Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

Hugging Face Daily Papers Papers
whisper hallucination-detection hallucination-mitigation sparse-autoencoder representation-steering speech-recognition model-interpretability

Summary

This paper demonstrates that Whisper's hallucination failures on silence, noise, or music can be detected and mitigated purely from internal activations using sparse autoencoders, achieving large reductions in hallucination rate without fine-tuning.

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.
Original Article
View Cached Full Text

Cached at: 06/09/26, 12:41 PM

Paper page - Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

Source: https://huggingface.co/papers/2606.07473 Whisper Hallucination Detection and Mitigation via Hidden Representation Steering and Sparse AutoEncoders

Whisper has a well-known failure mode: feed it silence, noise, or music, and it will often respond with confidently fabricated transcripts. This paper shows you can detect and mitigate these hallucinations purely from the model’s internal activations without fine-tuning.

We probe two representation spaces in Whisper’s audio encoder: raw activations and Sparse AutoEncoder (SAE) latents. Both turn out to encode linearly separable hallucination signals, concentrated in a sparse subset of features that strengthen in deeper layers. Steering activations away from these directions at inference yields large drops in hallucination rate on non-speech samples from different datasets:

  • Whisper small:**72.63% → 14.11%**hallucination rate on non-speech samples
  • Whisper large-v3:86.88% → 27.33%

WER on regular speech data barely budges, and the method reaches numbers competitive with fine-tuning approaches like Calm-Whisper, without touching any model weights. A finding worth highlighting: since steering only a handful of encoder-side SAE features is enough to suppress hallucinations, the hallucination signal is not purely a decoder-side generation issue, it is already encoded in Whisper’s encoder representations of non-speech audio.

Paper:https://arxiv.org/abs/2606.07473

Similar Articles

HalluSAE: Detecting Hallucinations in Large Language Models via Sparse Auto-Encoders

arXiv cs.CL

Researchers from Beihang University and other institutions propose HalluSAE, a framework using sparse autoencoders and phase transition theory to detect hallucinations in LLMs by modeling generation as trajectories through a potential energy landscape and identifying critical transition zones where factual errors occur.

Hallucination Detection via Activations of Open-Weight Proxy Analyzers

arXiv cs.CL

This paper introduces a proxy-analyzer framework that detects hallucinations in large language models by analyzing internal activations of small, open-weight models rather than the generator itself. The method achieves superior performance on benchmarks like RAGTruth compared to existing methods like ReDeEP, demonstrating that model size is less critical than the analysis approach.

From Architecture to Output: Structural Origins of Hallucination in Large Language Models and the Amplifying Role of Data

arXiv cs.AI

This paper analyzes hallucination in large language models as a structural consequence of three architectural decisions: self-attention's co-occurrence learning, maximum likelihood estimation training objective, and autoregressive decoding's left-to-right commitment. It maps each mechanism to specific hallucination types and argues that dataset pathologies amplify but do not cause these vulnerabilities.