Pith.

Networking and Internet Architecture · arXiv:2610.35344 v1 · submitted 26 Oct 2026 · 61 references

Uncertainty-Guided Counterfactual Sensor Selection and Safety-Gated Recognition for Autonomous Sensing Agents in IoT Systems

Shirong Guo, Hongming Zhou, Muhammad Attique Khan, Jinfeng Xu, Yonis Gulzar, Manar Mousa Altamimi, Kunfu Wang

The paper says

“Autonomous IoT systems rely on distributed sensing, but sensor faults and communication loss can leave multichannel evidence incomplete or unreliable. Their agents must therefore classify measurements, assess evidence sufficiency, select a sensor to reacquire, and hand off unsafe decisions.”

more of the abstract

We formulate counterfactual sensing: a degraded window defines the factual state, each candidate restoration defines a controlled alternative, and a pre-request utility model estimates loss reduction. Physics-Guided Invariant-Kinematic Fusion (PGIK-Fusion) constructs class evidence, while calibrated risk and cost-sensitive utility select acceptance, sensor request, or handoff from factual information. On the Logistic Activity Recognition Challenge dataset, PGIK-Fusion achieves 0.768 accuracy and a macro-averaged F1 score (Macro-F1) of 0.696, exceeding extreme gradient boosting (XGBoost)–one-dimensional convolutional neural network (1D-CNN) fusion by 2.49 and 2.00 percentage points, respectively. Across seven predefined degradation conditions on held-out subjects, the policy reduces unsafe acceptance from 33.0% to 13.8%; relative to an action-rate-matched random policy, it raises automation accuracy by 18.5 percentage points and lowers expected cost by 24.7%. In the auxiliary within-event wind-turbine study, PGIK-Fusion reaches 0.975 Macro-F1, improves on XGBoost–1D-CNN by 1.17 percentage points, and reduces unsafe acceptance from 20.1% to 8.0%. The evaluation characterizes the controlled responses of a frozen recognition model to specified sensor restorations under predefined fault families and severities. Within this scope, the results demonstrate that calibrated recognition, targeted evidence acquisition, and conservative deferral can form a unified perception–action loop that reduces unsafe automation under sensor degradation.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 27 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A solid, well-scoped systems paper whose safety claim mostly holds, but the piece they brand as the main novelty—counterfactual request benefit—is the weakest supported component and needs a targeted ablation before the paper can stand as written. This line is wrong

3 major objections · 5 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 61 references say.

The paper says, at Section IV-D / IV-E (Eqs. 16, 22)

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The request action exists only when the benefit model in Eq. (16) predicts positive net utility, yet that model is weakly predictive: Pearson r = 0.219 and MAE 0.087, the latter worse than the constant-zero predictor's MAE of about 0.060 given that 93.97% of targets are neutral (Section IV-D).

more of objection 1

The corresponding margin of the full policy over confidence-only gating is only 0.83pp unsafe acceptance and 0.93pp automation accuracy (Section IV-E), reported without an interval. The request-cost sweep shows the request rate falling to zero for costs of at least 0.20 but reports only expected cost (0.236–0.242), not unsafe acceptance or automation accuracy in that zero-request regime where the benefit model is effectively inert. An ablation that zeroes the benefit term or disables requests under the same risk gate, together with an uncertainty estimate for the 0.83pp difference, is needed to support the claim that targeted evidence acquisition, rather than the risk gate plus handoff, contributes to the measured safety reduction.

The paper says, at Section IV-D

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The 85.22% opportunity-capture rate is computed over all 46,984 degraded window–condition instances (5,018 of 5,888 windows containing at least one corrective restoration), whereas the policy actually requests a sensor for only 8.89% of windows; for most windows every candidate's net utility is negative and no request is made (Fig. 10).

more of objection 2

This metric therefore measures the model's ranking quality over the full counterfactual grid, not its realized request outcomes, and no random-candidate baseline conditioned on the same request decisions is provided. Reporting opportunity capture and correction rates restricted to the requested-window subset, with a random-candidate control on those same windows, would directly test whether the model selects the right sensor when a request is actually issued.

The paper says, at Section IV-E (Fig. 11)

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The action-rate-matched random policy preserves aggregate accept/request/handoff counts while randomizing both the allocation across windows and the requested sensor, so the reported 18.51pp automation-accuracy gain conflates the risk gate's window selection with the benefit model's candidate choice. A sharper comparator would fix the policy's accept/request/handoff decisions and randomize only the requested sensor, or compare against a fixed-candidate rule on the same request windows; without that, the distinct contribution of candidate-specific benefit estimation is not isolated.

more of objection 3

The same limitation applies to the auxiliary study, where the unsafe-acceptance estimate of 0.080 ± 0.057 makes the reported 3.76pp margin over confidence-only gating difficult to assess without paired or interval estimates.

5 minor objections

The paper says, at Section IV-A / Table I

Pith says: minor objection 1 of 5

The headline perception and policy comparisons (2.49pp accuracy, 2.00pp Macro-F1, 18.51pp automation accuracy) are point estimates without confidence intervals or significance tests; multi-seed results are reported for PGIK-Fusion only (0.694 ± 0.006 over three seeds), so the stability of the 2.00pp gain over XGBoost–1D-CNN is unclear.

The paper says, at Section III-C

Pith says: minor objection 2 of 5

Because q_t is condition-supplied rather than produced by field fault detectors, the policy evaluation assumes the quality metadata are exact; a sensitivity analysis that perturbs q_t or mislabels the degradation type would clarify behavior under imperfect metadata without requiring unseen-fault-family evaluation.

show 3 more

The paper says, at Data Availability

Pith says: minor objection 3 of 5

The repository URL 'https://github.com/shannongsr/IoTJ counterfactual sensor selection' contains an embedded space and should be replaced by a single valid hyperlink.

The paper says, at Section III-B / Eq. (10)

Pith says: minor objection 4 of 5

The 'physics-guided' framing rests on coordinate-frame invariance, which Eq. (10) states correctly but which is elementary; since the ablation in Table III shows the descriptor value empirically, the paper could temper the terminology or state explicitly that the physical guidance is limited to rotation/reflection invariance.

The paper says, at Section III-D / Abstract

Pith says: minor objection 5 of 5

The abstract's 'calibrated recognition' phrasing covers both temperature scaling and the conformal set, but the 0.929 empirical coverage is not a guaranteed 90% coverage because temperature selection and quantile estimation share S11–S12; the paper acknowledges this in Section III-D, and it would help to repeat the caveat wherever the 0.929 figure is cited.

What would settle it

On the LARa held-out set, using the restoration outcomes already computed for all 140,952 window-candidate pairs, run two controlled ablations and report bootstrap confidence intervals over windows. (A) Zero-out h_theta in Eq. (22) so no request is ever selected, keeping the risk model and costs fixed; this measures the value of having the request action at all. (B) Keep the full policy's accept/request/handoff decisions exactly as made, but replace the benefit model's chosen sensor with a uniformly random sensor among the three IMUs for every request; this isolates candidate-specific utility. Compare unsafe acceptance, automation accuracy, and expected cost of (A) and (B) against the full policy.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 27 Oct 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic partly agrees 1 min 43 s · 25,609 tokens read, 13,874 written
  3. Editor wrote the verdict above 23 s · 26,544 tokens read, 2,459 written
  4. Referee major revision: 3 major objections, 5 minor 2 min 52 s · 25,887 tokens read, 15,941 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

Short version: this is a competent, honestly scoped systems paper. The central result—a safety-gated accept/request/handoff loop that cuts unsafe acceptance from 33.0% to 13.8% on held-out subjects—is supported by a clean protocol. The weak spot is the component they market as the headline contribution: the counterfactual benefit model that drives sensor requests. It is least load-bearing, and the paper never isolates it properly.

What is actually new and good: reframing sensor reacquisition as a frozen model's controlled response to a specified restoration, with a pre-request utility model that estimates loss reduction without seeing restored observations or labels. The evaluation design is careful: S11–S12 for calibration and policy fitting, S13–S14 held out, Eq. (14) used only for target construction, and the deployment loop uses only pre-request evidence. The scope conditions are stated plainly: same seven fault families in development and test, condition-supplied quality metadata, one wind-turbine event. The latency story is also useful: sub-millisecond perception plus a 0.179 ms policy layer is a deployable pattern.

Now the soft spots, in proportion. The benefit model is weakly predictive by the paper's own numbers: Pearson r=0.219, MAE 0.087 versus roughly 0.060 for a constant-zero predictor on a target that is 93.97% neutral. Requests are only 8.89% of actions, and the full policy beats a confidence-only gate by 0.83 percentage points of unsafe acceptance, with no interval or significance test. The 85.22% opportunity-capture rate is computed over candidate rankings for all degraded windows, not over windows the policy actually requested, and the action-rate-matched random baseline cannot isolate the benefit model's candidate choice. So the stress-test concern lands: the aggregate safety reduction is dominated by the risk gate plus handoff, and "targeted evidence acquisition" as a distinct contributor is not yet demonstrated. This is fixable with a request-conditioned baseline and a candidate-choice analysis, but it should be fixed before publication.

Two smaller issues. The perception comparison, a 2.00 pp Macro-F1 gain over XGBoost–1D-CNN, has no variance or significance testing; the multi-seed result for PGIK-Fusion alone is not enough. And the code repository link in Data Availability is malformed, with a space in the URL and no commit hash, so the claimed artifact cannot currently be inspected.

Who gets value: people building IoT sensing agents who want a reusable decision-layer pattern and a careful evaluation template. I would not treat the counterfactual requests as validated yet, but the overall framework and protocol are worth engaging with.

Recommendation: send this to peer review, not desk reject. Ask for an isolation of the request mechanism, significance or intervals on the key comparisons, and a working code link. My own verdict is conditional rather than reject; the direction is sound, and the missing evidence is obtainable.

The strongest objection

The load-bearing pillar that is least secured is the request mechanism. Section IV-D reports that the benefit model h_theta in Eq. (16) has Pearson r=0.219 and MAE 0.087, worse than the ~0.060 constant-zero predictor on a target that is 93.97% neutral. This model is the only input that makes request(v_j) preferable to accept/handoff in Eq. (22). Yet Section IV-E reports just a 0.83 pp unsafe-acceptance gain and a 0.93 pp automation-accuracy gain over confidence-only gating, which takes no requests, and no interval is given for that difference. The 85.22% opportunity-capture rate is computed over candidate rankings for all degraded windows, not over windows the policy actually requests, and no random-candidate baseline conditions on the same request decisions. The action-rate-matched random comparator randomizes global action allocation, so it cannot isolate the benefit model's candidate choice. The measured safety reduction could therefore come almost entirely from the risk gate plus handoff, leaving 'targeted evidence acquisition' unsupported as a distinct contributor.

The reader's conditional verdict is sound; my stress-test does not change it. The paper is unusually disciplined about scope: Section III-C explicitly restricts generalization to unseen faults and severities, Section III-D discloses the shared calibration/conformal partition, and Section IV-F acknowledges single-event temporal dependence. Those are acknowledged limitations and I do not re-litigate them. The unresolved gap is internal: the safety improvement is dominated by the risk gate plus handoff, and the paper's distinctive claim, counterfactual sensor selection, is not isolated by any comparator whose request decisions and candidate choices are controlled. Section IV-D quantifies the benefit model's weakness: Pearson r=0.219, MAE 0.087 versus about 0.060 for constant zero, on a target that is 93.97% neutral. The benefit model only enters the action through Eq. (22), and requests amount to 8.89% of actions (Section IV-E). Against confidence-only gating, which takes no requests, the full policy gains 0.83 pp unsafe acceptance and 0.93 pp automation accuracy, but no uncertainty interval is reported, and the confidence gate uses a different accept/handoff threshold. The 85.22% opportunity-capture rate evaluates candidate ranking over all degraded windows, not over windows the policy actually requests, and no random-candidate baseline is reported for those same request decisions. The matched-random comparator randomizes global action allocation, so it cannot separate the benefit model's candidate choice from the risk gate's accept/handoff choice. Because the handoff mechanism alone already reaches 14.68% unsafe acceptance, the central claim's 'targeted evidence acquisition' component rests on this thin, unisolated increment. The proposed check uses the already computed restoration outcomes for all candidates, so it can be run without new experiments. If the benefit model's candidate choice performs no better than random candidate choice at the same request decisions, then the headline contribution reduces to a risk-gated handoff policy plus a small request effect, and the conditional verdict should be understood as conditional on supplying a missing central ablation. If it does beat random candidate choice, the conditional accept stands and the reader's deployment-transfer concerns remain the binding ones.

The referee's report

If the results hold, the paper supplies a cleanly specified, decision-theoretically coherent template for cost-aware reacquisition with deferral in multichannel recognition. The derivation in Eqs. (22)–(25) is internally consistent: the request condition U > max{0, r − C_h} follows correctly from the loss decomposition, and the cost ordering C_j < C_h < 1 produces the intended accept/request/handoff regimes without a hand-tuned confidence threshold. The held-out protocol is sound: risk and benefit models are fit on S11–S12 only, S13–S14 is untouched, and Algorithm 1 uses no restored observations or labels at decision time. The estimand is honestly bounded to the frozen model's controlled response to specified restorations, the within-event limitation of the auxiliary study is stated, and the authors report the benefit model's modest correlation rather than concealing it. The headline unsafe-acceptance reduction is large and is shown to be driven predominantly by the calibrated risk gate plus handoff.

Pith's reading of the paper (a machine's summary, not the authors')

This paper claims that an IoT sensing agent can manage degraded sensor data by treating each degraded window as a factual state and each possible one-sensor restoration as a controlled counterfactual, then deciding from pre-request evidence alone whether to accept the prediction, reacquire one physical sensor, or hand the decision to a supervisor. The claim is that calibrated risk and a cost-priced estimate of each candidate restoration's benefit put recognition, evidence acquisition, and deferral on one cost scale, forming a loop that reduces unsafe automation. On held-out subjects under seven predefined degradations the policy cuts unsafe acceptance from 33.0% to 13.8%, and the auxiliary wind-turbine study reports a drop from 20.1% to 8.0%. The authors state the scope: the same fault families and severities used to build the policy are used to test it, and quality metadata is supplied by the experiment rather than by a field fault detector.

What carries the argument. The load-bearing object is the counterfactual sensing formulation paired with a cost-symmetric safety gate. A degraded window $x_t^{(0)}$ defines the factual state; each candidate restoration $x_t^{(j)} = \mathcal{I}_j(x_t^{(0)})$ changes exactly one requestable sensor and defines the paired counterfactual whose realized value is the signed loss reduction $\Delta\ell_{tj}$. Because neither the restored window nor the true label is available before acting, a ridge model fits $\hat{b}_{tj}$ from the pre-request state $z_t$ (calibrated class probabilities, confidence, entropy, leading-class margin, and per-sensor quality), while a logistic model fits the factual error risk $\hat{r}_t$. The gate compares $\hat{r}_t$, $\mathrm{clip}(\hat{r}_t - \hat{b}_{tj}, 0, 1) + C_j$, and $C_h$ on one cost scale, so choosing which sensor to request is a matter of predicted net utility while releasing a label at all is controlled by factual risk and deferral cost. Perception is carried by PGIK-Fusion, whose rotation-invariant operators, vector magnitude $\rho$, first-order kinematic difference $\delta$, and directional consistency $\kappa$, supply the physical evidence that the ablation credits for most of the recognition gain.

Core claim. The central discovery is that a frozen recognition model's response to a specified single-sensor restoration can be learned as a counterfactual target, namely the signed reduction in 0-1 classification loss, $\Delta\ell_{tj} = \ell(y_t, \hat{y}_t^{(0)}) - \ell(y_t, \hat{y}_t^{(j)}) \in \{-1,0,1\}$, and that a model fitted from factual evidence alone predicts this target well enough to steer action selection on held-out subjects. A standardized ridge benefit model $\hat{b}_{tj} \approx \mathbb{E}[\Delta\ell_{tj} \mid z_t, v_j]$ and a logistic risk model $\hat{r}_t \approx \Pr(\hat{y}_t^{(0)} \ne y_t \mid z_t)$ feed a cost gate: acceptance pays $\hat{r}_t$, requesting sensor $v_j$ pays $\mathrm{clip}(\hat{r}_t - \hat{b}_{tj}, 0, 1) + C_j$, and handoff pays $C_h$. With request cost 0.10 and handoff cost 0.30, the policy requests the matching sensor for essentially all complete-sensor-loss windows, captures 85.22% of windows containing a corrective restoration, and at identical action rates beats a random allocation policy by 18.5 percentage points of automation accuracy and 24.7% lower expected cost. The paper's own reading is that calibrated recognition, targeted evidence acquisition, and conservative deferral form a unified perception-action loop that reduces unsafe automation under sensor degradation.

The premise it rests on. The whole request component stands or falls on the assumption that a benefit model trained on the same seven restoration families used for testing, and fed condition-supplied quality scores rather than outputs from a real fault detector, still predicts which sensor is worth reacquiring when the system is actually deployed.

If the paper is right (Pith's inferences, not the authors')

Autonomous IoT agents can operate on a finite evidence-management loop, accept, request one sensor, hand off, with the whole perception-plus-policy path running in about 0.70 ms per window on a CPU, so the safety layer adds only about a third of the perception latency.

Unsafe automation under sensor degradation is driven more by where the agent acts than by how often: at identical accept/request/handoff rates, targeted allocation lowers unsafe acceptance by 19.02 percentage points and expected cost by 24.7% relative to a random policy.

Pre-request evidence can identify the degraded sensor without seeing the restored signal: the selector matched the true degraded IMU for 98.5% of trunk-loss windows and all left- and right-loss windows, and captured 85.22% of windows that had any corrective restoration.

Fault severity and recoverability are distinct: the most damaging degradation (trunk loss, Macro-F1 down 33.17 points) is the one worth requesting, while under distributed noise and delay all candidate net utilities stay negative and the policy declines to request.

The same release rule re-instantiates across domains with different sensor semantics, wearable IMUs and wind-turbine load/strain/vibration, suggesting the abstraction is not tied to one sensing modality.

Because the benefit model is only weakly correlated with realized correction (Pearson $r = 0.219$) and requests form just 8.89% of actions, the measured safety gain is plausibly carried mostly by the confidence gate and handoff; a clean ablation that disables the request action entirely would isolate the counterfactual utility model's true contribution.

A natural stress test the paper leaves open is deployment under an unseen fault family or with quality metadata produced by a real fault detector rather than supplied by the experiment; if request precision collapses under that shift, the pre-request utility component would need re-training or a richer information state.

The ablation result that directional consistency alone recovered about 86% of the full physical-feature gain suggests the $\kappa$ operator is a cheap, transferable feature for any multi-axis wearable or structural monitoring task, independent of the rest of the framework.

The single-step request rule invites a multi-step extension in which the agent re-decides after each reacquisition whether another sensor is worth restoring; the paper explicitly lists long-horizon planning as beyond its present scope.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Autonomous IoT systems rely on distributed sensing, but sensor faults and communication loss can leave multichannel evidence incomplete or unreliable. Their agents must therefore classify measurements, assess evidence sufficiency, select a sensor to reacquire, and hand off unsafe decisions. We formulate counterfactual sensing: a degraded window defines the factual state, each candidate restoration defines a controlled alternative, and a pre-request utility model estimates loss reduction. Physics-Guided Invariant-Kinematic Fusion (PGIK-Fusion) constructs class evidence, while calibrated risk and cost-sensitive utility select acceptance, sensor request, or handoff from factual information. On the Logistic Activity Recognition Challenge dataset, PGIK-Fusion achieves 0.768 accuracy and a macro-averaged F1 score (Macro-F1) of 0.696, exceeding extreme gradient boosting (XGBoost)–one-dimensional convolutional neural network (1D-CNN) fusion by 2.49 and 2.00 percentage points, respectively. Across seven predefined degradation conditions on held-out subjects, the policy reduces unsafe acceptance from 33.0% to 13.8%; relative to an action-rate-matched random policy, it raises automation accuracy by 18.5 percentage points and lowers expected cost by 24.7%. In the auxiliary within-event wind-turbine study, PGIK-Fusion reaches 0.975 Macro-F1, improves on XGBoost–1D-CNN by 1.17 percentage points, and reduces unsafe acceptance from 20.1% to 8.0%. The evaluation characterizes the controlled responses of a frozen recognition model to specified sensor restorations under predefined fault families and severities. Within this scope, the results demonstrate that calibrated recognition, targeted evidence acquisition, and conservative deferral can form a unified perception–action loop that reduces unsafe automation under sensor degradation.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-27

The authors' main claim, as the classifier picked it out 1.00

“Within this scope, the results demonstrate that calibrated recognition, targeted evidence acquisition, and conservative deferral can form a unified perception--action loop that reduces unsafe automation under sensor degradation.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.04 · states limitation 0.83 · human participants 0.60 · controlled comparison 0.95 · randomized assignment 0.03 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.02

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Autonomous IoT systems rely on distributed sensing, but sensor faults and communication loss can leave multichannel evidence incomplete or unreliable. background 1.00 0.05 0.02 asserted 0.95
Their agents must therefore classify measurements, assess evidence sufficiency, select a sensor to reacquire, and hand off unsafe decisions. background 0.95 0.07 0.04 asserted 0.98
We formulate counterfactual sensing: a degraded window defines the factual state, each candidate restoration defines a controlled alternative, and a pre-request utility model estimates loss reduction. method 0.77 0.83 0.06 asserted 0.98
Physics-Guided Invariant-Kinematic Fusion (PGIK-Fusion) constructs class evidence, while calibrated risk and cost-sensitive utility select acceptance, sensor request, or handoff from factual information. method 0.99 0.82 0.09 asserted 1.00
On the Logistic Activity Recognition Challenge dataset, PGIK-Fusion achieves 0.768 accuracy and a macro-averaged F1 score (Macro-F1) of 0.696, exceeding extreme gradient boosting (XGBoost)--one-dimensional convolutional neural network (1D-CNN) fusion by 2.49 and 2.00 percentage points, respectively. result 1.00 0.95 0.08 asserted 1.00
Across seven predefined degradation conditions on held-out subjects, the policy reduces unsafe acceptance from 33.0% to 13.8%; relative to an action-rate-matched random policy, it raises automation accuracy by 18.5 percentage points and lowers expected cost by 24.7%. result 1.00 0.95 0.16 asserted 1.00
In the auxiliary within-event wind-turbine study, PGIK-Fusion reaches 0.975 Macro-F1, improves on XGBoost--1D-CNN by 1.17 percentage points, and reduces unsafe acceptance from 20.1% to 8.0%. result 1.00 0.95 0.08 asserted 1.00
The evaluation characterizes the controlled responses of a frozen recognition model to specified sensor restorations under predefined fault families and severities. method 0.49 0.86 0.06 asserted 0.95
Within this scope, the results demonstrate that calibrated recognition, targeted evidence acquisition, and conservative deferral can form a unified perception--action loop that reduces unsafe automation under sensor degradation. interpretation 1.00 0.88 0.89 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from analysis of existing data 0.45

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-27: 1 request, 7,325 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (13)
Figure 1 of the paper
Figure 1. Representative operating scenarios for evaluating the autonomous
Figure 2 of the paper
Figure 2. End-to-end framework for uncertainty-guided counterfactual sensor selection and safety-gated recognition. The upper, middle, and lower layers denote
Figure 3 of the paper
Figure 3. Counterfactual sensor benefit and finite action selection: (a) validation
Figure 4 of the paper
Figure 4. Test Macro-F1 versus single-window CPU latency on S13–S14;
Figure 5 of the paper
Figure 5. Deterministic t-SNE projections of sampled test windows using (a)
Figure 6 of the paper
Figure 6. Row-normalized test confusion matrices for (a) raw XGBoost–1D
Figure 7 of the paper
Figure 7. Per-class test recall change from raw fusion to full PGIK-Fusion.
Figure 8 of the paper
Figure 8. Selective test risk as retained coverage increases for temperature
Figure 9 of the paper
Figure 9. Test recognition performance under controlled missingness, noise,
Figure 11 of the paper
Figure 11. Safety-policy comparison across seven degraded test conditions: (a)
Figure 12 of the paper
Figure 12. Illustrative one-step decisions under IoT sensor degradation: (a)
Figure 13 of the paper
Figure 13. Finite-action policy comparison across seven degraded MDWTBM

Showing 12 of the paper's 13 figures. Figures are the authors', reproduced from the paper.

References (61)
  1. [1] M. Wooldridge and N. R. Jennings, “Intelligent agents: theory and practice,”The Knowledge Engineering Review, vol. 10, no. 2, pp. 115– 152, 1995 on Pith
  2. [2] T. Duan, Z. Zhang, Z. Lin, S. Guo, X. Guan, G. Wu, Z. Fang, H. Meng, X. Du, J.-Z. Zhouet al., “LLM-Driven Stationarity-Aware Expert Demonstrations for Multi-Agent Reinforcement Learning in Mobile Systems,”arXiv preprint arXiv:2511.19368, 2025 on PitharXiv
  3. [3] D. Huang, M. Du, J. M. Zhang, Z. Lin, M. Luo, Q. Zhang, and S.- K. Ng, “Nexus: Execution-grounded multi-agent test oracle synthesis,” arXiv preprint arXiv:2510.26423, 2025 on PitharXiv
  4. [4] J. Xu, Z. Chen, S. Yang, J. Li, P. Wu, Z. Liu, Z. Lin, J. Tang, J. Yang, W. Wanget al., “Agents as knowledge integrator and utilizer in multimodal recommendation,”arXiv preprint arXiv:2608.29410, 2026 on PithPith's reviewarXiv
  5. [5] Y . Zhang, M. Hu, Z. Lin, X. Fan, F. Xie, Z. Fang, J. Yang, W. Zhu, Z. Chen, C. Lvet al., “Hera: Learning long-horizon co- ordination for device-cloud collaborative llm agents,”arXiv preprint arXiv:2605.24598, 2026 on PithPith's reviewarXiv
  6. [6] H. Zhang, X. Mao, G. Dong, Z. Li, X. Su, K. Chen, J. Yang, and Z. Lin, “Memmark: State-evolution attribution watermarking for agent long-term memory systems,”arXiv preprint arXiv:2605.25002, 2026 on PithPith's reviewarXiv
  7. [7] T. Duan, Z. Zhang, S. Guo, D. Huang, Y . Zhao, Z. Lin, Z. Fang, D. Luan, H. Cui, and Y . Cui, “Leed: A highly efficient and scalable llm-empowered expert demonstrations framework for multi-agent rein- forcement learning,”arXiv preprint arXiv:2509.14680, 2025 on PitharXiv
  8. [8] A. Murad, F. A. Kraemer, K. Bach, and G. Taylor, “Uncertainty- aware autonomous sensing with deep reinforcement learning,”Future Generation Computer Systems, vol. 156, pp. 242–253, 2024. IEEE INTERNET OF THINGS JOURNAL 14 on Pith
  9. [9] V . Bianchi, M. Bassoli, G. Lombardo, P. Fornacciari, M. Mordonini, and I. De Munari, “IoT Wearable Sensor and Deep Learning: An Integrated Approach for Personalized Human Activity Recognition in a Smart Home Environment,”IEEE Internet of Things Journal, vol. 6, no. 5, pp. 855... on Pith
  10. [10] T. Yang and C. Lv, “A Secure Sensor Fusion Framework for Connected and Automated Vehicles Under Sensor Attacks,”IEEE Internet of Things Journal, vol. 9, no. 22, pp. 22 357–22 365, 2022 on Pith
  11. [11] Y . Ovadia, E. Fertig, J. Ren, Z. Nado, D. Sculley, S. Nowozin, J. V . Dillon, B. Lakshminarayanan, and J. Snoek, “Can you trust your model’s uncertainty? Evaluating predictive uncertainty under dataset shift,” in Advances in Neural Information Processing Systems, vol. 32. Cur... on Pith
  12. [12] H.-S. Lee, S.-H. Park, and S.-Y . Kim, “Cost-Efficient Sensing With Self- Improving Soft Sensor Based on Predictive Uncertainty Estimation in IoT Environments,”IEEE Internet of Things Journal, vol. 13, no. 2, pp. 3055–3068, 2026 on Pith
  13. [13] F. Niemann, C. Reining, F. Moya Rueda, N. R. Nair, J. A. Steffens, G. A. Fink, and M. ten Hompel, “LARa: Creating a Dataset for Human Activity Recognition in Logistics Using Semantic Attributes,”Sensors, vol. 20, no. 15, p. 4083, 2020 on Pith
  14. [14] F. Niemann, C. Reining, F. Moya Rueda, N. R. Nair, P. Oberdiek, H. Bas, R. Spiekermann, E. Altermann, J. A. Steffens, G. A. Fink, and M. ten Hompel, “Logistic Activity Recognition Challenge (LARa Version 03) – A Motion Capture and Inertial Measurement Dataset,” 2023 on Pith
  15. [15] T. Li, C. Li, Y . Qin, L. Tan, B. Jiang, and H. Deng, “Multimodal dataset for wind turbine blade monitoring during lightning strikes,”Scientific Data, vol. 12, no. 1, p. 1524, 2025 on Pith
  16. [16] Z. Lin, G. Qu, X. Chen, and K. Huang, “Split Learning in 6G Edge Networks,”IEEE Wirel. Commun., 2024 on Pith
  17. [17] Z. Fang, Q. Wang, H. An, Z. Lin, Y . Deng, X. Chen, and Y . Fang, “Aggregation alignment for federated learning with mixture-of-experts under data heterogeneity,”arXiv preprint arXiv:2603.21276, 2026 on PithPith's reviewarXiv
  18. [18] Z. Lin, Y . Zhang, Z. Chen, Z. Fang, C. Wu, X. Chen, Y . Gao, and J. Luo, “LEO-Split: A Semi-Supervised Split Learning Framework over LEO Satellite Networks,”IEEE Trans. Mobile Comput., 2025 on Pith
  19. [19] G. Qu, Q. Chen, W. Wei, Z. Lin, X. Chen, and K. Huang, “Mobile edge intelligence for large language models: A contemporary survey,”IEEE Communications Surveys & Tutorials, 2025 on Pith
  20. [20] Z. Lin, G. Qu, W. Wei, X. Chen, and K. K. Leung, “Adaptsfl: Adaptive Split Federated Learning in Resource-Constrained Edge Networks,” IEEE Trans. Netw., 2025 on Pith
show 41 more
  1. [21] J. Tan, X. Cai, Z. Lin, R. Zhu, Z. Fang, P. Chen, and W. Ni, “Exploit- ing label-aware channel scoring for adaptive channel pruning in split learning,”IEEE Communications Letters, 2026 on Pith
  2. [22] W. Wei, Z. Lin, X. Liu, H. Du, D. Niyato, and X. Chen, “Optimizing split federated learning with unstable client participation,”arXiv preprint arXiv:2509.17398, 2025 on PithPith's reviewarXiv
  3. [23] Z. Lin, W. Wei, Z. Chen, C.-T. Lam, X. Chen, Y . Gao, and J. Luo, “Hi- erarchical Split Federated Learning: Convergence Analysis and System Optimization,”IEEE Trans. Mobile Comput., 2025 on Pith
  4. [24] Z. Fang, M. Yang, Z. Lin, Z. Lin, Z. Fang, Z. Zhang, T. Duan, D. Huang, and S. Zhu, “Nsc-sl: A bandwidth-aware neural subspace compression for communication-efficient split learning,”arXiv preprint arXiv:2602.02696, 2026 on PitharXiv
  5. [25] N. Y . Hammerla, S. Halloran, and T. Pl ¨otz, “Deep, Convolutional, and Recurrent Models for Human Activity Recognition Using Wearables,” inProceedings of the Twenty-Fifth International Joint Conference on Artificial Intelligence. AAAI Press, 2016, pp. 1533–1540 on Pith
  6. [26] J. Lu, X. Zheng, M. Sheng, J. Jin, and S. Yu, “Efficient Human Activity Recognition Using a Single Wearable Sensor,”IEEE Internet of Things Journal, vol. 7, no. 11, pp. 11 137–11 146, 2020 on Pith
  7. [27] S. Li, T. Zhu, M. Nie, H. Ning, Z. Liu, and L. Chen, “P2LHAP: Wearable-Sensor-Based Human Activity Recognition, Segmentation, and Forecast Through Patch-to-Label Seq2Seq Transformer,”IEEE Internet of Things Journal, vol. 12, no. 6, pp. 6818–6830, 2025 on Pith
  8. [28] S. Li, T. Zhu, F. Duan, L. Chen, H. Ning, C. Nugent, and Y . Wan, “HARMamba: Efficient and Lightweight Wearable Sensor Human Ac- tivity Recognition Based on Bidirectional Mamba,”IEEE Internet of Things Journal, vol. 12, no. 3, pp. 2373–2384, 2025 on Pith
  9. [29] H. Yuan, Z. Chen, Z. Lin, J. Peng, Z. Fang, Y . Zhong, Z. Song, and Y . Gao, “SatSense: Multi-Satellite Collaborative Framework for Spectrum Sensing,”IEEE Trans. Cogn. Commun. Netw., 2025 on Pith
  10. [30] Z. Fang, Z. Lin, S. Hu, Y . Tao, Y . Deng, X. Chen, and Y . Fang, “Dynamic uncertainty-aware multimodal fusion for outdoor health monitoring,” arXiv preprint arXiv:2508.09085, 2025 on PithPith's reviewarXiv
  11. [31] H. Yuan, Z. Chen, J. Peng, F. Tian, P. Du, X. Xu, and Y . Gao, “Neural fitting for sparse radio map construction in leo satellite network,”IEEE Transactions on Mobile Computing, 2026 on Pith
  12. [32] S. Mekruksavanich, D. Tancharoen, and A. Jitpattanakul, “Human Activ- ity Recognition in Logistics Using Wearable Sensors and Deep Residual Network,” inTENCON 2023 – 2023 IEEE Region 10 Conference (TENCON). IEEE, 2023, pp. 194–198 on Pith
  13. [33] L. Breiman, “Random Forests,”Machine Learning, vol. 45, no. 1, pp. 5–32, 2001 on Pith
  14. [34] T. Chen and C. Guestrin, “XGBoost: A Scalable Tree Boosting System,” inProceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining. ACM, 2016, pp. 785–794 on Pith
  15. [35] G. Ke, Q. Meng, T. Finley, T. Wang, W. Chen, W. Ma, Q. Ye, and T.-Y . Liu, “LightGBM: A Highly Efficient Gradient Boosting Decision Tree,” inAdvances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017, pp. 3146–3154 on Pith
  16. [36] S. Hochreiter and J. Schmidhuber, “Long Short-Term Memory,”Neural Computation, vol. 9, no. 8, pp. 1735–1780, 1997 on Pith
  17. [37] K. Cho, B. van Merri ¨enboer, C ¸ . G¨ulc ¸ehre, D. Bahdanau, F. Bougares, H. Schwenk, and Y . Bengio, “Learning Phrase Representations using RNN Encoder–Decoder for Statistical Machine Translation,” inProceed- ings of the 2014 Conference on Empirical Methods in Natural Langua... on Pith
  18. [38] A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is All you Need,” inAdvances in Neural Information Processing Systems, vol. 30. Curran Associates, Inc., 2017, pp. 5998–6008 on Pith
  19. [39] S. Bai, J. Z. Kolter, and V . Koltun, “An Empirical Evaluation of Generic Convolutional and Recurrent Networks for Sequence Modeling,”arXiv preprint arXiv:1803.01271, 2018 on PithPith's reviewarXiv
  20. [40] Z. Wang and T. Oates, “Imaging Time-Series to Improve Classification and Imputation,” inProceedings of the Twenty-Fourth International Joint Conference on Artificial Intelligence. AAAI Press, 2015, pp. 3939– 3945 on Pith
  21. [41] H. Wu, T. Hu, Y . Liu, H. Zhou, J. Wang, and M. Long, “TimesNet: Temporal 2D-Variation Modeling for General Time Series Analysis,” in International Conference on Learning Representations, 2023 on Pith
  22. [42] D. Luo and X. Wang, “ModernTCN: A Modern Pure Convolution Structure for General Time Series Analysis,” inInternational Conference on Learning Representations, 2024 on Pith
  23. [43] X. Chen, P. Qiu, W. Zhu, H. Li, H. Wang, A. Sotiras, Y . Wang, and A. Razi, “TimeMIL: Advancing Multivariate Time Series Classification via a Time-aware Multiple Instance Learning,” inProceedings of the 41st International Conference on Machine Learning, ser. Proceedings of Mac... on Pith
  24. [44] T. N. Kipf and M. Welling, “Semi-Supervised Classification with Graph Convolutional Networks,” inInternational Conference on Learning Representations, 2017 on Pith
  25. [45] J. Willard, X. Jia, S. Xu, M. Steinbach, and V . Kumar, “Integrating Scientific Knowledge with Machine Learning for Engineering and Environmental Systems,”ACM Computing Surveys, vol. 55, no. 4, pp. 1–37, 2022 on Pith
  26. [46] C. Guo, G. Pleiss, Y . Sun, and K. Q. Weinberger, “On Calibration of Modern Neural Networks,” inProceedings of the 34th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 70. PMLR, 2017, pp. 1321–1330 on Pith
  27. [47] R. El-Yaniv and Y . Wiener, “On the Foundations of Noise-free Selective Classification,”Journal of Machine Learning Research, vol. 11, pp. 1605–1641, 2010 on Pith
  28. [48] Y . Geifman and R. El-Yaniv, “SelectiveNet: A Deep Neural Network with an Integrated Reject Option,” inProceedings of the 36th Interna- tional Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol. 97. PMLR, 2019, pp. 2151–2159 on Pith
  29. [49] L. F. Cattelan and D. Silva, “How to Fix a Broken Confidence Esti- mator: Evaluating Post-hoc Methods for Selective Classification with Deep Neural Networks,” inProceedings of the Fortieth Conference on Uncertainty in Artificial Intelligence, ser. Proceedings of Machine Learni... on Pith
  30. [50] G. Shafer and V . V ovk, “A Tutorial on Conformal Prediction,”Journal of Machine Learning Research, vol. 9, pp. 371–421, 2008 on Pith
  31. [51] A. N. Angelopoulos and S. Bates, “Conformal Prediction: A Gentle Introduction,”Foundations and Trends in Machine Learning, vol. 16, no. 4, pp. 494–591, 2023 on Pith
  32. [52] D. B. Rubin, “Estimating causal effects of treatments in randomized and nonrandomized studies,”Journal of Educational Psychology, vol. 66, no. 5, pp. 688–701, 1974. IEEE INTERNET OF THINGS JOURNAL 15 on Pith
  33. [53] J. Pearl, “The Seven Tools of Causal Inference, with Reflections on Machine Learning,”Communications of the ACM, vol. 62, no. 3, pp. 54–60, 2019 on Pith
  34. [54] B. Sch ¨olkopf, F. Locatello, S. Bauer, N. R. Ke, N. Kalchbrenner, A. Goyal, and Y . Bengio, “Toward Causal Representation Learning,” Proceedings of the IEEE, vol. 109, no. 5, pp. 612–634, 2021 on Pith
  35. [55] A.-H. Karimi, B. Sch ¨olkopf, and I. Valera, “Algorithmic Recourse: From Counterfactual Explanations to Interventions,” inProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency. ACM, 2021, pp. 353–362 on Pith
  36. [56] J. Janisch, T. Pevn ´y, and V . Lis´y, “Classification with Costly Features Using Deep Reinforcement Learning,”Proceedings of the AAAI Confer- ence on Artificial Intelligence, vol. 33, no. 01, pp. 3959–3966, 2019 on Pith
  37. [57] C. Ma, S. Tschiatschek, K. Palla, J. M. Hern ´andez-Lobato, S. Nowozin, and C. Zhang, “EDDI: Efficient Dynamic Discovery of High-Value Information with Partial V AE,” inProceedings of the 36th International Conference on Machine Learning, ser. Proceedings of Machine Learning R... on Pith
  38. [58] Y . Li and J. Oliva, “Active Feature Acquisition with Generative Surro- gate Models,” inProceedings of the 38th International Conference on Machine Learning, ser. Proceedings of Machine Learning Research, vol on Pith
  39. [59] A. Karpatne, G. Atluri, J. H. Faghmous, M. Steinbach, A. Banerjee, A. Ganguly, S. Shekhar, N. Samatova, and V . Kumar, “Theory-Guided Data Science: A New Paradigm for Scientific Discovery from Data,” IEEE Transactions on Knowledge and Data Engineering, vol. 29, no. 10, pp. 231... on Pith
  40. [60] F. J. Ord ´o˜nez and D. Roggen, “Deep Convolutional and LSTM Recur- rent Neural Networks for Multimodal Wearable Activity Recognition,” Sensors, vol. 16, no. 1, p. 115, 2016 on Pith
  41. [139] PMLR, 2021, pp. 6450–6459 on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment