SuperMimic Cuts Humanoid Whole-Body Training Compute by 1000x
A single differentiable world model lets one policy learn from large human-motion datasets on a single GPU, versus days on a cluster.
Research Digest
Preprints and papers, rewritten in plain English.
New Benchmark Tests LLMs' Ability to Spot Equivalent Code
A new benchmark called CETBench probes whether large language models can recognise functionally equivalent code after simple transformations, an ability the paper says is important for code rewriting.
Stage-Aware Evaluation Reveals Where VLA Policies Fail Under Shift
A new arXiv proposal argues that aggregate success rates mask stage-level failures in long-horizon household robotics.
Shared brain morphometry markers link Alzheimer's and Parkinson's
A transfer learning model finds common imaging-derived morphometric biomarkers across Alzheimer's and Parkinson's, improving diagnosis on small datasets.
SRA protocol metadata completeness declines as archive grows
A decade after the original audit, the share of Sequence Read Archive experiments with complete protocol annotation in the designated field has fallen, though additional terms appear in description fields.
DRIADA: open-source Python toolkit links neuron selectivity to behavior
A new Python toolkit provides a unified framework for relating single-neuron selectivity, population activity, and network dynamics to behavioral measures, validated on hippocampal recordings and simulations.
Whale calls expose gaps in audio-language models' listening skills
A new benchmark uses marine mammal vocalizations to test whether audio-language models can reason about unfamiliar sounds from fine-grained acoustic cues, and finds they lag far behind humans.
TTS-Q: 4-bit quantization keeps TTS voices on target
A new quantization scheme recovers speaker timbre lost by standard 4-bit compression of text-to-speech diffusion transformers, without retraining.
Snoring sounds reveal airway collapse sites in sleep apnea
A deep learning model can localize upper airway obstructions from snoring audio, offering a non-invasive alternative to endoscopic exams.
Compressing Conversational Audio Context for LLM-Based ASR
A learned 'abstract compression' replaces prior-turn audio with fixed latent tokens, recovering much of the accuracy gain of raw context at lower cost.
Curriculum-augmented GFlowNets improve mRNA sequence design
A new training strategy helps generative models explore synonymous codon sequences more efficiently, producing diverse mRNA candidates for target proteins.
Phonetic Expert Routing Improves Accented Speech Recognition
A new low-rank adaptation method organizes experts by phonetic categories rather than accent, improving zero-shot accented English recognition.
Speaker-anchored routing keeps full-duplex dialogue tracks stable
A small routing module that matches separated speech streams to fixed speaker identities before diffusion cuts wrong-track time from 43.9% to 4.1% on remote conversations.
Audio Representation Autoencoders Match Dedicated TTA Latent Spaces
A new study finds that generation in self-supervised audio representation spaces can match or beat dedicated text-to-audio latent spaces, challenging reconstruction quality as the key metric.
Cross-Modal Neurons Explain How Audio-Language Models Align Sound and Text
Researchers identify shared neurons that align audio and text, revealing how large audio-language models recognize acoustic attributes.
Audio-Centric Model Speeds Up Drone Referring Segmentation
A new audio-centric approach lets UAVs follow spoken commands to segment targets in aerial images faster and more accurately than text-based methods.
TRACE benchmark exposes unreliable reasoning in audio-language models
A new benchmark reveals that large audio-language models often locate relevant sounds but fail to interpret them reliably, especially on ordering and counting tasks.
Pulse physiology simulator gets first open reliability atlas
A 1,280-run sensitivity study finds skewed input influence and flags operator-actionable failure modes in an open-source adult physiology simulator.