Memory as transformation: LETHE, a self-referential gan-inspired architecture

Francesco Vitucci; Anthony Di Furia; Francesco Scagliola

cs.CLSep 3, 2026

Memory as transformation: LETHE, a self-referential gan-inspired architecture

Authors: Francesco Vitucci, Anthony Di Furia, Francesco Scagliola

Organizations: Conservatorio di Musica “N. Piccinni” di Bari

Abstract

LETHE (Latent-parameter Evolution with Temporal Hierarchical quasi-Equilibrium) is a self-referential sonic-oblivion system implemented in SuperCollider. It adopts the formal vocabulary of Generative Adversarial Networks in a closed configuration without external datasets or supervision after initialization. Audio is processed by a 3 x 3 mixing matrix built around two delay lines; its nine coefficients and two delay times evolve through the interaction of a five-feature linear discriminator and a random-perturbation optimizer analogous to single-sample REINFORCE. The discriminator compares current energy behavior with an archive of the initial state and guides parameter updates. Circular, fixed, and live sources can be mixed independently. Across fixed and circular sessions with an ablation control, the active generator is necessary for parametric evolution ( $Δc_{22}=0.000$ in all 15 ablation sessions). Situated in the tradition of self-referential electroacoustic music, LETHE delegates the sonic outcome to an adaptive closed loop whose parametric space is defined by the composer.

Figures & tables

Panel of Figure 1 — Figure 1: General scheme of LETHE. The closed loop receives no external signals after the initialization phase.

Parameter	Default	Effect
bufDur	10.0 s	Buffer and warm-up duration
feedbackGBuffer	0.75	Life of the system
warmUpCycles	auto	Initial observation cycles
archiveSize	100	DNA capacity
archiveThreshold	0.55	DNA selectivity
dLR	0.005	D learning rate

Table 1: Main configurable parameters of LETHE with default values.

Metric	Fixed GAN	Ablation	Circular GAN
$\mu$ score $D$	0.441	0.446	0.433
$\sigma$ reward (WN)	0.061	0.065	0.089
Corr base $c_{22}/c_{33}$ (WN)	$-$ 0.995	n.a.	$-$ 0.969
$\Delta$ base $c_{22}$	$+$ 0.102 $\pm$ 0.067	0.000	$+$ 0.101 $\pm$ 0.081

Table 2: Comparison among the three conditions.

\mu

score

D

and

\Delta

base

c_{22}

: aggregate mean over

n=15

sessions (5 signals

\times

3 repetitions), common window of 180 s.

\sigma

reward and correlation

c_{22}/c_{33}

: representative WN session.

Figure 2: $\Delta c_{22}$ : Fixed GAN vs. Fixed Ablation for all signals ( $n=3$ , error bars $=\pm 1$ std. dev.). The marker * indicates $p<0.05$ in the one-sample $t$ -test (df $=2$ ). Ablation at zero for all signals: the Fixed Ablation bars are not visually distinguishable from the axis because their height is exactly zero by construction ( $\eta_{G}=0$ ), not because data is missing. The inset reports the aggregate statistics ( $n=15$ ).

Figure 3: Evolution of $c_{22}$ and $c_{33}$ (distinguished by line shade, see in-figure legend) for the WN signal in the three conditions. The anti-correlation is present in both GAN conditions; the ablation condition remains flat at 0.600 for the entire session.

Music Generation

Music generation studies computational methods for composing, synthesizing, and editing music from symbolic, audio, textual, and multimodal inputs.

See all 107 papers

Libretto: Giving LLM Agents a Sense of Musical Structure
Jun 21, 2026Yichen XuControllable Music Generation Symbolic Music Generation
University of California, Berkeley
Architecture and Affordances of PLAUD: Performative Latents and Unsupervised DDSP
Aug 13, 2026Błażej Kotowski, Frederic FontAudio Generation
Music Technology Group Universitat Pompeu Fabra Barcelona
LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling and Progressive Post-Training
Jun 29, 2026Shun Lei, Huaicheng Zhang, Dapeng Wu +8Text-to-Music Generation Controllable Music Generation
Tsinghua-CUHK Joint Research Center for Media Sciences, Technologies and Systems, Shenzhen International Graduate School, Tsinghua University, Shenzhen, China · Wuhan University, Wuhan, China · Hong Kong Polytechnic University, Hong Kong, China +1

Jun 21, 2026J/K move · Enter open · S save

Libretto: Giving LLM Agents a Sense of Musical Structure

Yichen Xu

University of California, Berkeley

Controllable Music Generation Symbolic Music Generation Music Generation

Open paper AskPDF arXiv

Generative music systems can now produce impressive audio from text prompts, but audio outputs are difficult to inspect, edit, and diagnose as musical structure. We introduce Libretto, an agent-facing framework for symbolic music generation and revision. Libretto uses an LLM-native grammar with explicit onset slots, voices, and bar-level organization, then evaluates each piece in a corpus-calibrated statistical space over rhythm, harmony, melody, texture, form, and variation. The same structural axes support retrieval, diagnosis, copy-risk control, and iterative self-revision. Across gap filling, reference-guided full-piece generation, gradual morphing, and educational music generation, Libretto turns symbolic music from a raw token sequence into a measurable and editable object for language-model agents.

Abstract

Music Generation

Explore similar work

Libretto: Giving LLM Agents a Sense of Musical Structure

Architecture and Affordances of PLAUD: Performative Latents and Unsupervised DDSP

LeVo 2: Stable and Melodious Song Generation via Hierarchical Representation Modeling and Progressive Post-Training