Pith. sign in

REVIEW 3 major objections 5 minor 60 references

Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging

T0 review · 3 major / 5 minor · reviewed 2026-08-15 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read REAM merges fast- and slow-thinking recommenders at attention-head granularity, cutting reasoning length by up to 24.3 percent while preserving rating accuracy.

desk verdict Useful empirical study of head-level model merging for reasoning compression, but the fast model's task vector is built from a different base checkpoint, so the mechanism claim does not hold as stated. read the letter →

arxiv 2608.10447 v1 pith:XLOQYE3V submitted 2026-08-11 cs.IR cs.AI

Linh Dieu Le , Tong Chen , Shazia Sadiq , Hongzhi Yin , Ming Jin , Junliang Yu This is my paper · ORCID
classification cs.IRcs.AI
keywords modelmergingreasoningcompressionLLM-basedrecommendersystemsattentionheadstaskvectorsretrievalcriticalitydecisionfaithfulnessratingprediction
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

This paper tries to establish that you can make a slow-thinking LLM recommender cheaper without retraining by merging its weights with those of a fast-thinking counterpart that predicts ratings directly. The proposed method, REAM, does this at the level of individual attention heads: each head gets its own merge coefficient, chosen so that heads important for retrieving user–item evidence and for grounding the final rating are protected, while less critical heads absorb more of the concise model's update. If the claim holds, reasoning traces shrink by up to 24.3 percent while rating accuracy stays intact or improves across Amazon Book, Yelp, and Amazon Music. This matters because reasoning traces dominate inference cost in LLM recommenders, and existing compression routes require either extra training or brittle decoding-time tricks.

What carries the argument

The central object is the head-level task-vector allocation. REAM anchors on the slow-thinking model and adds a fraction of the fast-thinking model's task vector to each attention-head parameter group. The fraction comes from a constrained optimisation whose per-head cost combines three signals: retrieval criticality (how often a head retrieves segment-relevant user–item evidence), decision faithfulness (how much of the rating step's attention goes to the compatibility-match segment), and Fisher-weighted update sensitivity (loss curvature along the fast update). The optimisation is solved by a water-filling KKT solution that allocates the update budget across heads, and the merge applies head-specific coefficients while excluding the final feed-forward layers.

What would settle it

Recompute retrieval criticality and decision faithfulness on the merged model and compare the top head sets with the pre-merge calibration; if the sets diverge substantially and ablating the newly critical heads changes ratings, the fixed-signal assumption fails.

Watch

Extended reading notes

Core claim

REAM's central claim is that reasoning compression in slow-thinking recommenders is a selective merging problem: import conciseness from a direct-prediction model, but allocate the task-vector update per attention head according to how much that head supports the reasoning behind the rating. On three datasets, the merged model reduces mean generated tokens by 17.9–24.3 percent compared with the slow model, and produces lower MAE and RMSE than the slow model on all three while beating eight training-free merging baselines on the accuracy–efficiency trade-off. The method needs no training and no decoding changes.

Load-bearing premise

The load-bearing premise is that the importance and sensitivity signals measured on the slow model before merging still describe which heads matter after merging; if the merge reshuffles head roles, the protected heads may no longer be the ones protecting accuracy.

Editorial extensions

If this is right

  • On all three datasets, mean generated tokens drop by 17.9% to 24.3% relative to the slow model, and MAE and RMSE are lower than the slow model's.
  • REAM beats all eight training-free baselines on the accuracy–efficiency trade-off, and head-level allocation is necessary: layer-wise activation baselines leave longer traces and worse MAE.
  • The retrieval-critical and decision-faithful head sets are causally important and nearly disjoint, so both signals pull their weight; ablating either set raises held-out NLL and divergence far more than random head ablation.
  • The pipeline is training-free and cheap: about 0.69 GPU-hours per dataset, roughly 4% of the combined cost of training the two source models.
  • The approach transfers across model scale and family, improving on the slow model on all metrics under the RecOne reasoning backbone, though on Qwen2.5-7B it shortens traces without fully matching the slow model's accuracy.

Reading between the lines

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

  • If the head roles are stable across inputs, the same coefficient schedule could be computed once per domain and applied to new users without recalibration; the paper's cross-domain Jaccard stability hints at this.
  • The two near-disjoint head sets suggest a natural extension: compress by steering updates through faithfulness-neutral heads while keeping retrieval heads frozen, which could be tested by comparing REAM against a frozen-retrieval variant.
  • The perturbation-cost formulation could be reused to predict per-input token budgets, choosing how much fast behaviour to inject based on trace difficulty rather than using one merged model for all inputs.
  • Because REAM only needs a verbose model and a concise model sharing a base, the same recipe might compress reasoning in other generation tasks with structured traces, such as tool-use or planning, not just rating prediction.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

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 REAM, a model merging method for compressing chain-of-thought reasoning in LLM-based recommender systems without additional training. REAM merges a slow-thinking RecZero model with a fast-thinking TALLRec model at the granularity of attention heads, assigning each head a merge coefficient based on retrieval criticality, decision faithfulness, and Fisher-weighted update sensitivity. A constrained optimization allocates the fast-thinking update while limiting perturbation risk, and late-layer FFN updates are excluded based on validation. Experiments on Amazon Book, Yelp, and Amazon Music report token reductions of 17.9% to 24.3% with maintained or improved MAE/RMSE relative to the slow model, and the paper includes ablations, causal validation of the head-level signals, hyperparameter analyses, and a generalisability study across model scales and backbones.

Significance. If the central mechanism is established, REAM would be a useful contribution: it is, to the authors' knowledge, the first merging framework for reasoning compression in recommender systems, it operates at a finer granularity than most merging baselines, and it ships with detailed derivations, a stability analysis, causal head ablations, and public code. The empirical pattern is consistent across three datasets, and the paper is honest about the training-free caveats. However, the significance is currently conditional on resolving a load-bearing issue with the definition of the fast task vector, which affects the interpretation of all the main results.

major comments (3)
  1. [§5.1, Appendix A.1, and §3.2] The fast task vector Δ_F is not defined relative to the declared shared base θ_B. The paper states that θ_F (adapted TALLRec) is fully fine-tuned from Qwen2.5-3B-non-instruct, while θ_B is Qwen2.5-3B-Instruct and is used as the shared base for both task vectors. Consequently Δ_F = θ_F − θ_B = (θ_F − θ_{B,non-instruct}) + (θ_{B,non-instruct} − θ_B). The first term is the genuine TALLRec adaptation vector; the second is a checkpoint difference unrelated to fast reasoning behavior. Since Eq. (8) applies this Δ_F to θ_S, the reported token reduction and accuracy effects could be driven by injecting the non-instruct/instruct checkpoint delta rather than by selectively transferring concise reasoning. This contradicts the statement in §3.2 that both task vectors are represented relative to the same reference, and it undermines the central claim that REAM's reasoning-aware coefficients enable a grounded transfer of conciseness. I ask the authors to re-run the experiments with a fast model fine-tuned from Qwen2.5-3B-Instruct, and to report a control experiment that merges only the pure checkpoint delta (θ_{B,non-instruct} − θ_B) using REAM's coefficients.
  2. [§4.3 and §5.4.2] The merge coefficients are computed entirely from signals measured on the pre-merge slow model θ_S, and the causal validation in §5.4.2 ablates heads of θ_S. The method assumes that the heads that are reasoning-critical before merging remain the heads that support reasoning after the fast update is applied, but the paper does not test post-merge head roles. This assumption is load-bearing because the entire design protects certain heads from the fast update; if the merge shifts which heads are reasoning-critical, the coefficients stop serving their stated purpose. Please provide evidence on the merged model itself, for example by re-computing retrieval criticality or decision faithfulness on θ★, or by ablating the previously protected heads in θ★ and comparing the effect with random ablations.
  3. [Table 1] The accuracy gains over the slow model θ_S are within overlapping 95% confidence intervals on all three datasets. For example, on Book, REAM's MAE is 0.6338±0.0181 versus θ_S's 0.6650±0.0180; on Yelp, 0.7564±0.0237 versus 0.7769±0.0250; on Music, 0.5348±0.0351 versus 0.5433±0.0360. The RMSE intervals overlap similarly. The claim that REAM 'maintains' accuracy is therefore supported only in the weak sense that the observed degradation is not statistically significant; the stronger claim that REAM improves accuracy over θ_S is not established. The paper should state this explicitly and, if the stronger claim is intended, provide paired significance tests or a power analysis.
minor comments (5)
  1. [§5.4.3 and Appendix B.9] The FFN exclusion window (layers 30–35) and the budget fraction ρ=0.3 are selected on validation subsets and then used as fixed choices for the reported results; Appendix B.9 itself states that the exact window should be validated once for each new backbone. This should be acknowledged as validation-based hyperparameter selection rather than a parameter-free aspect of the method, and the authors should clarify how this selection affects the generality claims in Table 3.
  2. [§4.2 and Appendix B.4] The use of the diagonal empirical Fisher as a curvature surrogate is justified by reference to the generalised Gauss–Newton decomposition, but the paper should explicitly note the known limitations of the empirical Fisher approximation (e.g., reference [23]) and state why the diagonal assumption is acceptable at head granularity.
  3. [Appendix A.4] The method is described as training-free, which is reasonable in the sense that no model parameters are optimized, but the Fisher-sensitivity estimation requires backward passes to compute gradients. The authors should clarify this terminology, since readers may understand 'training-free' to mean 'no gradient computation at all.'
  4. [Table 12] The ordering of rows in the Music panel differs from the Yelp panel (Mass-based, √n-faith, Density-faith versus Mass-based, Density-faith, √n-faith); please make the ordering consistent.
  5. [A.7] The paper notes that fully unstructured traces are left to future work; this is a fair limitation, but it should also be mentioned in the main text's discussion of applicability, not only in the appendix.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: REAM's merge coefficients are fitted to calibration statistics and evaluated on held-out test data, with no target metric encoded in the derivation.

full rationale

REAM's derivation chain is self-contained against the reported results. The merge coefficients are produced by Eqs. (1)-(7) from retrieval criticality, decision faithfulness, and Fisher-weighted update sensitivity computed on the 500-example calibration set Dslow, with fixed hyperparameters (alpha-bar=1, gamma=3, rho=0.3) and an FFN exclusion window selected on validation subsets. The reported MAE, RMSE, and token counts (Table 1) are measured on held-out test sets, so no reported quantity is a fitted input renamed as a prediction. The causal validation (Sec. 5.4.2) further tests the head-level signals against random ablations rather than deriving them from the final accuracy metric. The only self-reference, citation [54], supports the background claim that slow-thinking recommenders overthink and is not load-bearing for REAM's mechanism. The skeptical concern that TALLRec was fine-tuned from Qwen2.5-3B-non-instruct while the declared shared base is Qwen2.5-3B-Instruct, contaminating Δ_F with a checkpoint difference, is a real validity or correctness risk for the task-vector interpretation, but it is not circularity: the method's conclusion is not defined into its premises, and the reported compression could in principle fail to support the mechanism story without making the derivation self-referential. Under the stated rules, no circular step can be exhibited, so the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The method introduces no new physical or ontological entities; it is an algorithmic procedure over existing model weights. The central claim rests on the standard task-vector additivity assumption, the diagonal empirical Fisher approximation, and a set of validation-tuned hyperparameters (gamma, rho, FFN exclusion window).

free parameters (3)
  • gamma (exponential multiplier strength) = 3
    Chosen by hand and fixed across datasets; scales the reasoning-importance exponent in the perturbation weight s_b (Section 4.3).
  • rho (budget fraction) = 0.3
    Selected on validation subsets (Figure 6); controls the permitted fraction of the perturbation budget in the constraint (Section 4.3).
  • FFN exclusion window = layers 30-35 for Qwen2.5-3B; re-selected for each other backbone
    Chosen by validation sweep (Figure 5 and Appendix B.9); final FFN layers are excluded from merging because they degrade the accuracy-efficiency trade-off.
assumptions (5)
  • domain assumption Task-vector arithmetic: a linear combination of fine-tuned weights relative to a shared base yields a model with blended behavior.
    Invoked in Section 3.2 (theta_merged = theta_B + sum lambda_m Delta_m) and Section 4.3 Eq. (8). All merging baselines share this assumption; REAM inherits it.
  • domain assumption The diagonal empirical Fisher of the teacher-forced loss approximates the loss curvature along the fast update direction.
    Used in Section 4.2 and Appendices B.4-B.5; the residual Hessian term is discarded and the empirical Fisher is not the model Fisher.
  • domain assumption The match segment is the dominant decision-critical region, justified by perturbation analysis.
    Section 4.1.2 and Appendix B.2: perturbing match yields 3.3 to 5.1 times larger JSD at rating tokens than user or item; decision faithfulness is defined as attention share to match.
  • domain assumption The retrieval-head criterion of Wu et al. transfers from general QA to recommendation reasoning.
    Section 4.1.1 adopts Cond.1-Cond.2 alignment of generated tokens with source positions; no new derivation, relies on prior work.
  • domain assumption Max-aggregation over GQA groups conservatively treats shared K/V projections as critical when any query head in the group is critical.
    Section 4.1.2; a modeling choice to handle grouped-query attention, not empirically validated.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging." pith.science (2026). https://pith.science/paper/XLOQYE3V

@misc{pith2026260810447,
  author       = {Pith},
  title        = {Pith review of: Towards Efficient Reasoning in LLM-Based Recommender Systems via Model Merging},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XLOQYE3V}},
  note         = {Machine review of arXiv:2608.10447}
}
read the original abstract

Large language model-based recommender systems are increasingly adopting slow-thinking models that generate step-by-step reasoning before making predictions, often achieving higher accuracy than fast-thinking models that predict directly. However, their reasoning traces are often unnecessarily verbose, increasing inference costs without commensurate accuracy gains. Existing training-based approaches to reasoning compression often incur substantial adaptation costs, while inference-time methods are brittle and difficult to scale. These limitations motivate model merging as a promising training-free direction for transferring specialised behaviours between models in a shared parameter space. In particular, merging a slow-thinking model with a fast-thinking counterpart provides a natural mechanism for balancing recommendation accuracy and reasoning conciseness. To this end, we propose, to our knowledge, the first model merging framework for reasoning compression in recommender systems. Unlike conventional merging methods that apply uniform merge coefficients across model components, our method performs fine-grained merging at the level of individual attention heads, capturing heterogeneous patterns in recommendation reasoning. Each attention head is assigned a distinct merge coefficient according to its contribution to critical reasoning evidence and its sensitivity to parameter change, enabling selective injection of the concise behaviour of the fast-thinking model into the slow-thinking model and reducing reasoning verbosity without compromising recommendation quality. Experiments on three benchmark datasets show that our method reduces reasoning length by up to 24.3% while outperforming competitive model merging baselines in maintaining recommendation accuracy. The code is available at https://github.com/linhledieu/REAM.

Figures

Figures reproduced from arXiv: 2608.10447 by the authors.

Figure 1
Figure 1. Motivation for merging fast- and slow-thinking [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework. later generation to drift from the user–item context and propagate errors to the final prediction. Concretely, we follow the retrieval-head criterion of Wu et al. [44], which identifies retrieval by aligning the generated token with the source position receiving a head’s strongest attention. Since slow￾thinking recommendation traces comprise distinct segments, we evaluate this cri… view at source ↗
Figure 3
Figure 3. Per-head merge coefficients 𝛼 ★ across layers and attention heads on Music, Yelp, and Book. MAE on Music. Although several baselines attain stronger results on isolated metrics, none sustains these gains across both accuracy measures and reasoning length. Specifically, AIM+TA achieves a lower RMSE on Music (0.8227 vs. 0.8324), but worsens MAE relative to REAM (0.5678 vs. 0.5348) and performs less favourably on the o… view at source ↗
Figures from the paper (6 more)
Figure 6
Figure 6. Figure 6: Hyperparameter analysis: MAE, RMSE, and Tok vs. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 4
Figure 4. Figure 4: NLL and JSD increase (log scale) from ablating [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FFN exclusion sweep on Music and Yelp subsets. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Retrieval criticality 𝑅ℓ,ℎ per head, across layers, on Music, Yelp, and Book. codebase [42] 1 ; AIM [31], ACM [50], and RAIN-Merging [16] use their official codebases.2 Average Merging directly averages 𝜃𝑆 and 𝜃𝐹 ’s parameters. TA adds task vectors (parameter differenc…
Figure 8
Figure 8. Figure 8: Decision faithfulness faithℓ,ℎ per head, across layers, on Music, Yelp, and Book. analysis can be applied to identifiable regions containing the in￾formation required at each reasoning stage, including retrieved, collaborative, or multimodal context. The region connect…
Figure 10
Figure 10. Figure 10: FFN exclusion sweep on Yelp for Llama-3.2-3B [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

60 extracted references · 12 canonical work pages

  1. [1]

    Pranjal Aggarwal and Sean Welleck. 2025. L1: Controlling How Long A Reasoning Model Thinks With Reinforcement Learning. arXiv:2503.04697 [cs.CL] https: //arxiv.org/abs/2503.04697

    arXiv 2025

  2. [2]

    Aytes, Jinheon Baek, and Sung Ju Hwang

    Simon A. Aytes, Jinheon Baek, and Sung Ju Hwang. 2025. Sketch-of- Thought: Efficient LLM Reasoning with Adaptive Cognitive-Inspired Sketching. arXiv:2503.05179 [cs.CL] https://arxiv.org/abs/2503.05179

    arXiv 2025

  3. [3]

    Keqin Bao, Jizhi Zhang, Wenjie Wang, Yang Zhang, Zhengyi Yang, Yancheng Luo, Chong Chen, Fuli Feng, and Qi Tian. 2023. A Bi-Step Grounding Paradigm for Large Language Models in Recommendation Systems. arXiv:2308.08434 [cs.IR] https://arxiv.org/abs/2308.08434

    arXiv 2023

  4. [4]

    Keqin Bao, Jizhi Zhang, Yang Zhang, Wenjie Wang, Fuli Feng, and Xiangnan He. 2023. 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). ACM, 1007–1014. doi:10.1145/ 3604915.3608857

    arXiv 2023

  5. [5]

    2009.Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit

    Steven Bird, Ewan Klein, and Edward Loper. 2009.Natural Language Processing with Python: Analyzing Text with the Natural Language Toolkit. O’Reilly Media, Inc

    work page 2009

  6. [6]

    Jundong Chen, Honglei Zhang, Chunxu Zhang, Fangyuan Luo, and Yidong Li. 2025. Breaking the Aggregation Bottleneck in Federated Recommendation: A Personalized Model Merging Approach. arXiv:2508.12386 [cs.DC] https: //arxiv.org/abs/2508.12386

    work page Pith review arXiv 2025

  7. [7]

    Yi Fang, Wenjie Wang, Yang Zhang, Fengbin Zhu, Qifan Wang, Fuli Feng, and Xiangnan He. 2025. Reason4Rec: Large Language Models for Recommendation with Deliberative User Preference Alignment. arXiv:2502.02061 [cs.IR] https: //arxiv.org/abs/2502.02061

    arXiv 2025

  8. [8]

    Jingtong Gao, Zeyu Song, Chi Lu, Xiaopeng Li, Derong Xu, Maolin Wang, Peng Jiang, Kun Gai, Qingpeng Cai, and Xiangyu Zhao. 2026. Rein- forced Preference Optimization for Reasoning-Augmented Recommendations. arXiv:2605.21967 [cs.IR] https://arxiv.org/abs/2605.21967

    work page Pith review arXiv 2026

Show all 60 references
  1. [9]

    Shijie Geng, Shuchang Liu, Zuohui Fu, Yingqiang Ge, and Yongfeng Zhang. 2023. Recommendation as Language Processing (RLP): A Unified Pretrain, Personalized Prompt & Predict Paradigm (P5). arXiv:2203.13366 [cs.IR] https://arxiv.org/abs/ 2203.13366

    2023 arXiv

  2. [10]

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. 2021. Transformer Feed-Forward Layers Are Key-Value Memories. arXiv:2012.14913 [cs.CL] https: //arxiv.org/abs/2012.14913

    2021 arXiv

  3. [11]

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Peiyi Wang, et al. 2025. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning. Nature645, 8081 (2025), 633–638. doi:10.1038/s41586-025-09422-z

    2025 doi

  4. [12]

    Tingxu Han, Zhenting Wang, Chunrong Fang, Shiyu Zhao, Shiqing Ma, and Zhenyu Chen. 2025. Token-Budget-Aware LLM Reasoning. arXiv:2412.18547 [cs.CL] https://arxiv.org/abs/2412.18547

    2025 arXiv

  5. [13]

    Bairu Hou, Yang Zhang, Jiabao Ji, Yujian Liu, Kaizhi Qian, Jacob Andreas, and Shiyu Chang. 2025. ThinkPrune: Pruning Long Chain-of-Thought of LLMs via Reinforcement Learning. arXiv:2504.01296 [cs.CL] https://arxiv.org/abs/2504. 01296

    2025 arXiv

  6. [14]

    Min Hou, Xin Liu, Le Wu, Chenyi He, Hao Liu, Zhi Li, Xin Li, and Si Wei

  7. [15]

    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. 2021. LoRA: Low-Rank Adaptation of Large Language Models. arXiv:2106.09685 [cs.CL] https://arxiv.org/abs/2106.09685

    2021 arXiv

  8. [16]

    Zhehao Huang, Yuhang Liu, Baijiong Lin, Yixin Lou, Zhengbao He, Hanling Tian, Tao Li, and Xiaolin Huang. 2026. RAIN-Merging: A Gradient-Free Method to Enhance Instruction Following in Large Reasoning Models with Preserved Thinking Format. arXiv:2602.22538 [cs.LG] https://arxiv...

    2026

  9. [17]

    Gabriel Ilharco, Marco Tulio Ribeiro, Mitchell Wortsman, Suchin Gururangan, Ludwig Schmidt, Hannaneh Hajishirzi, and Ali Farhadi. 2023. Editing Models with Task Arithmetic. arXiv:2212.04089 [cs.LG] https://arxiv.org/abs/2212.04089

    2023 arXiv

  10. [18]

    Jie Jiang, Yang Wu, Qian Li, Yuling Xiong, Hongbo Tang, Xun Liu, Haoze Wang, Jun Zhang, Huan Yu, and Hailong Shi. 2026. SCoTER: Structured Chain-of- Thought Transfer for Enhanced Recommendation. arXiv:2511.19514 [cs.IR] https://arxiv.org/abs/2511.19514

    2026

  11. [19]

    2011.Thinking, Fast and Slow

    Daniel Kahneman. 2011.Thinking, Fast and Slow. Farrar, Straus and Giroux, New York

    2011

  12. [20]

    Hyunsoo Kim, Jaewan Moon, Seongmin Park, and Jongwuk Lee. 2026. MergeRec: Model Merging for Data-Isolated Cross-Domain Sequential Recommendation. InProceedings of the 32nd ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1(Republic of Korea)(KDD ’26). Association...

    2026

  13. [21]

    Jieyong Kim, Hyunseo Kim, Hyunjin Cho, SeongKu Kang, Buru Chang, Jinyoung Yeo, and Dongha Lee. 2025. Review-driven Personalized Preference Reasoning with Large Language Models for Recommendation. InProceedings of the 48th International ACM SIGIR Conference on Research and Deve...

    2025

  14. [22]

    Xiaoyu Kong, Junguang Jiang, Bin Liu, Ziru Xu, Han Zhu, Jian Xu, Bo Zheng, Jiancan Wu, and Xiang Wang. 2025. Think before Recommendation: Autonomous Reasoning-enhanced Recommender. arXiv:2510.23077 [cs.IR] https://arxiv.org/ abs/2510.23077

    2025

  15. [23]

    Frederik Kunstner, Lukas Balles, and Philipp Hennig. 2020. Limita- tions of the Empirical Fisher Approximation for Natural Gradient Descent. arXiv:1905.12558 [cs.LG] https://arxiv.org/abs/1905.12558

    2020 arXiv

  16. [24]

    Shijun Li, Yu Wang, Jin Wang, Ying Li, Joydeep Ghosh, and Anne Cocos. 2026. LLM Reasoning for Cold-Start Item Recommendation. InProceedings of the ACM Web Conference 2026(United Arab Emirates)(WWW ’26). Association for Computing Machinery, New York, NY, USA, 8409–8412. doi:10....

    2026 doi

  17. [25]

    Jiayi Liao, Sihang Li, Zhengyi Yang, Jiancan Wu, Yancheng Yuan, Xiang Wang, and Xiangnan He. 2024. LLaRA: Large Language-Recommendation Assistant. arXiv:2312.02445 [cs.IR] https://arxiv.org/abs/2312.02445

    2024 arXiv

  18. [26]

    Jianghao Lin, Xinyi Dai, Yunjia Xi, Weiwen Liu, Bo Chen, Hao Zhang, Yong Liu, Chuhan Wu, Xiangyang Li, Chenxu Zhu, Huifeng Guo, Yong Yu, Ruiming Tang, and Weinan Zhang. 2024. How Can Recommender Systems Benefit from Large Language Models: A Survey. arXiv:2306.05817 [cs.IR] htt...

    2024 arXiv

  19. [27]

    Shuqi Liu, Han Wu, Bowei He, Xiongwei Han, Mingxuan Yuan, and Linqi Song

  20. [28]

    Xinyin Ma, Guangnian Wan, Runpeng Yu, Gongfan Fang, and Xinchao Wang. 2025. CoT-Valve: Length-Compressible Chain-of-Thought Tuning. arXiv:2502.09601 [cs.AI] https://arxiv.org/abs/2502.09601

    2025 arXiv

  21. [29]

    arXiv:2502.12420 [cs.CL] https://arxiv.org/abs/2502.12420

    Sens-Merging: Sensitivity-Guided Parameter Balancing for Merging Large Language Models. arXiv:2502.12420 [cs.CL] https://arxiv.org/abs/2502.12420

    arXiv

  22. [30]

    Michael Matena and Colin Raffel. 2022. Merging Models with Fisher-Weighted Averaging. arXiv:2111.09832 [cs.LG] https://arxiv.org/abs/2111.09832

    2022 arXiv

  23. [31]

    James Martens. 2020. New insights and perspectives on the natural gradient method. arXiv:1412.1193 [cs.LG] https://arxiv.org/abs/1412.1193

    2020 arXiv

  24. [32]

    Yein Park, Minbyul Jeong, and Jaewoo Kang. 2026. Thinking Sparks!: Emergent Attention Heads in Reasoning Models During Post Training. arXiv:2509.25758 [cs.AI] https://arxiv.org/abs/2509.25758

    2026 arXiv

  25. [33]

    Amin Heyrani Nobari, Kaveh Alim, Ali ArjomandBigdeli, Akash Srivastava, Faez Ahmed, and Navid Azizan. 2025. Activation-Informed Merging of Large Language Models. arXiv:2502.02421 [cs.CL] https://arxiv.org/abs/2502.02421

    2025

  26. [34]

    Yang Sui, Yu-Neng Chuang, Guanchu Wang, Jiamu Zhang, Tianyi Zhang, Jiayi Yuan, Hongyi Liu, Andrew Wen, Shaochen Zhong, Na Zou, Hanjie Chen, and Xia Hu. 2025. Stop Overthinking: A Survey on Efficient Reasoning for Large Language Models. arXiv:2503.16419 [cs.CL] https://arxiv.or...

    2025 arXiv

  27. [35]

    Jung Hyun Ryu, Jaeheyoung Jeon, Jewoong Cho, and Myungjoo Kang 1. 2023. Fisher-Weighted Merge of Contrastive Learning Models in Sequential Recom- mendation. arXiv:2307.05476 [cs.IR] https://arxiv.org/abs/2307.05476

    2023 arXiv

  28. [36]

    Elena Voita, David Talbot, Fedor Moiseev, Rico Sennrich, and Ivan Titov. 2019. Analyzing Multi-Head Self-Attention: Specialized Heads Do the Heavy Lifting, the Rest Can Be Pruned. arXiv:1905.09418 [cs.CL] https://arxiv.org/abs/1905. 09418 Le et al

    2019 arXiv

  29. [37]

    Zhongxiang Sun, Zihua Si, Xiaoxue Zang, Kai Zheng, Yang Song, Xiao Zhang, and Jun Xu. 2024. Large Language Models Enhanced Collaborative Filtering. arXiv:2403.17688 [cs.IR] https://arxiv.org/abs/2403.17688

    2024 arXiv

  30. [38]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Brian Ichter, Fei Xia, Ed Chi, Quoc Le, and Denny Zhou. 2023. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. arXiv:2201.11903 [cs.CL] https: //arxiv.org/abs/2201.11903

    2023 arXiv

  31. [39]

    Xinfeng Wang, Jin Cui, Yoshimi Suzuki, and Fumiyo Fukumoto. 2025. RDRec: Rationale Distillation for LLM-based Recommendation. arXiv:2405.10587 [cs.CL] https://arxiv.org/abs/2405.10587

    2025 arXiv

  32. [40]

    Willmott and Kenji Matsuura

    Cort J. Willmott and Kenji Matsuura. 2005. Advantages of the mean absolute error (MAE) over the root mean square error (RMSE) in assessing average model performance.Climate Research30 (2005), 79–82. https://api.semanticscholar. org/CorpusID:120556606

    2005

  33. [41]

    Tianjun Wei, Enneng Yang, Yingpeng Du, Huizhong Guo, Jie Zhang, and Zhu Sun. 2026. MMGRid: Navigating Temporal-aware and Cross-domain Generative Recommendation via Model Merging. arXiv:2601.15930 [cs.IR] https://arxiv.org/ abs/2601.15930

    2026

  34. [42]

    Han Wu, Yuxuan Yao, Shuqi Liu, Zehua Liu, Xiaojin Fu, Xiongwei Han, Xing Li, Hui-Ling Zhen, Tao Zhong, and Mingxuan Yuan. 2025. Unlocking Efficient Long-to-Short LLM Reasoning with Model Merging. arXiv:2503.20641 [cs.CL] https://arxiv.org/abs/2503.20641

    2025 arXiv

  35. [43]

    Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt

    Mitchell Wortsman, Gabriel Ilharco, Samir Yitzhak Gadre, Rebecca Roelofs, Raphael Gontijo-Lopes, Ari S. Morcos, Hongseok Namkoong, Ali Farhadi, Yair Carmon, Simon Kornblith, and Ludwig Schmidt. 2022. Model soups: averaging weights of multiple fine-tuned models improves accurac...

    2022 arXiv

  36. [44]

    Wenhao Wu, Yizhong Wang, Guangxuan Xiao, Hao Peng, and Yao Fu

  37. [45]

    Likang Wu, Zhi Zheng, Zhaopeng Qiu, Hao Wang, Hongchao Gu, Tingjia Shen, Chuan Qin, Chen Zhu, Hengshu Zhu, Qi Liu, Hui Xiong, and Enhong Chen. 2024. A Survey on Large Language Models for Recommendation. arXiv:2305.19860 [cs.IR] https://arxiv.org/abs/2305.19860

    2024 arXiv

  38. [46]

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of Draft: Thinking Faster by Writing Less. arXiv:2502.18600 [cs.CL] https://arxiv.org/abs/ 2502.18600

    2025 arXiv

  39. [47]

    Prateek Yadav, Derek Tam, Leshem Choshen, Colin Raffel, and Mohit Bansal. 2023. TIES-Merging: Resolving Interference When Merging Models. arXiv:2306.01708 [cs.LG] https://arxiv.org/abs/2306.01708

    2023 arXiv

  40. [48]

    Heming Xia, Chak Tou Leong, Wenjie Wang, Yongqi Li, and Wenjie Li

  41. [49]

    arXiv:2502.12067 [cs.CL] https://arxiv.org/abs/2502.12067

    TokenSkip: Controllable Chain-of-Thought Compression in LLMs. arXiv:2502.12067 [cs.CL] https://arxiv.org/abs/2502.12067

  42. [50]

    Yuxuan Yao, Shuqi Liu, Zehua Liu, Qintong Li, Mingyang Liu, Xiongwei Han, Zhijiang Guo, Han Wu, and Linqi Song. 2025. Activation-Guided Consensus Merging for Large Language Models. arXiv:2505.14009 [cs.CL] https://arxiv.org/ abs/2505.14009

    2025

  43. [51]

    Zixuan Yi and Iadh Ounis. 2025. A Multi-modal Recipe for Improved Multi- domain Recommendation. InAdvances in Information Retrieval (ECIR 2025) (Lucca, Italy). Springer-Verlag, Berlin, Heidelberg, 421–436. doi:10.1007/978- 3-031-88708-6_27

    2025 doi

  44. [52]

    Enneng Yang, Li Shen, Guibing Guo, Xingwei Wang, Xiaochun Cao, Jie Zhang, and Dacheng Tao. 2025. Model Merging in LLMs, MLLMs, and Beyond: Methods, Theories, Applications and Opportunities. arXiv:2408.07666 [cs.LG] https:// arxiv.org/abs/2408.07666

    2025 arXiv

  45. [53]

    Junyao Yang, Jianwei Wang, Huiping Zhuang, Cen Chen, and Ziqian Zeng. 2026. RCP-Merging: Merging Long Chain-of-Thought Models with Domain-Specific Models by Considering Reasoning Capability as Prior. arXiv:2508.03140 [cs.CL] https://arxiv.org/abs/2508.03140

    2026

  46. [54]

    Jiaqi Zhang, Junliang Yu, Zongwei Wang, Wei Yuan, Tong Chen, Quoc Viet Hung Nguyen, Bin Cui, and Hongzhi Yin. 2025. Towards Reasoning-Aware Recom- mender Systems: A Survey in the LLM Era.TechRxiv(2025). doi:10.36227/ techrxiv.176287939.92578520/v2

    2025

  47. [55]

    Yang Zhang, Wenxin Xu, Xiaoyan Zhao, Wenjie Wang, Fuli Feng, Xiangnan He, and Tat-Seng Chua. 2025. Reinforced Latent Reasoning for LLM-based Recommendation. arXiv:2505.19092 [cs.AI] https://arxiv.org/abs/2505.19092 A Implementation Details A.1 Fast-Thinking Model Adaptation TA...

    2025

  48. [56]

    Le Yu, Bowen Yu, Haiyang Yu, Fei Huang, and Yongbin Li. 2024. Language Models are Super Mario: Absorbing Abilities from Homologous Models as a Free Lunch. arXiv:2311.03099 [cs.CL] https://arxiv.org/abs/2311.03099

    2024 arXiv

  49. [57]

    Junjie Zhang, Ruobing Xie, Yupeng Hou, Wayne Xin Zhao, Leyu Lin, and Ji- Rong Wen. 2023. Recommendation as Instruction Following: A Large Language Model Empowered Recommendation Approach. arXiv:2305.07001 [cs.IR] https: //arxiv.org/abs/2305.07001

    2023 arXiv

  50. [576]

    Moreover, 96.6% of these heads reappear in the bootstrap top-5% set in at least 900 of the 1,000 draws

    have a median relative confidence-interval width of 14.95%, compared with 20.88% across all heads. Moreover, 96.6% of these heads reappear in the bootstrap top-5% set in at least 900 of the 1,000 draws. Together, these analyses show that retrieval-criticality estimates stabili...

  51. [2024]

    arXiv:2404.15574 [cs.CL] https://arxiv.org/abs/2404.15574

    Retrieval Head Mechanistically Explains Long-Context Factuality. arXiv:2404.15574 [cs.CL] https://arxiv.org/abs/2404.15574

    arXiv

  52. [2025]

    arXiv:2510.26546 [cs.IR] https://arxiv.org/abs/ 2510.26546

    WeaveRec: An LLM-Based Cross-Domain Sequential Recommendation Framework with Model Merging. arXiv:2510.26546 [cs.IR] https://arxiv.org/abs/ 2510.26546

Pith tools

Reviewed August 15, 2026 · model on record in the stance chip above.