DailyPapers

Tweeting interesting papers submitted at your own at and link models/datasets/demos to it!

[DailyPapers @HuggingPapers]

Explore StreamEval, the open benchmark for interactive streaming video generation. Evaluate your models, see the leaderboard, and contribute.

Paper: https://t.co/5tR9kkmPxT
Dataset: https://t.co/ROqquY01Bz
Leaderboard: https://t.co/nfWrWtV67A
Code: https://t.co/vJrfhafcda

[DailyPapers @HuggingPapers]

StreamEval: The first benchmark for interactive streaming video generation

It evaluates 37 models across 5 tasks with 300 scripted test cases, revealing that reliable request execution remains a central challenge.

[DailyPapers @HuggingPapers]

1M-token context, MIT license

HLE 42.6 · Terminal-Bench 42.3 · DeepSWE 77.0

Model: https://t.co/sIwQwuRGHB

[DailyPapers @HuggingPapers]

DeepSeek just released an upgraded Flash model on Hugging Face

Better reasoning, coding, agentic tasks and visuals, with faster execution and shorter outputs at lower cost.

[DailyPapers @HuggingPapers]

Paper: https://t.co/KEvyLZMrSf

Code: https://t.co/2iVLGJ6o4r
Project: https://t.co/KWJyAbkI9n

[DailyPapers @HuggingPapers]

AKV: Cache Engineering for Agents

A new serving system that lets LLM agents directly manage their KV cache — dropping unused history and repositioning retained keys without rewriting or re-prefilling. Reduces average cost by 27.7% while improving task accuracy by 1.7 points.

[DailyPapers @HuggingPapers]

Paper: https://t.co/Ft8gESfYQ5

Envs & tasks: https://t.co/VFGBblQuKS

Models: https://t.co/fi5q3vv69V

[DailyPapers @HuggingPapers]

Microsoft Research's WildAWM

A scalable pipeline that synthesizes executable computer-use agent environments and tasks using only open-source models, yielding 341 long-horizon env-task pairs and Qwen agents that generalize to 8 unseen benchmarks.

[DailyPapers @HuggingPapers]

Each history is a byte-preserved caption log with timestamps, episodes, and evidence labels.

Dataset: https://t.co/vB68h0PlRS

Code: https://t.co/5qORbBCC7E

[DailyPapers @HuggingPapers]

Meta just released MemProBench on Hugging Face

800 frozen memory histories from real-world Ego4D activity, built to study memory-driven proactive assistance: when should an assistant speak up, not just what to say?

[DailyPapers @HuggingPapers]

Paper: https://t.co/AlNGMsD81V

Model: https://t.co/wSxLBtdAtr

Project: https://t.co/nzyrRTcmfT

[DailyPapers @HuggingPapers]

Sage-A1.5

Shanghai AI Laboratory just released a 35B MoE agent on Hugging Face that reaches trillion-parameter performance: it orchestrates parallel subagents, verifies its own progress, and conducts open scientific discovery.

[DailyPapers @HuggingPapers]

Paper: https://t.co/yDzpvAv0HJ

Project: https://t.co/GdH1qQmq24
Code: https://t.co/XKgefRj5if

[DailyPapers @HuggingPapers]

S-Agent 2.0: Spatial reasoning with a general coding agent

A command-line perception toolkit closes the evidence loop — acquire, remember, interpret — boosting a coding agent by +20.5 on ReVSI and +13.9 on MMSI.

[DailyPapers @HuggingPapers]

Paper: https://t.co/Ax0wbIHnFd

Model: https://t.co/R5F7o0xVi7

Dataset: https://t.co/hMoYIU5gO7

[DailyPapers @HuggingPapers]

Ropedia W0

A world model that learns loco-manipulation from 10k hours of human egocentric video, then transfers across embodiments, from robot arms to humanoids and from grippers to dexterous hands.

[DailyPapers @HuggingPapers]

Paper: https://t.co/pENRuCMjC6

Models, weights & game assets (Medieval Castle, Ancient Egypt, City Racing): https://t.co/hc3BiNHQWT

[DailyPapers @HuggingPapers]

Skywork AI's Matrix-Game 4.0

A playable game world rendered by a video model in real time: 480p at 60 FPS on a single RTX 5090, with a key-press-to-display latency of just 112ms.

[DailyPapers @HuggingPapers]

Paper: https://t.co/8d9sRntK5Q

HF paper page: https://t.co/8d9sRntK5Q

From Ant International researchers. Code and models would be a great addition to the Hugging Face Hub.

[DailyPapers @HuggingPapers]

PAR: Progress-Aligned Recurrence for length generalization in looped LMs

Loop k learns to predict the state after k operations, and recurrent states are rescaled to first-iteration norms. A 0.5B Qwen2.5 trained on 32 steps hits 95.5% at 512 steps—16x extrapolation.