On-Device LLM for Context-Aware Wi-Fi Roaming

Ju-Hyung Lee; Klaus Doppler; Yanqing Lu

REVIEW 4 major objections 5 minor 12 references

On-Device LLM for Context-Aware Wi-Fi Roaming

T0 review · 4 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 The paper claims the first on-device LLM that reasons in the application layer and issues actions in the PHY/MAC stack, using a 3.2 GB quantized Llama-3.1-8B to balance Wi-Fi roaming stability and signal quality better than fixed…

desk verdict The on-device threshold-adjustment demo is a credible modest result, but the 'real-time actions in the PHY/MAC stack' headline is contradicted by the paper's own latency numbers. read the letter →

arxiv 2505.04174 v2 pith:YOHORHRF submitted 2025-05-07 cs.LG cs.AIcs.NIeess.SP

Ju-Hyung Lee , Yanqing Lu , Klaus Doppler This is my paper · ORCID

classification cs.LGcs.AIcs.NIeess.SP

keywords on-deviceLLMWi-Firoamingcross-layerwirelesscontroldynamicthresholdadjustmentBSSIDselectionquantizationLoRAfine-tuningchain-of-thoughtprompting

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 tries to establish that a compressed on-device large language model, running at the application layer, can manage Wi-Fi roaming decisions that normally live in the PHY/MAC stack, replacing fixed RSSI thresholds and heuristic handover rules. Concretely, it claims the LLM can select the best access point (BSSID) from structured context prompts (Task 1) and can periodically adjust the roaming trigger threshold on-device (Task 2), and that on indoor and outdoor logs this beats legacy heuristics and a PPO baseline on the roaming-stability/signal-quality tradeoff. The realistic core is the second task, because the paper itself reports LLM inference taking seconds versus the 10–100 ms budget for AP selection, so the near-real-time claim in Task 1 is presented as a future prospect rather than a demonstrated achievement. If the paper is right, edge devices can carry an LLM that reads context such as location and time and tunes lower-layer wireless parameters without retraining per environment.

What carries the argument

The load-bearing mechanism is the structured prompt as the interface between lower-layer Wi-Fi measurements and the LLM's application-layer reasoning. Chain-of-thought prompting with a few labeled examples turns raw RSSI-BSSID scans plus situational context into an interpretable decision trace; LoRA fine-tuning adapts the pretrained model to Wi-Fi-specific cues; DPO or ORPO aligns decisions with preference data; and Q2K quantization compresses the model to 2.56 bits per parameter, fitting the 16 GB unified memory of the test laptop. The control loop runs the LLM every 30 seconds to set the roaming threshold, so the fast per-scan decisions remain rule-based while the slow, context-aware reasoning loop adjusts the rules.

What would settle it

Measure the end-to-end latency on the target edge device, not a server GPU, from RSSI scan to LLM output for Task 1; if the median latency exceeds 100 ms under realistic load, the claim of real-time PHY/MAC actions is unsupported. Additionally, run a no-context version of the same prompt on the same test data: if the context-augmented prompt does not consistently beat the no-context baseline on the handover/RSSI tradeoff across indoor and outdoor environments, the claimed benefit of contextual reasoning fails.

Watch

Extended reading notes

Core claim

In the paper's own terms, the discovery is that a quantized Llama-3.1-8B LLM, trimmed to 3.2 GB by Q2K quantization, can run on a consumer-class Apple M-series laptop and, given structured prompts of RSSI values, location, time, and battery state, make roaming decisions that balance average signal strength against handover count better than fixed-threshold legacy logic, a random heuristic, and a PPO agent. For Task 1 (choosing which BSSID to roam to), the LLM reduced handovers to 93 versus 100 for Legacy and 107 for PPO while keeping average RSSI close to Legacy, though it did not match the global optimum. For Task 2 (when to roam), the LLM adjusts the roaming threshold every 30 seconds and outperforms any single fixed threshold across indoor and outdoor scenarios, with Q2K quantization cutting model size by about 2.7 times relative to Q8 with negligible loss in decision quality. The paper explicitly acknowledges that near-real-time BSSID selection within 10–100 ms is not met by current inference times, and positions the demonstrated on-device use case as the slower threshold-adjustment task.

Load-bearing premise

The paper's central 'real-time' claim rests on the premise that an on-device LLM can make a BSSID-selection decision within the 10–100 ms roaming budget, and the paper's own latency measurements (seconds on an A100 GPU) show that premise is not yet met.

Editorial extensions

If this is right

If a 3.2 GB on-device LLM can tune roaming thresholds, then similar slow-cadence wireless control tasks, such as channel selection or power-save policy, become candidates for on-device LLM reasoning without cloud connectivity.
The 30-second threshold-adjustment cadence generalizes into a two-timescale architecture: a slow LLM reasoning loop that periodically reconfigures a fast rule-based control loop, which could extend to other protocol parameters beyond roaming.
Because the LLM can absorb new context signals by extending the prompt format rather than retraining, the approach promises easier adaptation to new buildings, mobility patterns, or user preferences compared with DRL agents that need environment-specific retraining.
The paper's own latency measurements imply that for near-real-time lower-layer decisions, on-device LLMs are not yet ready, so the realistic near-term impact is on control tasks with decision intervals of tens of seconds or more.
The interpretable chain-of-thought reasoning output could provide a diagnostic trace for network operators, showing why a particular handover or threshold change was made.

Reading between the lines

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

If future edge AI accelerators close the 10–100 ms latency gap, the same prompt-and-quantization pipeline could be applied to other PHY/MAC control loops with natural slow cadences, such as beam management in mmWave systems or energy-saving mode selection in IoT devices, which the paper does not discuss.
The 30-second threshold cadence suggests a broader design pattern: pairing a slow, context-aware LLM with a fast, deterministic feedback loop. A testable extension would be to vary the cadence with mobility speed and measure where the LLM's benefit over a fixed threshold disappears.
The paper's context ablation shows that adding all available context is not always optimal (146 handovers with all context versus 142 with Time plus Battery), which implies that prompt feature selection, not just model capacity, drives performance; an editor-level inference is that attentional or learned feature gating could further improve the results.
If the Q2K result holds at scale, 2-bit quantization on edge LLMs could make 'LLM as MAC configurator' a realistic product pattern even before LLM inference itself becomes millisecond-fast, because the slow control loop hides the inference latency.

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.

Desk Editor's Note

read the letter

Colleague,

You should know two things about this one. The useful half is the on-device demo, not the real-time claim. The paper overreaches in its abstract and introduction; if you read only Section 4 and Appendix A, you get a plausible result: a compressed 8B LLM, Q2K-quantized, runs on a MacBook and adjusts the Wi-Fi roaming threshold every 30 seconds with reasonable effect. If you read Section 3, you get a BSSID-selection task that the authors themselves admit cannot meet the 10–100 ms budget—their Fig 5 shows seconds on an A100.

What is genuinely new is the application: as far as I can tell from their references, this is the first on-device LLM pointed at PHY/MAC roaming. The engineering pipeline is real: CoT, LoRA, DPO, ORPO, quantization, and an actual on-device demonstration with code on GitHub. They are also honest about the latency problem in Section 3.5, which is more than many such papers do.

The soft spots are real and proportionate. All performance numbers are single-run point estimates; no error bars, no seeds. The dataset is self-collected and not public. Table 2's 'accuracy' is agreement with opt-HO labels that are generated by the same #HO objective used to train and evaluate the model, so high agreement is not an external validation. The 'first' novelty claim is plausible within their reference list but not against a broad literature search. Even within the paper, the context ablation (Table 6) shows that adding all context features slightly hurts performance—that is fine, but it undercuts the 'context-aware' spin.

The bottom line: the narrow result—a slow, on-device threshold adjustment loop that beats fixed thresholds across varied scenarios—is credible. The headline 'real-time actions in the PHY/MAC stack' is not. This is a legitimate application-paper idea that needs a major rewrite to match the evidence.

For peer review: send it out. A serious referee can sort the claims, demand error bars and a public dataset, and push the authors to reframe the latency story. The paper deserves that engagement, but the current version would need heavy revision before publication.

Referee Report

4 major / 5 minor

Summary. The manuscript proposes using a compressed on-device LLM (Llama-3.1-8B with Q2K quantization, 3.2 GB) as a Wi-Fi roaming controller. It formulates two tasks: (T1) selecting the best BSSID from structured context (RSSI, location, time) and (T2) periodically adjusting the scanRSSI roaming threshold every 30 seconds on a MacBook. The authors combine chain-of-thought prompting, LoRA fine-tuning, DPO/ORPO preference optimization, and quantization, and compare against legacy heuristics, opt-HO/opt-RSSI oracles, and PPO on self-collected indoor/outdoor data, reporting that the LLM achieves a better #HO/AvgRSSI trade-off than heuristics and PPO. The paper includes an honest limitations section acknowledging that LLM inference latency is on the order of seconds on an A100 GPU, far above the 10–100 ms roaming budget, but the abstract and introduction nevertheless claim 'real-time actions executed in the PHY/MAC stack.' The demonstrated, narrower contribution is a context-driven threshold-adjustment loop, with the on-device demo in Appendix A covering only Task 2.

Significance. If the empirical claims held, the contribution would be notable as an early demonstration of application-layer LLM control of a lower-layer wireless parameter, with a credible edge-deployment pipeline and a clear comparison framework. The authors deserve credit for releasing code and a demonstration video, for including an explicit limitations section, and for separating the slow threshold-adjustment task from the harder AP-selection task. However, the current evidence does not support the headline real-time cross-layer claim: the paper's own latency measurements contradict it, the on-device demonstration is qualitative rather than quantitative, and most quantitative conclusions rest on single runs without variance information. The contribution is better framed as a feasibility study of a 30-second supervisory threshold-adjustment loop, not as real-time PHY/MAC control.

major comments (4)

[Abstract; §1; §3.5; Fig. 5] The central claim that the on-device LLM issues 'real-time actions executed in the PHY/MAC stack' is contradicted by the paper's own latency data: §3.5 states that Task 1 decisions must be made within 10–100 ms, while Fig. 5 reports inference times on the order of seconds on an A100 GPU and says these 'far exceed' the requirement. Since no on-device latency measurement for Task 1 is provided anywhere in the paper, the load-bearing claim of real-time AP selection is unsupported and should be removed or explicitly replaced by a Task-2-only claim.
[§3.4; Table 2] Table 2's 'accuracy' is the fraction of LLM-selected APs that match opt-HO labels, where opt-HO is generated by exhaustive search minimizing #HO over the same test sequence used for evaluation. Because the fine-tuning objective is likewise handover minimization, the reported 85% agreement partially measures the model's ability to reproduce the authors' own label-generation rule rather than an independent ground truth. The paper should also report agreement with opt-RSSI labels or with a separately labeled human-preference dataset, and should report accuracy with multiple seeds and confidence intervals.
[§3.4; §4.3; Fig. 4; Fig. 7; Tables 3–6] All quantitative results in Figs. 4, 6, 7 and Tables 2–6 are single point estimates with no error bars, no repeated trials, and no seeds, despite the fact that LLM generation is stochastic and LoRA fine-tuning involves random initialization. The claimed improvements over PPO (e.g., 93 vs. 107 handovers in Fig. 4) are small relative to the variability expected from such runs and cannot be assessed without statistical support. The dataset is also self-collected and not released, which further limits reproducibility.
[Appendix A; §4.3; §5 Q3] The on-device demonstration in Appendix A covers only Task 2 and reports no quantitative latency, throughput, memory, or power measurements; the statement that 'comprehensive performance metrics confirm the feasibility' is not supported by the presented evidence. Fig. 5's A100 measurements cannot serve as a proxy because the MacBook M3 has different performance characteristics, so the paper should provide a measured on-device inference-time distribution and resource-utilization numbers for the 30-second loop before claiming practical feasibility.

minor comments (5)

[§4.3] The paragraph beginning 'Effect of Quantization' contains an incomplete/typo-laden opening: 'Q4 K MTable 4 compares...' should be 'Table 4 compares...'.
[§3.4; Table 2] The phrase 'opt-HO-global optimal (exahasutive) results' contains a typo ('exahasutive' should be 'exhaustive'), and 'textsfLegacy' appears as a format artifact in §3.4.
[Fig. 5 caption] The caption should specify that the latency plot is for the A100 GPU only and should explicitly note that no on-device latency is reported elsewhere.
[Table 3] The statement that SFT+DPO 'lowers the error rate' should make the comparison base explicit: it lowers the error rate relative to No FT and SFT, but ORPO has a higher error rate (33.00%), which the text does acknowledge.
[Appendix B] The claim that the LLM's reasoning is 'validated through the subsequent successful handover' conflates a single successful handover with correctness of the reasoning; this should be framed as an anecdotal illustration rather than validation.

Circularity Check

1 steps flagged · score 2.0 of 10

No equation-level circularity; one mild metric-design issue where Table 2 'accuracy' is agreement with the authors' own opt-HO labels.

fitted input called prediction [Section 3.4, Table 2 caption and Section 3.3 baseline definitions]
"Accuracy is the percentage of LLM-selected APs matching test labels generated by opt-HO, which minimizes handovers over the time sequence (Base model: Llama3.1-8B (Grattafiori et al., 2024))."

The 'accuracy' metric is measured against opt-HO, which is defined by the authors as the exhaustive-search oracle that minimizes #HO over the test sequence. Since #HO is also the paper's primary roaming-stability objective and the quantity targeted during post-training (SFT/DPO), this accuracy score reports how well the LLM reproduces the authors' own label generator rather than an independent ground truth. This is a mild form of fitted input called prediction: the labels and the optimized objective are generated from the same criterion. It does not, however, force the central Figure 4 trade-off result, because #HO and AvgRSSI there are computed from the actual simulated association sequences and compared against external baselines (Heuristic, Legacy, PPO).

full rationale

The paper's central claims are not circular in an equation-level sense. The evaluation uses held-out test sequences, external baselines, and metrics (#HO, AvgRSSI) computed from the resulting association traces. The optimal oracles opt-HO and opt-RSSI are benchmarks, not fitted parameters smuggled into the LLM's predictions. The one mild circular component is the Table 2 'accuracy' definition: the LLM's accuracy is measured as agreement with opt-HO labels, and opt-HO is generated by minimizing the same #HO objective that the LLM is post-trained to optimize. This makes the reported accuracy partly a self-consistency check rather than an independent measure of roaming quality, but it is not load-bearing for the main claim, which rests on Figure 4 and the Task 2 on-device comparison. Self-citations (e.g., Lee et al. 2024) are background references only and are not used as load-bearing evidence. The latency discrepancy between the abstract's 'real-time actions' and the paper's own Section 3.5 measurements is a validity/overclaim concern, not a circularity concern, and is explicitly acknowledged in the paper. Overall, the derivation and evaluation are substantially self-contained, with only a minor metric-construction issue.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The ledger shows the central results are not derived; they are produced by fitting a LoRA adapter on self-collected logs and by manually selecting quantization, update interval, prompt format, and context features. The only new artifact is the trained policy itself, which is better described as a fitted software artifact than an invented physical entity. There are no independent falsifiable handles outside the demonstrations, and the dataset and trained adapter are not released.

free parameters (6)

scanRSSI threshold = -70 dBm
Legacy trigger used as baseline and as input to Task 2; chosen by hand, not learned.
Threshold adjustment interval = 30 s
Selected from a 10-300 s sweep in Figure 6 as the practical middle ground; directly changes #HO and AvgRSSI.
LoRA rank = 128
Selected from the Table 2 sweep (rank 32 vs 128); affects accuracy and memory of all post-training results.
Quantization scheme = Q2K, 2.56 bits per weight
Selected from the Table 4 sweep; yields a 3.2 GB model footprint used in all Task 2 deployments.
Context feature set = Location + Time (no battery)
Chosen after the Table 6 ablation; including all context worsens results, so the selection is manual.
LoRA adapter weights = Not released
The fitted policy itself; no checkpoint or hash is provided, so the trained behavior is not independently inspectable.

assumptions (4)

domain assumption Roaming can be correctly framed as prompt-conditioned BSSID or threshold selection.
Both tasks assume RSSI, location, time, and battery in a text prompt determine the right handover decision; this is stated in Sections 3.2 and 4.2.
domain assumption opt-HO labels generated by exhaustive search over the test sequence are a valid supervised target.
Table 2 defines accuracy as agreement with opt-HO labels that minimize #HO; if minimizing handovers is not the true objective, the accuracy and SFT/DPO training signal lose meaning.
domain assumption The self-collected indoor and outdoor logs represent the dynamic mobile Wi-Fi scenarios of interest.
Appendix D describes ten sessions at a campus and a corridor but gives no external validation or public dataset; generalization to automotive or D2D mobile-AP cases is asserted, not tested.
domain assumption LLM outputs are reliably parseable into executable actions.
Table 3 reports error rates up to 33% for invalid AP selections, so some fallback mechanism is implicitly required; the paper does not describe how invalid outputs are handled.

how reviews work

0 comments

Cite this review

Pith. "Pith review of On-Device LLM for Context-Aware Wi-Fi Roaming." pith.science (2026). https://pith.science/paper/YOHORHRF

@misc{pith2026250504174,
  author       = {Pith},
  title        = {Pith review of: On-Device LLM for Context-Aware Wi-Fi Roaming},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YOHORHRF}},
  note         = {Machine review of arXiv:2505.04174}
}

read the original abstract

Roaming in Wireless LAN (Wi-Fi) is a critical yet challenging task for maintaining seamless connectivity in dynamic mobile environments. Conventional threshold-based or heuristic schemes often fail, leading to either sticky or excessive handovers. We introduce the first cross-layer use of an on-device large language model (LLM): high-level reasoning in the application layer that issues real-time actions executed in the PHY/MAC stack. The LLM addresses two tasks: (i) context-aware AP selection, where structured prompts fuse environmental cues (e.g., location, time) to choose the best BSSID; and (ii) dynamic threshold adjustment, where the model adaptively decides when to roam. To satisfy the tight latency and resource budgets of edge hardware, we apply a suite of optimizations-chain-of-thought prompting, parameter-efficient fine-tuning, and quantization. Experiments on indoor and outdoor datasets show that our approach surpasses legacy heuristics and DRL baselines, achieving a strong balance between roaming stability and signal quality. These findings underscore the promise of application-layer LLM reasoning for lower-layer wireless control in future edge systems.

Figures

Figures reproduced from arXiv: 2505.04174 by the authors.

**Figure 1.** Cross-layer control via on-device LLM: Rule-based handover (legacy) vs. on-device LLM context-aware handover. all feature mobility on both the client and AP sides, creating highly dynamic topologies that demand instantaneous, context-aware decisions at the wireless edge. Despite decades of refinement, current roaming mechanisms still rely heavily on threshold-based triggers—such as scanning when RSSI drops below −… view at source ↗

**Figure 2.** Common failure modes of rule-based Wi-Fi roaming. Despite their simplicity, fixed RSSI thresholds and rulebased roaming logic frequently lead to suboptimal outcomes. As shown in [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** Impact of prompt engineering. Effect of Post-Training. To reduce computational overhead without compromising accuracy, we adopt PEFT via LoRA. As shown in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Figures from the paper (7 more)

**Figure 4.** Figure 4: Comparison of # HO (left) and AvgRSSI (right) for each method for best BSSID selection (Task 1). 3.5. Limitations For Task (1) BSSID Selection, the STA must make roaming decisions within 10–100 ms after a scan to meet near-realtime (near-RT) requirements. However, as …

**Figure 5.** Figure 5: Average LLM inference time (A100 GPU). 4. Task (2): Online Roaming Optimization by On-Device LLM Having established that an LLM can effectively choose APs, we now adapt it to the challenge of deciding when to roam in on-device setup. Our second objective is to evaluate…

**Figure 6.** Figure 6: Frequency of threshold adjustments and its effect on performance. Comparison on On-Device: Legacy vs. LLM. Finally, we compare our on-device LLM against four fixed thresholds (−50, −60, −70, −80 dBm) in on-device setup. As illustrated in [PITH_FULL_IMAGE:figures/ful…

**Figure 7.** Figure 7: Legacy vs. Context-Aware On-Device LLM: Comparison of # HO (top) and AvgRSSI (bottom) for each method for dynamic threshold selection (Task 2) in various roaming scenarios. 5. Discussion In conclusion, this section discusses key insights and practical considerations f…

**Figure 8.** Figure 8: Real-world demonstration of our on-device LLM for Task 2 (dynamic threshold adjustment). In each setting the laptop (Macbook Pro) first senses RSSI and context information (left), then the LLM reasons and outputs an adaptive roaming threshold (right), validating practi…

**Figure 9.** Figure 9: Reasoning output of the on-device LLM during an indoor roaming scenario. The model assesses AP candidates by eliminating those with declining RSSI values and prioritizes APs with stable signal strength, ultimately selecting BSSID 34:3A:20:79:C8:B2 for handover. In the …

**Figure 10.** Figure 10: Data-collection environments used for on-device LLM evaluation. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_10.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

12 extracted references · 2 canonical work pages

[3]

Hong, J., Lee, N., and Thorne, J

URL https:// arxiv.org/abs/2407.21783. Hong, J., Lee, N., and Thorne, J. ORPO: Mono- lithic preference optimization without reference model. arXiv preprint arXiv:2403.07691 [cs.CL] ,

arXiv
[4]

URL https://arxiv.org/abs/2403.07691. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., et al. LoRA: Low-rank adaptation of large language models. In Inter- national Conference on Learning Representations,

arXiv
[6]

Irpan, A., Herzog, A., Toshev, A

URL https:// arxiv.org/abs/2201.07207. Irpan, A., Herzog, A., Toshev, A. T., Zeng, A., et al. Do As I Can, Not As I Say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 [cs.RO],

arXiv
[7]

Lacava, A., Polese, M., Sivaraj, R., Soundrarajan, R., et al

URL https://arxiv.org/abs/ 2204.01691. Lacava, A., Polese, M., Sivaraj, R., Soundrarajan, R., et al. Programmable and customized intelligence for traffic steering in 5G networks using open RAN architectures. IEEE Trans. on Mobile Computing , 23(4):2882–2897,

arXiv
[8]

Lee, J.-H., Park, C., Park, S., and Molisch, A

doi: 10.1109/TMC.2023.3266642. Lee, J.-H., Park, C., Park, S., and Molisch, A. F. Han- dover protocol learning for LEO satellite networks: Ac- cess delay and collision minimization. IEEE Trans. on Wireless Commun. , 23(7):7624–7637,

arXiv 2023
[9]

Rafailov, R., Sharma, A., Mitchell, E., Manning, C

doi: 10.1109/TWC.2023.3342975. Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., et al. Direct preference optimization: Your language model is secretly a reward model. Advances in Neu- ral Information Processing Systems , 36:53728–53741,

arXiv 2023
[13]

Machine Learning & Wi-Fi: Unveiling the Path Towards AI/ML-Native IEEE 802.11 Networks

URL https:// arxiv.org/abs/2405.11504. Wu, D., Wang, X., Qiao, Y ., Wang, Z., et al. NetLLM: Adapting large language models for net- working. In Pro. ACM SIGCOMM 2024 , pp. 661–678,

work page Pith review arXiv 2024
[14]

1145/3651890.3672268

URL https://doi.org/10. 1145/3651890.3672268. Xu, J., Li, Z., Chen, W., Wang, Q., Gao, X., Cai, Q., and Ling, Z. On-device language models: A comprehensive review. arXiv preprint arXiv:2409.00088 [cs.CL],

arXiv

Show all 12 references

[15]

9 On-Device LLM for Context-Aware Wi-Fi Roaming A

URL https://arxiv.org/abs/2409.00088. 9 On-Device LLM for Context-Aware Wi-Fi Roaming A. Real-World Demonstration To validate on-device deployment, we ran our LLM-based roaming agent on a MacBook Pro (Apple M-series, 16 GB RAM, macOS). Figure 8 captures key moments from a live...

arXiv
[16]

Reasoning output of the on-device LLM during an indoor roaming scenario. The model assesses AP candidates by eliminating those with declining RSSI values and prioritizes APs with stable signal strength, ultimately selecting BSSID 34:3A:20:79:C8:B2 for handover. In the real-wor...

1944
[2017]

Wei, J., Wang, X., Schuurmans, D., Bosma, M., et al

URL https://arxiv.org/abs/1707.06347. Wei, J., Wang, X., Schuurmans, D., Bosma, M., et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information pro- cessing systems, 35:24824–24837,

arXiv
[2024]

Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., et al

URL https://arxiv.org/abs/ 2412.08905. Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., et al. The Llama 3 Herd of Models. arXiv preprint arXiv:2407.21783 [cs.AI] ,

arXiv

Pith tools

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

[1] [3]

Hong, J., Lee, N., and Thorne, J

URL https:// arxiv.org/abs/2407.21783. Hong, J., Lee, N., and Thorne, J. ORPO: Mono- lithic preference optimization without reference model. arXiv preprint arXiv:2403.07691 [cs.CL] ,

arXiv

[2] [4]

URL https://arxiv.org/abs/2403.07691. Hu, E. J., Shen, Y ., Wallis, P., Allen-Zhu, Z., et al. LoRA: Low-rank adaptation of large language models. In Inter- national Conference on Learning Representations,

arXiv

[3] [6]

Irpan, A., Herzog, A., Toshev, A

URL https:// arxiv.org/abs/2201.07207. Irpan, A., Herzog, A., Toshev, A. T., Zeng, A., et al. Do As I Can, Not As I Say: Grounding language in robotic affordances. arXiv preprint arXiv:2204.01691 [cs.RO],

arXiv

[4] [7]

Lacava, A., Polese, M., Sivaraj, R., Soundrarajan, R., et al

URL https://arxiv.org/abs/ 2204.01691. Lacava, A., Polese, M., Sivaraj, R., Soundrarajan, R., et al. Programmable and customized intelligence for traffic steering in 5G networks using open RAN architectures. IEEE Trans. on Mobile Computing , 23(4):2882–2897,

arXiv

[5] [8]

Lee, J.-H., Park, C., Park, S., and Molisch, A

doi: 10.1109/TMC.2023.3266642. Lee, J.-H., Park, C., Park, S., and Molisch, A. F. Han- dover protocol learning for LEO satellite networks: Ac- cess delay and collision minimization. IEEE Trans. on Wireless Commun. , 23(7):7624–7637,

arXiv 2023

[6] [9]

Rafailov, R., Sharma, A., Mitchell, E., Manning, C

doi: 10.1109/TWC.2023.3342975. Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., et al. Direct preference optimization: Your language model is secretly a reward model. Advances in Neu- ral Information Processing Systems , 36:53728–53741,

arXiv 2023

[7] [13]

Machine Learning & Wi-Fi: Unveiling the Path Towards AI/ML-Native IEEE 802.11 Networks

URL https:// arxiv.org/abs/2405.11504. Wu, D., Wang, X., Qiao, Y ., Wang, Z., et al. NetLLM: Adapting large language models for net- working. In Pro. ACM SIGCOMM 2024 , pp. 661–678,

work page Pith review arXiv 2024

[8] [14]

1145/3651890.3672268

URL https://doi.org/10. 1145/3651890.3672268. Xu, J., Li, Z., Chen, W., Wang, Q., Gao, X., Cai, Q., and Ling, Z. On-device language models: A comprehensive review. arXiv preprint arXiv:2409.00088 [cs.CL],

arXiv