DailyPapers

Paper: paperswithcode.co/paper/2607.114… Code: github.com/zjunlp/LightMe…

LightMem-Ego: Your AI Memory for Everyday Life A lightweight streaming multimodal memory system for smart glasses and smartphones that continuously captures egocentric visual and audio streams and organizes them into current, short-term, and long-term memory.

Accepted to SIGGRAPH 2026. Paper: paperswithcode.co/paper/2607.091… Project page: sisyphm.github.io/studiorecon-pa…

StudioRecon Reconstructs dynamic 4D human scenes from just 4 low-overlap cameras by decoupling background and humans — video diffusion densifies the background, SMPL constrains the humans, and a recursive module harmonizes both.

Paper: paperswithcode.co/paper/2607.053… Models on Hugging Face: huggingface.co/collections/By…

Direct-OPD: Weak-to-Strong Generalization via Direct On-Policy Distillation ByteDance Seed reuses RL exploration from a small model as a dense implicit reward. Boosts Qwen3-1.7B by +10 points on AIME24 in just 4 hours on 8 A100s.

Check it out here: huggingface.co/nvidia/Nemotro… It tops the multilingual RTEB leaderboard: huggingface.co/spaces/mteb/le…

NVIDIA just released the Nemotron-3 Embed model on Hugging Face State-of-the-art multilingual text embedding with 8B parameters, supporting 34 languages for retrieval and semantic search.

Paper: paperswithcode.co/paper/2607.103… Project page: amap-cvlab.github.io/ABot-Navigatio…

Alibaba's ABot-N1 A visual language navigation foundation model that decouples cognition from control, achieving a 35% gain in POI arrival to 77.3% and 95.4%/92.9% SR in complex indoor and outdoor scenes.

Paper: paperswithcode.co/paper/2607.103… Introduces EmbodiedWorldBench: 16 indoor/outdoor scenes, 4 difficulty levels, over 200 tasks for navigation, object search, NPC dialogue, and dynamic events.

ABot-AgentOS A general robotic Agent Operating System that provides scene-conditioned planning, context-isolated skill execution, multi-modal memory, and self-evolution for long-horizon embodied tasks.

Paper: paperswithcode.co/paper/2607.065… Code & pretrained models: huggingface.co/xmz111/ReChann…

ReChannel A new way to extract depth, normals, matting, and segmentation from a single image using a frozen text-to-image DiT — with only 33K trainable parameters per task.

Paper: paperswithcode.co/paper/2607.079… All 30+ quantized models (W2, W3, W4) are on the Hugging Face Hub: huggingface.co/donghyunli

KronQ A new post-training quantization framework that introduces gradient covariance into the quantization objective, achieving 7.93 perplexity on Llama-3-70B at 2-bit where GPTQ diverges

Paper: paperswithcode.co/paper/2607.047… Read more on Hugging Face Paper Pages

Trust Region Policy Distillation (TOP-D) Transforms unstable on-policy distillation into a stable training paradigm by dynamically constructing a proximal teacher, improving sample efficiency and final performance at zero additional computational cost.

Paper: paperswithcode.co/paper/2607.096…

Scalable Visual Pretraining for Language Intelligence Training language models directly on visual documents—figures, equations, layouts—outperforms text-only pretraining, challenging the default assumption that LMs need plain text.