DeCoda: Multifaceted Decoding Framework

• DeCoda is a multifaceted framework suite that spans neural program decompilation, JavaScript malware detection, disentangled audio coding, and dialogue summarization.
• Its neural decompiler employs specialized tree-LSTM encoders, grammar-constrained AST decoders, and an iterative error correction loop to achieve state-of-the-art program recovery.
• The hybrid models leverage LLM-assisted deobfuscation and clustering techniques to enhance security, speech processing, and call-center dialogue analyses.

DeCoda is a name shared by several rigorously defined frameworks and datasets in disparate technical domains: neural code decompilation (as in "Coda," the end-to-end program decompiler), cluster-aware hybrid models for JavaScript deobfuscation and malware detection, universal disentangled audio codecs (DeCodec), and the corpus foundational to French call-center dialogue summarization research. Each usage reflects different origins, methodologies, and technical contributions but is unified by a focus on disentangling, decoding, or deobfuscating complex data representations.

1. DeCoda as an End-to-End Neural-Based Program Decompiler

DeCoda (originally "Coda") is an end-to-end neural framework for decompilation that addresses fundamental limitations of traditional binary code decompilers, notably language pair inflexibility, semantic drift, and poor interpretability. It operates in two principal phases: (1) code-sketch generation via an instruction-type–aware sequence encoder and grammar-constrained AST decoder, and (2) iterative error correction leveraging an ensembled neural error predictor (EP) validated by Levenshtein edit distance against the binary assembly. This yields an iterative, semantics-driven approach capable of state-of-the-art program recovery accuracy far exceeding conventional tools (Fu et al., 2019).

Key mathematical structures include:

• Specialized N-ary Tree-LSTM encoders per instruction type (mem, art, br), maintaining operand/type integrity.
• A tree-structured decoder with dual attention (parent context and input-instruction weights), growing left-child/right-sibling ASTs and enforcing syntactic well-formedness.
• The iterative error correction loop, accepting only edits that monotonically decrease edit distance (Δ(ϕ,ϕ″)≤Δ′).

On four synthetic C-program benchmarks, DeCoda achieves sketch token-accuracy of ~96.8% (vs. 82% for seq2seq+attention) and full-program recovery rates of ~82% (where existing decompilers score 0%) (Fu et al., 2019). For complex real-world binaries (e.g., PyTorch-CPP instantiations), 100% program recovery is reported.

2. DeCoda in Cluster-Aware Hybrid Defense for Malicious JavaScript

In the domain of security, DeCoda denotes a hybrid LLM+graph pipeline for detecting malicious JavaScript under heavy obfuscation (Liang et al., 30 Jul 2025). Its multi-stage prompt learning pipeline guides an LLM (DeepSeek-R1) through progressive deobfuscation:

• Stage 1: String/payload decoding (hex, base64, eval).
• Stage 2: Semantic variable renaming and control-flow simplification.
• Stage 3: Dynamic invocation and closure restoration.

The clean code is parsed into normalized ASTs, which are enriched with control/data-flow for hierarchical graph learning. A METIS-based clustering coarsens node groups, and a Cluster-wise Graph Transformer employs dual node-to-cluster attention to capture both local and global code semantics. The joint loss optimizes for classification (malicious/benign), cluster regularization, and deobfuscation preservation.

Empirical results show F1-score gains of 10.74–13.85% over baselines like BERT, CodeBERT, and GCN, with pronounced true-positive rate improvements under extremely low FPR constraints (4.82–13.09× higher TPR at very low FPR) (Liang et al., 30 Jul 2025).

3. DeCodec (DeCoda) as a Universal Disentangled Audio Codec

DeCodec (also referenced as DeCoda [Editor’s term]), reconceptualizes the neural audio codec as a universal, task-agnostic, disentangled representation learner (Luo et al., 11 Sep 2025). It factorizes a mixed waveform $y = s + n$ into orthogonal subspaces:

• Semantic speech tokens ($Z_{(c)}$)
• Paralinguistic tokens ($Z_{(r)}$)
• Background-sound tokens ($Z_{(n)}$)

A convolutional encoder feeds into a Subspace Orthogonal Projection (SOP) block (enforcing $P_s + P_n = I$, $P_s P_n^T = 0$), followed by parallel residual vector quantizers (RVQs) for speech and noise. A semantic guidance (SG) loss aligns the top-level speech representation to pretrained HuBERT features. Representation Swap Training (RST) uses mixed pairs to drive strict disentanglement.

Quantitatively, DeCodec outperforms baselines (EnCodec, DAC, SpeechTokenizer) in clean and noisy codec reconstruction, achieves DNSMOS OVL≅3.39 and BAK≅4.13 for speech enhancement, and delivers effective voice conversion under high environmental noise, with robust downstream ASR and TTS performance (Luo et al., 11 Sep 2025). Discrete token recombination enables selective denoising, voice conversion, or background suppression in a task-adaptive manner.

4. The DECODA Corpus for Call-Center Dialogue Summarization Research

The DECODA corpus is a large-scale French spoken dialogue dataset, recorded from public-transport call-center exchanges. Its most widely used subcorpora (DECODA-1/2/3) contain over 1500 annotated telephone conversations, with DECODA-3 providing richly human-annotated, multi-reference abstractive synopses (Zhou et al., 2023, Pontes, 2016). This resource is pivotal for evaluating both extractive and abstractive dialogue summarization models and associated Spoken Language Understanding (SLU) tasks.

Corpus statistics (Akani et al., 2024, Pontes, 2016):

Split	#Dialogs	Conv. Len	Sum. Len
Hum. (v1)	200	545	55.3
Aug. (v2)	1390	470	47.9
Test	200	496	52.7

The corpus supports a variety of protocols, from extractive summarization via graph-based sentence scoring (Pontes, 2016), to NLG-based abstractive systems, and as a tuning/evaluation set for faithfulness metrics (call-type accuracy, NE F1) in ASR-degraded settings (Akani et al., 2024).

5. DECODA in Dialogue Summarization: Methods and Metrics

DECODA’s task formulation centers on succinct, structured recaps adhering to three communicative guidelines: main issue identification, sub-issue recognition, and explicit resolution reporting (Zhou et al., 2023). Benchmarking approaches range from graph-based extractive methods (LIA-RAG, TF-ISF, sentence centrality) (Pontes, 2016), to modern pretrained transformers (BARThez), and prompt-engineered LLMs (GPT-4/ChatGPT) (Zhou et al., 2023).

Faithfulness in abstractive summarization is assessed (beyond ROUGE/BERTScore) via:

• CT-Acc: call-type classification accuracy of summaries,
• NE F1: agreement on named entities between summary and input,
• KL-divergence of call-type distributions in generation selection,
• NEHR: hallucination risk w.r.t. named entities not present in the source (Akani et al., 2024).

Injecting SLU signals—predicted call types and entity constraints—into generation and selection mitigates semantic hallucination. NEHR + D_{KL}-based summary selection further improves semantic fidelity (CT-Acc=0.82, NE-F1=0.44 on ASR input) (Akani et al., 2024).

6. Distinct Roles and Contributions Across the "DeCoda" Landscape

Despite nominal convergence, the term "DeCoda" (or "Coda", "DeCodec", "DECODA corpus") marks unrelated, technically rigorous innovations:

• In binary code analysis, DeCoda formalizes the first end-to-end neural decompiler with superior recovery accuracy and strict syntactic/semantic preservation (Fu et al., 2019).
• In software security, DeCoda advances LLM-assisted, cluster-aware graph modeling for robust JavaScript malware detection under obfuscation (Liang et al., 30 Jul 2025).
• In speech/audio, DeCodec establishes a new paradigm: learnable, multiply disentangled codecs enabling modularity and control in nearly all main audio AI tasks (Luo et al., 11 Sep 2025).
• As a dataset, DECODA grounds dialogue summarization and SLU faithfulness evaluation, supporting both extractive and abstractive techniques under real-world noise and annotation constraints (Pontes, 2016, Zhou et al., 2023, Akani et al., 2024).

This proliferation underscores the evolutionary trajectory of "decoding"-centric research in both symbolic and sub-symbolic domains, as well as the necessity for precise context when referencing "DeCoda" in the scholarly literature.