Pith. sign in

REVIEW 4 major objections 6 minor 27 references

A locally deployed 3-billion-parameter model, routing only ambiguous commands through minimal reasoning and learning from a weighted memory of past interactions, can outperform uniform heavy-reasoning assistants on direct and indirect smart

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 16:16 UTC pith:JHKTWOM5

load-bearing objection A solid, useful system paper with a modest claim; the routing idea and preference memory work reasonably, but the missing intent-classifier confusion matrix leaves the safety story incomplete. the 4 major comments →

arxiv 2607.18034 v2 pith:JHKTWOM5 submitted 2026-07-20 cs.AI

AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models

Eu Jin Lim , Zhaoxing Li , Sebastian Stein This is my paper
classification cs.AI
keywords smart homesmall language modelsintent-aware planningChain-of-Draftpreference adaptationlocal deploymentpersonalizationlatency reduction
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

AdaHome is a smart-home assistant designed to run entirely on a 3-billion-parameter language model on a modest local computer. Its central claim is that cloud-based assistants seem necessary not because large models are required, but because reasoning is applied uniformly: direct commands get a heavyweight multi-step pipeline that slows them down and makes small models hallucinate. AdaHome classifies each command as direct, indirect, or ambiguous; direct commands go straight to a simple prompt, while only the other two get a minimal Chain-of-Draft reasoning pass. On top of this, a preference memory stores confirmed and corrected actions, weighting past interactions by semantic similarity and recency, so the system personalizes over time without retraining or growing its prompt. The paper reports that this beats three prior assistant designs on direct and indirect accuracy, cuts latency by up to 3x, and reaches 88% preference consistency versus 52.5% for prompt augmentation.

Core claim

AdaHome's discovery is that 'reasoning only when needed' is not just an engineering shortcut but an accuracy win for small models. Under a unified 3B-model setting, a straightforward prompt plus schema validation yields 86.7% exact-match success on direct commands, where mandatory reasoning pipelines scored 41.1–63.3% and produced spurious device activations; the same routing with compact Chain-of-Draft drafts reaches 86.7% on indirect commands. The second discovery is that continual personalization can be obtained from a kernel-regression estimator over a memory of past interactions: each past action contributes weight S^gamma * exp(-lambda dt), making the system stable under temporary devi

What carries the argument

The intent-aware planner: a classifier sends each command to a direct prompt or to a reasoning planner; the reasoning planner uses Chain-of-Draft prompting, which limits intermediate reasoning to at most eight words and emits a JSON action in a single pass. The preference module is a weighted kernel estimator: confirmed/corrected interactions are stored as embeddings, and for a new command the activation probability of device d is the similarity-and-recency weighted average of past binary states, with a preference-extraction step that normalizes paraphrases into a core intent.

Load-bearing premise

The claim that intent-classification errors only hurt speed, never correctness, rests on the unproven premise that an ambiguous or preference-dependent command, if mistakenly routed to the direct planner, is still executed correctly and safely without the confirmation step.

What would settle it

Run the published 90-command test set through the intent classifier and report a confusion matrix; specifically, count ambiguous commands labeled DIRECT. If even a handful of those produce an action plan a user would reject or that contradicts their preference, the 'misclassification affects efficiency rather than correctness' claim collapses.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Direct commands become cheap and reliable on resource-constrained hardware, enabling on-device assistants in homes where cloud round-trips are too slow or too private.
  • Personalization no longer requires prompt growth or fine-tuning; memory can keep updating indefinitely without hitting context limits.
  • The stability-plasticity trade-off can be controlled by two scalars (gamma and lambda), so behavior can be tuned for conservative or responsive users.
  • The same routing principle may transfer to other edge-agent domains: apply filtered reasoning only to the hard parts instead of running uniform chains.
  • If accuracy holds, privacy-preserving local deployment becomes a practical default for smart-home control, reducing data exposure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The paper's safety claim depends on an untested cell of the confusion matrix: if an ambiguous, preference-dependent command is misclassified as DIRECT, it skips both the reasoning planner and the user confirmation step; the reported 75.6% classifier accuracy without a confusion matrix leaves this gap open.
  • The memory formulation is restricted to binary device states; a natural testable extension is to generalize the kernel weighting to continuous attributes (brightness, temperature) via weighted means or quantile regression, which the paper itself flags as future work.
  • Because the preference extractor's normalization drives recall, the approach should transfer to other paraphrastic user expressions, but personalization quality then depends on the extractor's fidelity; a cheaper alternative like hashing intent templates could be compared.
  • If the 3x latency reduction holds across larger home schemas, the routing cost-benefit shifts: the direct-planner path is the one to optimize further, since a large share of real commands are direct.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 6 minor

Summary. The paper proposes AdaHome, a smart-home assistant for locally deployed 3B-parameter language models. It classifies commands into direct, indirect, and ambiguous categories; direct commands are handled by a prompt-only planner, indirect and ambiguous commands by a Chain-of-Draft reasoning planner, and ambiguous commands additionally receive user confirmation. A preference memory stores confirmed/corrected actions, embeds extracted intents, and predicts device activation using a time-decayed kernel-weighted average of semantically similar past interactions. In a single-turn evaluation with 90 commands against reimplementations of Sasha, SAGE, and Harmony under Llama 3.2-3B, AdaHome reports the highest direct-command accuracy (86.7%), highest indirect accuracy (86.7%), and near-best ambiguous accuracy (88.9%) with lower latency and token usage. In a multi-turn simulation, AdaHome is reported to achieve 87.5% preference consistency, 80% recovery, and 100% adaptation success, versus 52.5%, 10%, and 30% for a RAG baseline. The paper concludes that intent-aware routing with compact reasoning and memory-based personalization is a promising local-SLM alternative to cloud-based pipelines.

Significance. AdaHome addresses a timely and practical question: whether small local language models can support efficient, privacy-preserving, self-personalizing smart-home control without retraining or cloud inference. The system design is transparent, and the appendices provide prompts, the evaluation dataset, the device schema, and a longitudinal dataset, which substantially aids reproducibility. The evaluation uses a fixed device schema, a unified small model, greedy decoding, and an LLM-as-judge validated against human annotations (Cohen's κ = 0.834), which is a genuine strength. If the reported results are robust, the work demonstrates a favorable efficiency–accuracy trade-off for edge deployment and a lightweight mechanism for continual preference adaptation. However, the central comparative and robustness claims rest on a small, partly templated dataset, reimplemented baselines, and an intent classifier whose dangerous error direction is not analyzed. These uncertainties need to be addressed before the claims can be taken at face value.

major comments (4)
  1. [§4.1.2 (Impact of Intent Classification), §3 (User Confirmation)] The robustness claim is supported only for one error direction. The intent classifier is reported at 75.6% accuracy, and the discussion argues that direct commands misclassified as indirect preserve correctness and that indirect/ambiguous routing errors only affect confirmation. But an ambiguous command misclassified as DIRECT would be routed to the direct planner, bypassing both the reasoning planner and the user confirmation step — the exact safeguard the paper cites for safe execution of preference-dependent requests. The paper provides no confusion matrix or per-class accuracy, so the rate of this safety-relevant error direction is unknown. Please report the full confusion matrix and evaluate the success rate of ambiguous commands when routed to the direct planner. Without this, the statement 'misclassification affects execution efficiency rather than action correctness' is missing l
  2. [§4.2 (Preference Adaptation Analysis), Appendix G] The multi-turn preference experiment appears to inject the ground-truth preference at every turn. The longitudinal dataset is constructed from predefined scenarios with Preference A/B configurations, and each turn is assigned a target behavior; the paper does not state how user confirmation/correction is simulated. If the memory is updated with the ground-truth action at every turn, then the Consistency/Recovery/Adaptation metrics measure the kernel smoother's ability to track a label sequence rather than the full system's ability to interpret ambiguous commands and learn from feedback. This is a form of label leakage for the main personalization claim. Please specify the feedback-generation protocol; if memory is updated with true labels, present that as an oracle-feedback experiment and also report a variant using the system's own generated actions or noisy/corrective feedback.
  3. [§4.1.1 (Dataset), Table 2] The evaluation dataset is small (n=90, 30 per category), and the direct commands are largely templated variants such as 'turn on/off X,' 'lock/unlock the front door,' and 'set X to high/low.' With only three runs and no confidence intervals or significance tests, the claim of 'substantially higher accuracy' on direct commands (86.7% vs 41.1–63.3%) may not generalize to more naturalistic user instructions. Please provide per-command error analysis, bootstrap or other confidence intervals, and ideally expand the dataset with more varied, paraphrased direct commands.
  4. [§4.1.1 (Baselines)] Sasha, SAGE, and Harmony are reimplementations rather than official released systems. The paper does not report validation against the original systems' published performance, nor does it provide the exact baseline prompts and pipeline details in the appendices. Because the headline accuracy and latency comparisons are against these baselines, the authors should show that each reimplementation behaves as intended (e.g., sample intermediate outputs on a few canonical examples) or release the baseline prompts/configurations for independent verification.
minor comments (6)
  1. [§3.2, Eq. (1)] The cosine similarity threshold is mentioned but never specified; k=10, γ=3, and λ=0.1 are fixed without sensitivity analysis. Please report the threshold value and add a small sensitivity study (e.g., varying λ and γ) to show the preference results are not brittle.
  2. [Figure 6] Figure 6 does not include error bars or a numeric table of the metrics. Given that the longitudinal dataset has 30 sequences, reporting variance or per-sequence results would strengthen the stability–plasticity conclusions.
  3. [Table 2] The abstract claims latency reduction 'up to 3×,' but in Table 2 the largest speedup is 28.84s/10.48s ≈ 2.75× for direct commands. Please adjust the wording or provide the exact ratio. Also, the absolute direct-command latency of about 10.5s on the test hardware is high for interactive smart-home control; a brief discussion of real-time requirements would be useful.
  4. [§4.1.1 (Dataset construction)] The text says the dataset was derived from the Sasha dataset (40 commands) and SAGE dataset (50 commands), but the final dataset contains 90 commands. Please clarify how many commands came from each source and how many were manually constructed to achieve balanced coverage.
  5. [§4.3.1 (Ablation study)] The 'No Memory' configuration achieves only 7.5% stable accuracy on repeated stable commands. This is surprising for a deterministic greedy-decoding system. Please explain why repeated identical or semantically similar ambiguous commands fail without memory, or provide representative outputs for inspection.
  6. [Abstract vs §4.2.1] The abstract states '88% preference consistency,' while Section 4.2.1 reports 87.5% stable accuracy. Please make the numbers consistent.

Circularity Check

0 steps flagged

No significant circularity found: AdaHome's routing, CoD reasoning, and preference adaptation are empirically evaluated components with external grounding, and no load-bearing result reduces to its own inputs by construction.

full rationale

The paper has no derivation chain whose conclusion is secretly an input. The intent-aware router is an architectural choice implemented as a structured LLM prompt (Appendix A), not a parameter fitted to the reported accuracies; the classifier's 75.6% accuracy is reported as an independent measurement and its robustness claim ('misclassification affects execution efficiency rather than action correctness', §4.1.2) is a missing-support issue, not circularity. Chain-of-Draft is adopted from an external publication [20], with no self-citation or uniqueness theorem imported from the authors. The preference adaptation mechanism (Eqs. 1–2) transparently defines a weighted kernel-regression estimator over stored feedback; this is the mechanism under test, not a hidden reuse of the evaluation labels as a fitted prediction. The multi-turn experiment feeds confirmed/corrected actions into memory and then measures consistency and adaptation on synthetic sequences; this is a standard feedback-loop evaluation, and the parameters γ=3 and λ=0.1 are fixed a priori with external justification rather than fitted to the target metrics. The only self-reference is the provenance footnote identifying the paper as an extended version of itself, which is not load-bearing. Overall, the central claims are empirical and independently testable; the identified weaknesses (missing confusion matrix, synthetic preference sequences) concern evaluation validity, not circular reduction.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The preference mechanism depends on four hand-set constants (gamma, lambda, k, and an unspecified similarity threshold) and several domain assumptions about embeddings, feedback reliability, and temporal forgetting. These choices are load-bearing for the multi-turn results; changing them would change the stability-plasticity trade-off and the reported consistency, recovery, and adaptation numbers.

free parameters (4)
  • gamma = 3
    Exponent on cosine similarity in Eq. (1), set by hand 'to increase selectivity'; it directly shapes the weights in all preference predictions.
  • lambda = 0.1 per day
    Temporal decay rate in Eq. (1), set so influence halves after about 7 days; it controls how quickly old interactions are forgotten in preference estimates.
  • k = 10
    Number of nearest neighbors retrieved from preference memory before weighting; fixed in Section 3.2 and used in all multi-turn experiments.
  • cosine_similarity_threshold
    Threshold used to filter weak matches in preference retrieval; mentioned in Section 3.2 but its value is not given, so it is an unspecified free choice that affects whether memory is used at all.
axioms (4)
  • domain assumption Semantically equivalent user intents are close in the embedding space of the unspecified sentence-embedding model, making cosine similarity a valid kernel for preference generalization.
    Section 3.2: extracted intents are embedded and compared by cosine similarity, but no validation of embedding quality or threshold behavior is provided.
  • domain assumption User confirmations and corrections are accurate ground-truth preference labels and are immediately available for memory updates.
    Section 3.2: confirmed or corrected actions are stored as new preference tuples; there is no modeling of noisy, mistaken, or inconsistent user feedback.
  • domain assumption Recent behavior is exponentially more informative than older behavior, with a 7-day half-life.
    Section 3.2: Eq. (1) uses e^{-lambda*delta t} with lambda=0.1, motivated by weekly patterns in smart home usage [15]; this timescale is assumed, not fitted or validated.
  • standard math Nadaraya-Watson kernel regression with the cosine/exponential kernel gives a valid estimate of device-activation probability.
    Eq. (2) cites [12,18]; the estimator is classical and not proved in the paper, which is acceptable, but the kernel choice is an unvalidated modeling decision.

pith-pipeline@v1.3.0-alltime-deepseek · 16056 in / 12655 out tokens · 126369 ms · 2026-08-01T16:16:09.123947+00:00 · methodology

0 comments
read the original abstract

Smart home assistants interpret a wide range of user commands, from explicit device control to underspecified and preference dependent requests. While recent systems based on Large Language Models (LLMs) improve this capability, they often rely on heavyweight reasoning pipelines and cloud-based deployment, limiting their efficiency and suitability for resource-constrained environments, and raising privacy concerns. In addition, existing approaches provide limited support for stable long-term personalization. To address these issues, we present AdaHome, an adaptive smart home assistant designed for locally deployed small language models in smart home environments. Rather than applying complex reasoning uniformly, AdaHome introduces an intent-aware planning framework that dynamically routes commands either to straightforward prompt-based or lightweight reasoning-based components. For commands requiring interpretation, we adopt a Chain-of-Draft strategy to enable efficient and stable decision-making. To support personalization, we further propose a preference adaptation mechanism that learns from user feedback over time without requiring prompt augmentation or model retraining. We evaluate AdaHome against representative LLM-based baselines under a unified small model setting. AdaHome achieves substantially higher accuracy on direct commands (86.7%) while reducing latency by up to 3$\times$. Furthermore, it maintains competitive performance on ambiguous inputs with lower computational cost. In multi-turn scenarios, AdaHome achieves 88% preference consistency, compared to 52.5% for a prompt augmentation baseline.

Figures

Figures reproduced from arXiv: 2607.18034 by Eu Jin Lim, Sebastian Stein, Zhaoxing Li.

Figure 1
Figure 1. Figure 1: Overview of AdaHome system architecture, consisting of an intent-aware planning pipeline and a continual preference [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: An example of Chain-of-Draft reasoning generating [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of AdaHome’s preference adaptation mechanism. The system retrieves semantically similar past interactions, [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Success rate across Direct, Indirect, and Ambiguous [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Overall trade-off between average end-to-end la [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Multi-turn preference adaptation performance: Ada [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Ablation of AdaHome components: effect of mem [PITH_FULL_IMAGE:figures/full_fig_p008_7.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

27 extracted references · 7 linked inside Pith

  1. [1]

    Matthijs Douze, Alexandr Guzhva, Chengqi Deng, Jeff Johnson, Gergely Szilvasy, Pierre-Emmanuel Mazaré, Maria Lomeli, Lucas Hosseini, and Hervé Jégou. 2024. The Faiss library. (2024). arXiv:2401.08281 [cs.LG]

    Pith/arXiv arXiv 2024

  2. [2]

    Simone Gallo, Fabio Paternò, and Alessio Malizia. 2024. A conversational agent for creating automations exploiting large language models.Personal and Ubiqui- tous Computing28, 6 (2024), 931–946

    2024

  3. [3]

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Ab- hishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schel- ten, Alex Vaughan, et al . 2024. The llama 3 herd of models.arXiv preprint arXiv:2407.21783(2024). AdaHome: An Adaptive Smart Home Assistant using Local Small Language Models

    Pith/arXiv arXiv 2024

  4. [4]

    Dongge Han, Menglin Xia, Daniel Madrigal Diaz, Samuel Kessler, Ankur Mallick, Xuchao Zhang, Mirian Del Carmen Hipolito Garcia, Jin Xu, Victor Rühle, and Sar- avan Rajmohan. 2025. Enhancing reasoning capabilities of small language models with blueprints and prompt template search.arXiv preprint arXiv:2506.08669 (2025)

    Pith/arXiv arXiv 2025

  5. [5]

    Xinyu Huang, Leming Shen, Zijing Ma, and Yuanqing Zheng. 2026. Towards privacy-preserving and personalized smart homes via tailored small language models.IEEE Transactions on Mobile Computing(2026)

    2026

  6. [6]

    Get ready for a party

    Evan King, Haoxiang Yu, Sangsu Lee, and Christine Julien. 2023. " Get ready for a party": Exploring smarter smart spaces with help from large language models. arXiv preprint arXiv:2303.14143(2023)

    Pith/arXiv arXiv 2023

  7. [7]

    Evan King, Haoxiang Yu, Sangsu Lee, and Christine Julien. 2024. Sasha: creative goal-oriented reasoning in smart homes with large language models.Proceedings of the ACM on Interactive, Mobile, Wearable and Ubiquitous Technologies8, 1 (2024), 1–38

    2024

  8. [8]

    Song Lai, Zhe Zhao, Fei Zhu, Xi Lin, Qingfu Zhang, and Gaofeng Meng. 2025. Pareto continual learning: Preference-conditioned learning and adaption for dynamic stability-plasticity trade-off. InProceedings of the AAAI Conference on Artificial Intelligence, Vol. 39. 18008–18016

    2025

  9. [9]

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rock- täschel, et al. 2020. Retrieval-augmented generation for knowledge-intensive nlp tasks.Advances in neural information processing systems33 (2020), 9459–9474

    2020

  10. [10]

    Shuozhe Li, Daniel Yamakov, and Christine Julien. [n. d.]. Smart Homes Reimag- ined: Human-Guided LLMs for Efficient and Secure Home Automation. ([n. d.])

  11. [11]

    Yuetai Li, Xiang Yue, Zhangchen Xu, Fengqing Jiang, Luyao Niu, Bill Yuchen Lin, Bhaskar Ramasubramanian, and Radha Poovendran. 2025. Small models struggle to learn from strong reasoners. InFindings of the Association for Computational Linguistics: ACL 2025. 25366–25394

    2025

  12. [12]

    Elizbar A Nadaraya. 1964. On estimating regression.Theory of Probability & Its Applications9, 1 (1964), 141–142

    1964

  13. [13]

    Qwen, :, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li,...

    Pith/arXiv arXiv 2025

  14. [14]

    Dmitriy Rivkin, Francois Hogan, Amal Feriani, Abhisek Konar, Adam Sigal, Xue Liu, and Gregory Dudek. 2024. AIoT smart home via autonomous LLM agents. IEEE Internet of Things Journal12, 3 (2024), 2458–2472

    2024

  15. [15]

    Abdussalam Salama, Reza Saatchi, Maryam Bagheri, Mahpara Saleem, and Muhammad Usman Shad. 2025. Development and Evaluation of a Real-Time Home Monitoring Application Utilising Long Short-Term Memory Integrated in a Smartphone.Algorithms18, 12 (2025), 780

    2025

  16. [16]

    Yingtian Shi, Xiaoyi Liu, Chun Yu, Tianao Yang, Cheng Gao, Chen Liang, and Yuanchun Shi. 2024. Bridging the gap between natural user expression with com- plex automation programming in smart homes.arXiv preprint arXiv:2408.12687 (2024)

    Pith/arXiv arXiv 2024

  17. [17]

    Fali Wang, Zhiwei Zhang, Xianren Zhang, Zongyu Wu, Tzuhao Mo, Qiuhao Lu, Wanjing Wang, Rui Li, Junjie Xu, Xianfeng Tang, et al. 2025. A comprehensive survey of small language models in the era of large language models: Techniques, enhancements, applications, collaboration with llms, and trustworthiness.ACM Transactions on Intelligent Systems and Technolog...

    2025

  18. [18]

    Geoffrey S Watson. 1964. Smooth regression analysis.Sankhy ¯a: The Indian Journal of Statistics, Series A(1964), 359–372

    1964

  19. [19]

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. 2022. Chain-of-thought prompting elicits reason- ing in large language models.Advances in neural information processing systems 35 (2022), 24824–24837

    2022

  20. [20]

    Silei Xu, Wenhao Xie, Lingxiao Zhao, and Pengcheng He. 2025. Chain of draft: Thinking faster by writing less.arXiv preprint arXiv:2502.18600(2025)

    Pith/arXiv arXiv 2025

  21. [21]

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. 2022. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations

    2022

  22. [22]

    Ziqi Yin, Mingxin Zhang, and Daisuke Kawahara. 2025. Harmony: A Human- Aware, Responsive, Modular Assistant with a Locally Deployed Large Language Model. InCompanion of the 2025 ACM International Joint Conference on Pervasive and Ubiquitous Computing. 126–130

    2025

  23. [23]

    Chaerin Yu, Chihun Choi, Sunjae Lee, Hyosu Kim, Steven Y Ko, Young-Bae Ko, and Sangeun Oh. 2026. Leveraging LLMs for Efficient and Personalized Smart Home Automation.arXiv preprint arXiv:2601.04680(2026)

    arXiv 2026

  24. [24]

    Zeyu Zhang, Quanyu Dai, Xiaohe Bo, Chen Ma, Rui Li, Xu Chen, Jieming Zhu, Zhenhua Dong, and Ji-Rong Wen. 2025. A survey on the memory mechanism of large language model-based agents.ACM Transactions on Information Systems 43, 6 (2025), 1–47

    2025

  25. [25]

    Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric Xing, et al. 2023. Judging llm-as-a-judge with mt-bench and chatbot arena.Advances in neural information processing systems36 (2023), 46595–46623

    2023

  26. [26]

    Wanjun Zhong, Lianghong Guo, Qiqi Gao, He Ye, and Yanlin Wang. 2024. Memo- rybank: Enhancing large language models with long-term memory. InProceedings of the AAAI conference on artificial intelligence, Vol. 38. 19724–19731

    2024

  27. [27]

    turn on the living room light

    Yutao Zhu, Huaying Yuan, Shuting Wang, Jiongnan Liu, Wenhan Liu, Chenlong Deng, Haonan Chen, Zheng Liu, Zhicheng Dou, and Ji-Rong Wen. 2025. Large language models for information retrieval: A survey.ACM Transactions on Information Systems44, 1 (2025), 1–54. Eu Jin Lim, Zhaoxing Li, and Sebastian Stein A Intent Classification Prompt Intent Classification P...

    2025