Source-linked AI summary
CSI-Agent: LLM-Assisted Few-Shot Adaptation for Cross-Domain Wi-Fi CSI Sensing
Tianya Zhao, Chuan Liu, Xuyu Wang
TL;DR
Cross-domain CSI sensing is brittle, and fixed adaptation may not suit scarce target labels or class-specific failures. CSI-Agent uses structured sensing evidence and an LLM planner to make verified deployment-time adaptation decisions; it improves target-domain performance across varied domain shifts.
Problem
Cross-domain shifts can make CSI sensing models brittle, while existing adaptation approaches use fixed procedures despite scarce target labels and heterogeneous class behavior.
Source: §I · p. 1Method
CSI-Agent uses sensing tools to form structured evidence, an LLM to plan class-conditioned adaptation, and a verifier to replace unreliable actions with a safer fallback.
Source: §I · p. 2Results
Experiments show that CSI-Agent consistently improves target-domain performance across various domain shifts.
Source: §VII · p. 9Takeaways & Limitations
The results support class-level action selection, query-aware evidence, and constrained planning for deployment-time cross-domain CSI adaptation.
Source: §VII · p. 9
Abstract
from arXiv · showhide
Wi-Fi channel state information (CSI) has enabled device-free sensing applications such as human activity recognition. However, CSI sensing models remain brittle in cross-domain deployment, where changes in users or environments can produce incorrect predictions. Existing solutions usually treat this problem as an offline model-design problem, by pretraining a stronger representation or applying one fixed adaptation method to the entire target domain. In practice, labeled target data are scarce and different classes may fail in different ways under the same domain shift. To address this, we propose CSI-Agent, an evidence-seeking LLM agent that reformulates cross-domain CSI adaptation as a deployment-time decision-making problem. Rather than processing raw CSI or making sample-level predictions, CSI-Agent summarizes target-domain behavior into sensing-grounded class-level evidence. It establishes a strong target-adaptive default from complementary CSI views and uses an LLM planner to determine whether each class should retain the default or invoke a specialized action. Deterministic verification and bounded execution further reduce unreliable interventions. We evaluate CSI-Agent on four public datasets using five cross-domain splits covering device, user, environment, and compositional shifts. Under 1-shot adaptation, CSI-Agent achieves the best target-domain performance across all splits and improves the average Macro-F1 by about 16\% compared to the strongest baseline method.
I. INTRODUCTION
CSI sensing can be fragile across deployment domains, while fixed few-shot adaptation may affect target domains and activity classes differently. CSI-Agent reframes adaptation as a deployment-time decision, using sensing-grounded evidence and an LLM planner to select class-conditioned actions subject to verification and fallback.
I. INTRODUCTION
- CSI distributions can change with environment, user, and transceiver configuration, making the same activity produce different patterns and activities less distinguishable in unseen domains.CSI is jointly shaped by human motion, environmental geometry, transceiver configuration, hardware, and multipath propagation.Source: §I · p. 1
- Fixed adaptation overlooks class-heterogeneous transfer effects, and applying a mismatched mechanism across a target domain can introduce negative transfer.Cross-domain models may remain highly confident even when predictions are incorrect.Source: §I · p. 1
- The paper frames adaptation as deployment-time diagnosis and control of complementary sensing actions under limited target-domain evidence, rather than fixed model optimization or a predetermined action.The system determines whether and how recovery should occur, while rejecting actions that may degrade sensing performance.Source: §I · pp. 1–2
- CSI-Agent converts raw CSI into structured diagnostics, plans class-conditioned adaptation with an LLM, and verifies proposed actions against reliability and no-harm criteria before execution.The LLM does not directly classify CSI; it coordinates sensing tools and produces auditable adaptation decisions.Source: §I · p. 2
- CSI-Agent was evaluated on four public Wi-Fi CSI datasets across five cross-domain splits, and the results show improved performance over fixed adaptation baselines.The splits cover device, user, environment, and compositional shifts; the system also reduces negative transfer and produces auditable decisions.Source: §I · p. 2
II. BACKGROUND AND MOTIVATION
The section introduces CSI as a sensing modality and motivates deployment-time adaptation through preliminary evidence that fixed methods vary across domains and classes. It also shows why LLM planning needs structured sensing evidence rather than raw confidence or CSI inputs.
A. Channel State Information
- CSI estimates the complex channel response across subcarriers and antenna links; human motion changes propagation paths and creates temporal CSI variations that encode activity information.Consecutive CSI packets form a time-frequency representation used for activity recognition.Source: A. Channel State Information · p. 2
B. Preliminary Study
- A fixed 1-shot PTN improves macro-F1 on Dev. 2 and Comp. 1 but reduces it on other targets, showing that its effect depends on the target domain.Dev. 2 is AmazonPlug and Comp. 1 is U02/E01 in CSI-Bench [24].Source: B. Preliminary Study · p. 3, Figure 1 · p. 3
- Within the same cross-device target domain, 1-shot ProtoNet can substantially improve some classes and considerably degrade others, making a single domain-level adaptation decision too coarse.The study motivates diagnosing individual classes and allowing different adaptation actions.Source: B. Preliminary Study · p. 3
- Among source predictions with confidence above 0.9, 49.6% to 69.2% are incorrect; serializing a target query set requires 530M to 1.4B tokens, motivating structured sensing evidence.One CSI sample alone requires approximately 112K tokens, and raw CSI does not explicitly expose class-level behavior needed for adaptation.Source: B. Preliminary Study · p. 3, Figure 2 · p. 3
III. RELATED WORK
Prior cross-domain Wi-Fi sensing work improves representations or adapts with few labeled target examples, generally using predetermined procedures; LLM sensing work explores prompting and tool use. The paper instead frames adaptation as evidence-driven deployment-time selection among predictors, under a closed-set setting with limited labeled support and unlabeled target observations.
A. Cross-Domain Wi-Fi Sensing
- Cross-domain sensing studies learn more invariant or robust representations, while few-shot methods use labeled target examples but generally follow predetermined post-deployment procedures.Examples include domain-invariant representations, physical features, semi-supervised adaptation, and one-shot or few-shot adaptation.Source: A. Cross-Domain Wi-Fi Sensing · p. 3
B. LLM-Based Sensing
- LLM-based sensing and signal-processing work includes physical-knowledge prompting, natural-language program generation, and agentic workflow construction with external tools.CSI-Agent distinguishes itself by grounding an LLM planner in target-domain diagnostics and using deterministic tools for processing, verification, and bounded execution.Source: B. LLM-Based Sensing · pp. 3–4
IV. PROBLEM FORMULATION
- The formulation assumes a known label space that remains unchanged across source and deployment domains, despite CSI distribution shifts from users, environments, or devices.At deployment, the system has a small labeled support set and an unlabeled target set whose labels must be inferred.Source: §IV · p. 4
- Rather than learn one adapted model, the system selects a predictor from a constrained family using support behavior and unlabeled-query statistics to maximize target-domain macro-F1.The hidden query labels are reserved for offline evaluation.Source: §IV · p. 4
V. SYSTEM DESIGN
CSI-Agent turns deployment-time adaptation into class-wise action selection: deterministic tools create diagnostic evidence, an LLM proposes actions, and a verifier and bounded executor constrain interventions. Its support-weighted combination of complementary CSI views serves as the target-adaptive default against which specialized actions are considered.
A. Overview
- The framework builds a deterministic sensing toolbox, summarizes class-level behavior in diagnostic cards, asks an LLM for class-wise actions, then verifies proposals and executes accepted actions relative to a common default.The overview figure presents this deployment-time pipeline.Source: A. Overview · p. 4, Figure 3 · p. 4
B. Sensing Action Toolbox
- An action is a fixed sensing scorer, not a free-form program, and each action produces a class-probability vector.The source action applies the frozen source classifier; it can work when source representations transfer well but may degrade under severe shifts.Source: B. Sensing Action Toolbox · p. 4
- CSI-Agent uses full, lowpass, and quarter-band CSI views; the lowpass view smooths temporal variation, while quarter-band retains a contiguous subcarrier quarter.A shared feature shift provides coarse global alignment, but may not align every class equally; reliable class-specific shifts are difficult to estimate from few labels.Source: B. Sensing Action Toolbox · pp. 4–5
- For each view, shifted-prototype distances yield support-side diagnostic probabilities, and a ridge classifier combines limited labeled target fitting with retained source class structure.The support probabilities assess each view's reliability on labeled examples.Source: B. Sensing Action Toolbox · p. 5
- The support-weighted action combines view-specific scorers by their labeled-support reliability, emphasizing more reliable views while retaining complementary evidence as a robust target-adaptive default.A small positive constant prevents any view from receiving exactly zero weight.Source: B. Sensing Action Toolbox · p. 5
C. Sensing-Grounded Diagnostic Evidence
CSI-Agent builds per-class, per-action diagnostic cards from labeled support and unlabeled target behavior, combining recognition, activation, risk, and viability evidence.
C. Sensing-Grounded Diagnostic Evidence
- Labeled support diagnostics measure class recall, mean class probability, mean rank, false-activation rate, and a one-vs-rest margin.Together, these capture class recognition and interference from competing classes.Source: C. Sensing-Grounded Diagnostic Evidence · p. 5
- Query activation rate measures each action’s unlabeled target predictions and is compared with the support-weighted default.A coarse uniform reference flags severe under- or over-activation, but is never sufficient by itself to accept an action.Source: C. Sensing-Grounded Diagnostic Evidence · p. 5
- Each diagnostic card records Support Fit, Activation Fit, Risk Flags, and a viability, weakness, or risk Verdict for each action.The fields summarize support separation, target activation changes, and risks such as increased false activation or conflicting evidence.Source: C. Sensing-Grounded Diagnostic Evidence · pp. 5–6
D. LLM Agent Planning and Verification
For each class, an LLM proposes a sensing action from its diagnostic card, and a deterministic verifier checks the proposal before execution.
D. LLM Agent Planning and Verification
- The LLM plans from class-specific diagnostic cards and candidate actions, without receiving raw CSI or producing sample-level predictions.It proposes a deterministic scorer and returns a structured diagnosis, failure mode, action, and evidence-grounded reason.Source: D. LLM Agent Planning and Verification · p. 6
- A verifier checks whether each proposed action matches the diagnosed failure and shows improvement without substantial false activation or reduced support separation.Rejected proposals fall back to the support-weighted action; candidate filtering and final verification are separate steps.Source: D. LLM Agent Planning and Verification · p. 6
E. Bounded Action Execution
Accepted actions modify a shared support-weighted default through bounded class-specific adjustments; rejected actions retain the default score.
E. Bounded Action Execution
- Accepted actions adjust the common support-weighted score with intervention strength β = 0.35, rather than necessarily replacing it.Rejected actions retain the support-weighted score, limiting negative transfer from uncertain proposals.Source: E. Bounded Action Execution · p. 6
- CSI-Agent predicts the class with the highest final score, preserving a common scoring reference across classes.The bounded adjustment limits negative transfer from uncertain action proposals.Source: E. Bounded Action Execution · p. 6
A. Experiment Setup
Across five cross-domain splits, CSI-Agent leads target-domain performance; its class-level action selection is more consistent than fixed or domain-wide choices.
A. Experiment Setup
- Experiments span four public CSI datasets and five device, user, environment, and compositional splits, with source-side, unlabeled-adaptation, and few-shot baselines.Few-shot methods use the same k-shot support set as CSI-Agent.Source: A. Experiment Setup · pp. 6–7
B. Cross-Domain Performance
- 0.4890 average Macro-F1, 0.1657 above FewSense, and the best target-domain performance on all five evaluation splits.CSI-Agent also surpasses the strongest baseline by 0.1121 on WiAR, 0.5051 on SignFi, and 0.0226 on WiMANS.Source: Table II · p. 7, B. Cross-Domain Performance · p. 7
C. Selective Action Analysis
- Fixed sensing actions vary in effect across target domains; no candidate action consistently improves on the support-weighted default.The source action helps one CSI-Bench device target but severely degrades SignFi; Low-pass helps some WiAR targets but reduces performance on both WiMANS targets.Source: Figure 4 · p. 7, C. Selective Action Analysis · p. 7
- CSI-Agent improves 12 of 13 target domains through class-level selection, with its only degradation limited to 0.2 percentage points.Domain-level planning is neutral on many targets and noticeably degrades two cases.Source: Figure 4 · p. 7, C. Selective Action Analysis · p. 8
- Fig. 5 examines support-size sensitivity, while Fig. 6 compares one-shot ablations against full CSI-Agent.The supplied figure captions identify the measures and comparisons, but do not state the outcomes.Source: Figure 5 · p. 8, Figure 6 · p. 8
- Class-specific action combinations yield positive gains on targets where none of the individual candidate actions improves over the default.CSI-Agent does not match the best fixed action on every individual domain, but is more consistent than any fixed action.Source: C. Selective Action Analysis · p. 8
D. Support-Size Sensitivity Analysis
CSI-Agent benefits from additional target support, query-aware class-level planning, and safety mechanisms, while showing similar performance across the tested LLM planners.
D. Support-Size Sensitivity Analysis
- Performance improves across all five evaluation cases as support grows from 1 to 7 shots, with the largest gains generally from 1-shot to 3-shot and diminishing returns at larger sizes.SignFi approaches saturation, and no split shows unstable or non-monotonic behavior as support increases.Source: D. Support-Size Sensitivity Analysis · p. 8
E. Ablation Study
- 45.96% average performance (∆= −2.93) without the LLM planner and 45.65% average MacroF1 (∆= −3.25) without query statistics show both support joint planning.Without query statistics, WiMANS drops by ∆= −12.55; support-only action selection is insufficient under heterogeneous target shifts.Source: E. Ablation Study · p. 8
- Removing the verifier degrades WiAR (∆= −2.3) and SignFi (∆= −1.3), while removing bounded execution degrades CSI-Bench Compositional (∆= −1.8) and WiMANS (∆= −1.4).Verification and bounded residual execution complement the planner by limiting the impact of unreliable interventions.Source: E. Ablation Study · p. 8
F. LLM Planner Robustness
- Average Macro-F1 ranges only from 48.57% to 48.90% at 1-shot and from 62.65% to 62.77% at 5-shot across tested LLM planners, with no planner consistently dominating all splits.The planners use the same prompt, evidence, and execution protocol; the approach uses class-level decisions over structured evidence and a constrained action space.Source: F. LLM Planner Robustness · pp. 8–9
G. Overhead Analysis
CSI-Agent uses class-level planning to reduce the cost of LLM-assisted adaptation relative to sample-level reasoning, and its decision traces illustrate class recovery across several shifts. The conclusion highlights consistent target-domain improvements and constrained, evidence-grounded planning.
G. Overhead Analysis
- 8.8×–547.0× lower latency and 3.3×–661.9× lower input-token cost than estimated sample-level reasoning make sensing-grounded aggregation more scalable.The complete pipeline takes 20.59–91.02 s per target domain; SignFi has the highest cost and a 47.67K-input-token class-level prompt.Source: G. Overhead Analysis · p. 9
H. Illustrative Decision Traces
- Three selected non-default interventions recover classes under device, user, and environment shifts: QBand on CSI-Bench Cross-Device, and Low-pass on WiAR cross-user and SignFi cross-environment.Outcomes use hidden query labels only after actions are fixed; the condensed rationales are auditable references, not faithful causal explanations.Source: Table V · p. 9, H. Illustrative Decision Traces · p. 9
VII. CONCLUSION
- CSI-Agent consistently improves target-domain performance across shifts, with benefits from class-level action selection and query-aware evidence, robustness across LLM backbones, and avoidance of costly sample-level inference.The conclusion suggests that practical cross-domain sensing can use LLM agents when decisions are grounded in structured sensing evidence and constrained by reliable execution mechanisms.Source: §VII · p. 9