Confucius4-R2T2: A Stable-Prefix Streaming ASR Model for Real Real-Time Transcription
Confucius4-R2T2 adapts a full-context ASR model for append-only streaming output, using Longest Stable Prefix learning to balance accuracy, latency, and stability.
ls ./papers
Trending AI research papers from Hugging Face, summarized in plain language: what each paper does, why it matters and where to read more.
140 papers
Confucius4-R2T2 adapts a full-context ASR model for append-only streaming output, using Longest Stable Prefix learning to balance accuracy, latency, and stability.
Weight-Space Programming lets reasoning agents edit model parameters directly. Results show striking localized gains, uneven transfer, and serious reliability challenges.
EvoWorld gives generative video models a reasoning layer for events that happen off-screen, improving subject memory, causal reactions, and revisit consistency.
Agent evolution can raise benchmark scores while losing previously solved tasks. This study measures reliability, coverage, cost, transfer, and the mechanisms behind durable improvement.
HeLM gives robot policies an explicit language memory for counts, sequences, and cross-episode context, improving long-horizon manipulation while exposing key limits.
A new SAE-based analysis finds that MoE experts contribute to distinct latent features, revealing fine-grained roles such as Java, web, geometry, and arithmetic specialization.
Muir redesigns linear attention by allocating gate rank according to sensitivity and using matched-filter readout, improving memory retrieval, language modeling, and long-context efficiency.
SwarmEvolve replaces verbose agent reasoning with latent KV-cache communication, enabling large evolutionary swarms that search mathematical and scientific programs faster.
DR E A M FR E E distills a video diffusion model's knowledge into current-observation features, delivering faster robot control and promising video-only task transfer.
TB-MathModel evaluates AI agents on 18 multimodal mathematical-modeling tasks with terminal execution, multi-reference scoring, and Lean-checked feasibility gates.
PLatour shows how large models can guide smaller decoders during prefill, while PACT-Check reveals why benchmark gains do not always prove genuine information transfer.
Hao Fei proposes cognitive symbionts: lifelong, private AI companions that close the gap between what people know and what they can access in the moment.