{"id":"f9f41be6-46a7-4fd2-a7a2-041992f1a1f0","arxiv_id":"2608.07427","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"When time-series data is rendered as plots, vision-language models use 3.6 to 10.4x fewer input tokens and 1.8 to 2.5x less measured inference energy than text-only LLMs, with equal or better anomaly detection accuracy.","lead":"This paper measures whether vision-language models that read time-series plots can use less energy than text-only LLMs that read raw numbers, while keeping or improving accuracy. It reports 1.8 to 2.5x lower measured inference energy and better anomaly detection on public and telecom data, but some headline savings figures do not match the paper's own tables.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Accuracy claim confounds fine-tuning with modality: Table IV compares a fine-tuned VLM against a zero-shot text LLM, so the 220.7% precision gain may largely reflect fine-tuning rather than the visual representation.","rationale":"The paper's stated contribution is that VLM encoding is both cheaper and more accurate for time-series anomaly detection. The energy savings are backed by direct NVML measurements and are directionally consistent across models, though the 7.2 MJ/day operational figure is about 100x smaller than a naive multiplication of Table V's per-query savings by the stated query volume, and token counts are inconsistent between tables. The accuracy claim is the load-bearing part because it upgrades the result from 'efficiency trade-off' to 'strict win,' and it rests on a comparison that is not matched by training status. The reader identified the same weakest assumption, and I agree that it is the most serious issue. The proposed test is feasible: fine-tune the text-only model with the same recipe and rerun the evaluation. Until that is done, the central accuracy claim should be treated as unverified, and the CONDITIONAL verdict is appropriate. I do not recommend REJECT because the zero-shot VLM already outperforms the zero-shot text baseline, and the energy reductions are measured directly, so the overall direction may hold even if the headline accuracy margin shrinks.","tokens_in":8850,"tokens_out":2960,"duration_ms":28326,"concrete_test":"Fine-tune the text-only Llama-3.2-90B with the identical LoRA recipe used for the SFT VLM (same training dates, same ground-truth labels, same epochs and hyperparameters), then evaluate on the same 4 test dates with the same per-day temporal protocol. If the fine-tuned text model's precision and F1 approach or exceed the SFT VLM's 0.465/0.464, the accuracy advantage is not due to modality. If the text model remains near 0.145/0.185, the visual-modality claim is supported. Also re-measure Qwen vision tokens at 150 DPI to reconcile Table III (5,442) and Table VI (17,303).","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that VLMs improve accuracy while cutting energy. The energy part is directionally plausible from measured per-query joules, but the accuracy part rests on Table IV, where the SFT VLM (precision 0.465, F1 0.464) is compared against 'LLM text-only' (precision 0.145, F1 0.185) with no indication that the text model was fine-tuned. The paper also reports a zero-shot VLM (precision 0.298) that beats zero-shot text (0.145), suggesting a real modality advantage, but the headline 220.7% figure uses the fine-tuned VLM, conflating training with modality. If a text-only Llama-3.2-90B were fine-tuned with the same LoRA recipe and data, its precision and F1 could rise substantially, shrinking or erasing the claimed accuracy edge. The abstract and conclusion lean on this number to support 'accuracy-superior modality,' so the claim is not robust as stated. A secondary internal inconsistency reinforces the concern: Table IV lists 42,565 tokens for the text LLM while Table III lists 46,101 tokens for the same 8-KPI, 381-point input, and Table VI shows Qwen at 150 DPI using 17,303 vision tokens versus Table III's 5,442, indicating token accounting is not consistently reproduced. These do not invalidate the energy direction, but they worsen confidence in the quantitative claims surrounding the central result.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper argues that encoding numerical time-series as 2D plots and feeding them to vision-language models (VLMs) reduces input token counts by 3.6–10.4x and measured inference energy by 1.8–2.5x relative to text-only LLM inputs, while also improving anomaly-detection accuracy. Experiments cover Llama-3.2-90B-Vision, Qwen2.5-VL-72B, and Pixtral-12B on both a public AWS CloudWatch benchmark and a live telecom KPI dataset, with GPU energy measured via NVML counters. The paper also studies image-resolution compression as an additional energy lever and claims operational savings of about 7.2 MJ/day at a 209-cell monitoring scale.","tokens_in":9094,"tokens_out":5405,"duration_ms":50704,"significance":"If the central claims held, the paper would provide useful empirical evidence that modality choice, not just model size or quantization, is a practical lever for inference-energy reduction in time-series workloads. Strengths include direct per-query energy measurement with hardware energy counters, evaluation across three architecturally distinct VLMs, and the inclusion of both a public benchmark and a real operational telecom dataset. The energy-reduction direction is plausible and consistent with the reported token reductions. However, the paper's headline accuracy claim is weakened by a fine-tuning confound, the reported token and energy numbers contain internal inconsistencies, and the operational-savings arithmetic appears incorrect; these issues need resolution before the quantitative conclusions can be accepted.","major_comments":[{"comment":"The central accuracy claim confounds modality with fine-tuning. Table IV compares a LoRA fine-tuned Llama-3.2-90B-Vision model (precision 0.465, F1 0.464) against a zero-shot text-only LLM (precision 0.145, F1 0.185), and the abstract's 220.7% precision improvement is cited as evidence that VLMs are 'accuracy-superior.' Because fine-tuning and modality are varied simultaneously, this improvement cannot be attributed to the visual representation. The zero-shot VLM result (F1 0.360) partially mitigates the concern, but the paper must either add a text-only Llama fine-tuned with the same LoRA recipe and data, or explicitly restrict the accuracy-superiority claim to the zero-shot comparison.","section":"§IV.C, Table IV"},{"comment":"There are unresolved inconsistencies in the token and energy accounting. Table III reports 46,101 text tokens for the Llama 8-KPI, 381-time-point telecom input, while Table IV lists 42,565 tokens for the same experimental setting. More seriously, Table VI reports Qwen2.5-VL-72B using 17,303 vision tokens at 150 DPI, whereas Table III reports 5,442 vision tokens for the same model and the same 8-KPI input; the corresponding image-mode energies are 8,818 J in Table VI versus 2,397 J in Table V. These discrepancies affect both the token-reduction ratios and the energy-savings claims, and they must be reconciled or clearly explained as different configurations.","section":"§IV.B, Tables III, IV, V, VI"},{"comment":"The claimed operational saving of ~7.2 MJ/day does not match the reported per-query measurements. Using Table V's Llama numbers, the per-query saving is 59,751 J - 23,557 J = 36,194 J. With 209 cells and 96 intervals per day, the paper itself states there are ~20,064 queries/day, giving 36,194 × 20,064 ≈ 726 MJ/day, not 7.2 MJ/day. If a different query count or per-cell basis is intended, it should be stated explicitly; as written, the headline saving is off by two orders of magnitude.","section":"§VI.C, Abstract"},{"comment":"The public-benchmark evaluation does not specify how the 17 candidate signals were reduced to 10 retained signals per model, and the text later refers to 'five signals' in Figure 3 and Table II. The retention rule appears nowhere in the methodology, making it impossible to assess selection bias or to reproduce the mean F1 and J/F1 numbers. The authors should state the exact selection criterion, list the retained signals, and report per-signal results for both modalities.","section":"§IV.A, §V.A (Figures 3–4, Table II)"},{"comment":"The text-only comparison is further confounded by truncation: Table V states that text inputs were truncated to each model's maximum context length (40K for Llama, 32K for Qwen/Pixtral), while Table III shows the full text representation requires 46,101–59,803 tokens. Thus the accuracy comparison in Table IV and the energy comparison in Table V evaluate a truncated text input, not the full time series. A text-only model given the full series through chunking or a longer-context model should be evaluated, or the paper should clearly state that its conclusion applies only to the truncated-text regime.","section":"§IV.B, Table V and §V.C"}],"minor_comments":[{"comment":"The table header 'NVIDIA A100/A6000' does not indicate which hardware was used for each model; the text gives deployment configurations in §IV, but the table itself should list the GPU per row.","section":"Table V header"},{"comment":"Energy values are reported as point estimates over three runs; the paper should report the spread (e.g., standard deviation or min–max) because the 1.8–2.5x savings ratios are central to the conclusions.","section":"§III.B"},{"comment":"The J/F1 metric excludes F1 = 0 cases, but the exclusion rule is only stated parenthetically; please specify how many signals were excluded and whether this affects the reported 20.6x improvement.","section":"§V.A, Table II"},{"comment":"The statement 'mean F1 image 0.72 vs. text 0.49 on public data' is not directly traceable to Figures 3 or Table II; please clarify whether this is an average over all models or a specific configuration.","section":"§VI.A"},{"comment":"The compression experiment uses only Qwen2.5-VL-72B and zero-shot mode; the conclusion about resolution reduction should be framed as architecture-specific rather than general.","section":"§V.F, Table VI"}],"recommendation":"major_revision","confidential_remarks":"The paper's energy-direction result is interesting and potentially useful for practitioners, but the quantitative claims currently contain internal inconsistencies and a confounded accuracy comparison. I would encourage the editor to require the authors to reconcile Table III/V/VI token and energy numbers, correct the 7.2 MJ/day arithmetic, and add a fine-tuned text baseline before considering publication. The paper may also benefit from a data-availability statement, since reproducibility of the public-benchmark selection is currently impossible."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. The energy measurements are the real asset here: direct NVML-based per-query joules for three VLMs versus text-only LLMs on time-series anomaly detection, on both public and telecom data. The direction is plausible and consistent: image modality cuts token count 3.6-10.4x and energy 1.8-2.5x. That is worth engaging with. But the paper's headline accuracy claim is built on a confounded comparison, and the operational-scale number is wrong by two orders of magnitude.\n\nThe strongest new piece is the stacked-subplot multivariate rendering, which extends VLM4TS [9] from univariate to multivariate without losing the token-reduction advantage. The measurement methodology is sound in principle: NVML energy counters, multiple runs, warmup, and per-model native tokenizers. Credit where due.\n\nSoft spots, in order of severity.\n\nFirst, Table IV compares the fine-tuned Llama VLM (precision 0.465) against a zero-shot text-only LLM (0.145) and calls the difference a modality effect. It is at least partly a fine-tuning effect. The zero-shot VLM also beats zero-shot text (0.298 vs 0.145), so the modality advantage is real, but the 220.7% figure in the abstract and conclusion should not be sold without a fine-tuned text baseline. Second, token accounting is inconsistent: Table III gives Qwen 5,442 vision tokens, Table VI gives 17,303 at the same 150 DPI; Table III gives Llama 46,101 text tokens, Table IV gives 42,565. These matter because the token ratios are the basis for the energy-savings argued at scale. Third, the operational estimate is off. If Llama saves roughly 36 kJ per query (59,751 minus 23,557 J) and there are about 20,000 queries/day, that is ~720 MJ/day, not 7.2 MJ. The paper needs to correct this or explain the unit conversion. Fourth, the 10-of-17 signal retention rule is unexplained; stratified retention can hide selection bias.\n\nNone of this kills the central energy-direction claim, which is well supported by the raw measured joules. But the paper overstates accuracy and fumbles the arithmetic in a way that invites distrust. It reads like a conference submission that got in before the numbers were reconciled.\n\nFor a reader working on efficient VLM inference or telecom analytics, the per-query energy tables and the scaling analysis (Fig. 5) are worth a look. I would not cite the paper as-is, but I would send it to a serious referee with a clear brief to require a fine-tuned text baseline, consistent token counts, and a corrected scaling estimate. The underlying idea deserves scrutiny; the current version does not.","headline":"Useful energy measurements undercut by a confounded accuracy claim and a hundredfold arithmetic error; worth peer review but needs major revision.","tokens_in":9674,"tokens_out":5035,"would_cite":false,"duration_ms":42636,"reading_group":"maybe","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 argues that rendering time-series data as 2-D plots instead of text tokens makes vision-language models both cheaper and more accurate for anomaly detection.","keywords":["vision-language models","time-series anomaly detection","inference energy efficiency","token compression","telecom network analytics","multimodal LLM","GPU energy measurement"],"falsifier":"Run the same LoRA fine-tuning recipe on the text-only Llama-3.2-90B and compare precision/F1 on the same telecom test dates; if it matches or beats the SFT VLM (0.465 precision, 0.464 F1), the accuracy-superiority claim fails. Separately, an end-to-end energy measurement that includes image rendering and CPU-GPU transfer could shrink or reverse the 1.8–2.5x reduction if rendering overhead is large.","tokens_in":8616,"feed_emoji":"⚡","tokens_out":8939,"duration_ms":76297,"temperature":0.7,"pith_summary":"The paper tries to establish that for numerical time-series analysis, the cheapest and most accurate way to feed a large language model is not the raw numbers but a picture of them. It measures token counts and GPU energy for three vision-language models processing identical KPI windows as text and as 2-D plots, finding 3.6–10.4x fewer input tokens and 1.8–2.5x less inference energy for the visual form. The same visual encoding also improves anomaly detection, with a fine-tuned Llama-3.2-90B-Vision reaching 220.7% higher precision than its text-only counterpart and F1 0.464 versus 0.185. If correct, this makes image rendering a deployment prerequisite rather than an optimization for high-dimensional telecom monitoring, since 24-KPI text windows exceed common 128K context limits. It also gives edge operators a free lever: dropping image resolution to 75 DPI cuts another 24% of energy without measurable accuracy loss.","feed_headline":"Plots beat text: 2.5x less energy, better anomaly detection","feed_subtitle":"Visual time-series inputs cut inference cost 1.8-2.5x while lifting F1 on telecom and public benchmarks.","key_machinery":"The mechanism is the token-energy mismatch: LLM inference energy scales with input token count, while raw numerical KPI windows tokenize into tens of thousands of floating-point tokens. The paper replaces that representation with 2-D plot images—univariate line charts and vertically stacked subplots sharing a common time axis—so the vision encoder compresses the same signal into 5,442–16,800 visual tokens depending on the architecture. The stacked-subplot composition is motivated as a visual cross-section of the d-dimensional manifold traced by a d-variate time series, preserving inter-variable alignment while allowing attention to co-occurring waveform morphologies. Structured prompts that name spike, hump, and oscillatory-deviation primitives steer the model's attention toward waveform geometry. The argument is carried by precise token accounting per model tokenizer and hardware-counter energy readings per query.","core_discovery":"The paper's central claim is that for numerical time-series workloads, the visual modality dominates the text modality on both cost and accuracy. Across three vision-language architectures, encoding KPI windows as 2-D plots rather than serialized floating-point numbers reduces input tokens by 3.6–10.4x and measured inference energy by 1.8–2.5x per query, with energy savings measured by hardware counters on A6000/A100 GPUs. On the public realAWSCloudwatch benchmark, image inputs raise mean F1 from 0.58–0.66 (text) to 0.70–0.88 (image); on live 4G/5G telecom data, a fine-tuned Llama-3.2-90B-Vision reaches precision 0.465 versus 0.145 for text and F1 0.464 versus 0.185, while beating ARIMA and LSTM by over 144%. The paper further claims Pixtral-12B achieves a 20.6x improvement in energy-normalized efficiency (J/F1: 2,538 image vs 52,356 text), and that at 24 KPIs text inputs exceed typical 128K context windows while visual inputs remain within limits. Together these results are offered as empirical grounding for treating energy as a first-class constraint in AI inference systems.","pith_inferences":["Because the token-compression ratio grows with window length and KPI count, the energy advantage should widen for longer histories; a scaling experiment across window sizes would test this directly.","The accuracy comparison's confound suggests a same-recipe fine-tuned text LLM is the decisive follow-up; until that runs, the 'accuracy-superior' claim is only as strong as the modality-versus-tuning assumption.","The same visual encoding could transfer to other high-dimensional monitoring domains such as financial tick data, industrial sensors, or power-grid telemetry, but the paper only measures telecom and cloudwatch data.","Because the paper measures GPU inference only, full-system accounting including image rasterization, memory transfer, and CPU preprocessing could change the net savings; the reported 1.8–2.5x should be read as a GPU-bound upper bound."],"forward_implications":["At 24 KPIs, text representations exceed the 128K context window of production LLMs, making visual input the only feasible modality without lossy truncation.","Deploying the visual modality at 209 cells monitored every 15 minutes saves about 7.2 MJ/day per model, roughly the daily residential energy of 2.4 US households.","Reducing rendering resolution from 150 to 75 DPI cuts vision tokens by 70% and inference energy by 24% with no measured F1 drop, adding a free tuning lever for edge hardware.","Architecture choice tunes the trade: Llama's fixed 6,404-token budget gives predictable energy, Qwen's dynamic tiling gives maximum compression (10.4x), and Pixtral's patch tokenization gives the best energy-normalized accuracy (J/F1 2,538 vs 52,356)."],"supporting_citations":[{"why":"Establishes that LLM inference energy scales linearly with input token count, the lever the paper pulls.","marker":"[6]"},{"why":"Supplies the visual time-series encoding method and the realAWSCloudwatch public benchmark used for evaluation.","marker":"[9]"},{"why":"Provides prior evidence that visual time-series representations outperform numeric ones for reasoning and anomaly localization.","marker":"[14]"},{"why":"Defines the single-GPU edge memory ceiling that makes text-mode token counts infeasible for deployment.","marker":"[15]"},{"why":"Supplies the hardware-counter-based energy measurement method used for per-query GPU energy readings.","marker":"[16]"},{"why":"Defines the Llama-3.2-90B architecture whose fixed-budget vision encoder is one of the three measured models.","marker":"[17]"},{"why":"Defines the Qwen2.5-VL architecture whose dynamic-resolution patching gives the largest token compression.","marker":"[18]"},{"why":"Defines the Pixtral-12B architecture whose patch tokenization yields the best J/F1 efficiency.","marker":"[19]"},{"why":"Provides the causal-anomaly ground-truth labels for the live telecom KPI evaluation.","marker":"[20]"}],"fun_headline_variants":["Time-series as images: 2.5x less energy, 220% better precision","Visual VLMs: 10x fewer tokens, 2.5x less energy, higher F1","Energy-efficient AI: plots cut inference cost 2.5x, boost accuracy","Plots slash AI energy 2.5x and beat text on anomaly detection"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the accuracy gain is caused by the visual modality itself, because the main comparison pits a fine-tuned vision model against a zero-shot text model and un-fine-tuned ARIMA and LSTM baselines.","fun_headline_variants_meta":{"raw":{"variants":["Time-series as images: 2.5x less energy, 220% better precision","Visual VLMs: 10x fewer tokens, 2.5x less energy, higher F1","Energy-efficient AI: plots cut inference cost 2.5x, boost accuracy","Plots slash AI energy 2.5x and beat text on anomaly detection"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00068,"raw_usage":{"total_tokens":3192,"prompt_tokens":1151,"completion_tokens":2041,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":767,"completion_tokens_details":{"reasoning_tokens":1947}},"tokens_in":767,"tokens_out":2041,"duration_ms":12982,"temperature":1.0,"reasoning_tokens":1947,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T04:47:19.622051+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same LoRA fine-tuning recipe on the text-only Llama-3.2-90B and compare precision/F1 on the same telecom test dates; if it matches or beats the SFT VLM (0.465 precision, 0.464 F1), the accuracy-superiority claim fails. Separately, an end-to-end energy measurement that includes image rendering and CPU-GPU transfer could shrink or reverse the 1.8–2.5x reduction if rendering overhead is large.","supporting_citations":[{"cited_title":"Edge-Deployable LLM Fine-Tuning on a Single GPU for Telecom Network Troubleshooting","cited_arxiv_id":"2607.02523","evidence_quote":"Defines the single-GPU edge memory ceiling that makes text-mode token counts infeasible for deployment."},{"cited_title":"Zeus: Understanding and optimizing GPU energy consumption of DNN training,","cited_arxiv_id":null,"evidence_quote":"Supplies the hardware-counter-based energy measurement method used for per-query GPU energy readings."},{"cited_title":"Causal inter- vention sequence analysis for fault tracking in radio access networks,","cited_arxiv_id":null,"evidence_quote":"Provides the causal-anomaly ground-truth labels for the live telecom KPI evaluation."}],"review_version":1}