Pith. sign in

REVIEW 2 major objections 6 minor 47 references

A dependency-free native runtime runs the full Stable Audio 3 pipeline on ordinary GPUs, CPUs, and a Raspberry Pi 5, with eight-bit precision showing no measurable quality loss against seed variation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · grok-4.5

2026-07-10 05:57 UTC pith:GRZ4MY3K

load-bearing objection Solid systems paper: first full-pipeline native SA3 runtime with careful in-place quantization and honest multi-oracle steering; main gap is automatic oracles only, which the authors already flag. the 2 major comments →

arxiv 2607.08526 v1 pith:GRZ4MY3K submitted 2026-07-09 cs.SD cs.PF

A Quantized Native Runtime for On-Device Semantic Audio Generation

Matteo Spanio , Antonio Rod\`a This is my paper
classification cs.SD cs.PF
keywords efficient inferencemodel quantizationedge computingaudio diffusionmusic generationactivation steeringon-devicesonic seasoning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Semantic audio tools need to run on hardware people own, not only on framework-heavy cloud stacks. This paper builds aria, a small C/CUDA runtime that executes Stable Audio 3’s entire text-to-music pipeline with no Python or deep-learning framework underneath. Its central study is quantization treated as an in-place memory replacement: once weights are packed to eight- or four-bit, the full-precision copy is freed. Eight-bit stays inside the ordinary variation between random seeds on three independent automatic checks—prompt adherence, distributional audio quality, and taste preservation—while cutting memory and becoming the fastest GPU mode; four-bit adds a bounded cost yet fits the 1.2-billion-parameter model on an 8 GB Pi. Because the runtime owns every tensor, activation steering is a built-in, zero-overhead control, demonstrated on taste associations (sonic seasoning) with genuine but limited success. The result is a practical basis for on-device controllable music generation.

Core claim

A compact quantized native runtime can deploy a full state-of-the-art latent-diffusion music model on commodity and embedded hardware so that eight-bit precision is indistinguishable from half precision under three independent automatic quality checks, matches or exceeds the official stack on warm generation, cold-starts roughly seven times faster, and exposes activation steering as a free in-graph primitive.

What carries the argument

The aria runtime: a ~7.7k-line dependency-free C/CUDA engine that owns the full SA3 pipeline, packs weights from half precision down to 4-bit while releasing the original so lower precision reduces resident memory in place, optionally runs 8-bit arithmetic on GPU integer tensor cores and ARM, and injects steering directions into residual stream, latent, or text conditioning inside a captured graph at zero measurable overhead.

Load-bearing premise

That staying inside the ordinary seed-to-seed variation on three automatic metrics is enough to call eight-bit “no measurable quality loss” for music audio without human listening confirmation.

What would settle it

A pairwise listening study in which listeners systematically prefer the half-precision clips over matched eight-bit clips for the same prompts and seeds, or rate eight-bit outputs as lower quality or less prompt-faithful beyond chance.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Eight-bit weight and activation quantization can be treated as effectively free for SA3-class music diffusion transformers under the three automatic checks used here.
  • The 1.2-billion-parameter medium model becomes runnable on an 8 GB Raspberry Pi 5 at four-bit precision.
  • Interactive and edge services gain most from a warm-resident native binary because cold start, GPU-context setup, and per-length recompilation dominate short generations.
  • Activation steering can ship as a built-in runtime feature rather than a trained adapter or external Python patch.
  • Genuine semantic control of hard-to-lexicalize attributes such as taste exists but is confined to a narrow strength window and only a subset of axes when independent oracles are required.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same in-place quantization plus dependency-free runtime pattern is likely to unlock other DiT-based audio and image generators currently locked behind heavy Python stacks.
  • Any evaluation that optimizes a learned regressor for activation steering should adopt multi-oracle gating (target plus independent semantic check plus degradation metric) as standard practice.
  • If listening studies later confirm the automatic floors, on-device sonic seasoning becomes feasible for real-time gastronomy or assistive audio products without cloud round-trips.
  • Releasing the full-precision weights after packing is a simple systems habit other edge quantizers can copy to avoid the usual dual-copy memory tax.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

2 major / 6 minor

Summary. The paper presents aria, a dependency-free C/CUDA runtime that executes the full Stable Audio 3 text-to-music pipeline (tokenizer, T5Gemma encoder, DiT denoiser, SAME autoencoder) on commodity GPUs, CPU-only hosts, and a Raspberry Pi 5. Its central systems claim is a deployment-oriented quantization study: weights stored from fp16 down to 8- and 4-bit, with an optional W8A8 arithmetic path, releasing the full-precision copy so lower precision reduces rather than adds memory. Quality is gated against fp16 by three automatic metrics (CLAP prompt adherence, CLAP-embedding FAD, wav2taste L2) scaled to re-seed noise floors; 8-bit stays inside every floor and is the fastest GPU mode, while 4-bit enables the 1.2B medium model on an 8 GB Pi at a bounded cost. Against the official PyTorch stack (commit cited), aria matches or slightly exceeds warm generation and cold-starts ~7× faster (Table I). Because the runtime owns every tensor it also exposes in-graph activation steering; a sonic-seasoning case study shows genuine but narrow multi-oracle control for a subset of taste axes.

Significance. If the measurements hold, the work supplies a practical, open, single-binary path for on-device latent-diffusion music generation—an under-served niche relative to LLM and image runtimes. Strengths that should be credited explicitly include: (i) a clean warm/invocation/cold protocol with direct comparison to a pinned official commit, (ii) memory-replacement quantization rather than additive compression, (iii) three independent automatic fidelity checks scaled by re-seed floors (Table II, Fig. 4), (iv) a multi-oracle steering protocol that actively surfaces metric-gaming rather than hiding it, and (v) a released dependency-free codebase. These make the systems contribution reproducible and the control interface immediately usable for Internet-of-Sounds prototypes.

major comments (2)
  1. Section IV-A0e / Table II / Fig. 4: the strongest claim—“8-bit shows no measurable quality loss on any measure”—rests entirely on three automatic metrics staying inside fp16 re-seed floors. The paper itself flags perceptual confirmation as future work (V-C) and reports wav2taste’s imperfect held-out correlations (r≈0.59–0.82). Without at least a small listening study (or an explicit, narrower re-statement of the claim as “within automatic re-seed noise”), the certification of “no measurable quality loss” remains under-supported for a systems paper whose headline result is quality-preserving quantization.
  2. Section IV-B / Table I: the efficiency comparison cites a tuned official baseline (faster of default vs. options-enabled+compiled), yet the official path still uses the repository’s fallback attention and incurs per-length recompilation (14.5–48 s). A short sensitivity paragraph quantifying how much of the 7× cold-start and warm-parity advantage survives against a fully optimized reference (FlashAttention-class kernels, persistent server) would strengthen the systems claim; the current numbers are already useful but risk overstating the gap relative to a production-tuned stack.
minor comments (6)
  1. Figure 1 caption and abstract: “7× faster cold start” is clear; ensure the same factor is consistently reported for both model sizes in the body (I gives 7.2–7.7×).
  2. Table II: the re-seed floors (ΔCLAP ±0.004, FAD 44.0, Δtaste 0.153) are crucial; state the exact seed sets and sample size used to compute them in the caption or IV-A0e.
  3. Section III-B / Eq. (2): α is defined in units of mean residual norm; a one-sentence reminder that this makes α comparable across blocks and model sizes would help readers of the dense-window tables.
  4. Table V and IV-C0g: LoRA is a fair training-based comparator, but the trigger-token objective and 800-step budget could be briefly justified against a longer or caption-matched alternative so the negative result is not read as under-training.
  5. Typographical: “Rod `a” and “Fr ´echet” appear with stray spaces/accents in several places; normalize author names and FAD spelling throughout.
  6. Outlook (VI-A): the planned Bradley–Terry listening study is the right next step; stating the planned sample size or attribute set would make the limitation more concrete.

Circularity Check

1 steps flagged

No derivation-by-construction circularity; systems claims are direct measurements. Only minor residual self-citation risk in the scoped taste case study, already gated by independent oracles.

specific steps
  1. self citation load bearing [Section III-B / IV-C (taste case study); refs [40], [41]]
    "Directions are grounded in norm-sonic-seasoning [40] (377 human-rated clips). ... Our target is wav2taste [41], a learned audio-to-taste regressor. The effect on axis i is ∆i(α)=τ̂i(xα)−τ̂i(x0) against the matched α=0 baseline."

    The case-study attribute (basic tastes) and the optimized target oracle both come from prior work by the same authors. This is a mild self-citation dependency for the steering narrative only. It is not load-bearing for the paper's central systems/quantization claims, and the paper gates 'genuine' control on independent CLAP/FAD/drift rather than on wav2taste alone, so the reduction is partial and non-central rather than definitional.

full rationale

This is an engineering/systems paper whose load-bearing claims are empirical measurements, not first-principles derivations that could collapse into their inputs. Runtime efficiency (warm/invocation/cold latency, VRAM, Pi RSS, W8A8 speed) is measured against the official SA3 PyTorch stack under an explicit multi-regime protocol (Table I); nothing is fitted then re-reported as a prediction. Quantization fidelity is gated against fp16 re-seed noise floors on three metrics, deliberately leading with CLAP prompt adherence and CLAP-embedding FAD that share no training path with the taste oracle, with wav2taste L2 only as a narrow corroborator (Section IV-A0e, Table II, Fig. 4). The steering case study optimizes wav2taste but treats success as genuine only when independent CLAP rises with it at bounded FAD/drift, and the paper surfaces the failure mode where the target keeps climbing while checks collapse (Fig. 2, Table III). Self-citations to the authors' sonic-seasoning dataset [40] and wav2taste [41] supply the case-study attribute and one oracle, but they are not uniqueness theorems, do not force the systems results, and are not the sole evidence for the claimed genuine window. The paper itself scopes steering as a bounded case study and flags perceptual listening as future work. Under the circularity criteria, this is at most a minor non-load-bearing self-citation (score 1), not a reduction of prediction to input.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 2 invented entities

The systems claims rest mainly on engineering measurements and standard PTQ/runtime assumptions, not on new physics-like entities. Load-bearing soft assumptions are that automatic audio metrics within seed noise certify deployment quality, that difference-in-means residual directions encode the intended semantic attributes, and that the official SA3 comparison configuration is a fair baseline. Free parameters are mostly experimental knobs (α, layer, bit-width, LoRA rank) rather than fitted universal constants. Invented entities are essentially the runtime and the operational “genuine window” concept—not new physical mediators.

free parameters (6)
  • steering strength α (per axis / site)
    Hand-swept continuous scale in residual-norm units; operating points (e.g. α=0.1–0.3) are selected from multi-oracle windows, not derived.
  • injection layer ℓ per taste axis
    Nominated by target-only scan then retained only if multi-oracle panel agrees; discrete free choice of site.
  • quantization bit-width and W8A8 mode
    Deployment choices (fp16/q8/W8A8/q4) selected for memory/speed tradeoffs; not predicted from theory.
  • denoising step count (default 8; fast preset 6)
    Sampler hyperparameter fixed for benchmarks; fast preset claimed below seed noise but still a free schedule choice.
  • LoRA rank and training steps (r=8, 800 steps; larger r=32 ablations)
    Training-based comparator hyperparameters chosen by authors for cost comparison.
  • top/bottom-k contrast set size for audio-side directions
    Difference-in-means estimator depends on which rated clips define high/low taste poles.
axioms (6)
  • domain assumption Linear representation hypothesis: taste and related attributes are approximately linear directions in DiT residual (or latent) space, so additive injection shifts the attribute.
    Invoked throughout Section III-B and related-work steering citations; not proved for SA3 taste.
  • domain assumption CLAP prompt similarity, CLAP-embedding FAD, and wav2taste vector distance are adequate proxies for prompt adherence, distributional quality, and taste preservation when compared to re-seed floors.
    Section IV-A0e / Table II gate all quantization claims on these automatic metrics.
  • domain assumption Difference-in-means over contrastive sets yields transferable steering directions (audio-side preferred).
    Eq. (1) and extraction protocol; justified by prior steering literature, not re-derived.
  • domain assumption Official SA3 PyTorch path (cited commit, half precision, 8-step sampler, faster of default vs tuned config) is a fair efficiency baseline.
    Section IV-A0d efficiency protocol; fallback attention and compile-per-length behavior affect interpretation.
  • standard math Standard floating-point / integer matmul and graph-capture semantics preserve model behavior up to documented precision effects.
    Background for claiming bit-identical α=0 steering and windowed decoder byte-identity.
  • domain assumption Crossmodal sonic-seasoning literature correctly maps acoustic structure to basic-taste associations used as external ground.
    Section II-D / III-B; grounds the case-study semantics.
invented entities (2)
  • aria runtime (dependency-free SA3 C/CUDA engine) independent evidence
    purpose: Owns full pipeline tensors to enable in-place quantization, edge deployment, and in-graph steering without a framework.
    Primary artifact; independent evidence is the public code release and reported benchmarks, not a new physical object.
  • multi-oracle “genuine window” operating regime for taste steering no independent evidence
    purpose: Operational definition separating genuine semantic control from metric-gaming degradation.
    Methodological construct defined by joint rise of wav2taste and CLAP at low FAD/drift; useful but paper-defined.

pith-pipeline@v1.1.0-grok45 · 22432 in / 4081 out tokens · 57830 ms · 2026-07-10T05:57:07.424586+00:00 · methodology

0 comments
read the original abstract

Semantic audio applications increasingly require controllable generation on commodity and embedded hardware rather than through framework-heavy datacenter stacks. We present \textit{aria}, a dependency-free native runtime that runs the complete text-to-music pipeline of Stable Audio~3 (SA3) on ordinary GPUs, CPU-only machines, and a Raspberry~Pi~5, with no Python or deep-learning framework underneath. Our main contribution is a study of quantization: running the model at lower numerical precision to fit tight memory budgets, saving memory in place rather than adding to it. Because the runtime owns every internal tensor, it also exposes activation steering, a low-cost way to steer what the model generates. We judge the quality cost with three independent measures of the output (prompt adherence, overall audio quality, taste preservation), each compared against the ordinary variation between random seeds. Eight-bit precision shows no measurable quality loss on any measure while sharply cutting memory, and it is the fastest mode on the GPU; four-bit adds a small, bounded cost but shrinks the footprint enough to run the $1.2$-billion-parameter model on an $8$\,GB Pi. Against the official implementation, aria matches or exceeds generation speed and starts about seven times faster. A case study of the steering interface generates music carrying taste associations (\emph{sonic seasoning}), with genuine but bounded control for a subset of attributes. These results make a compact, quantized runtime with built-in control a practical basis for on-device semantic audio in Internet-of-Sounds settings. The \textit{aria} runtime is released at https://github.com/matteospanio/aria.

Figures

Figures reproduced from arXiv: 2607.08526 by Antonio Rod\`a, Matteo Spanio.

Figure 1
Figure 1. Figure 1: With aria, the same Stable Audio 3 checkpoints move from a framework-bound serving stack (top) to a single dependency-free binary on hardware people own (bottom): 7× faster cold start, warm parity, and both model variants on a Raspberry Pi 5. Owning the pipeline also makes activation steering a runtime feature, injected at 1 the DiT residual stream, 2 the latent, or 3 the text conditioning. as ONNX Runtime… view at source ↗
Figure 2
Figure 2. Figure 2: The degradation demonstration (SOUR, small-music, its best layer L19), as two stacked panels sharing α (no dual axis). Top: the steering target (wav2taste ∆, with a 95% bootstrap CI band over the paired per-clip deltas) and the independent CLAP ∆ on one axis; both rise inside the shaded low-α genuine window, then diverge as wav2taste stays high while CLAP collapses. Bottom: FAD explodes past the window. Th… view at source ↗
Figure 4
Figure 4. Figure 4: Quantization performance in one space (small-music; all GPU device [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

47 extracted references · 47 canonical work pages · 10 internal anchors

  1. [1]

    Stable audio open,

    Z. Evans, J. D. Parker, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, “Stable audio open,” inICASSP 2025 - 2025 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), April 2025, pp. 1–5

    work page 2025

  2. [2]

    Highly Optimized Kernels and Fine-Grained Codebooks for LLM Inference on Arm CPUs

    D. Gope, D. Mansell, D. Loh, and I. Bratt, “Highly optimized kernels and fine-grained codebooks for llm inference on arm cpus,” 2024. [Online]. Available: https://arxiv.org/abs/2501.00032

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  3. [3]

    Onnx runtime,

    O. R. developers, “Onnx runtime,” https://onnxruntime.ai/, 2021, ver- sion: 1.27.0

    work page 2021

  4. [4]

    Dwarfstar,

    S. Sanfilippo and contributors, “Dwarfstar,” https://github.com/antirez/ ds4, 2026, deepSeek 4 Flash and PRO local inference engine for Metal, CUDA and ROCm

    work page 2026

  5. [5]

    Ace- step 1.5: Pushing the boundaries of open-source music generation,

    J. Gong, Y . Song, W. Zhao, S. Wang, S. Xu, J. Guo, and X. Yang, “Ace- step 1.5: Pushing the boundaries of open-source music generation,”

    work page

  6. [6]

    Available: https://arxiv.org/abs/2602.00744

    [Online]. Available: https://arxiv.org/abs/2602.00744

    work page arXiv

  7. [7]

    Stable Audio 3

    Z. Evans, J. D. Parker, M. Rice, C. Carr, Z. Zukowski, J. Taylor, and J. Pons, “Stable audio 3,” 2026. [Online]. Available: https: //arxiv.org/abs/2605.17991

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  8. [8]

    A multimodal symphony: integrating taste and sound through generative ai,

    M. Spanio, M. Zampini, A. Rod `a, and F. Pierucci, “A multimodal symphony: integrating taste and sound through generative ai,”Frontiers in Computer Science, vol. V olume 7 - 2025, 2025

    work page 2025

  9. [9]

    Towards emotionally aware ai: Challenges and opportunities in the evolution of multimodal generative models,

    M. Spanio, “Towards emotionally aware ai: Challenges and opportunities in the evolution of multimodal generative models,” inProceedings of the AIxIA Doctoral Consortium 2024 co-located with the 23nd International Conference of the Italian Association for Artificial Intelligence (AIxIA 2024), ser. CEUR Workshop Proceedings. http://CEUR-WS.org, 2024. [Onlin...

    work page 2024

  10. [10]

    SAME: A Semantically-Aligned Music Autoencoder

    J. D. Parker, Z. Evans, C. Carr, Z. Zukowski, J. Taylor, M. Rice, and J. Pons, “Same: A semantically-aligned music autoencoder,” 2026. [Online]. Available: https://arxiv.org/abs/2605.18613

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  11. [11]

    Scalable diffusion models with transformers,

    W. Peebles and S. Xie, “Scalable diffusion models with transformers,” in 2023 IEEE/CVF International Conference on Computer Vision (ICCV), Oct 2023, pp. 4172–4182

    work page 2023

  12. [12]

    Gemma 2: Improving Open Language Models at a Practical Size

    G. Team, M. Riviere, S. Pathak, P. G. Sessa, C. Hardin, S. Bhupatiraju, L. Hussenot, T. Mesnard, B. Shahriari, A. Ram ´eet al., “Gemma 2: Improving open language models at a practical size,” 2024. [Online]. Available: https://arxiv.org/abs/2408.00118

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  13. [13]

    Flow Matching for Generative Modeling,

    Y . Lipman, R. T. Q. Chen, H. Ben-Hamu, M. Nickel, and M. Le, “Flow Matching for Generative Modeling,” inInternational Conference on Learning Representations, 2023. [Online]. Available: https://mlanthology.org/iclr/2023/lipman2023iclr-flow/

    work page 2023

  14. [14]

    Live music models,

    A. Caillon, B. McWilliams, C. Tarakajian, I. Simon, I. Manco, J. Engel, N. Constant, Y . Li, T. I. Denk, A. Lalama, A. Agostinelli, C.-Z. A. Huang, E. Manilow, G. Brower, H. Erdogan, H. Lei, I. Rolnick, I. Grishchenko, M. Orsini, M. Kastelic, M. Zuluaga, M. Verzetti, M. Dooley, O. Skopek, R. Ferrer, Z. Borsos, A. van den Oord, D. Eck, E. Collins, J. M. Ba...

    work page 2025

  15. [15]

    Obsidian-neural: Stable audio 3 medium running locally on cpu,

    innermost47, “Obsidian-neural: Stable audio 3 medium running locally on cpu,” https://github.com/innermost47/ai-dj, 2026, local edition; re- ports≈11s/generation on a laptop CPU (Apple Silicon). Verify before camera-ready

    work page 2026

  16. [16]

    whisper.cpp: Port of openai’s whisper model in c/c++,

    G. Gerganov and contributors, “whisper.cpp: Port of openai’s whisper model in c/c++,” https://github.com/ggml-org/whisper.cpp, 2022, ggml- based, dependency-free speech recognition on CPU and GPU

    work page 2022

  17. [17]

    Open-Source Acceleration of Stable-Diffusion.cpp Deployable on All Devices

    J. Ng, C. Lv, P. Zhao, W. Niu, J. Lin, M. Pan, Y . Liang, and Y . Wang, “Open-source acceleration of stable-diffusion.cpp deployable on all devices,” 2025. [Online]. Available: https://arxiv.org/abs/2412.05781

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  18. [18]

    ExecuTorch -- A Unified PyTorch Solution to Run AI Models On-Device

    M. Nachin, D. Desai, S. S. Jia, C. Lai, M. Liu, J. Szwejbka, R. Alvarez, R. Ascani, D. Bort, M. Candaleset al., “ExecuTorch - a unified PyTorch solution to run AI models on-device,”arXiv preprint arXiv:2605.08195,

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    Available: https://github.com/pytorch/executorch

    [Online]. Available: https://github.com/pytorch/executorch

    work page

  20. [20]

    MLX: Efficient and flexible machine learning on apple silicon,

    A. Hannun, J. Digani, A. Katharopoulos, and R. Collobert, “MLX: Efficient and flexible machine learning on apple silicon,” 2023. [Online]. Available: https://github.com/ml-explore

    work page 2023

  21. [21]

    (2018) Tensorrt

    NVIDIA Corporation. (2018) Tensorrt. https://developer.nvidia.com/ tensorrt. Per-model engine compilation for datacenter GPU inference. [Online]. Available: https://github.com/NVIDIA/TensorRT

    work page 2018

  22. [22]

    SmoothQuant: Accurate and efficient post-training quantization for large language models,

    G. Xiao, J. Lin, M. Seznec, H. Wu, J. Demouth, and S. Han, “SmoothQuant: Accurate and efficient post-training quantization for large language models,” inInternational Conference on Machine Learn- ing (ICML), 2023

    work page 2023

  23. [23]

    Q-Diffusion: Quantizing diffusion models,

    X. Li, Y . Liu, L. Lian, H. Yang, Z. Dong, D. Kang, S. Zhang, and K. Keutzer, “Q-Diffusion: Quantizing diffusion models,” inIEEE/CVF International Conference on Computer Vision (ICCV), 2023

    work page 2023

  24. [24]

    Post-training quantization for audio diffusion transformers,

    T. Khandelwal and M. Fuentes, “Post-training quantization for audio diffusion transformers,” 2025

    work page 2025

  25. [25]

    Steering Language Models With Activation Engineering

    A. M. Turner, L. Thiergart, D. Udell, G. Leech, U. Mini, and M. MacDiarmid, “Activation addition: Steering language models without optimization,”CoRR, vol. abs/2308.10248, 2023. [Online]. Available: https://doi.org/10.48550/arXiv.2308.10248

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2308.10248 2023

  26. [26]

    Representation Engineering: A Top-Down Approach to AI Transparency

    A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, R. Ren, A. Pan, X. Yin, M. Mazeika, A.-K. Dombrowski, S. Goel, N. Li, M. J. Byun, Z. Wang, A. Mallen, S. Basart, S. Koyejo, D. Song, M. Fredrikson, J. Z. Kolter, and D. Hendrycks, “Representation engineering: A top-down approach to ai transparency,” 2025. [Online]. Available: https://arxiv.org/abs/2310.01405

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  27. [27]

    Analysing the generalisation and reliability of steering vectors,

    D. Tan, D. Chanin, A. Lynch, B. Paige, D. Kanoulas, A. Garriga-Alonso, and R. Kirk, “Analysing the generalisation and reliability of steering vectors,” inProceedings of the 38th International Conference on Neural Information Processing Systems, ser. NIPS ’24. Red Hook, NY , USA: Curran Associates Inc., 2024

    work page 2024

  28. [28]

    Mood vectors in audio diffusion: Steering stable audio 3,

    G. Camporese, “Mood vectors in audio diffusion: Steering stable audio 3,” May 2026. [Online]. Available: https://guglielmocamporese.github. io/blog/audio-mood-steering.html

    work page 2026

  29. [29]

    Activation Patching for Interpretable Steering in Music Generation

    S. Facchiano, G. Strano, D. Crisostomi, I. Tallini, T. Mencattini, F. Galasso, and E. Rodol `a, “Activation patching for interpretable steering in music generation,” 2025. [Online]. Available: https: //arxiv.org/abs/2504.04479

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  30. [30]

    TADA! tuning audio diffusion models through activation steering,

    Ł. Staniszewski, K. Zaleska, M. Modrzejewski, and K. Deja, “TADA! tuning audio diffusion models through activation steering,” inICLR 2026 Workshop on Representational Alignment (Reˆ4-Align), 2026. [Online]. Available: https://openreview.net/forum?id=OUux4upENk

    work page 2026

  31. [31]

    Steering autoregressive music generation with recursive feature machines,

    D. Zhao, D. Beaglehole, J. McAuley, T. Berg-Kirkpatrick, and Z. Novack, “Steering autoregressive music generation with recursive feature machines,” inThe Fourteenth International Conference on Learning Representations, 2026. [Online]. Available: https: //openreview.net/forum?id=NaHzPMaCY9

    work page 2026

  32. [32]

    Discovering and steering interpretable concepts in large generative music models,

    N. Singh, M. Cherep, and P. Maes, “Discovering and steering interpretable concepts in large generative music models,” inAI for Music Workshop, 2025. [Online]. Available: https://openreview.net/ forum?id=jVSlJk5qNA

    work page 2025

  33. [33]

    Learning interpretable features in audio latent spaces via sparse autoencoders,

    N. Paek, Y . Zang, Q. Yang, and R. Leistikow, “Learning interpretable features in audio latent spaces via sparse autoencoders,” inMechanistic Interpretability Workshop at NeurIPS 2025, 2025. [Online]. Available: https://openreview.net/forum?id=5fsYFQzzMX

    work page 2025

  34. [34]

    LoRA: Low-rank adaptation of large language models,

    E. J. Hu, yelong shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen, “LoRA: Low-rank adaptation of large language models,” inInternational Conference on Learning Representations, 2022. [Online]. Available: https://openreview.net/forum?id=nZeVKeeFYf9

    work page 2022

  35. [35]

    Adding conditional control to text-to-image diffusion models,

    L. Zhang, A. Rao, and M. Agrawala, “Adding conditional control to text-to-image diffusion models,” inProceedings of the IEEE/CVF International Conference on Computer Vision (ICCV), October 2023, pp. 3836–3847

    work page 2023

  36. [36]

    Concept sliders: Lora adaptors for precise control in diffusion models,

    R. Gandikota, J. Materzy ´nska, T. Zhou, A. Torralba, and D. Bau, “Concept sliders: Lora adaptors for precise control in diffusion models,” inComputer Vision – ECCV 2024, A. Leonardis, E. Ricci, S. Roth, O. Russakovsky, T. Sattler, and G. Varol, Eds. Cham: Springer Nature Switzerland, 2025, pp. 172–188

    work page 2024

  37. [37]

    As bitter as a trombone: Synesthetic correspondences in nonsynesthetes between tastes/flavors and musical notes,

    A.-S. Crisinel and C. Spence, “As bitter as a trombone: Synesthetic correspondences in nonsynesthetes between tastes/flavors and musical notes,”Attention, Perception, & Psychophysics, vol. 72, no. 7, pp. 1994– 2002, 2010. [Online]. Available: https://doi.org/10.3758/APP.72.7.1994

    work page doi:10.3758/app.72.7.1994 1994

  38. [38]

    The taste of music,

    B. Mesz, M. A. Trevisan, and M. Sigman, “The taste of music,” Perception, vol. 40, no. 2, pp. 209–219, 2011

    work page 2011

  39. [39]

    Crossmodal correspondences between sounds and tastes,

    K. Kn ¨oferle and C. Spence, “Crossmodal correspondences between sounds and tastes,”Psychonomic Bulletin & Review, vol. 19, pp. 992– 1006, 2012

    work page 2012

  40. [40]

    “what’s your taste in music?

    Q. J. Wang, A. T. Woods, and C. Spence, ““what’s your taste in music?” a comparison of the effectiveness of various soundscapes in evoking specific tastes,”i-Perception, vol. 6, no. 6, 2015

    work page 2015

  41. [41]

    Spence,Gastrophysics: The New Science of Eating

    C. Spence,Gastrophysics: The New Science of Eating. Viking, 2017

    work page 2017

  42. [42]

    Multimodal Dataset Normalization and Perceptual Validation for Music-Taste Correspondences

    M. Spanio, V . Frezzato, and A. Rod`a, “Multimodal dataset normalization and perceptual validation for music-taste correspondences,” 2026. [Online]. Available: https://arxiv.org/abs/2604.10632

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  43. [43]

    Taste-aware music retrieval from audio embeddings,

    M. Spanio and A. Rod `a, “Taste-aware music retrieval from audio embeddings,” inProceedings of the 23rd IEEE International Conference on Content-Based Multimedia Indexing (CBMI). Toulouse, France: IEEE, Oct. 2026, accepted for publication

    work page 2026

  44. [44]

    Clap learning audio concepts from natural language supervision,

    B. Elizalde, S. Deshmukh, M. A. Ismail, and H. Wang, “Clap learning audio concepts from natural language supervision,” inICASSP 2023 - 2023 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), June 2023, pp. 1–5

    work page 2023

  45. [45]

    Natural language supervision for general-purpose audio representations,

    B. Elizalde, S. Deshmukh, and H. Wang, “Natural language supervision for general-purpose audio representations,” inICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), April 2024, pp. 336–340

    work page 2024

  46. [46]

    Fr ´echet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms,

    K. Kilgour, M. Zuluaga, D. Roblek, and M. Sharifi, “Fr ´echet Audio Distance: A Reference-Free Metric for Evaluating Music Enhancement Algorithms,” inInterspeech 2019, 2019, pp. 2350–2354

    work page 2019

  47. [47]

    Rank analysis of incomplete block designs: I. the method of paired comparisons,

    R. A. Bradley and M. E. Terry, “Rank analysis of incomplete block designs: I. the method of paired comparisons,”Biometrika, vol. 39, no. 3/4, pp. 324–345, 1952

    work page 1952