Georgii Aparin, Vadim Popov, Tasnima Sadekova, Assel Yermekova
5 min
Abstract
Whisper, a widely adopted ASR model, is known to suffer from hallucinations - coherent transcriptions generated for non-speech audio entirely disconnected from the input. We investigate whether hallucinations can be detected and mitigated through Whisper's internal representations. We extract audio encoder activations and evaluate two representation spaces: raw Whisper activations and Sparse AutoEncoder (SAE) latents. We show that both spaces encode linearly separable hallucination-related information, with discriminative power concentrated in a sparse feature subset and increasing toward deeper encoder layers. We propose two steering strategies: activation-space steering and SAE latent-space steering. SAE-based steering reduces hallucination rate from 72.63% to 14.11% for Whisper small and from 86.88% to 27.33% for Whisper large-v3 on the full non-speech test set, with small WER degradation on speech data, approaching the performance of fine-tuning-based methods.
Alex: But if we're surgically removing these features, don't we risk damaging the model's ability to handle quiet or muffled speech? The kind that might look a lot like silence?
Sam: That is the central trade-off, and the researchers took it seriously. They tracked what's called the Word Error Rate — essentially, how often the model gets real words wrong — across a range of different audio conditions. The goal was to confirm that the intervention only targets the hallucination behaviour, not the model's core ability to understand speech. The evidence suggests they achieved that balance, though the researchers are careful not to overstate it.
Alex: So the "surgery" is precise enough to remove the bad behaviour without causing collateral damage.
Sam: That's the finding. And what makes it particularly notable is how few features were involved. The researchers found that a relatively small number of internal features — on the order of 50 to 100 — were sufficient to control this behaviour in a model that contains millions of parameters. That points to something important about how these errors are structured.
Alex: Why would such a small number of features have that much influence?
Sam: It comes down to a property called *polysemanticity* — the idea that individual units inside a neural network often carry multiple meanings at once. One unit might be doing several jobs simultaneously, which is efficient but messy. When hallucination-prone behaviour is tangled up with other functions, it's hard to address. The Sparse AutoEncoder untangles that, and what you find is that the hallucination tendency is actually quite concentrated. It's not spread evenly across the whole model — it clusters in specific places. Find those places, and you have a precise target.
Alex: It's a bit like discovering that a persistent leak in a large building is actually coming from one cracked pipe, not the whole plumbing system.
Sam: That's a fair analogy. And the broader implication is this: we don't always need more data or a bigger model to fix a reliability problem. Sometimes, we need better visibility into how the system is actually thinking. This research is a step toward that kind of transparency — understanding not just *what* a model does wrong, but *where inside it* the mistake originates.
Alex: That's a useful place to leave it. Thanks for listening to ResearchPod.