aiqbench analysis
This research demonstrates that hallucinations in the Whisper ASR model can be effectively mitigated by steering internal representations using Sparse AutoEncoders, achieving significant error reduction without the computational overhead of fine-tuning. The study reveals that hallucination-related information is linearly separable within the model's deeper layers, offering a promising direction for improving reliability in non-speech audio scenarios. For practitioners, this approach provides a viable alternative to traditional retraining methods for enhancing model robustness against coherent but incorrect transcriptions.
Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders
arXiv:2606.07473v1 Announce Type: cross 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.
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.
Sound (cs.SD); Artificial Intelligence (cs.AI)
arXiv:2606.07473 [cs.SD]
(or arXiv:2606.07473v1 [cs.SD] for this version)
https://doi.org/10.48550/arXiv.2606.07473 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 226 words
Related coverage
Reflections on unlearning and inoculation
LessWrong · Sep 20, 2026
Beyond the Transcript: Detecting Covert Co ordination in Latent Multi-Agent Communication
arXiv · Aug 21, 2026
ROT: Rotating Hidden States towards Contextual Vectors for Hallucination Mitigation in LVLMs
arXiv · Oct 7, 2026
Preference Instability in Reward Models: Detection and Mitigation via Sparse Autoencoders
arXiv · Oct 1, 2026