Structural and Representational Agency in LLM-Driven Timbre Generation and Interpolation

arXiv:2610.12532 · cs.SD, cs.HC, cs.LG, cs.MM, eess.AS · Submitted 2026-10-08 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Today's paper: "Structural and Representational Agency in LLM-Driven Timbre Generation and Interpolation".

Jane: The gist The user can cause, what they can cause proves separable from what the interface reports; structural constraints survive model failure,

Tom: First, who's behind it and why it matters.

Paper summary: Jane: So, wrapping this up, what’s the big picture here with the paper Structural and Representational Agency in LLM-Driven Timbre Generation and Interpolation? It really seems to be about separating what a user can do from what the interface tells them they can do.

Tom: Exactly. The main takeaway is that structural constraints survive model failure, but guarantees resting on the model’s own report don't hold up > We've shown that effective agency—what you can cause—is separate from epistemic agency—whether you can know it > That’s the core finding.

Lu: It means we don't have to throw out the interface just because the model makes a mistake about your position; we just need to be more careful about what we use that report for > The structure of what's possible is real, even if the model misreports where you are within it.

Meng: For us in production, it means when we build systems like this, we have to focus on building checks against the actual output rather than just trusting the AI’s self-assessment of its performance > We need those objective measurements they suggested.

Jane: It suggests that for creative tools using LLMs, we can start designing interfaces where control and knowledge of control are fundamentally separable > That separation is a very important distinction for how we think about building these kinds of systems moving forward.

Conclusion: Tom: So, we’ve looked at how this paper handles agency in these LLM systems, and now we’re getting to the conclusion about what this whole thing means for us > It boils down to separating what a user can actually cause from what the interface tells them they can do.

Jane: Right. The authors are saying that effective agency—what you can make happen—is totally different from epistemic agency, which is just whether you know it's possible > That separation is key here for building reliable tools.

Lu: Exactly. They show that structural constraints survive even when the model fails and misreports its own state > So even if the AI gets confused about where you are in the audio space, there are still hard limits on what it can reach.

Meng: From a deployment standpoint, that means we don't have to obsess over every tiny detail of the AI’s internal math if we build our safety around these structural rules > We focus on making sure the physical limits of what the system does are sound.

Lalam: I see this as a way for us to design interfaces where the control you exert is independent of the model's current confidence level about that control > It suggests a more robust way to structure user interaction.

Tom: It means when we look at these LLM-driven timbre generation tools, we need to ask if we’re building agency based on real capability or just on what the model *claims* it can do > That's the mechanical difference here.

Jane: And that leads us right into the practical problems they found, like how their display errors were way off by three times the size of where you could actually go > It’s a big gap between what’s on screen and what’s actually happening in the audio space.

Lu: That error decomposition is important because it separates model-dependent things, like those vertex terms, from things that are just interpolation steps, which are more stable across different models > The stability versus the uncertainty is a major theme.

Meng: And that uncertainty is what kills my confidence in shipping something based purely on the model's report if we don't have better ways to measure that gap > We need concrete metrics for that unreliability.

Tom: So, to sum it up, they’re arguing that structural constraints are solid even when the model is lying about its location, but the guarantees it gives you about its own state are not trustworthy on their own > That’s a fair summary of their stance.

Jane: It really pushes us to think more like engineers about how we design these interfaces, focusing on what's structurally possible rather than just what the AI describes > It shifts the focus from model accuracy to system design itself.

Lu: And that opens up questions about how we can engineer that separation into the very architecture of these generative systems > How do you bake this structural understanding in from the start?

Meng: That’s where I think we need to start thinking about measuring those reachable sets more directly, instead of relying on the model’s description of them > We need to measure what's actually there.

Joseph M. Cameron, Alan F. Blackwell

University of Cambridge

cs.SD, cs.HC, cs.LG, cs.MM, eess.AS

Submitted: 2026-10-08

Updated: 2026-10-08

Comments: 10 pages, 2 figures, 4 tables. Accepted at NeurIPS 2026 (Main Conference, Creative AI Track)

Project page: https://audiocommons.github.io/assets/files/AC-WP5-SURREY-D5.8%20Release%20of%20timbral%20characterisation%

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 86/100

The gist: The gist The user can cause, what they can cause proves separable from what the interface reports; structural constraints survive model failure, whereas guarantees resting on the model’s own report

Key concepts

Separation of Agency: This concept distinguishes between 'effective agency,' which is what a user can actually cause (the physical output), and 'epistemic agency,' which is whether the user can know that they caused it. The paper proves these two are independently engineered, meaning you can still cause things even if the knowledge about that causation is unreliable.
Reachable Set as Zonotope: The set of all possible sounds a user can author is mathematically characterized as a zonotope. This means the space of reachable audio is a precisely defined convex region with clear geometric boundaries, rather than an imprecise cloud. Each new sound adds exactly one dimension to this reachable space.
Fidelity and Error Decomposition: The error between what the interface displays and the actual rendered audio splits into two parts: a 'vertex term' (model-dependent) and a 'linearity term' (model-independent). The vertex term shows where model predictions are inaccurate, while the linearity term relates to how parameters are interpolated, which is consistent regardless of model quality.

Summary

The gist The user can cause, what they can cause proves separable from what the interface reports; structural constraints survive model failure, whereas guarantees resting on the model’s own report do not.

Separation of Agency

Effective agency and epistemic agency are shown to be separable and independently engineered through the system design; structural constraints survive model failure while guarantees resting on the model’s own report do not The first is exactly characterisable: the reachable set is a zonotope whose dimension increases by one per sound placed, so the space the model may act within is demonstrably the user’s ownThe first is exactly characterisable: the reachable set is a zonotope whose dimension increases by one per sound placed, so the space the model may act within is demonstrably the user’s own. The interface's display differs from measurement by approximately three times the width of that region, and this error has two independent sources The displayed position differs from measurement by approximately three times the width of that region, and the error has two independent sources: model-predicted coordinates, whose accuracy varies twofold between the two models evaluated, and the interpolation step, which arises from combining synthesiser parameters and is unaffected by model qualityThe displayed position differs from measurement by approximately three times the width of that region, and the error has two independent sources. Effective agency is what a user can cause, while epistemic agency is whether they can know it Effective agency (what a user can cause) and epistemic agency (whether they can know it) are therefore separable and independently engineered.

Characterization of the Reachable Set

The reachable set is characterised exactly as a zonotope, yielding a closed-form measure of the space a user has authored First, we characterise the interface’s reachable set exactly as a zonotope, yielding a closed-form measure of the space a user has authored. The audible perceptual vector admits a closed form that is given by equation (1) ac(t) = 1/E X e E

fixed base point + X e te acb(e) − aca(e) E. This set is a precisely characterised convex region rather than an approximate cloud The reachable set is consequently a precisely characterised convex region rather than an approximate cloud. The affine dimension of the region is min(N − 1, 7) for N placed sounds, meaning each additional sound contributes exactly one dimension of reach First, the affine dimension of the region is min(N − 1, 7) for N placed sounds, so each additional sound contributes exactly one dimension of reach.

Fidelity and Error Decomposition

The display error relative to rendered audio decomposes into two independent components: a vertex term and a linearity term The displayed position is in error by approximately three times the width of the region being navigated (Table 1); the error thus exceeds the extent of the territory the display purports to locate the user within. The vertex term arises because stored coordinates are model predictions rather than measurements The vertex term arises because stored coordinates are model predictions rather than measurements, and it is model-dependent, and this dependence is not signalled The vertex term is model-dependent, and the dependence is not signalled. The linearity term arises because the audible mix interpolates synthesiser parameters whereas the display interpolates perceptual vectors The linearity term arises because the audible mix interpolates synthesiser parameters whereas the display interpolates perceptual vectors, and it is model-independent The linearity term is model-independent.

End-to-End Outcomes and Silent Failure Mode

When a natural-language refinement instruction fails to be parsed successfully, the routine returns the current position unchanged along two distinct fallback paths The target is therefore identical to the present state, the descent correctly reports convergence after zero steps, no parameter is altered, and no error surfaced. This silent failure mode presents complete failure of the facility as success Complete failure of the facility is thus presented to the user as success. The system seldom moves the audio in the wrong direction; more commonly it does not move it at all Across both models, 26 of 36 instructions produced no change exceeding the noise floor.

Implications for Practice

Six measures are suggested to remove error, including deriving displayed quantities from the artefact rather than from the model’s description of it Measure, do not ask: derive displayed quantities from the artefact rather than from the model’s description of it. Substituting measured coordinates for the model’s predictions at each vertex halved the display error, reducing it from roughly three times the width of the reachable region to 1.7 times it Measure the composition, not only its parts: substituting measured coordinates for the model’s predictions at each vertex halved the display error, from roughly three times the width of the reachable region to 1.7 times it. Establishing a noise floor by re-running identical inputs and thresholding every claimed effect against observed variation removed a false positive from analysis Establish a noise floor: re-run identical inputs and threshold every claimed effect against the variation observed, a step that removed a false positive from our own analysis.

Conclusion

The position this interface displays is wrong by roughly three times the width of the region a user can reach, and most refinement instructions produced no audible change while reporting success The position this interface displays is wrong by roughly three times the width of the region a user can reach, and most refinement instructions produced no audible change while reporting success. Effective agency and epistemic agency are therefore separable because structural constraints survive model failure where guarantees resting on the model’s own report do not Effective and epistemic agency are therefore separable: structural constraints survive model failure whereas guarantees resting on the model’s own report do not. The reliability of a model-derived display is an unmeasured free parameter of system configuration, latent in any interface rendering model judgement as instrumentation We claim the mechanism generally: the reliability of a model-derived display is an unmeasured free parameter of system configuration, latent in any interface rendering model judgement as instrumentation. The study involves no human participants, but it establishes that the display is unfaithful The study involves no human participants: we establish that the display is unfaithful.

Limitations

The study's sample sizes are small throughout, involving three canvases and nine vertices, and concern one system, one synthesiser and one feature space Sample sizes are small throughout (three canvases, nine vertices, 36 outcome trials, 52 concepts, two models), and concern one system, one synthesiser and one feature space. The reference values are regression-model outputs rather than perceptual judgements The reference values are regression-model outputs rather than perceptual judgements, though they remain the appropriate internal reference because the system’s optimiser targets them The reference values are regression-model outputs rather than perceptual judgements, though they remain the appropriate internal reference because the system’s optimiser targets them. The paper claims that audio is unusual in admitting a check, which is why the practice remains unexamined elsewhere Audio is unusual in admitting the check, which is why the practice remains unexamined elsewhere.

Reproducibility Details

The system depends on a commercial synthesiser plugin that cannot be redistributed, so bit-exact reproduction is not available to readers without a licence The system depends on a commercial synthesiser plugin that cannot be redistributed, so bit-exact reproduction is not available to readers without a licence. This appendix specifies the measurement procedure in enough detail for the study to be repeated on another system This appendix therefore specifies the measurement procedure in enough detail for the study to be repeated on another system. The models were used with a temperature setting of 0 The models were used with a temperature setting of 0.

References

[1] Eric Horvitz. Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’99), pages 159–166. ACM, 1999 Eric Horvitz. Principles of mixed-initiative user interfaces. In Proceedings of the SIGCHI Conference on Human Factors in Computing Systems (CHI ’99), pages 159–166. ACM, 1999.

Improvements for AI systems

Improved UI/UX by decoupling effective agency from epistemic agency. This allows for structural constraints to enforce predictable outcomes regardless of model quality, ensuring that structural constraints survive model failure, thereby preventing silent failures like silence presented as success.
Enhanced fidelity through explicit measurement protocols. Implement a mandatory procedure where users are instructed to derive displayed quantities from the artefact rather than from the model’s description of it and to establish a noise floor by thresholding changes against observed variation, which has been shown to remove false positives.
Improved control via coordinate substitution in refinement loops. Instead of relying solely on model-predicted coordinates, substitute them with measured ones: Substituting measured vertex coordinates for the model’s predictions at each vertex halved the display error.
Robustness against silent failure through explicit error reporting. Implement a protocol to Never return the input on a parse failure: if the model’s output cannot be parsed, report that, preventing the system from exhibiting Complete failure of the facility is thus presented to the user as success.
Configuration-aware reliability settings. Introduce an auditable configuration setting that explicitly controls display accuracy, recognizing that the reliability of an instrument is determined by a configuration setting presented to the user as a preference.

Abstract

Generative creative interfaces routinely report the user's position within a space of possibilities. Such displays imply that the quantities shown were measured from the artefact; in a growing class of systems they are instead predicted by a language model describing its own output, and the two are indistinguishable on screen. Most creative domains lack an agreed measurement against which such a self-report can be checked; audio admits one. In a measurement study of a working LLM-driven timbre interface, what a user can cause proves separable from what the interface reports. The first is exactly characterisable: the reachable set is a zonotope whose dimension increases by one per sound placed, so the space the model may act within is demonstrably the user's own. The second is inaccurate. The displayed position differs from measurement by approximately three times the width of that region, and the error has two independent sources: model-predicted coordinates, whose accuracy varies twofold between the two models evaluated, and the interpolation step, which arises from combining synthesiser parameters and is unaffected by model quality. End to end, 26 of 36 refinement instructions produced no perceptible change, and under the weaker model the facility performed no action while reporting convergence. Effective agency (what a user can cause) and epistemic agency (whether they can know it) are therefore separable and independently engineered: structurally enforced constraints survive model failure, whereas guarantees resting on the model's own report do not. The failure mode that remains is silence presented as success. We close with a short protocol for building and auditing such displays.