{"id":"cb0aa940-eebe-4595-9421-362a628df0a9","arxiv_id":"2505.15835","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"Fine-tuned decoder-only LLaMA models predict WiFi transmitter-receiver distance from raw RSSI, FTM, and CSI tokens, with reported sub-meter median errors, though the evaluation lacks baselines and has split-design concerns.","lead":"This paper fine-tunes LLaMA language models on WiFi signal measurements to predict the distance between transmitters and receivers, reporting sub-meter errors on public and custom datasets. A smart generalist would read it to see whether large language models can replace specialized signal-processing pipelines for indoor positioning.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"CSI few-shot evaluation randomly splits 5-minute per-position recordings, so temporally correlated near-duplicate samples leak into the test set; reported centimeter-level errors reflect memorization rather than localization generalization.","rationale":"I focused on the CSI split because it is the only experimental result that produces the 'centimeter-level precision' claimed in the abstract. The FTM/RSSI results in Table IV report MAEs around 0.9-1.3 meters, which already weakens the 'sub-meter' claim and are not compared against any baseline, but those numbers at least follow the dataset's intended train/test structure. The CSI result, by contrast, is internally suspect: a random split of a 5-minute stationary recording is not a valid evaluation protocol for spatial generalization. The reader's weakest_assumption identifies exactly this leakage, and I agree. There are additional independent problems, including the absence of baselines, the exclusion of misaligned outputs from metrics, and the evaluation of distance rather than position, any of which would also prevent acceptance. Together these issues support the reader's REJECT verdict, so no change is needed. The proposed temporal-block rerun is cheap, uses data already described in the paper, and would settle whether the 6.9 cm MAE reflects genuine generalization or memorization of near-duplicate inputs.","tokens_in":16028,"tokens_out":4081,"duration_ms":44019,"concrete_test":"Reconstruct the ESP32-S2 CSI dataset and re-run the LLaMA-3.2-3B fine-tuning with a temporal block split: for each 1-meter position, use the first ~60 seconds for training, the next ~60 seconds for validation, and the remaining ~3 minutes for testing; also run leave-one-position-out. Report MAE, MSE, R2, and misalignment count. If MAE increases from 6.9 cm to well above 1 meter, or if a simple kNN amplitude fingerprinting baseline matches or beats the LLM under the same split, the random split is confirmed as the source of the reported centimeter-level precision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The headline centimeter-level CSI result depends entirely on the evaluation split in Section V.A. The CSI dataset was collected as roughly 5-minute continuous recordings at fixed 1-meter positions (Section IV.A.1), yet the paper splits it 'by class' into 10% training, 10% validation, and 80% testing. A random per-sample split over a continuous recording places adjacent, highly correlated CSI packets from the same physical location in both training and test sets. With 51,500 test samples, the model can effectively interpolate near-duplicate inputs, so the 6.9 cm MAE reported in Table III is not evidence of generalization to unseen positions or environments. The paper also skips misaligned or unparseable outputs when computing error metrics and provides no non-LLM baseline, but the temporal leakage alone is sufficient to undermine the central claim. If the split were temporal or leave-one-position-out, the error would likely rise substantially.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes WiFiGPT, a system that fine-tunes decoder-only LLaMA-3 models with LoRA adapters to regress transmitter-receiver distance from raw WiFi telemetry (CSI, FTM, RSSI) represented as token sequences. Telemetry arrays are embedded in a prompt template and the model is trained with next-token prediction to output a distance in meters; inference uses greedy decoding for deterministic outputs. Experiments are reported on a custom ESP32 CSI hallway dataset and a public FTM/RSSI dataset covering lecture theater, office, and corridor environments. The headline results are sub-meter MAE for FTM/RSSI (around 0.9-1.3 m) and centimeter-level MAE for CSI (about 6.9 cm), from which the authors conclude that LLM-based localization can match or surpass specialized techniques without handcrafted signal processing or calibration.","tokens_in":16180,"tokens_out":5329,"duration_ms":55910,"significance":"If the reported results held, the paper would support a genuinely unified, schema-less LLM-based approach to WiFi localization across heterogeneous telemetry types and environments. The deterministic greedy-decoding inference, the LoRA-based fine-tuning for efficiency, and the direct use of raw telemetry tokens are sensible design choices, and the use of a public FTM/RSSI dataset is a strength. However, the empirical evidence as presented does not establish the central claims: the CSI evaluation leaks temporally correlated samples between train and test, no baseline algorithms are run on the same data, and the 'without calibration' claim is contradicted by the per-environment fine-tuning protocol. The idea is worth exploring, but the manuscript in its current form does not provide a valid demonstration of the stated contribution.","major_comments":[{"comment":"The CSI evaluation splits a continuous ~5-minute recording made at each fixed 1-meter position randomly by sample (10% train, 10% validation, 80% test). Adjacent CSI packets recorded at the same stationary location are near-duplicates of one another, so the model can effectively memorize input patterns rather than learn a distance mapping. The reported 6.9 cm MAE on 51,500 test samples is therefore not evidence of generalization to unseen positions or environments. The evaluation must be redone with a temporal split (e.g., training on an early segment of each recording and testing on a later segment) or a leave-one-position-out split, with per-position errors reported.","section":"Section V.A, Table III"},{"comment":"The abstract and conclusion claim the system operates 'without handcrafted signal processing or calibration,' but Section III.C.2 states that the model is trained separately for each dataset and environment. No cross-environment or cross-session experiment is reported, so the system has not been shown to be calibration-free; it is fitted per deployment. Demonstrating the 'no calibration' claim would require evaluating an adapter trained in one environment on another environment or on unseen hardware without additional fine-tuning.","section":"Section III.C.2 and Abstract"},{"comment":"No baseline algorithm is evaluated on the same data and splits in the experiments. The statement that WiFiGPT 'matches and often surpasses' state-of-the-art methods is therefore unsupported; it appears to rely on comparisons to previously published numbers obtained with different datasets, protocols, and metrics. At minimum, the authors should report KNN, LGBM, and a simple linear or MLP regressor on the same telemetry features and splits, and for CSI should include a classical method such as SpotFi-style processing or a standard fingerprinting approach.","section":"Section V, Tables III and IV"},{"comment":"The 8B model is labeled 'LLaMA-3.1-8B' in the results tables, but reference [49] is an MLX conversion of DeepHermes-3-Llama-3-8B-Preview, a chat-tuned variant rather than the base LLaMA-3.1-8B model. Comparing this model with LLaMA-3.2-1B and LLaMA-3.2-3B confounds model size with model family and instruction tuning, so the claimed analysis of how accuracy varies with LLaMA model size is not valid as presented.","section":"References [47]-[49], Tables III and IV"},{"comment":"Misaligned or unparseable model outputs are skipped in the error metrics rather than counted as errors. Table III reports six misalignments for the 3B model and two for the 8B model; if these outputs correspond to degenerate or high-error predictions, the reported MAE/MSE values are optimistically biased. The paper should report the metrics with misaligned outputs included (e.g., counted at the maximum possible error) or otherwise demonstrate that their exclusion does not affect the conclusions.","section":"Section IV.B, Table III"}],"minor_comments":[{"comment":"The text says in Section III.C that MAE and MSE are used as custom loss metrics for training, but Section IV.B later states they are auxiliary monitoring tools and are not used for gradient updates; this inconsistency should be resolved and the training objective stated precisely.","section":"Section III.C and Section IV.B"},{"comment":"The references for the base models point to MLX-community checkpoint conversions rather than the original LLaMA-3 model releases; the paper should cite the original model descriptions and state the exact base checkpoints used, especially for the 8B variant.","section":"References [47]-[49]"},{"comment":"The CSI experiment is called 'Few-Shot Learning,' but 10% of 64,375 samples is roughly 6,400 training samples, which is not a few-shot regime; the labeling should be corrected or the training set size stated explicitly.","section":"Section V.A"},{"comment":"The text says 'as shown in Figure 2a/2b' when referring to trilateration geometry, but the cited figure is Figure 3; the cross-reference should be fixed.","section":"Section VI"},{"comment":"The high-percentile column is labeled '100th Iteration' and contains inconsistently formatted entries (e.g., '10.0100' versus '10.05'); the column label and numeric precision should be made uniform.","section":"Table IV"},{"comment":"There are several typos, including 'RToF' in Section II.A, 'course grained' in Section V.C.2, and 'eletrocardiogram' in Section II.B; these should be corrected.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The paper is a preliminary empirical study whose main quantitative claim, centimeter-level CSI accuracy, is invalidated by the random per-sample split of continuous recordings at fixed positions. The absence of any baseline comparison on the same data further weakens the comparative claims, and the per-environment fine-tuning protocol contradicts the 'no calibration' narrative. The novelty claim of being the first decoder-only LLM system for indoor localization is also questionable given existing transformer-based localization work such as LocGPT and CSI-BERT. I would not encourage resubmission in its current form; a major rework with proper evaluation splits, baselines, and cross-environment experiments would be needed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a legitimate new application — fine-tuning a decoder-only LLM on raw WiFi CSI, FTM, and RSSI token sequences for distance regression. That hasn't been done before, as far as I know. The prompt/delimiter design is simple, the greedy decoding choice is reasonable, and the FTM vs RSSI ablation is a useful sanity check. The paper also honestly says it measures distance rather than position and admits per-environment fine-tuning is needed. So the basic idea deserves a look.\n\nThe problem is the evidence. The headline CSI result comes from a few-shot split that is almost certainly leaking. They record about five minutes of CSI at each 1-meter position, then randomly split the samples 10/10/80 into train/validation/test. Adjacent CSI packets from the same continuous recording are near-duplicates, so the test set is full of samples that are effectively the same as training samples. The 6.9 cm MAE in Table III is more likely memorization than generalization. A temporal split — e.g., train on the first part of each recording and test on the last part, or leave-one-position-out — would be the right protocol. Without that, the abstract's \"centimeter-level precision\" claim is unsupported.\n\nThe FTM/RSSI results are more plausible: median errors around or below 1 meter in three environments. But there are no non-LLM baselines anywhere in the paper, so the \"matches and often surpasses conventional approaches\" claim is asserted rather than demonstrated. A simple fingerprinting baseline or a nonlinear regressor like LGBM on the same features would clarify whether the LLM is adding anything. Also, misaligned or unparseable outputs are skipped when computing metrics; the counts are small for CSI but reach 296 in one FTM-only ablation. Excluding them biases the errors downward, and the paper should report both inclusive and exclusive numbers.\n\nThe \"without calibration\" phrase is an overclaim. Since a separate LoRA adapter is trained for each environment and dataset, the reported errors are per-environment fits, not evidence of a single model generalizing across environments. That is a softer issue than the leakage, but it should be corrected in the framing.\n\nCitation pattern is fine: the related work covers the relevant LLM-regression and transformer-localization papers, and the dataset is public. No code or data is released for the custom CSI collection, which limits reproducibility.\n\nBottom line: the idea is plausible and worth a serious referee, but the current evaluation does not support the central claims. I would send it to peer review rather than desk-reject, because the flaws are fixable and the topic is timely. The authors need to redo the CSI split, add baselines, and handle misalignments transparently before the results can be trusted.","headline":"Plausible first use of LoRA-finetuned LLaMA for WiFi telemetry distance regression, but the headline CSI results are leaky and the paper lacks baselines.","tokens_in":16722,"tokens_out":2604,"would_cite":false,"duration_ms":27682,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that a decoder-only large language model fine-tuned with LoRA can map raw WiFi telemetry tokens to distance, reaching sub-meter RSSI/FTM errors and centimeter-level CSI errors without handcrafted signal processing or…","keywords":["indoor localization","WiFi telemetry","decoder-only transformer","LLM regression","CSI","FTM","RSSI","LoRA fine-tuning"],"falsifier":"Re-run the CSI experiment with a temporal split: train on the first minutes of each position's recording and test on the last minutes, or hold out entire positions. Centimeter-level errors that survive this split would confirm spatial generalization, while errors jumping to meter scale would show that the current accuracy came from memorization.","tokens_in":15815,"feed_emoji":"📡","tokens_out":6539,"duration_ms":58375,"temperature":0.7,"pith_summary":"The paper claims that a decoder-only large language model, fine-tuned with LoRA on raw WiFi telemetry serialized as text, can estimate transmitter-receiver distances accurately enough for indoor localization. The reported numbers are sub-meter median errors for RSSI and FTM measurements in corridor, office, and lecture-theatre environments, and centimeter-level mean absolute errors for CSI in a multipath hallway. The point is that one model, without handcrafted signal processing, calibration, data scaling, or a fixed input schema, can handle telemetry from different vendors and environments. If true, this makes LLM-based regression a practical alternative to specialized fingerprinting and ranging pipelines, especially on consumer hardware.","feed_headline":"A fine-tuned LLaMA maps WiFi signals to sub-meter positions","feed_subtitle":"Raw RSSI, FTM, and CSI tokens predict distance directly, with no calibration or handcrafted features.","key_machinery":"The load-bearing object is the prompt-based next-token regression interface: each WiFi packet is formatted as a JSON-style instruction with special delimiters, the telemetry values are plain tokens, and the model must emit a numeric answer in the form '{answer}m' as the final token. LoRA adapters freeze the base weights and train only a small fraction of parameters, while auxiliary MAE and MSE monitors track convergence without changing the underlying cross-entropy objective. Greedy decoding with temperature zero and a fixed seed makes the output deterministic. This machinery lets a language model treat distance as the next word, removing the need for separate encoders or handcrafted features.","core_discovery":"WiFiGPT is presented as a repurposing of a decoder-only LLM for indoor localization. The central discovery is that a causal language model can be re-tooled as a distance regressor by framing distance as the next token: after supervised fine-tuning with LoRA on prompt-formatted telemetry, the model outputs a single numeric value in meters. On the custom CSI hallway dataset, LLaMA-3.2-3B reaches MAE 0.069 m and MSE 0.168 m; on the public RSSI/FTM dataset, median errors stay near or below 1 m across all three environments and model sizes. The paper also shows that the model performs implicit data imputation, since FTM-only and RSSI-only ablation runs still produce usable estimates, with fused features performing best.","pith_inferences":["Editorial inference: the CSI few-shot evaluation is the weakest link, because the random sample-wise split of five-minute recordings at fixed positions likely lets temporally adjacent near-duplicates leak across the train/test boundary, making the centimeter-level numbers an overstatement of true spatial generalization.","Editorial inference: a stronger claim would test cross-environment transfer with a single adapter trained on multiple rooms and evaluated on an unseen layout, since the paper currently fine-tunes separately for each environment.","Editorial inference: the same serialization trick could extend to other single-antenna sensing modalities such as BLE, UWB, or inertial data, because nothing in the architecture depends on WiFi-specific features.","Editorial inference: prompt sensitivity, which the paper lists as a limitation, can be probed directly by paraphrasing the instruction while keeping the telemetry fixed; predicted distances should not change if the model learned geometry rather than surface text."],"forward_implications":["A single prompt-template system can replace separate preprocessing pipelines for RSSI, FTM, and CSI, since all telemetry types are treated as serialized text.","Deployment on resource-constrained devices becomes plausible because LoRA updates only a small fraction of weights and inference is a single autoregressive pass.","The system degrades gracefully under missing features: ablation results show that RSSI-only and FTM-only inputs still yield useful distance estimates.","Practitioners can smooth tail-heavy single-packet errors with a sliding window of 5 to 10 packets at a cost of only a few milliseconds of additional delay."],"supporting_citations":[{"why":"Supplies the transformer architecture, specifically attention, that WiFiGPT reuses for autoregressive next-token prediction.","marker":"[12]"},{"why":"Named as the source of the LLaMA decoder-only base models that are fine-tuned in the experiments.","marker":"[37]"},{"why":"Introduces LoRA, the low-rank adaptation method that makes fine-tuning efficient on consumer hardware.","marker":"[52]"},{"why":"Provides the public WiFi RSSI/FTM dataset covering lecture-theatre, office, and corridor environments.","marker":"[57]"},{"why":"Documents the dynamic-model-switching baseline and the environment layout for the RSS/FTM dataset.","marker":"[59]"},{"why":"Defines the JSONL instruction-tuning format used to structure telemetry prompts and the parseable distance answer.","marker":"[45]"},{"why":"Backs the claim that additional telemetry data improves generalization more than increasing model size.","marker":"[56]"},{"why":"Demonstrates that LLMs can perform regression from in-context examples, motivating the direct fine-tuning approach.","marker":"[43]"}],"fun_headline_variants":["LLM repurposed as WiFi regressor: sub-meter indoor positioning","Fine-tuned LLaMA turns WiFi telemetry into sub-meter positions","WiFiGPT: no calibration or handcrafting for sub-meter positions","Decoder-only GPT regresses distance from WiFi telemetry directly"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The weakest assumption is that randomly splitting the roughly five-minute recordings taken at each fixed position into train and test sets measures spatial generalization rather than the model recognizing near-identical, temporally adjacent samples from the same spot.","fun_headline_variants_meta":{"raw":{"variants":["LLM repurposed as WiFi regressor: sub-meter indoor positioning","Fine-tuned LLaMA turns WiFi telemetry into sub-meter positions","WiFiGPT: no calibration or handcrafting for sub-meter positions","Decoder-only GPT regresses distance from WiFi telemetry directly"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001451,"raw_usage":{"total_tokens":5848,"prompt_tokens":952,"completion_tokens":4896,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":568,"completion_tokens_details":{"reasoning_tokens":4820}},"tokens_in":568,"tokens_out":4896,"duration_ms":34304,"temperature":1.0,"reasoning_tokens":4820,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:51:34.816324+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the CSI experiment with a temporal split: train on the first minutes of each position's recording and test on the last minutes, or hold out entire positions. Centimeter-level errors that survive this split would confirm spatial generalization, while errors jumping to meter scale would show that the current accuracy came from memorization.","supporting_citations":[{"cited_title":"WiFi RSS & RTT dataset with different LOS conditions for indoor positioning,","cited_arxiv_id":null,"evidence_quote":"Provides the public WiFi RSSI/FTM dataset covering lecture-theatre, office, and corridor environments."},{"cited_title":"A wifi rss-rtt indoor positioning system using dynamic model switching algorithm,","cited_arxiv_id":null,"evidence_quote":"Documents the dynamic-model-switching baseline and the environment layout for the RSS/FTM dataset."}],"review_version":1}