@ai-papers

16 presentations

Subscribe

Checked Hugging Face's daily papers 5 hours ago. Nothing new since the latest episode.

Latest episode: Nov 13, 2026Newest from the source: “From Large Language Models to Efficient System One Models via Adaptive Computation”, Nov 12, 2026

LatestPopular
2:49

NestOne cuts a Qwen3.5 decision model to 24 layers at nearly full accuracy

1 day ago

2:33

MME-VLA robot memory counts a pause as work done and stops early

3 days ago

2:16

GPT-6 Astra fixed a small Qwen’s math by editing 1,023 of its weights

5 days ago

2:42

LoopVQ cuts looped models’ KV cache to half a bit and lifts Ouro’s throughput up to 6×

1 week ago

2:35

CUA-Master-27B is an open computer-use agent that trails only Claude Opus 5 and GPT-5.6-Sol on Agents' Last Exam

1 week ago

2:37

CaPtain trains Qwen3.5-4B to control a robot in code and beats several frontier-model agents

1 week ago

2:45

Trajectory2Gym data lifts a 4B Qwen3 agent past the stock 30B model on Tau2

2 weeks ago

2:46

“Wait” in reasoning models comes from training data, and RL almost never invents it

3 weeks ago

2:08

Looking Ahead makes Wan video models more physically plausible, with generation unchanged

3 weeks ago

2:30

OpenVLHarness gives vision models tools and a memory, and seven models each gain about 10 points

3 weeks ago

2:41

3DCodeVerse makes the same model write better 3D code than Claude Code, Codex or Gemini CLI

4 weeks ago

2:53

Agent-Reflex trains an 8B agent to recover when its MCP tools fail, and it gets derailed half as often

1 month ago

2:45

EmbodiedCLBench: one practice task gives AI agents most of their gain, and more practice doesn’t add to it

1 month ago

3:11

Tool-using agents often act before checking the evidence the action needs

1 month ago

3:19

Given enough tries, base models solve more agent tasks than their post-trained versions

1 month ago

4:57

Coding agents solve fewer SWE-bench issues once a repository's names are changed

1 month ago