DailyPapers
@HuggingPapers
Tweeting interesting papers submitted at
Submit your own at and link models/datasets/demos to it!
4 Following 21.6K Followers
Alibaba releases a new training framework for software engineering agents
It develops category experts via iterative Refresh–Repair–Expand (RRE) and consolidates them into one deployable model with label-routed multi-teacher distillation (MOPD). Achieves 58.04% on Pro-618 and 59.00% on SWE-bench Multilingual.
Show more
SpeakerMem-R1
A speaker-centered dual-track memory for multi-party dialogue that stores verbatim messages and organizes derived states by person and group, tackling attribution and state reconstruction.
Show more
Microsoft just released rho-roboeval on Hugging Face
A new model for robotics evaluation.
RoboDawn: Transferring the Intelligence of VLMs to Robotic Control
Tsinghua and Tencent Hunyuan show a frozen VLM can drive robots through simple move, rotate and gripper commands, reaching 73.6% one-shot success on RoboTwin 2.0 C2R.
Show more
Tencent ARC Lab just released GameHorizon Suite
A unified data and evaluation suite measuring AAA gameplay capabilities across multiple temporal horizons for VLMs, UMMs, GUI, coding, and game agents.
Show more
Alibaba's Qwen team just released RecreationWorld on Hugging Face
A five-platform sandbox where hybrid computer-use agents explore real apps, implement them, and visually verify their own builds.
Show more
Code2Skill
A fully automated pipeline that transforms 19,769 open-source repositories into over 1 million verified, reusable procedural skills for AI agents.
NVIDIA's SoL-Pi, now on Hugging Face paper pages
A token-efficient agent harness built by recursively scaling auto-research loops. It cuts token traffic by 44.7-49.0% and API cost by about one third while preserving performance.
Show more
Coding agent harnesses, taken apart
A new study isolates planning, action space, and context management across 176 settings. Key finding: context management matters most when budgets are tight, and planning flips from accuracy to efficiency as models improve.
Show more
NVIDIA just released FoundationPose on Hugging Face
A unified foundation model for 6-DoF object pose estimation and tracking that works on novel objects without fine-tuning.
PhysBrain 1.5: Open-source SOTA for Embodied AI
A single 8B model that understands scenes, generates robot actions, and predicts future frames — all as tokens. 72.5 on 28 benchmarks, beating all open models.
Show more
DataFlex-RL
A controlled evaluation platform for RLVR data policies, comparing rollout selection, reweighting, and domain-mixture choices under a shared GRPO recipe.
SWE-Bench Pro Verified
OpenCompass released a verified version of SWE-Bench Pro that fixes reward hacking and task-quality issues, revealing frontier models score far lower than previously reported.
Show more
AgentGrad: Intervention-guided Prompt Optimization
Improves multi-agent prompt optimization via sequential intervention and semantic textual gradient abstraction, achieving SOTA on 5 MAS benchmarks and 2.5x faster optimization.
Show more
Tencent releases AuK, a unified speech generation and editing model
It handles zero-shot TTS, instruction-driven content/acoustic/paralinguistic editing, plus enhancement and separation through one natural-language interface. The distilled AuK-Flash version runs 4.5x faster.
Show more
FlowBalance: verifier-grounded self-improvement for reasoning models
Improves math reasoning by +2.12 avg over GRPO on Qwen3-8B, with faster training, better stability, and higher solution diversity.
Show more
Why the recurrent half of a hybrid LLM is actually easy to quantize
Minima quantized all 496 linear layers of Qwen3.8-27B — including Gated DeltaNet — to NVFP4 W4A4, matching BF16 performance while cutting size ~2.9×
Show more
NVIDIA just released the NVFP4 quantized Qwen3.8-Flash-Next on Hugging Face
125B MoE with hybrid attention, now 63% smaller with minimal accuracy loss.
Compile by Training
Turn natural-language specs into small neural programs that run locally on a compact interpreter. Compilation takes about a minute, and on FuzzyBench-Hard it hits 83.6% semantic accuracy.
Show more
Terminal-Universe
A framework that reconstructs realistic terminal environments from agent trajectories, yielding 37.3k task-sufficient workspaces for post-training and evaluation.
Show more
