the ground-truth item and why the threshold is six rather than five.
  4. [Table 1 and §4.2.2] Since results are claimed to be averages of three independent runs, please report the per-run values or at least the standard deviation and the random seeds used; this is necessary for the reader to assess the stability of the reported differences.
  5. [Table 2, §4.3.1] The upper block of Table 2 uses the row label 'PromptFull,' while the text refers to the full prompt; renaming the row to 'Full' would improve consistency with the ablation narrative.

Circularity Check

2 steps flagged · score 6.0 of 10

Prompt construction puts the held-out target inside the input (most recent positive = one of the 'five recently liked items'), so the reported HitRatio@1 gain may be a direct-match artifact rather than reasoning.

  1. self definitional [Sec. 4.1.1 (Datasets) and Sec. 4.1.2 (Implementation Details)]
    "The most recent positively rated item is used as the ground-truth. ... For each user, the prompt includes demographic features (if available), five recently liked items, five sampled interaction chains, and a 20-item candidate set."

    Under the paper's own definitions the ground-truth is a positively rated (i.e., liked) item, and the prompt is specified to contain five recently liked items. The paper never states that the ground-truth is excluded from those five items, so by construction the target can appear in the input with its rating. A model can then rank it first by direct text match, which would inflate every HitRatio@1 number in Table 1 and the average 10.48% improvement over baselines that did not receive this signal. The claimed causal role of interaction-of-thought reasoning is therefore not established unless this overlap is ruled out.

  2. other [Sec. 3.2.1 (Interaction Chain) and Sec. 4.1.2 (Implementation Details)]
    "Given a target user u0, we first collect their interacted item set Iu0 = i1, i2, ..., in. ... we determine the shared item set Ishared = Iu0 ∩ Iu1, from which two items i0, i1 are selected to form the interaction chain: u0 → i0 → u1 → i1 → u0."

    The ground-truth is the target user's most recent positively rated item, so it is in I_u0. The chain items are selected from I_shared = I_u0 ∩ I_u1, and the paper does not state that the ground-truth is excluded from i0 or i1 or from the five sampled chains. If it is selected, the prompt contains the target item and the target user's rating of it, again permitting direct-match ranking. This makes the reported evaluation not a clean test of reasoning, because the same construction that defines the chain also defines the prediction target.

full rationale

Self-citation is not load-bearing here: references [9], [10], [19], etc. are related-work or survey citations, and no uniqueness theorem or ansatz is imported from the authors' prior work. The SFT traces are generated with the gold answer injected (Sec. 3.2.2, Appendix A.1), which weakens the interpretability claim but does not by itself make the ranking evaluation circular, since the final ranking is scored against held-out targets. The core problem is the prompt/test construction: the ground-truth is defined as the most recent positively rated item while the prompt is defined to include five recently liked items and chains sampled from I_u0, with no stated exclusion of the ground-truth. If the target appears in the prompt, HitRatio@1 can be solved by direct match, so the average 10.48% gain is unverified. This is a partial circularity in the evaluation rather than in the training objective, and it should be fixed by explicitly excluding the ground-truth from the liked items and chains and reporting overlap statistics.

Assumptions & free parameters 8 free parameters · 5 assumptions · 2 invented entities

The method relies on several domain assumptions about rating semantics, the faithfulness of LLM-generated rationales, and the adequacy of the hand-designed reward. The free parameters are mostly prompt and reward design choices. No new physical entities are introduced; the new constructs are prompt representations.

free parameters (8)
  • reward_weight_step = 1
    Weight for the reasoning step reward in Eq. 7; hand-chosen, no sensitivity analysis.
  • reward_weight_correctness = 2
    Weight for the ranking correctness reward in Eq. 7; hand-chosen, no sensitivity analysis.
  • step_target = 4
    reward_step = min(1, NumSteps/4) encourages at least 4 reasoning steps; chosen by hand.
  • correctness_thresholds = 1.0, 0.7, 0.5, 0.2, 0.0
    Graded reward for the rank of the ground-truth item (Eq. 6); chosen by hand.
  • num_interaction_chains = 5
    Five sampled interaction chains are included in each prompt (Section 4.1.2).
  • num_liked_items = 5
    Five recently liked items are included in each prompt (Section 4.1.2).
  • candidate_size = 20
    Each query ranks 20 candidate items, one positive and 19 randomly sampled negatives (Section 4.1.1).
  • training_samples = 500
    SFT and RL each use 500 training samples (Section 4.1.2).
assumptions (5)
  • domain assumption Explicit 1-5 ratings reflect user preference and can be used to construct meaningful interaction chains.
    The method builds reasoning on rating values in the interaction chain (Section 3.2.1, Appendix A.1).
  • domain assumption LLM-generated reasoning traces conditioned on the correct answer are faithful enough to serve as SFT supervision.
    The masked prompting strategy injects the answer and asks the model to ignore it; the paper assumes the resulting traces are valid (Section 3.3.1, Eq. 3).
  • domain assumption The progressive, masked prompting strategy prevents the model from using information after each hop, so the reasoning is genuinely stepwise.
    The prompt instructs the model to rely only on previous steps (Appendix A.1).
  • domain assumption The reward function combining step count and ranking of the ground-truth item is a sufficient proxy for recommendation quality.
    The RL stage uses only this reward; no reasoning-quality reward is used (Appendix A.2).
  • standard math GRPO is a valid policy optimization method and the group-relative advantage estimator is applied correctly.
    The paper invokes the standard GRPO objective (Eq. 4) from prior literature.
invented entities (2)
  • Interaction-of-thought (IoT)
    purpose: Structured stepwise reasoning traces over user-item interaction chains, used as SFT data and as model output for explainability.
    The paper introduces this representation and demonstrates its value only through its own experiments; there is no external benchmark or falsifiable prediction tied to IoT outside the paper.
  • Interaction chain with closure constraint
    purpose: A graph-sampled sequence u0->i0->u1->i1->u0 that grounds reasoning in behavioral data.
    This specific sampling scheme is new to the paper; its effectiveness is shown only in the paper's ablations, not against external criteria.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation." pith.science (2026). https://pith.science/paper/KR43MMWC

@misc{pith2026250605069,
  author       = {Pith},
  title        = {Pith review of: Reason-to-Recommend: Using Interaction-of-Thought Reasoning to Enhance LLM Recommendation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/KR43MMWC}},
  note         = {Machine review of arXiv:2506.05069}
}
abstract

Driven by advances in Large Language Models (LLMs), integrating them into recommendation tasks has gained interest due to their strong semantic understanding and prompt flexibility. Prior work encoded user-item interactions or metadata into prompts for recommendations. In parallel, LLM reasoning, boosted by test-time scaling and reinforcement learning, has excelled in fields like mathematics and code, where reasoning traces and correctness signals are clear, enabling high performance and interpretability. However, directly applying these reasoning methods to recommendation is ineffective because user feedback is implicit and lacks reasoning supervision. To address this, we propose $\textbf{R2Rec}$, a reasoning-enhanced recommendation framework that samples interaction chains from the user-item graph and converts them into structured interaction-of-thoughts via a progressive masked prompting strategy, with each thought representing stepwise reasoning grounded in interaction context. This allows LLMs to simulate step-by-step decision-making based on implicit patterns. We design a two-stage training pipeline: supervised fine-tuning teaches basic reasoning from high-quality traces, and reinforcement learning refines reasoning via reward signals, alleviating sparse explicit supervision. Experiments on three real-world datasets show R2Rec outperforms classical and LLM-based baselines with an average $\textbf{10.48%}$ improvement in HitRatio@1 and $\textbf{131.81%}$ gain over the original LLM. Furthermore, the explicit reasoning chains enhance interpretability by revealing the decision process. Our code is available at: https://anonymous.4open.science/r/R2Rec-7C5D.

Figures

Figures reproduced from arXiv: 2506.05069 by the authors.

Figure 1
Figure 1. Overview of Reason-to-Recommend(R2Rec) Framework. [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. The construction of the interaction chain and the progressive and masked prompt strategies [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. The two-stage training pipeline for improving recommendation performance. [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Case Study To illustrate the interpretability of our approach, we present a case study involving a female user under 18 years old, as shown in [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: Reward comparison between RL training with and without SFT initialization. [PITH_FULL_IMAGE:figures/full_fig_p015_5.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. From Understanding to Action: Feedback-Grounded Policy Discovery for Generative Recommendation

    cs.IR 2026-07 conditional novelty 6.0 of 10

    A feedback-grounded framework discovers recommendation policies by their measured advantage over intent-only baselines and distills them into two latent tokens of a lightweight Semantic-ID recommender.

  2. Think2Go: Generative Next POI Recommendation with LLM Reasoning

    cs.IR 2026-07 conditional novelty 4.0 of 10

    Think2Go couples SFT and RL-based reasoning in one LLM, with KDE- and reward-gap-based advantage calibration, and reports state-of-the-art Acc@1 on NYC, Tokyo, and California check-in data.

Reference graph

Works this paper leans on

65 extracted references · 39 canonical work pages · cited by 2 Pith papers

  1. [1]

    Tallrec: An effective and efficient tuning framework to align large language model with recommendation

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. Tallrec: An effective and efficient tuning framework to align large language model with recommendation. InProceedings of the 17th ACM Conference on Recommender Systems, RecSys ’23, page 1007–1014. ACM, September 2023

  2. [2]

    Uncovering chatgpt’s capabilities in recommender systems

    Sunhao Dai, Ninglu Shao, Haiyuan Zhao, Weijie Yu, Zihua Si, Chen Xu, Zhongxiang Sun, Xiao Zhang, and Jun Xu. Uncovering chatgpt’s capabilities in recommender systems. InProceedings of the 17th ACM Conference on Recommender Systems, RecSys ’23, page 1126–1132. ACM, September 2023

  3. [3]

    Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5), 2023

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. Recommendation as language processing (rlp): A unified pretrain, personalized prompt & predict paradigm (p5), 2023

  4. [4]

    Llara: Large language-recommendation assistant, 2024

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. Llara: Large language-recommendation assistant, 2024

  5. [5]

    Is chatgpt a good recommender? a preliminary study, 2023

    Junling Liu, Chao Liu, Peilin Zhou, Renjie Lv, Kang Zhou, and Yan Zhang. Is chatgpt a good recommender? a preliminary study, 2023

  6. [6]

    Improving llm-powered recommendations with personalized information, 2025

    Jiahao Liu, Xueshuo Yan, Dongsheng Li, Guangping Zhang, Hansu Gu, Peng Zhang, Tun Lu, Li Shang, and Ning Gu. Improving llm-powered recommendations with personalized information, 2025

  7. [7]

    Let me do it for you: Towards llm empowered recommendation via tool learning

    Yuyue Zhao, Jiancan Wu, Xiang Wang, Wei Tang, Dingxian Wang, and Maarten de Rijke. Let me do it for you: Towards llm empowered recommendation via tool learning. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2024, page 1796–1806. ACM, July 2024

  8. [8]

    Collaborative retrieval for large language model-based conversational recommender systems, 2025

    Yaochen Zhu, Chao Wan, Harald Steck, Dawen Liang, Yesu Feng, Nathan Kallus, and Jundong Li. Collaborative retrieval for large language model-based conversational recommender systems, 2025

Show all 65 references
  1. [9]

    Agentrecbench: Benchmarking llm agent-based person- alized recommender systems.arXiv preprint arXiv:2505.19623, 2025

    Yu Shang, Peijie Liu, Yuwei Yan, Zijing Wu, Leheng Sheng, Yuanqing Yu, Chumeng Jiang, An Zhang, Fengli Xu, Yu Wang, et al. Agentrecbench: Benchmarking llm agent-based person- alized recommender systems.arXiv preprint arXiv:2505.19623, 2025

  2. [10]

    Agentsociety challenge: Designing llm agents for user modeling and recommendation on web platforms

    Yuwei Yan, Yu Shang, Qingbin Zeng, Yu Li, Keyu Zhao, Zhiheng Zheng, Xuefei Ning, Tianji Wu, Shengen Yan, Yu Wang, et al. Agentsociety challenge: Designing llm agents for user modeling and recommendation on web platforms. InCompanion Proceedings of the ACM on Web Conference 202...

  3. [11]

    Think before recommend: Unleashing the latent reasoning power for sequential recommenda- tion, 2025

    Jiakai Tang, Sunhao Dai, Teng Shi, Jun Xu, Xu Chen, Wen Chen, Wu Jian, and Yuning Jiang. Think before recommend: Unleashing the latent reasoning power for sequential recommenda- tion, 2025

  4. [12]

    Got4rec: Graph of thoughts for sequential recommendation, 2025

    Zewen Long, Liang Wang, Shu Wu, Qiang Liu, and Liang Wang. Got4rec: Graph of thoughts for sequential recommendation, 2025

  5. [13]

    Cot4rec: Revealing user preferences through chain of thought for recommender systems.Proceedings of the AAAI Conference on Artificial Intelligence, 39(12):13142–13151, Apr

    Weiqi Yue, Yuyu Yin, Xin Zhang, Binbin Shi, Tingting Liang, and Jian Wan. Cot4rec: Revealing user preferences through chain of thought for recommender systems.Proceedings of the AAAI Conference on Artificial Intelligence, 39(12):13142–13151, Apr. 2025

  6. [14]

    Chain-of-thought prompting empowered generative user modeling for personalized recommendation.Neural Computing and Applications, 36(34):21723–21742, 2024

    Fan Yang, Yong Yue, Gangmin Li, Terry R Payne, and Ka Lok Man. Chain-of-thought prompting empowered generative user modeling for personalized recommendation.Neural Computing and Applications, 36(34):21723–21742, 2024

  7. [15]

    s1: Simple test-time scaling, 2025

    Niklas Muennighoff, Zitong Yang, Weijia Shi, Xiang Lisa Li, Li Fei-Fei, Hannaneh Hajishirzi, Luke Zettlemoyer, Percy Liang, Emmanuel Candès, and Tatsunori Hashimoto. s1: Simple test-time scaling, 2025. 10

  8. [16]

    Light-r1: Curriculum sft, dpo and rl for long cot from scratch and beyond, 2025

    Liang Wen, Yunke Cai, Fenrui Xiao, Xin He, Qi An, Zhenyu Duan, Yimin Du, Junchen Liu, Lifu Tang, Xiaowei Lv, Haosheng Zou, Yongchao Deng, Shousheng Jia, and Xiangzheng Zhang. Light-r1: Curriculum sft, dpo and rl for long cot from scratch and beyond, 2025

  9. [17]

    Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling, 2025

    Runze Liu, Junqi Gao, Jian Zhao, Kaiyan Zhang, Xiu Li, Biqing Qi, Wanli Ouyang, and Bowen Zhou. Can 1b llm surpass 405b llm? rethinking compute-optimal test-time scaling, 2025

  10. [18]

    Limo: Less is more for reasoning, 2025

    Yixin Ye, Zhen Huang, Yang Xiao, Ethan Chern, Shijie Xia, and Pengfei Liu. Limo: Less is more for reasoning, 2025

  11. [19]

    Towards large reasoning models: A survey of reinforced reasoning with large language models.arXiv preprint arXiv:2501.09686, 2025

    Fengli Xu, Qianyue Hao, Zefang Zong, Jingwei Wang, Yunke Zhang, Jingyi Wang, Xiaochong Lan, Jiahui Gong, Tianjian Ouyang, Fanjin Meng, et al. Towards large reasoning models: A survey of reinforced reasoning with large language models.arXiv preprint arXiv:2501.09686, 2025

  12. [20]

    Prompt learning for news recommendation

    Zizhuo Zhang and Bang Wang. Prompt learning for news recommendation. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’23, page 227–237. ACM, July 2023

  13. [21]

    Towards understanding and mitigating unintended biases in language model-driven conversational rec- ommendation.Information Processing & Management, 60(1):103139, 2023

    Tianshu Shen, Jiaru Li, Mohamed Reda Bouadjenek, Zheda Mai, and Scott Sanner. Towards understanding and mitigating unintended biases in language model-driven conversational rec- ommendation.Information Processing & Management, 60(1):103139, 2023

  14. [22]

    Improving conversational recommenda- tion systems’ quality with context-aware item meta information, 2021

    Bowen Yang, Cong Han, Yu Li, Lei Zuo, and Zhou Yu. Improving conversational recommenda- tion systems’ quality with context-aware item meta information, 2021

  15. [23]

    Recommendation as instruction following: A large language model empowered recommendation approach, 2023

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji-Rong Wen. Recommendation as instruction following: A large language model empowered recommendation approach, 2023

  16. [24]

    Personalized prompt learning for explainable recommen- dation, 2023

    Lei Li, Yongfeng Zhang, and Li Chen. Personalized prompt learning for explainable recommen- dation, 2023

  17. [25]

    Bridging items and language: A transition paradigm for large language model-based recommendation, 2024

    Xinyu Lin, Wenjie Wang, Yongqi Li, Fuli Feng, See-Kiong Ng, and Tat-Seng Chua. Bridging items and language: A transition paradigm for large language model-based recommendation, 2024

  18. [26]

    Dapo: An open-source llm reinforcement learning system at scale, 2025

    Qiying Yu, Zheng Zhang, Ruofei Zhu, Yufeng Yuan, Xiaochen Zuo, Yu Yue, Tiantian Fan, Gaohong Liu, Lingjun Liu, Xin Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Guangming Sheng, Yuxuan Tong, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Jinhua Zhu, Jiaze Chen, Jiangjie Chen, Chengyi Wa...

  19. [27]

    Vapo: Efficient and reliable reinforcement learning for advanced reasoning tasks, 2025

    Yu Yue, Yufeng Yuan, Qiying Yu, Xiaochen Zuo, Ruofei Zhu, Wenyuan Xu, Jiaze Chen, Chengyi Wang, TianTian Fan, Zhengyin Du, Xiangpeng Wei, Xiangyu Yu, Gaohong Liu, Juncai Liu, Lingjun Liu, Haibin Lin, Zhiqi Lin, Bole Ma, Chi Zhang, Mofan Zhang, Wang Zhang, Hang Zhu, Ru Zhang, X...

  20. [28]

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y . K. Li, Y . Wu, and Daya Guo. Deepseekmath: Pushing the limits of mathematical reasoning in open language models, 2024

  21. [29]

    Reinforce++: An efficient rlhf algorithm with robustness to both prompt and reward models, 2025

    Jian Hu, Jason Klein Liu, and Wei Shen. Reinforce++: An efficient rlhf algorithm with robustness to both prompt and reward models, 2025

  22. [30]

    Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

    Aaron Jaech, Adam Kalai, Adam Lerer, Adam Richardson, Ahmed El-Kishky, Aiden Low, Alec Helyar, Aleksander Madry, Alex Beutel, Alex Carney, et al. Openai o1 system card.arXiv preprint arXiv:2412.16720, 2024

  23. [31]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 11

  24. [32]

    Reft: Reasoning with reinforced fine-tuning, 2024

    Trung Quoc Luong, Xinbo Zhang, Zhanming Jie, Peng Sun, Xiaoran Jin, and Hang Li. Reft: Reasoning with reinforced fine-tuning, 2024

  25. [33]

    Maxwell Harper and Joseph A

    F. Maxwell Harper and Joseph A. Konstan. The movielens datasets: History and context.ACM Trans. Interact. Intell. Syst., 5(4), December 2015

  26. [34]

    Bridging language and items for retrieval and recommendation.arXiv preprint arXiv:2403.03952, 2024

    Yupeng Hou, Jiacheng Li, Zhankui He, An Yan, Xiusi Chen, and Julian McAuley. Bridging language and items for retrieval and recommendation.arXiv preprint arXiv:2403.03952, 2024

  27. [35]

    Neural graph collabo- rative filtering

    Xiang Wang, Xiangnan He, Meng Wang, Fuli Feng, and Tat-Seng Chua. Neural graph collabo- rative filtering. InProceedings of the 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR ’19, page 165–174. ACM, July 2019

  28. [36]

    Hop-rec: high- order proximity for implicit recommendation

    Jheng-Hong Yang, Chih-Ming Chen, Chuan-Ju Wang, and Ming-Feng Tsai. Hop-rec: high- order proximity for implicit recommendation. InProceedings of the 12th ACM Conference on Recommender Systems, RecSys ’18, page 140–144, New York, NY , USA, 2018. Association for Computing Machinery

  29. [37]

    Hamilton, Rex Ying, and Jure Leskovec

    William L. Hamilton, Rex Ying, and Jure Leskovec. Inductive representation learning on large graphs, 2018

  30. [38]

    Lightgcn: Simplifying and powering graph convolution network for recommendation, 2020

    Xiangnan He, Kuan Deng, Xiang Wang, Yan Li, Yongdong Zhang, and Meng Wang. Lightgcn: Simplifying and powering graph convolution network for recommendation, 2020

  31. [39]

    Graph-based embedding smoothing for sequential recommendation.IEEE Transactions on Knowledge and Data Engineering, 35(1):496–508, 2023

    Tianyu Zhu, Leilei Sun, and Guoqing Chen. Graph-based embedding smoothing for sequential recommendation.IEEE Transactions on Knowledge and Data Engineering, 35(1):496–508, 2023

  32. [40]

    Sequential recommendation with graph neural networks, 2023

    Jianxin Chang, Chen Gao, Yu Zheng, Yiqun Hui, Yanan Niu, Yang Song, Depeng Jin, and Yong Li. Sequential recommendation with graph neural networks, 2023

  33. [41]

    Dynamic graph neural networks for sequential recommendation, 2021

    Mengqi Zhang, Shu Wu, Xueli Yu, Qiang Liu, and Liang Wang. Dynamic graph neural networks for sequential recommendation, 2021

  34. [42]

    Knowledge-enhanced graph neural networks for sequential recommendation.Information, 11(8), 2020

    Baocheng Wang and Wentao Cai. Knowledge-enhanced graph neural networks for sequential recommendation.Information, 11(8), 2020

  35. [43]

    Invisible walls in cities: Leveraging large language models to predict urban segregation experience with social media content.arXiv preprint arXiv:2503.04773, 2025

    Bingbing Fan, Lin Chen, Songwei Li, Jian Yuan, Fengli Xu, Pan Hui, and Yong Li. Invisible walls in cities: Leveraging large language models to predict urban segregation experience with social media content.arXiv preprint arXiv:2503.04773, 2025

  36. [44]

    Large language model-driven meta-structure discovery in heterogeneous information network

    Lin Chen, Fengli Xu, Nian Li, Zhenyu Han, Meng Wang, Yong Li, and Pan Hui. Large language model-driven meta-structure discovery in heterogeneous information network. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 307–318, 2024

  37. [45]

    Hlm-cite: Hybrid language model workflow for text-based scientific citation prediction.arXiv preprint arXiv:2410.09112, 2024

    Qianyue Hao, Jingyang Fan, Fengli Xu, Jian Yuan, and Yong Li. Hlm-cite: Hybrid language model workflow for text-based scientific citation prediction.arXiv preprint arXiv:2410.09112, 2024

  38. [46]

    KCTS: Knowledge- constrained tree search decoding with token-level hallucination detection

    Sehyun Choi, Tianqing Fang, Zhaowei Wang, and Yangqiu Song. KCTS: Knowledge- constrained tree search decoding with token-level hallucination detection. In Houda Bouamor, Juan Pino, and Kalika Bali, editors,Proceedings of the 2023 Conference on Empirical Methods in Natural Lang...

  39. [47]

    Language agent tree search unifies reasoning acting and planning in language models, 2024

    Andy Zhou, Kai Yan, Michal Shlapentokh-Rothman, Haohan Wang, and Yu-Xiong Wang. Language agent tree search unifies reasoning acting and planning in language models, 2024

  40. [48]

    Tenenbaum, and Chuang Gan

    Shun Zhang, Zhenfang Chen, Yikang Shen, Mingyu Ding, Joshua B. Tenenbaum, and Chuang Gan. Planning with large language models for code generation, 2023

  41. [49]

    Don’t throw away your value model! generating more preferable text with value-guided monte-carlo tree search decoding, 2024

    Jiacheng Liu, Andrew Cohen, Ramakanth Pasunuru, Yejin Choi, Hannaneh Hajishirzi, and Asli Celikyilmaz. Don’t throw away your value model! generating more preferable text with value-guided monte-carlo tree search decoding, 2024. 12

  42. [50]

    A simple model of inference scaling laws, 2024

    Noam Levi. A simple model of inference scaling laws, 2024

  43. [51]

    Self-consistency improves chain of thought reasoning in language models, 2023

    Xuezhi Wang, Jason Wei, Dale Schuurmans, Quoc Le, Ed Chi, Sharan Narang, Aakanksha Chowdhery, and Denny Zhou. Self-consistency improves chain of thought reasoning in language models, 2023

  44. [52]

    Le, Christopher Ré, and Azalia Mirhoseini

    Bradley Brown, Jordan Juravsky, Ryan Ehrlich, Ronald Clark, Quoc V . Le, Christopher Ré, and Azalia Mirhoseini. Large language monkeys: Scaling inference compute with repeated sampling, 2024

  45. [53]

    Kimi Team, A Du, B Gao, B Xing, C Jiang, C Chen, C Li, C Xiao, C Du, C Liao, et al. Kimi k1. 5: Scaling reinforcement learning with llms, 2025.URL https://arxiv. org/abs/2501.12599

  46. [54]

    M6-rec: Gener- ative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084, 2022

    Zeyu Cui, Jianxin Ma, Chang Zhou, Jingren Zhou, and Hongxia Yang. M6-rec: Gener- ative pretrained language models are open-ended recommender systems.arXiv preprint arXiv:2205.08084, 2022

  47. [55]

    Large language models are zero-shot rankers for recommender systems

    Yupeng Hou, Junjie Zhang, Zihan Lin, Hongyu Lu, Ruobing Xie, Julian McAuley, and Wayne Xin Zhao. Large language models are zero-shot rankers for recommender systems. In European Conference on Information Retrieval, pages 364–381. Springer, 2024

  48. [56]

    Chat-rec: Towards interactive and explainable llms-augmented recommender system, 2023

    Yunfan Gao, Tao Sheng, Youlin Xiang, Yun Xiong, Haofen Wang, and Jiawei Zhang. Chat-rec: Towards interactive and explainable llms-augmented recommender system, 2023

  49. [57]

    Learning vector-quantized item representation for transferable sequential recommenders, 2023

    Yupeng Hou, Zhankui He, Julian McAuley, and Wayne Xin Zhao. Learning vector-quantized item representation for transferable sequential recommenders, 2023

  50. [58]

    Towards universal sequence representation learning for recommender systems, 2022

    Yupeng Hou, Shanlei Mu, Wayne Xin Zhao, Yaliang Li, Bolin Ding, and Ji-Rong Wen. Towards universal sequence representation learning for recommender systems, 2022

  51. [59]

    Where to go next for recommender systems? id- vs

    Zheng Yuan, Fajie Yuan, Yu Song, Youhua Li, Junchen Fu, Fei Yang, Yunzhu Pan, and Yongxin Ni. Where to go next for recommender systems? id- vs. modality-based recommender models revisited, 2023

  52. [60]

    Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

    Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. Lora: Low-rank adaptation of large language models, 2021

  53. [61]

    Session-based recommendations with recurrent neural networks, 2016

    Balázs Hidasi, Alexandros Karatzoglou, Linas Baltrunas, and Domonkos Tikk. Session-based recommendations with recurrent neural networks, 2016

  54. [62]

    Personalized top-n sequential recommendation via convolutional sequence embedding, 2018

    Jiaxi Tang and Ke Wang. Personalized top-n sequential recommendation via convolutional sequence embedding, 2018

  55. [63]

    Self-attentive sequential recommendation, 2018

    Wang-Cheng Kang and Julian McAuley. Self-attentive sequential recommendation, 2018. 13 A Appendix A.1 The progressive, masked prompting strategy to generate Interaction-of-Thoughts The specific prompt to generate Interaction-of-thoughts is as follows(Using the Movielens datase...

  56. [64]

    The reasoning process must comply with the following requirements: 1

    Your task is to pretend to reason step by step about this interaction chain without knowing the given answer in advance, independently analyze the information of each jump in the interaction chain, and gradually obtain the final answer. The reasoning process must comply with t...

  57. [65]

    The answers provided are only to guide your reasoning process

    Your reasoning process should not contain any information about the final answer. The answers provided are only to guide your reasoning process. 4. The answer obtained based on your reasoning process should be consistent with the correct one. Next, I will provide you with a us...

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.