Filer's CoAT and GAPO papers accepted to NeurIPS 2026
Filer said both papers will be presented on December 10 in Sydney; CoAT lifts speech emotion recognition accuracy on Qwen2.5-Omni from about 55% to about 79%.
Filer, a company working on video-understanding AI, announced that two of its research papers have been accepted to the NeurIPS 2026 AI conference. The papers are CoAT (continuous speech-of-thought) and GAPO (geometric anchor preference optimization), according to digitaltoday.co.kr.
Both papers will be presented on December 10 in Sydney, Australia. CoAT addresses emotion, intonation and pitch information that is lost when speech is converted to text, feeding it back into AI inference. Applied to three speech AI models, the method improved 38 of 42 performance metrics. On the Qwen2.5-Omni model, speech emotion recognition accuracy rose from about 55% to about 79%.
GAPO targets preference-based AI training, where incorrect or uncertain data can skew results. The method measures the stability of outcomes and lowers the training weight of unstable data.
Filer compared GAPO on four AI models against eight existing training methods and reported better overall answer quality and user preference scores, with the method remaining stable when 20% errors were injected into the training data.
The company said it plans to apply the research results to its Brand Safety and content moderation technologies.
Quick answers
When and where will Filer present its NeurIPS 2026 papers?
Both CoAT and GAPO will be presented on December 10 in Sydney, Australia.
What results did CoAT show on Qwen2.5-Omni?
Speech emotion recognition accuracy rose from about 55% to about 79% on that model.