Research Papers
RLE-Bench Tests Coding Agents as Robot Learning Engineers
Hugging Face paper introduces RLE-Bench, a benchmark that evaluates coding agents across four robotics development workflows and a...
ScholarCatalyst Benchmark Tests Whether AI Can Find the Papers That Spark New Research
A new benchmark built from the judgments of 184 lead authors shows that today's retrieval systems and LLM agents fall well short o...
AgSpec Boosts Coding-Agent Throughput Up to 4.76x via Retrieval-Based Speculative Decoding
Hugging Face paper AgSpec introduces corpus and draft-length policies that let retrieval-based speculative decoding work inside co...
OpenTumorBoard Benchmark Tests LLMs on Real Tumor Board Discussions
A new benchmark built from 219 publicly recorded multidisciplinary tumor board meetings evaluates how well large language models c...
OmniSeek Turns Omni-LLMs into Active Multi-Turn Audio-Visual Reasoners
A new agentic framework, OmniSeek, teaches an Omni-LLM to decide when to look or listen and over which time window, retrieving spa...
DataMagic Turns Raw Tables into Data Videos via Multi-Agent Orchestration
A new Hugging Face paper introduces DataMagic, a declarative multi-agent system that authors data videos directly from raw tabular...
Hugging Face Paper Proposes RouteFM: A Foundation Model That Learns to Route LLM Queries Once, Then Adapts Anywhere
A new paper introduces RouteFM, a foundation-model approach to LLM routing that infers the capabilities of anonymous candidate mod...
Hugging Face Paper Introduces HIDE Benchmark and SEEK Framework for Robotic Manipulation Memory
A new paper proposes HIDE, a 15-task benchmark for evaluating manipulation memory under partial observability, and SEEK, a framewo...
Hugging Face Paper Introduces Neighborhood OPSD for Stronger Math Reasoning Distillation
A new Hugging Face paper proposes Neighborhood On-Policy Self-Distillation (N-OPSD), which builds a pool of complementary teacher...
Hugging Face Paper Unveils ATR: A Multilingual Lip-Sync Judge for Dubbing Quality Control
Researchers introduce Align Then Reason (ATR), a reference-free multimodal judge that aligns lip motion with phonetic units before...
Hugging Face Paper Exposes 'Intent Drift' as a Measurable Multi-Turn Failure in LLM Agents
A new Hugging Face paper introduces IntentFlux, a benchmark showing that LLM agents degrade when user intent changes mid-conversat...
Hugging Face Paper 'Where-OPD' Uses Synthetic Scenes to Sharpen Multimodal AI Perception
A new paper proposes Where-OPD, an on-policy self-distillation method that gives a teacher model textual spatial guidance derived...