acceptodds
Under review as a conference paper at ICLR 2027

EM-Lens: Towards Observation-Level Electromagnetic Signal Understanding via Agentic Zoom-and-Measure

Abstract

Existing EM multimodal large language models (MLLMs) predominantly study isolated, pre-segmented signal instances, whereas real-world receivers continuously observe long streams containing multiple heterogeneous signal events. Bridging this gap toward real-world observation-level EM understanding is impeded by two critical bottlenecks: (1) the limited availability of observation-level EM data, and (2) the perceptual degradation that occurs when models reason directly over entire observations, as prolonged heterogeneous backgrounds dilute target signals. To address data scarcity, we introduce longEM-IQA and longEM-Bench, a training and evaluation suite comprising 11 in-phase/quadrature (I/Q)–language tasks spanning signal localization, identification, parameter estimation, and composite-entity understanding. To mitigate perceptual degradation, we propose EM-Lens, a framework that actively localizes target signal segments within complex observations and extracts their physical parameters via explicit signal-processing tools, while retaining end-to-end perception of individual segments. Furthermore, we devise a hybrid training strategy to reconcile the optimization conflict between active analysis and end-to-end perception. Extensive experiments demonstrate that EM-Lens achieves state-of-the-art performance on longEM-Bench, showing that active acquisition of relevant signal segments coupled with explicit measurements enables more reliable observation-level EM signal understanding.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Electromagnetic Signal Understanding, Multimodal Large Language Models, Signal Parameter Estimation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_D6yBOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
1
Read review
Paper summary

This paper presents EM-Lens for question answering over long I/Q observations containing communication, radar, and jamming signals. It introduces the synthetic longEM-IQA training set and longEM-Bench, which cover 11 identification, parameter-estimation, and composite-entity tasks. EM-Lens combines a Mamba2 signal encoder with Qwen3.5-4B and uses an agentic runtime to localize and crop relevant intervals and optionally invoke DSP tools. The model is trained with mixed long- and short-observation supervision, shallow temporal and deep decision LoRA routers, and GRPO over tool-interaction trajectories. EM-Lens obtains an average score of 71.3 on longEM-Bench and 71.7% accuracy on the short-observation EM-Bench subset.

Critical strengths

The paper addresses an important gap between isolated signal perception and mixed-observation analysis. Moving from pre-segmented snippets to query-conditioned localization and analysis within longer I/Q observations is practically relevant. The benchmark covers identification, physical parameter estimation, and composite entities, while the appendix provides explicit scoring rules, tool interfaces, and representative DSP estimators. This task formulation is a useful step toward more complete EM reasoning systems.

The framework performs strongly on the proposed benchmark while retaining short-segment perception. EM-Lens reaches 71.3 on longEM-Bench and leads the listed methods on all 11 tasks. More importantly, the internal comparisons show that routed mixed-data training raises the long-observation average from 65.9 to 69.5 and short-observation accuracy from 62.4% to 71.8%, while GRPO further improves the long-observation score to 71.3. These results show that the model can jointly support long-observation analysis and short-segment perception within the evaluated setting.

Critical weaknesses

The observation-level generalization claim is evaluated only within one synthetic data pipeline. Both longEM-IQA and longEM-Bench come from the authors’ simulator, while the separation of waveform templates, parameter ranges, and channel conditions between training and evaluation is unclear. The paper provides no evaluation on measured captures, an independent simulator, or EMRB, which it cites as a full-length I/Q benchmark. The results therefore establish performance on the proposed synthetic distribution, but do not yet support the broader motivation of real-world receiver analysis.

The core zoom-and-measure and routed-specialization mechanisms are not cleanly isolated. Localization is central to the method but is not evaluated separately. The 33.0-to-66.8 comparison adds CoT supervision and tool interaction together, while the appendix examples use cropping followed by direct perception rather than the DSP tools. It is therefore unclear how much the localization and measurement tools contribute. Routing is also introduced as a complete package: Figure 3 provides only correlational motivation, without a parameter-matched shared LoRA baseline or separate shallow and deep router ablations. The evidence does not yet isolate the paper’s main technical contributions.

Questions for authors
  1. How are the training, validation, and benchmark examples separated? In particular, do they share the same waveform templates and parameter ranges? Could you also evaluate transfer to EMRB or measured data?
  2. Can the authors report standalone localization results and separate the effects of cropping, DSP tools, and CoT supervision? A comparison between predicted and oracle crops and a classical localization-plus-DSP pipeline would be helpful.
  3. Does routed specialization help beyond adding adapter capacity? Could you compare it with a parameter-matched shared LoRA model and report the effects of the shallow and deep routers separately?
  4. How do the proprietary MLLMs receive the raw I/Q sequence, and do they use the same tools? Was MERLIN fine-tuned on longEM-IQA after its patch size was changed, or was only the pretrained model evaluated?
4 comments

We thank the reviewer for the detailed comments on data generalization, the Zoom-and-Measure mechanism, routing, and baseline fairness. We provide the following clarifications and new results.

Q1. Data splits, OOD transfer, and EMRB. The original data comprise 27,135/561 longEM-IQA training/validation QA pairs, 3,300 longEM-Bench questions (11 tasks × 300), and 131,432/2,675 short-signal training/validation examples. The original longEM-Bench shares the simulator, nominal waveform categories, and parameter ranges with training, and is therefore an in-distribution (ID) test. We added OOD evaluations. For unseen radar bandwidths, PRIs, and duty cycles (1,000 questions), mixed SFT scored 71.12 (ID) → 41.33 (OOD) and EM-Lens 76.99 → 46.93. For new multipath profiles and time-varying channels (1,100 questions; normalized Doppler products 0.05/0.20), scores were 64.81 → 53.44 and 66.83 → 56.72, respectively. We also collected real continuous I/Q recordings (>1 million samples each) with a USRP B220 mini and constructed 800 questions across eight tasks from 30,000-sample windows. EM-Lens scored 33.69, compared with 28.43 for mixed SFT. We will clarify observation-level splits and QA-to-observation mappings in the revision.

Why not report a direct EMRB evaluation? We appreciate this suggestion. After inspecting EMRB's generation mechanism and conducting preliminary transfer experiments, we found that its evaluation requires a different source-separation capability. EM-Lens is designed for temporally localizable signal events: it identifies temporal boundaries, extracts a target interval via zoom_in, and performs local perception or DSP measurement. EMRB's long I/Q observations are instead programmatically synthesized with multiple transmitters often coexisting in time, distinguished primarily by spectral characteristics. Temporal cropping cannot isolate an individual source in this setting, even with oracle boundaries; subsequent DSP estimation can reflect the mixed spectrum rather than the target transmitter. A direct score would therefore mix temporal event understanding with spectral source separation, beyond the current method's intended scope. Our OOD and USRP tests more directly assess the claimed capability; extending EM-Lens to frequency-selective separation remains future work.

Q2. Localization and CoT/Zoom/DSP contributions. Independent signal-segment detection tIoU × 100 was 96.74 for long-signal-only SFT, 91.33 for mixed SFT, 95.35 for EM-Lens (SFT), and 94.96 for EM-Lens (SFT+RL). Start/end MAE were 1,423.50/1,025.44 samples for mixed SFT versus 245.09/1,040.47 for EM-Lens (SFT). Progressive ablations scored 33.0 → 48.5 → 49.0 → 69.5 for no CoT/Zoom/DSP, CoT only, CoT+Zoom, and CoT+Zoom+DSP, respectively; DSP adds 20.5 points. It was actually invoked on 1,860/3,300 questions (2,188 calls), including 83.94% of six parameter-estimation tasks. For radar bandwidth, the invocation/non-invocation groups scored 87.63/2.96 (descriptive, not causal).

Q3. Routing versus adapter capacity. longEM-Bench/EM-Bench scores were 64.55/62.60 (selected-layer shared LoRA), 61.89/63.00 (all-layer shared LoRA), 52.42/62.80 (Temporal Router only), 61.66/62.97 (Decision Router only), and 69.49/71.83 (full dual routing). The full model improves on selected-layer shared LoRA by 4.94 points / 9.23 percentage points. We will report exact trainable parameter counts to clarify capacity controls.

Q4. Closed-source MLLMs and MERLIN. Closed-source MLLMs receive serialized I/Q floating-point sequences in original sample order plus the question. They are not given EM-Lens's zoom_in or DSP tools, although they may write and execute their own DSP code. In the original comparison, MERLIN's patch size was changed from 8 to 48, and the pretrained model was evaluated without fine-tuning (4.90). We additionally fine-tuned MERLIN using the same data and number of epochs as EM-Lens (SFT): it scored 58.61, versus 69.49 for EM-Lens. We will distinguish these baselines in the revision.

We appreciate the reviewer's constructive suggestions.

The authors have addressed my concern and I will change my score accordingly.

Thank you very much for your positive feedback. We are glad that our rebuttal has addressed your concerns. We sincerely appreciate your careful review and constructive suggestions.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

Thank you sincerely for your thoughtful engagement throughout the rebuttal discussion. We greatly appreciate the time and effort you have invested in evaluating our work, and we are encouraged by your indication that you are considering updating your score.

As the discussion period is drawing to a close, we would like to respectfully bring to your attention that the rating has not yet been updated in the system. We fully understand that you are busy, and we hope this serves merely as a gentle reminder. If our responses have adequately addressed your concerns, we would be very grateful if you could kindly reflect that in your updated rating before the deadline.

We remain happy to provide any further clarification should that be helpful.

Thank you again for your valuable time and consideration.

Reviewer_JvakOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces a benchmark and a baseline method for EM multimodal LLMs for long observations that contain multiple communication, radar, and jamming signals at once. In particular, the authors introduce longEM-IQA for training and longEM-Bench for evaluation. Both are simulated, with observations of about 25K to 30K I/Q samples, and the benchmark covers 11 tasks and composite-entity understanding. The training data are agent trajectories built by a large LLM rewriting the trace. The baseline model, EM-Lens, uses Qwen3.5-4B with a pretrained Mamba2 EM encoder. Based on this, the authors add routed LoRA, train on a mix of long observations and short segments, and finish with GRPO using real tool calls.

Critical strengths

The problem setting is well motivated. Real receivers see long streams with several overlapping signals, not clean snippets. Therefore, advancing EM LLMs to observation-level inputs is a straightforward step. The benchmark covers identification, parameter estimation, and composite reasoning in one setting, which existing EM benchmarks for long observations lack. If released, the data and benchmark could serve as a common testbed for this direction.

Critical weaknesses

The evaluation currently lacks real-world validation. Specifically, the training and test data come from the same simulator, and the paper provides no results on real long captures, whereas related benchmarks such as EMRB include raw, full-length I/Q captures. As a result, it is unclear whether the gains would hold on real signals, or whether the large margin over baselines mostly reflects training and testing on the same simulated distribution.

Questions for authors
  1. How often does EM-Lens call each DSP tool at inference, per task?
  2. Are there any results on real long captures, for example, the raw I/Q in EMRB, or at least on simulated data with settings not seen in training? This would help judge how far the gains transfer beyond the simulator.
  3. For Table 4, what is the result with LoRA at the same layers but without the routers, and with the same number of LoRA parameters spread across all layers?
4 comments

We thank the reviewer for recognizing the value of observation-level EM understanding and the longEM datasets. We address the three questions below.

Q1. Actual DSP invocation frequencies. Across 3,300 longEM-Bench questions, EM-Lens (+RL) invoked DSP for 1,860 (56.36%), making 2,188 calls: bandwidth 866, PRF 289, pulse width 439, duty cycle 251, and pulse count 343. Invocation counts (out of 300 questions per task) were: communication bandwidth 255, radar bandwidth 267, duty cycle 261, pulse count 250, PRF 223, pulse width 255, communication composite entity 274, and protocol identification 75; CJR, RJR, and jammer composite entity tasks had zero calls. Across the six parameter-estimation tasks, the invocation rate was 83.94%. Adding DSP to CoT+Zoom increased the score from 49.0 to 69.5 (+20.5), confirming the practical contribution of executed measurements.

Q2. Generalization and why EMRB was not used as a direct benchmark. On unseen radar parameters (bandwidth, PRI, duty cycle; 1,000 questions), EM-Lens (SFT) scored 46.93 vs. 41.33 for mixed SFT; on new multipath and time-varying channels (1,100 questions), 56.72 vs. 53.44. We also collected continuous I/Q recordings (>1 million samples each) in a real wireless environment using a USRP B220 mini. Segmenting them into 30,000-sample windows yielded 800 questions across eight tasks: EM-Lens scored 33.69 vs. 28.43, with improvements in protocol identification (73 vs. 59), communication bandwidth (23.54 vs. 2.80), and communication composite entity understanding (44.93 vs. 31.64).

We appreciate the suggestion to evaluate on EMRB and agree that external benchmarks are valuable. After examining its data-generation mechanism and conducting preliminary transfer experiments, we identified a mismatch in the required signal-separation mechanism. EM-Lens focuses on temporally localizable events: it predicts a target's temporal boundaries, uses zoom_in to extract the interval, and then performs fine-grained perception or physical measurement. In contrast, EMRB's long I/Q observations are programmatically synthesized and often contain multiple transmitters active simultaneously, whose signals must be distinguished spectrally. Even with perfect temporal boundaries, cropping such an observation cannot isolate one transmitter; a bandwidth estimator may measure the combined spectrum, not the target source. Thus, a direct EMRB score would conflate temporal localization and spectral source separation, which is outside EM-Lens's current scope. We therefore report OOD and real-USRP evaluations aligned with our task formulation, while recognizing spectral multi-source separation as an important future extension.

Q3. Shared-LoRA controls. Under the same SFT protocol, longEM-Bench/EM-Bench scores were: selected-layer shared LoRA without routing 64.55/62.60; all-layer shared LoRA 61.89/63.00; Temporal Router only 52.42/62.80; Decision Router only 61.66/62.97; and full dual routing 69.49/71.83. The full model outperformed selected-layer shared LoRA by 4.94 points / 9.23 percentage points. We will provide exact trainable parameter counts for each configuration to clarify capacity controls.

We thank the reviewer for the constructive feedback.

I appreciate the authors' rebuttal, which has addressed my major concerns. Therefore, I would like to raise my rating.

Thank you for your encouraging feedback. We are glad that our rebuttal has addressed your major concerns. We sincerely appreciate your careful consideration and constructive suggestions.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

Thank you sincerely for your thoughtful engagement throughout the rebuttal discussion. We greatly appreciate the time and effort you have invested in evaluating our work, and we are encouraged by your indication that you are considering updating your score.

As the discussion period is drawing to a close, we would like to respectfully bring to your attention that the rating has not yet been updated in the system. We fully understand that you are busy, and we hope this serves merely as a gentle reminder. If our responses have adequately addressed your concerns, we would be very grateful if you could kindly reflect that in your updated rating before the deadline.

We remain happy to provide any further clarification should that be helpful.

Thank you again for your valuable time and consideration.

Reviewer_UCdzOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates observation-level electromagnetic (EM) signal understanding. The authors introduce longEM-IQA and longEM-Bench, focusing on long I/Q observations containing heterogeneous signal events. To address the challenges of long-observation understanding, they propose EM-Lens, an agentic framework that localizes relevant signal segments and leverages signal-processing tools for detailed analysis. Additionally, a depth-aware routed specialization strategy is introduced to improve joint learning from long observations and short signal segments. Experimental results on longEM-Bench demonstrate improvements over several general-purpose and domain-specific baselines.

Critical strengths
  1. Well-motivated. The paper identifies a meaningful research gap, as existing EM understanding benchmarks primarily focus on isolated short signal segments. The proposed benchmark addresses the challenges of understanding long observations containing heterogeneous signals.
  2. Intuitive and effective framework. The proposed localization-and-measurement pipeline is straightforward and well aligned with the task requirements. Experimental results demonstrate its effectiveness on longEM-Bench.
Critical weaknesses
  1. Insufficient details on benchmark construction and quality control. Although the authors describe the general data generation process and mention several quality checks, the detailed procedures remain unclear. For example, how are structural validity, reference consistency, and numerical correctness verified? Is there any manual verification, or are all checks performed automatically? What proportion of generated samples passes these checks? Furthermore, since the benchmark is constructed using simulated signals, its representativeness of real-world EM observations remains uncertain. More details and validation would help establish the reliability and practical relevance of the proposed benchmark.
  2. Limited validation of the generalizability of depth-aware specialization. In Section 4, the authors analyze three backbone models (Qwen3.5-4B, Llama-3.2-3B, and Gemma-3-4B) and observe similar patterns of shallow-layer temporal sensitivity and deep-layer task divergence. These observations motivate the proposed depth-aware routed specialization strategy. However, the effectiveness of this strategy is evaluated only on Qwen3.5-4B. Additional experiments on Llama-3.2-3B or Gemma-3-4B would help determine whether the proposed design generalizes across different model architectures.
  3. Limited performance improvements on short-segment perception. As shown in Table 3, EM-Lens achieves an average accuracy of 71.7% on EM-Bench, slightly below Qwen3-VL-4B-Instruct (72.9%) and MERLIN (71.8%). Considering that the proposed training strategy incorporates substantial short-segment training data, this result raises questions about its effectiveness in preserving or improving fine-grained signal perception. Further analysis of the performance gap and the potential trade-off between long-observation understanding and short-segment perception would strengthen the paper.
  4. Insufficient ablation studies. Although Table 4 demonstrates the overall effectiveness of routed specialization and reinforcement learning, the individual contributions of the proposed components remain unclear. In particular, the authors should investigate the separate effects of the shallow Temporal Router and deep Decision Router and the sensitivity to router placement across different layers. Moreover, additional ablations on the zoom-in mechanism and signal-processing tools would help clarify their respective contributions to the overall performance improvements.
  5. Missing conclusion section. The paper lacks a dedicated conclusion section. The authors should summarize the main findings and contributions while discussing the limitations of the proposed approach and potential directions for future work.
  6. Figure 1 is not referenced in the main text. Although Figure 1 provides an overview of the proposed framework, it is not explicitly referenced or discussed in the main text. The authors should better integrate this figure into the introduction to improve the clarity and readability of the paper.
Questions for authors

See the weaknesses.

5 comments

We thank the reviewer for recognizing our motivation and localization-and-measurement framework. We address all six comments below.

Q1. Dataset quality control. The simulator first generates I/Q observations and structured ground truth for entities, temporal spans, physical parameters, and relationships. A teacher execution environment then constructs trajectories, actually runs zoom_in and applicable DSP tools, and checks numerical measurements against simulation ground truth. Qwen3.5-397B rewrites these trajectories into natural-language reasoning under factual constraints. Samples undergo rule-based checks for structural validity, referential consistency, and numerical correctness, followed by an LLM Judge assessing reasoning consistency, tool appropriateness, and coherence. Samples failing the judge receive additional manual review. We will document the validation criteria and filtering statistics in the revised appendix.

Q2. Cross-backbone generalization. With the same framework, conventional mixed SFT / EM-Lens SFT / EM-Lens+RL scored 51.19/54.04/67.41 on Llama-3.2-3B, and 51.45/53.24/60.94 on Gemma-3-4B. Routing-specialized SFT improves both backbones, and RL yields further gains, supporting cross-architecture applicability.

Q3. Short-segment perception and training balance. Short-signal supervision is intended to improve local semantic perception after Zoom, not merely maximize isolated EM-Bench accuracy. The original model obtained 71.7% on EM-Bench. Although short-signal examples constituted approximately 83% of training samples, they accounted for only 9.18% of signal tokens. Without changing the architecture, increasing their signal-token share to 50%/70% raised accuracy to 73.33%/74.86%, above Qwen3-VL-4B-Instruct (72.9%). The initial gap therefore largely reflects training-resource allocation rather than an inherent local-perception limit.

Q4. Component ablations. On longEM-Bench/EM-Bench, selected-layer shared LoRA scored 64.55/62.60, Temporal Router only 52.42/62.80, Decision Router only 61.66/62.97, and full dual routing 69.49/71.83, indicating complementary roles. Progressive CoT/Zoom/DSP ablations yielded 33.0 → 48.5 → 49.0 → 69.5; DSP adds 20.5 points beyond CoT+Zoom. Layers L0–L3 and L28–L31 were selected based on the temporal-sensitivity and task-divergence analyses in Section 4; we will clarify this design rationale in the revision.

Q5–Q6. Presentation. We will add a dedicated Conclusion discussing contributions, limitations, and future work, and explicitly reference and explain Figure 1 in the Introduction.

We appreciate the reviewer's valuable feedback.

Thank you for the response and additional experiments, which have addressed most of my concerns.

However, I still have a few questions:

  1. The filtering statistics and pass rates during benchmark construction remain unclear, as these details have not yet been incorporated into the manuscript. Additionally, how significant are the differences between simulated and real-world EM signals, and how might these differences affect the benchmark's practical relevance?

  2. In the component ablation study, the improvement from adding the Zoom mechanism appears relatively small (48.5 → 49.0). Could the authors further clarify the contribution and necessity of this component?

We thank the reviewer for the constructive follow-up. We clarify the dataset statistics and the task-dependent role of Zoom below.

Q1. Dataset filtering, quality control, and practical relevance.

Construction and selection. We generated 20,422,561 simulated I/Q records and selected 37,500 observations across five signal families (Radar, Bluetooth, 5G, Wi-Fi, and LoRa): 7,500 per family, comprising 2,500 signal–jammer mixtures, 2,500 signal-only observations, and 2,500 jammer-only observations. Their structured annotations yielded 902,895 QA plans. We formed separate pools of 90,000 training candidates and 7,800 evaluation candidates, with training and evaluation partitioned at the original I/Q observation level to prevent leakage.

From the 7,800 evaluation candidates, we selected 3,300 questions for longEM-Bench (42.31% selection ratio). Selection combined validity requirements and distributional balancing: each question had to be supported by source annotations, and attribute-based references had to identify a unique target. We then checked task counts, question-type distributions, and split isolation. The final benchmark contains 11 tasks with 300 questions each, evenly divided into 150 standard and 150 referring questions. Referring subtypes (ordinal, temporal, attribute-based, and relational) were balanced within applicable tasks; answer values were also balanced where relevant. For example, the 150 standard communication-bandwidth questions cover four bandwidth values in counts of 37/37/38/38. The 42.31% figure is a selection ratio, not a single-stage quality-check pass rate, because quota balancing also affects inclusion.

Trajectory validation. Qwen3.5-397B cumulatively generated 59,736 CoT candidates. Rule-based checks assessed structural validity, entity/reference consistency, and numerical correctness; an LLM Judge checked agreement of answers with executed tool outputs, reasoning logic, and appropriate tool use. Human inspection covered both Judge-accepted and Judge-rejected examples, informing iterative revisions of the generation and judging rules. Ultimately, 27,135 CoTs were retained, or 45.42% of cumulative CoT candidates. This is an overall retention ratio, not the Judge's first-pass acceptance rate. We will report these definitions and selection criteria in the appendix.

Simulation scope. The simulator models heterogeneous signals, overlaps, interference, variable SNR, multipath fading, carrier-frequency offsets, and I/Q imbalance. These controlled conditions allow precise labels for boundaries, physical parameters, and inter-signal relations, supporting reproducible evaluation of observation-level reasoning.

Q2. Why does Zoom yield a small aggregate gain?

The overall CoT-to-CoT+Zoom improvement masks strong task dependence. On protocol identification (PI), Zoom raises accuracy from 48.33% to 72.67% (+24.34 percentage points) without DSP; this is comparable to the full pipeline's 72.00%. Thus, Zoom independently benefits tasks requiring localized signal perception: it presents the selected segment for re-examination rather than forcing the model to infer its protocol amid unrelated events in a long observation.

In contrast, Zoom alone does not provide a numerical estimator. For bandwidth, PRF, and pulse-width estimation, a localized view may still leave the model relying on approximate learned predictions; some such tasks show little gain or declines, offsetting PI's improvement in the overall average. DSP instead performs explicit physical measurement on the extracted segment. After adding DSP to CoT+Zoom, radar-bandwidth performance rises from 4.41% to 77.62%, radar PRF from 32.33% to 74.80%, and radar pulse width from 34.13% to 79.19%.

Therefore, Zoom has two distinct roles: improving target-specific re-perception (as directly evidenced by PI) and supplying target-specific I/Q input to subsequent measurement tools. It excludes unrelated events outside the chosen interval.

Reviewer_vkYNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes EM-Lens, an agentic multimodal framework for observation-level electromagnetic (EM) signal understanding. Unlike existing EM models that primarily operate on pre-segmented signal instances, EM-Lens aims to understand long I/Q observations containing heterogeneous signal events. The paper introduces two datasets, longEM-IQA and longEM-Bench, covering 11 tasks in signal identification, parameter estimation, and composite-entity understanding.

The proposed framework combines active temporal localization, signal segment cropping, local signal perception, and explicit signal-processing tools for physical parameter estimation. To address the conflict between long-observation localization and fine-grained segment perception, the authors introduce a depth-aware routed specialization strategy, using shallow temporal routers and deep decision routers over a shared backbone. They further apply GRPO to optimize multi-turn tool interactions.

Experiments show that EM-Lens achieves an average score of 71.3 on longEM-Bench, substantially outperforming the reported baselines. Ablation studies suggest that routed specialization improves both long-observation understanding and short-segment perception, while reinforcement learning provides additional improvements in long-observation tasks.

Critical strengths
  1. An important and practical problem formulation.

The paper addresses observation-level EM signal understanding, which is more challenging and practically relevant than conventional recognition of isolated, pre-segmented signal instances. The proposed active localization and zoom-in mechanism provides a reasonable approach to handling sparse target signals embedded in long and heterogeneous observations.

  1. An interesting architecture and promising experimental results.

The depth-aware routed specialization strategy is well motivated by the layer-wise analysis of temporal sensitivity and task divergence. The experiments demonstrate substantial improvements over the reported baselines, particularly on long-observation tasks. The combination of long-observation supervision and short-segment perception is also a promising direction for developing more capable EM models.

Critical weaknesses
  1. Limited evaluation of generalization to real-world EM observations.

The proposed longEM-IQA and longEM-Bench datasets are constructed using controllable simulation of communication, radar, and jamming signals. Although this provides convenient access to ground-truth temporal intervals and physical parameters, it raises concerns about the generalization of the proposed approach to real-world EM environments, where signals may exhibit more complex interference, channel effects, hardware imperfections, and distribution shifts.

The paper does not provide experiments on real-world long-duration I/Q observations. Consequently, it is difficult to determine whether the reported improvements translate into practical observation-level EM understanding beyond the simulated setting.

  1. Insufficient evaluation and ablation of localization and tool utilization.

Localization is a central component of EM-Lens, but the paper states that localization is supervised through trajectories and evaluated within composite-entity scores rather than as a standalone task. This makes it difficult to assess whether the proposed localization mechanism actually identifies signal boundaries accurately.

Additionally, although the framework emphasizes explicit DSP tools for physical parameter estimation, the illustrative trajectories in Appendix D often show direct perception of parameters from cropped signals rather than explicit calls to the corresponding estimation tools. More detailed analysis of when and how these tools are utilized would strengthen the claims.

The ablation study also does not sufficiently isolate the contributions of the temporal router, decision router, and individual tool-based measurement components.

Questions for authors
  1. Generalization to real-world signals: Have the authors evaluated EM-Lens on any real-world long-duration I/Q observations? If not, what evidence supports the generalization of the proposed framework beyond the controllable simulation environment?

  2. Localization accuracy: Since localization is a major contribution, could the authors provide standalone localization metrics, such as temporal IoU or boundary detection accuracy, rather than evaluating localization only indirectly through composite-entity scores?

  3. Tool utilization: How frequently does EM-Lens invoke explicit DSP estimation tools compared with directly perceiving physical parameters from cropped I/Q segments? Could the authors provide a task-wise comparison to quantify the actual contribution of explicit signal-processing tools?

  4. Router ablation: Could the authors provide separate ablations for the shallow Temporal Router and deep Decision Router to better establish the contribution of each component?

3 comments

We thank the reviewer for recognizing the importance of observation-level EM understanding and the motivation for depth-aware routing. We provide four additional analyses.

Q1. Real-world and OOD generalization. We collected continuous I/Q recordings in a real wireless environment using a USRP B220 mini (over one million samples per recording) and divided them into windows of at most 30,000 samples, yielding 800 questions across eight tasks. EM-Lens (SFT) scored 33.69 versus 28.43 for mixed SFT. Protocol identification improved to 73.00 vs. 59.00, communication bandwidth estimation to 23.54 vs. 2.80, and communication composite entity understanding to 44.93 vs. 31.64. On unseen radar parameters (1,000 questions), EM-Lens achieved 46.93 vs. 41.33; on new multipath and time-varying channel conditions (1,100 questions), 56.72 vs. 53.44. Thus, the gains are not restricted to the original simulation distribution.

Q2. Independent localization accuracy. We evaluated signal-segment detection (SD) using temporal IoU (tIoU × 100) and start/end boundary mean absolute error (MAE, samples):

ModeltIoUStart MAEEnd MAE
Long-signal-only SFT96.74176.01718.06
Conventional mixed SFT91.331,423.501,025.44
EM-Lens (SFT)95.35245.091,040.47
EM-Lens (SFT+RL)94.96362.311,045.00

Relative to conventional mixed SFT, EM-Lens improves tIoU by 4.02 points and substantially reduces start-boundary error, demonstrating that routing specialization mitigates localization degradation during joint training.

Q3. Actual DSP usage and contribution. EM-Lens (+RL) invoked DSP tools for 1,860/3,300 questions (56.36%), making 2,188 calls: bandwidth 866, PRF 289, pulse width 439, duty cycle 251, and pulse count 343. Across six parameter-estimation tasks, DSP was used for 1,511/1,800 questions (83.94%). Progressive ablations scored 33.0 without CoT/Zoom/DSP, 48.5 with CoT only, 49.0 with CoT+Zoom, and 69.5 with CoT+Zoom+DSP. Explicit measurement thus adds 20.5 points beyond CoT+Zoom, while jammer-related classification tasks rely mainly on direct perception, reflecting task-adaptive tool use.

Q4. Separate router ablations. Under the same SFT protocol, longEM-Bench / EM-Bench scores were: shared LoRA at selected layers 64.55/62.60; Temporal Router only 52.42/62.80; Decision Router only 61.66/62.97; all-layer shared LoRA 61.89/63.00; and full dual-router EM-Lens 69.49/71.83. The full model gains 4.94 points / 9.23 percentage points over selected-layer shared LoRA. This supports the complementary roles of shallow temporal representation specialization and deep task-decision specialization.

We appreciate the reviewer's constructive suggestions.

AuthorsOpenReview ↗

Gentle Reminder

Dear Reviewer,

We hope this message finds you well. We sincerely appreciate your initial review and the constructive feedback you provided on our work.

We would like to take this opportunity to kindly remind you that the rebuttal period is approaching its end. As we have not yet had the chance to receive your feedback on our responses, we were wondering whether our rebuttal has adequately addressed the concerns raised in your original review.

We would be grateful if you could take a moment to revisit our responses at your earliest convenience, and kindly consider whether an update to your rating would be warranted. Of course, we also remain happy to provide any additional clarification if needed.

Thank you very much for your time and consideration.

I appreciate the authors’ detailed response and additional experiments. The real-signal and OOD tests, together with the localization results and component ablations, address most of my original concerns. Some uncertainty remains about broader generalization and adapter capacity controls, but the new evidence makes the case more convincing, and I raise my rating to 3.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.