Tencent's WeMM-Video-9B is now the best open model on DailyOmni: 83.6% accuracy, just 1 point behind closed Qwen3.5-Omni-Plus at 84.6%! 🎬
It's a 9B omni model (text, image, audio and video), with weights on Hugging Face under Apache 2.0.
It comes from WeMM-Video, Tencent and Tsinghua's new family of omni-modal models from 2B to 35B-A3B, trained on about 1.2 trillion tokens.
1
99
DailyOmni tests audio-visual reasoning on real-world videos of daily life. Weights: tencent/WeMM-Video-9B
paperswithcode.co/benchmark/…
52
Seyed Mahmoud Sajjadi Mohammadabadi just released a sobering meta-analysis for anyone building multi-agent LLM systems! 🤖
"More Agents, More Samples, or More Thinking? Untangling Coordination, Ensembling, and Test-Time Compute in Multi-Agent LLMs"
This paper shows that multi-agent gains largely disappear once you compare against a single model given the same number of samples or the same compute budget.
It covers 516 multi-agent configurations from nine evaluations, an average accuracy change of -0.042 against matched baselines, multi-agent setups at or below the baseline in 74% of comparable contrasts, when coordination can pay off (partitioned evidence, external verification, error diversity across models), and much more!
Read this excellent paper by Seyed Mahmoud Sajjadi Mohammadabadi: paperswithcode.co/paper/2611…
1
192
The paper also ships a reviewer checklist and accounting utilities for fair multi-agent comparisons.
Code: github.com/mahmoudsajjadi/pa…
105
Maksym Andriushchenko and collaborators at ELLIS Institute Tübingen just released a fascinating case study: an autonomous agent running its own open-model organisation on Hugging Face! 🤗
"Autonomous agents can run successful open-model organisations. Evidence from an 8-week case study"
This paper shows what happens when you give an agent a blank Hugging Face account, compute, and one goal: earn real adoption.
It covers Horizon, an agent built on Opus 5.5 running 24/7 in Claude Code, 600+ GPU-hours, 38 multilingual models fine-tuned from mmBERT (prompt-injection guards, PII redactors, rerankers), 50,000+ downloads, use in 100+ GitHub repos, and much more!
Read this excellent paper by @AWuhrmann, Maksym Andriushchenko and collaborators: paperswithcode.co/paper/1334…
External review finds Horizon's models often on the size/performance Pareto frontier, and the team releases the full agent trace.
Code: github.com/horizon-ai-labs/a…
1
235
HKUST, CUHK and Zhejiang University researchers, with Ser-Nam Lim, just released WorldDraft, a sharp test of a big question for video generation: if code builds the world, can video models just render it? 🎬
"WorldDraft: If Code Builds the World, Can Video Models Just Render It?"
This paper checks whether video models can keep prescribed 4D motion and camera paths while rendering the appearance a text prompt asks for.
It covers 432 paired test cases across 48 executable 4D world families, 12 video models evaluated, why richer drafts anchor structure but hurt appearance fidelity, source leakage of up to 64% under conflicting cues, WorldDraft-Primer, a plug-and-play visual prompt that improves both, and much more!
Read this excellent paper by Zhiyu Hou, Wen-Jie Shu and collaborators: paperswithcode.co/paper/2611…
1
204
CHORD is the new SOTA on Waterbirds: 93.11% worst-group accuracy! 🐦
It just edges out HSFM (93.1%), with DaC-C (92.6%) next.
Waterbirds crosses land and water birds with land and water backgrounds, so a model that leans on the background shortcut fails on the rare combinations.
CHORD's idea: a bias hidden in one feature space may be clearer in another. It debiases a ResNet-50 through consensus across several representation spaces, with no bias labels.
1
137
From Sichuan University and UESTC (Weiwei Li, Wen Li and collaborators). It improves worst-group accuracy on seven benchmarks across vision, language, speech and medical imaging, and code and checkpoints are released.
Full leaderboard: paperswithcode.co/benchmark/…
82
KAIST researchers, with Jaegul Choo, just released eDIAR, speaker diarization for hour-long meetings with no cap on how many people speak! 🎙️
"eDIAR: Training-Free Open-Speaker Streaming Diarization for Long Meetings and Large Archives via e-means"
This paper removes the fixed speaker limit of slot-based diarizers like Streaming Sortformer and Nemotron-3, without any retraining.
It covers e-means clustering inside a slot once every slot is used, bounded-window re-clustering so cost per update doesn't grow with the meeting, far better accuracy than Streaming Sortformer on hour-long meetings, archive-scale clustering several times faster than Lloyd's k-means, and much more!
Read this excellent paper by Kyudan Jung, JiHwan Kim and collaborators: paperswithcode.co/paper/2611…
1
94
The same recipe turns the four-slot Streaming Sortformer into an open-speaker system.
Project page: kyudan1.github.io/eDIAR/
49
M3OPD-Reasoner-9B is the new SOTA on LogicVista: 79.6% accuracy! 🧠
That's 11.1 points ahead of the previous best, MMVistaReason-9B (68.5%).
The recipe: train separate teachers for image STEM, perception and video, distill them into one Qwen3.5-9B student with on-policy distillation, then add RL on top.
Data, training code and every checkpoint are released, with weights on Hugging Face under Apache 2.0.
1
66
LogicVista: 448 multiple-choice questions testing logical reasoning in visual contexts. Paper by Nguyen Quang Trung, Hoang Phan, Shuo Sun, Longyin Zhang and AiTi Aw.
Full leaderboard: paperswithcode.co/benchmark/…
40
Congrats @AIatMeta, KAUST and Princeton: AVTR-9B is the new SOTA on CharadesSTA video temporal grounding! 🎬
It reaches 64.7 mIoU, ahead of VideoAuto-R1 (63.7) and TempSamp-R1 (62.1).
AVTR adds "visual imagination" to reasoning: instead of only thinking in text, the model can revisit what it saw through generative visual reflection, and learns when to use each.
The weights are on Hugging Face.
1
81
It's also #1 on VSI-Bench spatial reasoning (71.0), and across 16 benchmarks it adds 4.5 points over its Qwen3.5 base while cutting inference latency by 33%.
Full leaderboard: paperswithcode.co/benchmark/…
44
HKU, HKUST (Guangzhou) and NUS researchers just released MEM-SQL, which lets text-to-SQL systems reuse the verified queries they've already written! 🗄️
"MEM-SQL: Memory-Empowered Text-to-SQL"
This paper shows that earlier verified queries over the same database can make new SQL both more accurate and cheaper to generate.
It covers HARE, a new benchmark of 1,600 history-aware requests, memory of whole-query templates, reusable SQL blocks and concept-to-schema mappings, on-demand schema linking, execution accuracy up from 37.1% to 50.0% with 76.6% fewer input tokens and 68.8% lower latency, and much more!
Read this excellent paper by Xiaoqi Wang, Wenqi Pei and collaborators: paperswithcode.co/paper/2611…
1
54
The open problem: requests that reuse several verified queries, where accuracy falls from 56.7% to 29.7%.
paperswithcode.co/paper/2611…
20
ZERITH V2.5 is the new SOTA on LIBERO-Plus zero-shot transfer: 93.7% success! 🤖
That puts it ahead of InternW0-Δ (92.8%) and Qwen-RobotManip-Context (91.4%), out of 69 entries.
LIBERO-Plus tests robot policies trained on standard LIBERO demos under layout, camera, lighting, language and sensor-noise shifts they never trained on.
ZERITH V2.5 is a video-action model whose robot policy reads the features of a video generator.
1
65
It also hits 92.3% on Meta-World MT50. The report digs into why adapting a video generator at many noise levels helps the policy.
Full leaderboard: paperswithcode.co/benchmark/…
21










