DiSC-AMC: Efficient LLM-Based AMC

• DiSC-AMC is a hybrid symbolic-neural framework that reformulates automatic modulation classification as in-context reasoning over discretized higher-order signal statistics.
• The method employs a lightweight neural prefilter to prune exemplar candidates, drastically reducing prompt length while maintaining competitive accuracy.
• Experimental results show that DiSC-AMC achieves efficient, training-free AMC with reduced tokens and parameters compared to traditional plug-and-play LLM pipelines.

DiSC-AMC, short for Discretized Statistics In-Context Automatic Modulation Classification, is a prompt-engineered, LLM-based framework for automatic modulation classification that targets the practical bottlenecks of prior in-context AMC systems: long prompt contexts, large model sizes, and unstable exemplar selection. It reformulates AMC as a symbolic in-context reasoning problem by discretizing higher-order signal statistics and cumulants, pruning exemplars with a lightweight neural prefilter, filtering misleading or low-impact features using rationales extracted from prior LLM responses, and enforcing label-only predictions through a calibrated prompt template. In the reported synthetic ten-class setting, this combination reduces both input/output tokens and the model parameter footprint by more than half while maintaining competitive accuracy; a 7B DeepSeek-R1-Distill-Qwen baseline attains 5.2% accuracy, whereas DiSC-AMC with an approximately 5B-parameter Gemini-2.5-Flash model attains 45.5% accuracy (Rostami et al., 30 Sep 2025).

1. Conceptual setting and problem formulation

Automatic modulation classification seeks to identify the modulation scheme of a received complex baseband signal, written as $x[n] = I[n] + jQ[n]$. In the formulation adopted by DiSC-AMC, the central challenge is not only ordinary closed-set recognition, but open-set and training-free recognition with LLMs: the system is meant to classify signals without LLM fine-tuning and to remain practical for in-the-loop deployment despite the numerical and contextual burden of prompt-based inference (Rostami et al., 30 Sep 2025).

The method is explicitly positioned against prior plug-and-play LLM-based AMC pipelines that serialize handcrafted signal features as floating-point text and rely on long exemplar-rich prompts. That earlier regime preserves the attractive property of in-context learning, but it incurs approximately $2.9$K-token prompts in the reported setup and tends to require very large models to obtain usable performance. DiSC-AMC addresses these bottlenecks by reducing both the symbolic representation length and the candidate space presented to the LLM, while preserving the core idea that AMC can be performed as reasoning over examples rather than as task-specific weight adaptation (Rostami et al., 30 Sep 2025).

A common misconception is that DiSC-AMC is an end-to-end neural signal classifier over raw I/Q sequences. It is instead a hybrid symbolic-neural pipeline: handcrafted higher-order statistics provide the textual substrate for the LLM, a separate neural prefilter operates on a constellation diagram, and the final classifier is an in-context prompted LLM rather than a fine-tuned AMC backbone (Rostami et al., 30 Sep 2025).

2. Discretized statistical representation

The first stage of DiSC-AMC computes higher-order signal statistics and cumulants from the complex I/Q sequence. The paper describes the generic moments as

$m_{p,q} = \mathbb{E}[x^p (x^*)^q]$

and the higher-order cumulants as

$C_{p,q} = \kappa\big(x^p (x^*)^q\big).$

These descriptors inherit a long AMC tradition: different modulation families induce different symmetry and cumulant patterns, and the LLM is asked to reason over those patterns rather than over raw samples (Rostami et al., 30 Sep 2025).

The distinctive design choice is discretization. Instead of feeding floating-point values directly into the prompt, each scalar feature is mapped to one of $B$ bins. For a feature $f$ with range $[f_{\min}, f_{\max}]$, the bin width is

$\Delta = \frac{f_{\max} - f_{\min}}{B},$

with bin $b$ defined as

$[f_{\min} + (b-1)\Delta,\, f_{\min} + b\Delta).$

Each observed feature value is then encoded as a symbolic token such as a letter or an integer-coded symbol. The paper describes this as mapping each scalar to a symbolic token corresponding to one of $2.9$0 discrete bins. It also reports that the feature set is reduced from 21 continuous values to 17 discretized ones, and that SNR itself is discretized into a short symbolic code (Rostami et al., 30 Sep 2025).

This design has three concrete effects. First, it compresses prompt length by replacing verbose decimal strings with compact categorical symbols. Second, it normalizes feature scale, since the LLM no longer sees raw magnitude differences across heterogeneous statistics. Third, it encourages the model to reason over relative statistical regimes rather than over noisy numeric precision. A plausible implication is that DiSC-AMC reframes AMC as categorical pattern matching over higher-order-statistics signatures rather than as approximate numerical interpolation (Rostami et al., 30 Sep 2025).

3. Exemplar pruning and candidate shortlisting

In-context AMC depends strongly on which exemplars are present in the prompt. DiSC-AMC therefore introduces a k-top neural prefilter that restricts the prompt to a small candidate set before LLM inference. The shortlisting model takes the signal’s constellation diagram as input, uses a DINOv2 ViT-Base encoder, and predicts probabilities over ten modulation classes: 4ASK, 4PAM, 8ASK, 16PAM, CPFSK, DQPSK, GFSK, GMSK, OOK, and OQPSK (Rostami et al., 30 Sep 2025).

The prefilter is trained with standard cross-entropy loss,

$2.9$1

and returns a sorted class list $2.9$2. DiSC-AMC retains only the top-$2.9$3 candidates,

$2.9$4

and inserts only exemplars from those shortlisted classes into the prompt (Rostami et al., 30 Sep 2025).

This shortlisting step is critical both computationally and statistically. With $2.9$5, the paper reports 99.83% top-$2.9$6 accuracy, meaning the true class is almost always retained in the candidate set. At the same time, the prompt no longer needs to carry examples from all ten classes, which roughly halves exemplar burden. The paper’s ablations show that naive exemplar strategies are weak: deterministic centroid-based selection yields 8.63% accuracy, and random exemplar selection yields 16.47%, whereas the shortlist-driven strategy supports accuracies above 45% on the same overall task setting (Rostami et al., 30 Sep 2025).

4. Prompt template, rationale-based filtering, and inference protocol

The prompting stage is deliberately structured. The template contains three major segments: an instruction block with ROLE, OBJECTIVE, and CONTEXT; an exemplar block containing discretized statistics and labels for the shortlisted classes; and a final task block that presents the query signal’s discretized statistics together with the candidate label list. The exemplars use compact strings such as “snr: C, skewness: B, ...”, followed by an explicit answer label (Rostami et al., 30 Sep 2025).

A further refinement is rationale-based feature filtering. The paper states that rationales extracted from prior LLM responses are analyzed to identify features that the model rarely uses or interprets incorrectly. On that basis, low-impact fields such as nobs/min/max/mean/variance are removed, while higher-order statistics, cumulants, and discretized SNR are retained. This is not presented as a formal optimization algorithm; it is a practical prompt-design step guided by the LLM’s own explanatory behavior (Rostami et al., 30 Sep 2025).

The output side is also tightly controlled. DiSC-AMC enforces label-only predictions using a calibrated prompt template with explicit response rules. The model is instructed to place reasoning inside > ... </think> tags and, after ``, to emit only a single label from the provided classification options. This reduces output tokens to approximately one label token and stabilizes decoding. The inference regime is therefore neither unconstrained free-form generation nor supervised classification in the usual sense; it is constrained in-context multiple-choice reasoning over discretized AMC descriptors (Rostami et al., 30 Sep 2025).

5. Experimental setting and reported performance

The reported experiments use a synthetic AMC dataset with 10 digital modulation types, with 20 samples per class across SNR from $2.9$7 to $2.9$8. The shortlisting classifier is trained on synthetic constellation images across $2.9$9 to $m_{p,q} = \mathbb{E}[x^p (x^*)^q]$0. The overall evaluation protocol is one-shot ICL: for each query, the shortlisting model selects the top-$m_{p,q} = \mathbb{E}[x^p (x^*)^q]$1 candidate classes, one exemplar per candidate class is retrieved, and the LLM predicts one final label from the candidate list (Rostami et al., 30 Sep 2025).

The main reported comparisons are as follows.

Setting	Tokens	Accuracy
DeepSeek-R1-Distill-Qwen-7B baseline	~2.9K	5.20%
DiSC-AMC with DeepSeek-R1-Distill-Qwen-7B	~1.3K	33.50%
DiSC-AMC with Gemini-2.5-Flash (~5B)	~1.3K	45.50%
DiSC-AMC with Gemini-2.5-Pro	~0.9K	51.00%

These numbers support two distinct claims. First, the method is token-efficient: prompt length is reduced from more than 2,853 tokens to approximately 785–1,315 tokens, depending on $m_{p,q} = \mathbb{E}[x^p (x^*)^q]$2. Second, it is parameter-efficient: Gemini-2.5-Flash at approximately 5B parameters reaches 45.5% accuracy, which the paper presents as competitive with a 32B DeepSeek baseline at 47.8%, implying an approximately 84% reduction in parameters for similar overall performance (Rostami et al., 30 Sep 2025).

The ablations further indicate that prompt size and discretization granularity matter. For Gemini-2.5-Flash with 5 bins, the paper reports 44.50% at $m_{p,q} = \mathbb{E}[x^p (x^*)^q]$3, 45.50% at $m_{p,q} = \mathbb{E}[x^p (x^*)^q]$4, and a drop to 29.50% at $m_{p,q} = \mathbb{E}[x^p (x^*)^q]$5, where the prompt returns to baseline-like length near 2.9K tokens. Increasing bin granularity is not uniformly helpful: for Gemini-2.5-Flash, the best result is reported at 5 bins, whereas Gemini-2.5-Pro shows a non-monotonic profile with a peak around 10 bins in one ablation (Rostami et al., 30 Sep 2025).

6. Relation to adjacent AMC paradigms, scope, and limitations

The exact term DiSC-AMC is explicitly defined in "DiSC-AMC: Token- and Parameter-Efficient Discretized Statistics In-Context Automatic Modulation Classification" (Rostami et al., 30 Sep 2025). It should be distinguished from other AMC lineages in the literature. In particular, diffusion-based representation-learning approaches such as "ModFus-DM: Explore the Representation in Modulated Signal Diffusion Generated Models" use unsupervised generative pretraining on raw I/Q sequences and diffusion-aware feature fusion for downstream classification; that paradigm is a two-stage diffusion feature extractor plus lightweight classifier, not an LLM prompt-based in-context system (Tan et al., 3 Aug 2025).

This distinction matters because DiSC-AMC does not attempt to learn universal signal embeddings in the style of self-supervised or diffusion-based AMC. Its premise is instead that carefully chosen signal statistics, discretized into compact symbolic form and paired with dynamic exemplar selection, are sufficient for an LLM to perform useful modulation reasoning without task-specific fine-tuning. This suggests a different trade-off: weaker absolute accuracy than strong supervised AMC baselines, but much greater flexibility in prompt editing, candidate-set modification, and nominal open-set deployment (Rostami et al., 30 Sep 2025).

The paper also states several limitations. The experiments are conducted on synthetic data only; only 10 digital modulation types are considered; performance in extremely low SNR or under heavy fading is not deeply characterized; the results depend on the chosen LLM family, particularly the Gemini 2.5 models; open-set behavior with an explicit “unknown” label is discussed but not experimentally validated; and the rationale-based feature filtering procedure remains qualitative rather than automated (Rostami et al., 30 Sep 2025).

Within those limits, DiSC-AMC occupies a specific niche in AMC research: it is a training-free, prompt-centric, symbolically compressed, and candidate-pruned approach that treats modulation recognition as an in-context reasoning problem rather than as a conventional discriminative learning problem.