{"id":"3456515f-07ff-4466-b5d8-963bfa1457ec","arxiv_id":"2412.09289","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A quantized transformer and a compact Mamba model can classify indoor location with moderate accuracy within 32-64 KB model sizes, but on-device RAM usage is not measured.","lead":"This paper tests whether transformer and Mamba neural networks can be shrunk to run indoor location tracking on tiny, low-power microcontrollers. It reports that a quantized transformer fits under 64 KB and a compact Mamba model works under 32 KB, using public radio-signal datasets.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"RAM-feasibility claims are based on static model size, not measured on-device memory; the paper itself defers hardware validation, so the 32/64 KB claims remain unverified.","rationale":"The reader's weakest assumption correctly identifies the load-bearing concern: static model size is treated as equivalent to fitting within an MCU's RAM budget, but the paper never measures on-device memory. I agree with this assessment and with the CONDITIONAL verdict. The paper is honest about the gap—it explicitly lists profiling and testing as essential and defers physical-device validation to future work—but the abstract makes unconditional feasibility claims ('performs well within a 64 KB RAM constraint', 'under even tighter constraints, such as a 32 KB of RAM'), so the central claim is not yet established. The strongest support in the paper is the systematic empirical comparison and the reproducible code release, which make the F1 and size measurements credible as static-model comparisons. However, because the headline claim is about RAM feasibility and not just model size, the missing on-device profiling is a genuine correctness risk rather than a minor omission. The concrete test I propose would settle the issue by measuring peak RAM on real or faithfully simulated MCU hardware. If the measurements confirm the budgets, the claim stands; if not, the abstract's feasibility statements must be revised. Since the reader already conditioned acceptance on this issue, the verdict should remain CONDITIONAL (i.e., UNCHANGED).","tokens_in":18717,"tokens_out":3263,"duration_ms":37048,"concrete_test":"Compile the quantized MDCSA H16L1 and Mamba H8L1/H16L1 checkpoints into TensorFlow Lite Micro or CMSIS-NN and run them on an STM32L4 (64 KB RAM) target with the in-home input shape (4 s x 8 APs) and on a 32 KB-class target with the smallest Mamba variants; additionally run the UJIIndoorLoc input shape (520 APs) through the 44 KB Mamba model. Measure peak RAM via the interpreter arena size and linker map, and compare against the static 'Size (KB)' values. The abstract's RAM claims hold only if peak RAM, including activations and framework overhead, is actually at or below 64 KB and 32 KB respectively.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The abstract's central claims—quantized MDCSA fits 64 KB RAM and Mamba fits 32 KB—are supported only by the 'Size (KB)' columns in Tables 4-8. The paper never specifies how these sizes are computed; they appear to be parameter-count-based static checkpoint sizes. Actual MCU RAM usage additionally includes activation tensors, CRF/Viterbi buffers, Mamba selective-SSM scan state, input windows, and the inference runtime/interpreter arena, none of which are measured. The paper's own 'Complexity and Practical Challenges of Model Deployment' section states that 'profiling and testing are essential to evaluate memory usage, inference time, and power consumption on the target hardware' and concludes with 'Future work will include validating these models on physical edge devices.' Moreover, the 'static quantization' method used is LLM.int8(), a GPU-oriented mixed FP16/int8 scheme with outlier handling; it is unclear whether this maps to integer-only MCU kernels or a deployable MCU runtime. Thus the 'within 64 KB RAM' and 'within 32 KB RAM' statements are estimates based on a proxy, not demonstrated measurements. This does not invalidate the comparative F1 results, but it leaves the headline deployment claim unverified.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper develops TinyML versions of a transformer-based model (MDCSA) and a Mamba-based state-space model for RSSI indoor localisation, targeting low-power MCUs with 32-64 KB memory budgets. It applies post-training quantization (a claimed static LLM.int8() scheme and PyTorch dynamic quantization), knowledge distillation, and a hybrid of both, sweeping hidden size and layer count on four in-home BLE houses and on UJIIndoorLoc. The main reported results are F1, accuracy, and model size in KB for each configuration; the central claim is that the quantized MDCSA model fits in 64 KB RAM and the compact Mamba model fits in 32 KB RAM without compression.","tokens_in":18901,"tokens_out":6243,"duration_ms":67013,"significance":"If the deployment claim were substantiated, the paper would make a useful empirical contribution: it is, to my knowledge, the first systematic comparison of Mamba versus transformer architectures for indoor localisation under TinyML constraints, it covers two public datasets, and it reports a consistent sweep of model sizes with a clear comparison of quantization, distillation, and their combination. The repository link is a welcome element. However, the headline feasibility claim is currently supported only by static model size, not by measured memory on hardware, so the significance remains conditional; the comparative F1 results are useful, but the deployment claim needs either direct measurement or substantial qualification.","major_comments":[{"comment":"The paper's central claim that the quantized transformer 'performs well within a 64 KB RAM constraint' and that Mamba 'has strong performance under ... 32 KB of RAM' is not established by the reported evidence. The only support is the 'Size (KB)' columns in Tables 4-8, but no definition of this size is given and no on-device measurement is reported. Static checkpoint size omits activation tensors, CRF/Viterbi buffers, the Mamba selective-SSM scan state, the input window, and the inference interpreter or runtime arena. The paper's own 'Complexity and Practical Challenges of Model Deployment' section states that 'profiling and testing are essential to evaluate memory usage, inference time, and power consumption on the target hardware' and concludes that 'Future work will include validating these models on physical edge devices.' The RAM feasibility claims should be replaced by 'model storage size' claims, or supported by actual peak-memory measurements on a representative MCU with a deployment framework.","section":"Abstract; Tables 4-8; 'Complexity and Practical Challenges of Model Deployment'"},{"comment":"The method used for the 'static quant' columns is LLM.int8(), cited as Dettmers et al. (2022); this is a mixed FP16/int8 outlier-decomposition inference method developed for GPU transformer workloads, not a static integer-only quantization scheme for MCU kernels. The paper does not explain how FP16 outlier values, scale, and zero-point are stored, nor how the Mamba selective scan and CRF layers would be mapped to integer or otherwise MCU-compatible operations. Consequently, the 'static quant' size values do not by themselves establish compatibility with low-power MCUs. Please clarify the intended deployment path or adjust the TinyML feasibility claim accordingly.","section":"Methodology, 'Quantization'; 'Limitations of Quantization'"},{"comment":"The model-size metric is not reproducible as reported. The Evaluations section states only that model size is measured in KB and that the number of parameters is reported, without stating whether the value is the FP32 parameter count, the state-dict byte size, the serialized artifact size, or the size after framework conversion. Since the central claims in Tables 4-8 hinge on this number, the exact computation or measurement tool must be specified, and the relationship between this number and actual RAM usage should be stated explicitly.","section":"Evaluations, 'Model Size'"}],"minor_comments":[{"comment":"No random seed or number of repeated runs is reported; F1 differences of 1-2% between configurations should be interpreted with caution, and adding variance estimates or seeds would strengthen the comparisons.","section":"Experimental Setting"},{"comment":"The MDCSA H16L1 static-quant row shows accuracy collapsing from 63.37% to 19.05% while F1 remains around 46%; the text notes the drop, but the magnitude deserves an explanation, such as calibration sensitivity or the interaction of outlier handling with majority classes.","section":"Table 8"},{"comment":"The distillation loss in Equations (4)-(5) uses the teacher's hard class predictions rather than soft probability targets; this is a legitimate choice given the CRF output, but the method should be described as hard-label distillation to avoid confusion with conventional soft-label KD.","section":"Methodology, 'Knowledge Distillation'"},{"comment":"The asterisk notation is confusing: in Table 4, MDCSA H16L1 has a baseline size of 64 KB, which does not strictly exceed a 64 KB constraint, yet it is marked with an asterisk indicating that the size exceeds the memory constraint; the caption wording should be clarified.","section":"Tables 4-7"},{"comment":"MDCSA, used as the 'state-of-the-art' transformer baseline, is the authors' own prior model; including at least one external transformer baseline would make the state-of-the-art claim more convincing.","section":"Related Work"}],"recommendation":"major_revision","confidential_remarks":"The primary concern is that the abstract and conclusion make RAM feasibility claims that the experiments do not measure. This is fixable within the manuscript's scope by either adding real hardware memory measurements or by carefully rewriting the claims as storage-size claims. The paper would also be strengthened by an external transformer baseline and by reporting repeated-run variability. The self-citation of MDCSA is not, by itself, a problem, but the 'state-of-the-art' claim would carry more weight with an independent comparison."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Honest take: this is a solid empirical comparison, but the headline claims about fitting in 32/64 KB of RAM are estimates from static model size, not measurements on any device. The authors admit as much in their deployment section, so the abstract overstates what is shown. The real contributions are: first Mamba-based model for RSSI indoor localisation; a systematic comparison of a transformer (MDCSA) and Mamba under quantization, distillation, and hybrid compression on two public datasets; and a concrete, reproducible finding that static quantization can leave already-tiny models the same size or even larger because of scale/zero-point overhead. That last point is genuinely useful for TinyML practitioners.\n\nWhat the paper does well: the grid over hidden size and layer count is thorough, the two datasets are well chosen—small in-home BLE and large multi-building WiFi—and every configuration reports parameter count and KB size. Code is available. The limitation sections are candid: they acknowledge KD gave only marginal gains, note that their distillation uses hard teacher predictions rather than soft logits, and explicitly flag the quantization overhead issue. This is honest engineering work.\n\nThe soft spots, in proportion. The main one is the RAM-feasibility claim. The 'Size (KB)' columns appear to be parameter-based checkpoint sizes, not on-device memory footprints. Actual MCU RAM includes activation tensors, CRF/Viterbi buffers, Mamba's SSM scan state, and runtime/interpreter arena, none of which are measured. Moreover, LLM.int8() is a GPU-oriented mixed FP16/int8 scheme with outlier handling; it is not obviously deployable to integer-only MCU kernels. So 'within 64 KB RAM' and 'within 32 KB RAM' are unverified proxy statements. The paper says so itself, which is good, but the abstract does not. Second, all results are single runs with no error bars; for the Mamba-vs-transformer comparisons, seed variance matters. Third, because KD distills from hard categorical predictions rather than softened logits, the negative KD result should be read as 'this particular KD variant did not help,' not as a general statement about KD. One table oddity: in Table 8, static quantization of MDCSA H16L1 drops accuracy from 63% to 19% while F1 stays around 46%—worth checking, but not load-bearing.\n\nWho this is for: researchers working on TinyML indoor localisation, or anyone comparing state-space and transformer models under extreme size constraints. It deserves a serious referee: the study is reproducible, the comparisons are sensible, and the findings are useful even if the headline needs rewording.","headline":"A systematically useful TinyML comparison whose RAM-feasibility headline is a proxy, not a measurement.","tokens_in":19474,"tokens_out":3070,"would_cite":true,"duration_ms":30219,"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":"Indoor localisation models can be compressed down to 32–64 KB of microcontroller memory.","keywords":["TinyML","Indoor localisation","Quantization","Knowledge distillation","Mamba","Transformer","Edge devices","Microcontrollers"],"falsifier":"Measure peak RAM usage on an actual low-power MCU with 64 KB of RAM while running the quantized MDCSA H16L1 model, and on a 32 KB MCU while running the Mamba H8L1 and H16L1 models on the same in-home and UJIIndoorLoc test sets; if peak usage exceeds the respective budgets during inference, the feasibility claim is falsified.","tokens_in":18482,"feed_emoji":"📍","tokens_out":6686,"duration_ms":54308,"temperature":0.7,"pith_summary":"The paper sets out to show that advanced sequence models for indoor localisation can be compressed enough to run on low-power microcontrollers with only 32–64 KB of RAM. It takes a state-of-the-art transformer-based model (MDCSA), a compact Mamba state-space model, and applies post-training quantization plus knowledge distillation to shrink them. The central claim is that the quantized transformer performs well within a 64 KB budget for in-home room-level localisation, while a small Mamba model does well under a 32 KB budget without needing compression. On a large multi-building Wi-Fi dataset, the Mamba model becomes the better choice under 64 KB. The authors argue this makes on-device indoor localisation viable for wearable health monitoring, with lower latency and better privacy than sending data to a server.","feed_headline":"Indoor localisation shrinks to fit 32–64 KB microcontrollers","feed_subtitle":"A quantized transformer handles home-level room tracking at 64 KB; a compact Mamba model works down to 32 KB.","key_machinery":"The machinery is the size–performance trade-off curve obtained by varying hidden size $H$ and number of layers $L$ for two architectures: the Multihead Dual Convolutional Self-Attention (MDCSA) transformer, and a Mamba structured state-space model with linear embedding and a conditional random field (CRF) classification head. Post-training quantization — both static LLM.int8() and dynamic weight-only quantization — converts FP32 linear layers to int8, and knowledge distillation trains small students against the best-performing teacher using a weighted cross-entropy loss with $\\alpha=0.1$. The model size in kilobytes is the quantity that ties the whole argument together: it is what must fall below the 64 KB or 32 KB memory budget while the F1 score stays competitive.","core_discovery":"On the paper's own terms, the discovery is a sizing map: sweeping hidden size and layer count produces transformer and Mamba models whose static sizes land under 64 KB and 32 KB, and post-training quantization brings the best transformer variant (MDCSA H16L1) down to about 44 KB with F1 scores of 73.84–84.36 across four homes, while a one-layer Mamba with hidden size 8 or 16 fits in 12–21 KB and outperforms the transformer in the stricter 32 KB band. For the larger UJIIndoorLoc dataset, the one-layer Mamba with hidden size 8 and knowledge distillation is the strongest model under 64 KB (44 KB, F1 64%), while the transformer does not fit even after quantization. The paper also reports that quantization can inflate the size of already-tiny models because of the 32-bit scale and zero-point overhead, and that simple output-level knowledge distillation yields only marginal gains. The concluding recommendation is architecture-dependent: transformer with quantization for 64 KB in-home use, unquantized compact Mamba for 32 KB in-home use, and Mamba with distillation for large-building localisation under 64 KB.","pith_inferences":["Editorial extension: the 32–64 KB numbers are static model sizes, not measured on-device RAM; validating the same models on a physical MCU (with activations and runtime buffers) would be the direct test of the feasibility claim the paper leaves for future work.","Editorial extension: on datasets with hundreds of access points, the input dimension itself may become the dominant memory cost, so input-side reduction (AP selection or feature compression) could push Mamba below 32 KB for large buildings — a path the paper does not explore.","Editorial extension: the reported quantization overhead suggests a general rule for TinyML: below some model-size threshold, architecture shrinkage beats precision reduction, and the threshold can be identified by comparing the cost of scale/zero-point parameters against the weight savings.","Editorial extension: the finding that final-layer distillation helps less than expected points to layer-wise or feature-based distillation as a testable follow-up, and the paper itself names this direction."],"forward_implications":["If the claims hold, a wearable with 64 KB of RAM can run room-level localisation in homes using a quantized transformer, keeping location data on the device.","Under a 32 KB budget, a compact Mamba model (hidden size 8 or 16, one layer) is the recommended choice for in-home use, without requiring quantization.","For multi-building, high-dimensional Wi-Fi settings, the Mamba architecture with knowledge distillation is the viable option under 64 KB; the transformer baseline is not.","Quantization should be applied selectively: on models already under roughly 12 KB it can increase size rather than reduce it.","Simple final-layer knowledge distillation does not by itself deliver large gains, so its value in this setting is limited to cases like the large dataset where it nudges performance."],"supporting_citations":[{"why":"Supplies the MDCSA transformer model, the state-of-the-art baseline the paper compresses to under 64 KB.","marker":"[11]"},{"why":"Provides the Mamba selective state-space architecture used to build the compact alternative model.","marker":"[27]"},{"why":"Supplies the LLM.int8() static quantization method used to compress the linear layers.","marker":"[32]"},{"why":"Provides the knowledge distillation framework (teacher–student, weighted cross-entropy loss) used for the student models.","marker":"[9]"},{"why":"Provides the in-home BLE RSSI dataset used for the four-house evaluation.","marker":"[29]"},{"why":"Provides the UJIIndoorLoc multi-building Wi-Fi dataset used for the large-building evaluation.","marker":"[30]"}],"fun_headline_variants":["TinyML indoor tracking fits in 32KB with Mamba, 44KB with transformer","Mamba model outperforms transformer under 32KB for indoor localisation","Quantized transformer squeezes indoor localisation into 44KB","Edge indoor localisation: transformer at 44KB, Mamba at 12-21KB","Transformer and Mamba models fit indoor localisation on MCUs"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The claim that the models fit in 32–64 KB of RAM rests on comparing static model size in kilobytes against the MCU's RAM budget, without measuring what the device actually consumes at runtime (activations, buffers, framework overhead), and the paper's own future-work section says validation on physical devices remains to be done.","fun_headline_variants_meta":{"raw":{"variants":["TinyML indoor tracking fits in 32KB with Mamba, 44KB with transformer","Mamba model outperforms transformer under 32KB for indoor localisation","Quantized transformer squeezes indoor localisation into 44KB","Edge indoor localisation: transformer at 44KB, Mamba at 12-21KB","Transformer and Mamba models fit indoor localisation on MCUs"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000365,"raw_usage":{"total_tokens":2034,"prompt_tokens":1088,"completion_tokens":946,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":704,"completion_tokens_details":{"reasoning_tokens":844}},"tokens_in":704,"tokens_out":946,"duration_ms":7647,"temperature":1.0,"reasoning_tokens":844,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T17:05:11.094751+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure peak RAM usage on an actual low-power MCU with 64 KB of RAM while running the quantized MDCSA H16L1 model, and on a 32 KB MCU while running the Mamba H8L1 and H16L1 models on the same in-home and UJIIndoorLoc test sets; if peak usage exceeds the respective budgets during inference, the feasibility claim is falsified.","supporting_citations":[{"cited_title":"Multimodal indoor localisation in parkinson’s disease for detecting medication use: Observational pilot study in a free- living setting,","cited_arxiv_id":null,"evidence_quote":"Supplies the MDCSA transformer model, the state-of-the-art baseline the paper compresses to under 64 KB."},{"cited_title":"Gpt3. int8 (): 8-bit matrix multiplication for transformers at scale,","cited_arxiv_id":null,"evidence_quote":"Supplies the LLM.int8() static quantization method used to compress the linear layers."},{"cited_title":"Residential wearable rssi and accelerometer measurements with detailed location annotations,","cited_arxiv_id":null,"evidence_quote":"Provides the in-home BLE RSSI dataset used for the four-house evaluation."},{"cited_title":"Ujiindoorloc: A new multi-building and multi-floor database for wlan fingerprint-based indoor localization problems,","cited_arxiv_id":null,"evidence_quote":"Provides the UJIIndoorLoc multi-building Wi-Fi dataset used for the large-building evaluation."}],"review_version":1}