Picture a pair of smart earbuds that switches modes depending on where you are. It might cancel noise on a train, pass through announcements at an airport, or boost speech in a café. The earbuds learn what you like from your thumbs-up and thumbs-down. When you reject a choice, the system faces a hidden question. Did your taste change, or did it misjudge where you are? Many real places sound almost the same. A metro platform and a tram stop, or an airport concourse and a shopping mall, share the same crowd murmur and ventilation hum. So the sound-based guess about the location is often wrong in exactly the cases where the user's wishes differ. A system that reads every rejection as a change of taste will keep unlearning good settings. A system that treats rejections as noise will miss real changes.
We propose to make this credit assignment explicit. The assistant keeps two beliefs at once. The first is a calibrated probability over which acoustic scene it is in. The second is a per-user preference that is allowed to change at unknown moments. Each piece of feedback is explained as a mixture over the possible scenes. The consequence is simple. A surprising rejection heard in an acoustically ambiguous clip is weak evidence of preference change. The same rejection in an acoustically clear clip is strong evidence. Feedback also flows backward and sharpens the scene belief. We will build a benchmark from real scene recordings to test whether this attribution beats simpler forgetting-based learners. It should detect true preference changes faster, and it should falsely declare fewer of them.
Core Claim
We claim that an assistant which models "scene misclassified" and "preference changed" as competing explanations of each feedback event will detect genuine preference change with lower delay at a fixed false-alarm rate than learners that forget uniformly. We further predict where the gain comes from. It should concentrate on scene pairs the classifier confuses, and it should vanish when the classifier is replaced by ground-truth scene labels. If the gain appears without this pattern, or does not appear at all, the claim is wrong.
Grounding in Prior Work
The acoustic side motivates the problem and supplies the perception module.
Two papers show the confusion is structural. Kong's DCASE baseline notes that up to 95% of a clip's spectral energy can be shared background such as wind or traffic hum (Deep Neural Network Baseline for DCASE Challenge 2016). Bear et al. show that acoustic distance between cities predicts classification error (City classification from multiple real-world sound scenes). Together they establish that confusions between certain scene pairs are a property of the sounds, not a training accident.
The strongest classifiers reduce this confusion but do not remove it. Hou et al. couple scene and event towers through event co-occurrence statistics and reach 77.4% on TUT Urban Acoustic Scenes 2018 (Relation-guided acoustic scene classification aided with event embeddings). Heo et al. distill from a long-clip teacher using softened labels (a temperature of 5) and reach 77.36% (Acoustic Scene Classification Using Teacher-Student Learning with Soft-Labels). Roughly a quarter of clips therefore remain misclassified, concentrated in pairs like metro and tram. We reuse these classifiers as they are. What we change is how their output is consumed downstream: as a calibrated distribution rather than an argmax.
One paper acts as a constraint on our design. Vosoughi et al. show that pushing audio embeddings away from counterfactual captions helps fine-grained retrieval but collapses zero-shot UrbanSound8K accuracy from 0.798 to 0.475 (Learning Audio Concepts from Counterfactual Natural Language). We read this as a warning against sharpening the perception model to force crisp scene decisions on coarse labels. We keep the classifier soft and resolve ambiguity with interaction evidence instead.
The preference side supplies the failure modes.
Liu et al. find that users' preferences evolve in response to the system's actions (Evaluating Rule-based Programming and ReinforcementLearning for Personalising an Intelligent System). This makes reinforcement-learning feedback inconsistent and stressful: satisfaction was a median of 4–5/7 for reinforcement learning versus 7/7 for rules. Their environment, however, is a symbolic simulator in which context is known exactly. We remove that assumption.
AcCoRD separates several kinds of preference dynamics, including "triggered" preferences that only appear after the agent surfaces something (AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics). It reports under 30% perfect outcomes for frontier models and triggered-preference scores as low as 0.09. It is entirely text-based.
Çelikok et al. name the core difficulty: an unexpected user response is unidentifiable between changing goals and bounded behavior (Modeling needs user modeling). Our contribution is to show that in sensing systems, a third explanation enters the tangle, namely perception error. Unlike the other two, this one comes with a measurable prior, because the classifier's confusion structure can be estimated on held-out data.
Finally, we borrow one idea from P-ShareLoRA (A Shared Low-Rank Adaptation Approach to Personalized RLHF). That method shares one factor of a low-rank update across users and keeps another per user, and it beats a pooled global model on Reddit TL;DR. It assumes each user's preference is static. We keep the shared-versus-individual split as a prior for new users, and we add change over time.
Research Program
The assistant observes a clip $x_t$ at each step. It chooses an action $a_t$, one of a small set of listening modes, and may receive binary feedback $y_t$.
The scene classifier provides $q(s \mid x_t)$, a probability over scenes $s$. It is temperature-calibrated on held-out data, and a device-specific confusion correction is estimated on a validation split. The user's preference is a parameter vector $\theta$ with one component $\theta_s$ per scene, giving that scene's action utilities.
The feedback likelihood marginalizes over the scene:
$$p(y_t \mid a_t, x_t, \theta) = \sum_{s} q(s \mid x_t)\, p(y_t \mid a_t, \theta_s).$$
Preference change is modeled as unknown change points with a small per-step hazard. The assistant maintains a posterior over "time since the last change" for each scene's preference, updated online with this mixture likelihood. The key behavior follows directly. When $q$ is spread across two scenes whose preferred actions differ, a rejection can be explained by the other scene. The posterior mass on "a change just happened" then rises only slightly. When $q$ is peaked, the same rejection can only be explained by a change.
The same update gives a posterior over the current scene given feedback. A user who keeps rejecting noise cancellation in a clip labelled "mall" shifts belief toward "airport." This backward flow is the part a cascaded system cannot express.
A shared prior across users initializes each $\theta$, in the spirit of P-ShareLoRA's split between shared and per-user parameters. Actions are chosen by posterior sampling. As a secondary component, we add an optional one-tap clarifying question ("Are you on a platform?") whose cost is a tunable penalty. The agent may ask only when attribution between the two explanations is near even. This speaks to Liu et al.'s finding that users value legibility.
Since no benchmark for this problem exists, we build one: AmbiDrift. We assemble day-long trajectories by stitching real clips from TUT Urban Acoustic Scenes 2018 and DCASE 2020 Task 1A along plausible scene transitions, for example home to street to metro to airport. Using DCASE 2020 also brings in its device mismatch.
Simulated users have scene-conditioned mode preferences drawn from a population. Three kinds of change are injected:
- Abrupt changes at random times.
- Gradual changes.
- Triggered changes tied to real acoustic events located by a pretrained audio tagger, in the manner of AcCoRD's triggered preferences. An example is a user who wants pass-through mode after hearing a public announcement.
Feedback is sparse and flipped with a small probability, so that noise does not come only from the classifier. Critically, we control where changes land: in scenes the classifier confuses, or in scenes it recognizes well. We also control how often feedback arrives.
We state the main limitation plainly. The users are simulated. We mitigate it with a replay study in which about 20 participants hear the same stitched trajectories over headphones, choose modes, and give feedback. Their traces are then scored offline against all methods.
Baselines
Naive reference. An argmax scene classifier feeds a per-scene preference table that is learned from the first simulated day and then frozen. This measures the cost of preference inertia and gives the floor.
Oracle-scene learner. The same change-point learner receives ground-truth scene labels. This is an upper bound. It also defines the headroom attributable to perception error.
Discounted contextual bandit (strongest generic competitor). A contextual bandit uses the full calibrated scene probability vector as its features and forgets old evidence through a discount factor or sliding window. This is the most serious threat, for two reasons. It already consumes soft scene beliefs. And discounting is the standard, well-tuned answer to non-stationarity. We give it the same classifier, the same trajectories, and the same number of hyperparameter trials, spent on discount factor and exploration rate. It is tuned on separate validation trajectories.
Cascade with change detection. The argmax classifier feeds the same Bayesian change-point preference learner, without the mixture over scenes. This is exactly the design a reviewer would build by chaining two good components.
Prompted language model. A frontier model receives recent scene tags, the classifier's top-3 probabilities, and feedback history, and is prompted ReAct-style. It is included because it is the default a practitioner would try. AcCoRD's under-30% perfect-outcome rate and 0.09 triggered-preference score make us expect it to be weak.
Ablations isolating the key ingredient. We run three: our model with uncalibrated classifier probabilities; our model with the mixture replaced by the single top scene; and our model without the backward scene update.
Strongest reported numbers on the perception side. RGASC reports 77.4% accuracy on TUT 2018 (Relation-guided acoustic scene classification aided with event embeddings). Teacher-student reports 77.36% on DCASE 2018 Task 1-a (Acoustic Scene Classification Using Teacher-Student Learning with Soft-Labels). We will verify that our reimplementation lands near these figures before building on it.
Criterion for beating the baselines. No reported result was found on Lacuna for preference-change tracking under acoustic ambiguity, so beating the baselines is defined internally. Across seeds, the method must show lower cumulative regret and shorter change-detection delay than the discounted bandit, at a matched false-change rate, with non-overlapping 95% intervals. The improvement must be larger for changes placed in confusable scene pairs than for changes in well-recognized scenes.
Why not just…?
Why not just use a discounted contextual bandit on soft scene probabilities? This is the real threat, and it may work comparably when feedback is dense. Our argument is that a single forgetting rate faces a forced trade-off. Forgetting fast enough to catch real changes also discards good settings whenever classifier errors produce rejections. Forgetting slowly causes inertia. Attribution escapes this by forgetting selectively. The experiment that settles it is a sweep over the bandit's discount, plotting detection delay against false-change rate. If the bandit's curve matches ours in confusable-pair conditions, the explicit attribution buys nothing, and we would report that.
Why not just build a better scene classifier? If perception were perfect, the problem would reduce to ordinary change tracking. But the best systems in the file still sit near 77%. Vosoughi et al. show that forcing sharper distinctions can badly damage accuracy on coarsely labelled data. We test this directly by swapping in progressively stronger classifiers, from the DCASE baseline to fine-tuned PANNs to RGASC-style, and finally the oracle. We predict our advantage shrinks toward zero as accuracy rises. Measuring that slope tells practitioners when the added modeling is worth it.
Why not just prompt a language model with the history and scene tags? AcCoRD reports under 30% perfect outcomes for frontier models on text-only preference dynamics, falling to 0.09 on triggered preferences. Our setting adds perceptual uncertainty on top of those dynamics. We still run the prompted model as a baseline, so this is measured rather than assumed.
Why not let users write rules per scene? Liu et al. find that rules are preferred for transparency but leave users facing a blank canvas. Rules keyed to a scene label also fire wrongly whenever the classifier confuses two scenes. We approximate the rules condition by giving simulated users perfect knowledge of their own initial preferences, then freezing them. In the replay study, participants write rules before listening.
Decisive Checks
Three results would confirm the idea, and each has a clear failure pattern:
- Advantage localized to ambiguity. Our advantage over the discounted bandit and the cascade should be concentrated where changes occur in confusable scene pairs. It should disappear with oracle scenes. A uniform gain would suggest we are merely better tuned.
- Backward scene inference helps perception. Scene accuracy after incorporating feedback should exceed the classifier's own accuracy on the metro/tram and mall/airport pairs.
- Robustness to misspecified calibration. Performance should degrade gracefully when the confusion correction is estimated on one DCASE 2020 device and tested on another.
The replay study then checks whether rankings from simulated users hold for real feedback traces. If the ordering against the discounted bandit reverses on human data, the simulator's noise model is wrong. That would be the result we would report first.
References
- Deep Neural Network Baseline for DCASE Challenge 2016
- City Classification from Multiple Real-World Sound Scenes
- Relation-guided acoustic scene classification aided with event embeddings
- Acoustic Scene Classification Using Teacher-Student Learning with Soft-Labels
- Learning Audio Concepts from Counterfactual Natural Language
- Evaluating Rule-based Programming and ReinforcementLearning for Personalising an Intelligent System
- AcCoRD: Evaluating User-Agent Collaboration Under Realistic User Preference Dynamics
- Modeling needs user modeling
- A Shared Low-Rank Adaptation Approach to Personalized RLHF