Xiamen University

Predicting experimental vibrational spectra from molecular structure has been caught between theory and measurement: quantum chemistry is accurate but slow, while end-to-end neural networks are fast but bypass explicit vibrational information. HARMONY assembles harmonic spectra from atomic forces and spectroscopic tensors predicted directly by equivariant neural networks, then uses HarmoNet to map them to experimental IR and Raman spectra conditioned on molecular topology and measurement metadata. On the HarmoSpec test set, the full workflow reaches mean cosine similarities of 0.912 for infrared and 0.907 for Raman, with a mean of 2.12 seconds per molecule, roughly 600-fold faster than the evaluated DFT workflow.

Predicting experimental vibrational spectra from molecular structure has been caught between theory and measurement: quantum chemistry is accurate but slow, while end-to-end neural networks are fast but bypass explicit vibrational information. HARMONY assembles harmonic spectra from atomic forces and spectroscopic tensors predicted directly by equivariant neural networks, then uses HarmoNet to map them to experimental IR and Raman spectra conditioned on molecular topology and measurement metadata. On the HarmoSpec test set, the full workflow reaches mean cosine similarities of 0.912 for infrared and 0.907 for Raman, with a mean of 2.12 seconds per molecule, roughly 600-fold faster than the evaluated DFT workflow.

Latent mixture-of-experts decouples expert input dimension from residual width, making expert shape a tunable design choice; at 3.4B and 8B scales, near-square experts with capacity allocated by layerwise intrinsic dimension reach lower validation loss than matched uniform configurations.

Latent mixture-of-experts decouples expert input dimension from residual width, making expert shape a tunable design choice; at 3.4B and 8B scales, near-square experts with capacity allocated by layerwise intrinsic dimension reach lower validation loss than matched uniform configurations.

The authors introduce Duplex-MPE, a benchmark of 2,000 paired scenarios with three or four human speakers plus an assistant named Aria, contrasting explicit naming with implicit addressing, and evaluate five open-weight full-duplex speech systems on continuous audio without transcripts, speaker labels or turn boundaries using four scores—fresh response initiation, conditional answer accuracy, silence preservation and answering-window yield—finding that MiniCPM-o 4.5 leads on three scored capabilities while frequent speech from other systems often coexists with inaccurate answers or failures to remain silent.

The authors introduce Duplex-MPE, a benchmark of 2,000 paired scenarios with three or four human speakers plus an assistant named Aria, contrasting explicit naming with implicit addressing, and evaluate five open-weight full-duplex speech systems on continuous audio without transcripts, speaker labels or turn boundaries using four scores—fresh response initiation, conditional answer accuracy, silence preservation and answering-window yield—finding that MiniCPM-o 4.5 leads on three scored capabilities while frequent speech from other systems often coexists with inaccurate answers or failures to remain silent.

The work presents onPanda: while reading a model response, the annotator locates the first inappropriate token, either picks a substitute from the model's candidate tokens or types the correct text, and the system truncates everything after that position and continues generation from the corrected prefix, repeating this loop; a small controlled study suggests a 51.5% reduction in median annotation time over manual post-editing, 97.0% of tokens in the resulting data are model-generated, and the data carries position-precise, naturally paired positive-negative token-level correction supervision, alongside the released Chinese vision-language Panda-CVL dataset and a token-level correction benchmark.

The work presents onPanda: while reading a model response, the annotator locates the first inappropriate token, either picks a substitute from the model's candidate tokens or types the correct text, and the system truncates everything after that position and continues generation from the corrected prefix, repeating this loop; a small controlled study suggests a 51.5% reduction in median annotation time over manual post-editing, 97.0% of tokens in the resulting data are model-generated, and the data carries position-precise, naturally paired positive-negative token-level correction supervision, alongside the released Chinese vision-language Panda-CVL dataset and a token-level correction benchmark.