DuSAR: Dual-Strategy & Dual-Frequency Advances

• DuSAR is a dual framework that fuses co-adaptive LLM reasoning with dual-frequency radar techniques for enhanced decision making and lunar imaging.
• The AI component employs a holistic planning and local policy loop that outperforms retrieval-based methods by achieving up to 37.1% success on OOD tasks and reducing token usage by 3–9×.
• The radar system leverages simultaneous L-band and S-band polarimetric modes to deliver meter-scale resolution and radiometric accuracy within <2 dB, advancing lunar remote sensing.

DuSAR (Dual-Strategy Agent with Reflecting or Dual-Frequency Synthetic Aperture Radar, context-dependent) refers to two unrelated but notable advances in their respective fields: a framework for reasoning-driven LLM agents in AI research, and the Chandrayaan-2 Dual-Frequency Synthetic Aperture Radar (DFSAR) system for planetary remote sensing. Both are documented in the technical literature and share a focus on leveraging dual, co-adaptive strategies—cognitive or spectral—to achieve capabilities or insights not accessible via single-strategy or single-frequency approaches (Zhang et al., 9 Dec 2025, Bhiravarasu et al., 2021).

1. DuSAR in LLM-Based Agent Decision Making

DuSAR is a demonstration-free framework enabling a single frozen LLM to perform co-adaptive reasoning using two complementary strategies: a high-level holistic planning strategy and a context-grounded local policy (Zhang et al., 9 Dec 2025). The architecture internalizes human-like metacognitive planning by fusing global decomposition with local, adaptive action selection and continuous progress assessment via a lightweight reflection loop.

1.1 Motivation and Conceptual Foundation

Prevailing approaches, such as retrieval-augmented planning (e.g., Synapse, TRAD), are brittle under environmental shift, heavily dependent on long prompt demonstrations (1.5–3.6 k tokens/step), and primarily tied to proprietary APIs, limiting open-weight deployment. DuSAR is inspired by hierarchical and self-monitoring aspects of human cognition, embedding "two voices" (holistic and local strategies) in the agent's reasoning process.

1.2 Formalization

Let $I$ denote the task instruction and $E_{<t}$ the explore trace up to step $t-1$.

• Holistic Strategy ($H_t$): Maintains an ordered list of sub-goals; updated only on agent stagnation ($s_{t-1}=0$) or after milestone achievement ($50 \leq s_{t-1} \leq 99$), otherwise preserved. Formally,

$$H_t = \begin{cases} \operatorname{Ref}_h(I, E_{<t}, H_{t-1}), & \text{if } s_{t-1}=0 \text{ or } 50 \leq s_{t-1} \leq 99 \ H_{t-1}, & \text{if } 1 \leq s_{t-1} \leq 49 \ \text{Terminate}, & \text{if } s_{t-1}=100 \end{cases}$$

• Local Strategy ($L_t$): Conditioned on $o_t$, $H_t$, and $E_{<t}$0, proposes context-aligned actions and emits reasoning logs.
• Strategy Fitness Score ($E_{<t}$1): Scalar $E_{<t}$2, elicited solely via LLM prompting, encodes progress as:
  • $E_{<t}$3: Stagnation/invalid
  • $E_{<t}$4: Within sub-goal
  • $E_{<t}$5: Sub-goal completed
  • $E_{<t}$6: Task completion

1.3 Algorithmic Workflow

A co-adaptive loop evaluates and modifies $E_{<t}$7 or $E_{<t}$8 guided by $E_{<t}$9, as provided in explicit pseudocode. Efficiency is mandated through a trace window ($t-1$0), restricted token decoding, and deterministic generation (temperature=0).

2. DuSAR as Dual-Frequency SAR: System Design and Capabilities

The Chandrayaan-2 DFSAR ("DuSAR") is the first planetary SAR to operate two polarimetric frequencies (L-band 1.25 GHz, S-band 2.5 GHz) in simultaneous, meter-scale imaging, supporting both compact and full-polarimetric modes (Bhiravarasu et al., 2021). This enables unique lunar near-surface characterization unavailable to previous single-band systems.

2.1 Architecture

Band	Frequency (GHz)	Wavelength (cm)	Max. Bandwidth (MHz)	Antenna	Mass	Polarimetric Modes
L-band	1.25	~24	Up to 75	1.4 x 1.1 m microstrip	~20kg	Single, dual, compact, full
S-band	2.5	~12	Up to 75	Shared (with L-band)		as above

Independent transmit/receive chains for H and V polarization and onboard range compression yield a ~70% downlink reduction.

2.2 Polarimetric Modes

• Single-pol (HH or VV)
• Dual-pol (HH+HV or VV+VH)
• Compact-pol: Circular polarization transmit/linear receive
• Full-pol: Pulse-to-pulse H/V transmit, concurrent dual-pol receive (Sinclair matrix $t-1$1)

3. Calibration and Theoretical Foundations

3.1 Radiometric Calibration

Power calibration is achieved via the formula

$t-1$2

with normalized radar cross-section:

$t-1$3

Cross-validation with Mini-RF S-band data yields radiometric agreement within $t-1$4dB.

3.2 Polarimetric Calibration

Measured Sinclair matrices are corrected by

$t-1$5

where $t-1$6, $t-1$7 are gain/cross-talk matrices. An auto-calibration procedure achieves co- and cross-pol phase bias stability to $t-1$8 and amplitude imbalance $t-1$9 dB post-calibration.

3.3 Key Polarimetric Metrics

• Circular Polarization Ratio (CPR):

$H_t$0

Derived from linear-pol quantities and complex products, probing block-size and scattering regime.

• Entropy ($H_t$1) and Mean Alpha ($H_t$2):

$H_t$3

From eigen-decomposition of the coherency matrix, jointly characterizing randomness (H) and dominant scattering mechanism ($H_t$4).

4. Performance Characterization and Empirical Benchmarks

4.1 LLM-Based Agent (AI DuSAR)

• ALFWorld (134 OOD tasks):
  • Llama3.1-70B: 37.1% success vs. Synapse 13.0%, TRAD 9.9%
  • Llama3.1-8B: 11.7% vs. 0.0% (baselines)
• Mind2Web web environments:
  • Llama3.1-70B: 4.02% success vs. TRAD 1.96%, Synapse 0.37%
• Token efficiency: 335–564 tokens/step vs. 1.5–3.7k in retrieval-based baselines (3–9× reduction) (Zhang et al., 9 Dec 2025).

4.2 Chandrayaan-2 DuSAR

• Resolution: L-band CP mode achieves range/azimuth resolution of 1.9/2.3 m at 75 MHz.
• NESZ: Pre- vs post-launch (L-band FP) measured at −27.7 vs −27.9 dB (30° incidence).
• Radiometric accuracy: $H_t$52 dB difference in Mini-RF cross-comparison.
• Polarimetric accuracy: Phase bias stability $H_t$610°, amplitude imbalance $H_t$70.1 dB (Bhiravarasu et al., 2021).

5. Scientific and Applied Results

5.1 LLM Agent Applications

• Ablations reveal that the co-adaptive (holistic↔local) mechanism is essential; naive fusion or single-strategy variants perform $H_t$830% as well as the full model.
• Optional expert demonstration integration (HT/LT/BT) yields further improvement, particularly significant at smaller LLM scales.

5.2 Lunar Sciences: DFSAR DuSAR

• Permanently Shadowed Region Craters: L-band CPR$H_t$90.65–0.70 interior, 0.30 exterior; volume-dominated entropy/alpha signatures.
• Byrgius C: CPR$s_{t-1}=0$01.17, CPR$s_{t-1}=0$11.15 interior; ejecta CPR$s_{t-1}=0$20.45, CPR$s_{t-1}=0$30.70.
• Manzinus C crater trio: Fresh, partially degraded, and highly degraded craters exhibit monotonic interior/exterior CPR decline at both bands, evidencing co-evolution of block population and apparent crater age.
• Comparison with theoretical models constrains block/crack size distributions and the fraction of volume vs. surface/backscatter mechanisms.

6. Prospective Developments and Broader Significance

In the LLM domain, DuSAR establishes demonstration-free, token-efficient co-adaptive planning as a general solution for compositional and OOD generalization tasks, amenable to further augmentation by expert data or external knowledge (Zhang et al., 9 Dec 2025). In planetary radar, DFSAR’s dual-frequency, meter-resolution, and polarimetric versatility enable quantitative decompositions of lunar surface properties, advancing the state of planetary remote sensing for geology, age-dating, and volatile prospecting (Bhiravarasu et al., 2021). Ongoing work targets global mapping, inversion for dielectric properties, and exploitation of variable angle and frequency-polarization intercomparisons for granular material science.