Research Papers
EN ZH

Rankings

Best LLM Papers

Large language model papers from the last 90 days, ranked by three independent signals: how fast the official implementation is gaining stars on GitHub, how often the paper is cited, and how much attention it drew on Hugging Face. A paper that scores on only one of the three does not make the list. Rebuilt every day.

Ranking window: 2026-07-30 – 2026-10-27 Rankings last rebuilt 2026-10-28

1

Frontis-MA1: an open 35B meta-evolution agent for recursive self-improvement in ML engineering

Frontis-MA1 ships with OpenMLE, an open full-stack system (task gyms, RL operator learning, evolutionary search) for studying recursive self-improvement. Its 35B meta-evolution agent, post-trained around Draft, Improve, Debug and Crossover program-evolution operators, claims MLE-Bench Lite results the authors say exceed GPT-5.5 plus Codex, with components transferring to held-out NatureBench Lite.

  • 16 stars/7d
  • 3 citations
  • 186 upvotes

arXiv · FrontisAI/OpenRSI

3

YuE2: one Mixture-of-Transformers that writes a readable score first, then renders full-song audio from it

Symbolic music models plan composition explicitly but never sound like a finished recording, while audio models produce full songs that keep the composition hidden. YuE2 combines both in a single AR-NAR Mixture-of-Transformers that plans symbolically, expands the plan into semantic music tokens, and renders the full song, using two new representation models (MERT2, SheetSage2) to learn from recordings that lack aligned scores. The authors report a SongBench Global Avg of 6.73 on WildSongBench (6.96 at best-of-8), and in expert listening the best-of-8 samples were preferred over Suno v4.5 and roughly tied with Suno v5; the readable score also makes the output editable by external language-model agents.

  • 372 stars/7d
  • 0 citations
  • 216 upvotes

arXiv · multimodal-art-projection/YuE

4

WeVisOne-VL: one open VLM that detects, segments, counts, and reads text in a single model, down to pixel-level masks

Fine-grained perception usually means stitching together separate specialist models for detection, segmentation, counting, and OCR. WeVisOne-VL (from the WeChat Vision team) unifies object-level and pixel-level tasks in one VLM: a lightweight pixel decoder plugged in via a special <OBJ> token lets the same model output boxes, points, masks, and even alpha mattes at whatever granularity the user asks for. Trained with three-stage curriculum SFT plus mixed-task RL, the authors report leading results across 30+ benchmarks while still refusing queries where the requested object is absent, a failure mode they explicitly target for real-world deployment.

  • 335 stars/7d
  • 0 citations
  • 229 upvotes

arXiv · Tencent/WeVisOne-VL

5

Edge0 streams 35B MoE experts from SSD, predicting each layer's routing one token ahead

Edge0 is an open-source inference engine that serves a 35B mixture-of-experts model from a single 24GB GPU by loading expert weights from SSD on demand; a trained per-layer prerouter predicts the next layer's expert choices one token ahead so SSD reads overlap with compute, and a recovery LoRA pays back the quality lost to int4 quantization. The authors report 20 tok/s at about 3GiB of peak active memory, within a few points of the fp16 teacher on five public benchmarks.

  • 1,634 stars/7d
  • 0 citations
  • 8 upvotes

arXiv · Edge0-AI/edge0

6

How far can a VLM coding agent get with almost no robot scaffolding? A minimal harness reaches a reported RoboDojo state of the art

HexaShell is a deliberately thin robot-control harness: it only converts a coding agent's commands into motion, giving the agent three calibrated camera views, one command line, failure logs, and the ability to write reusable position-estimation tools based on camera-ray triangulation and ray-plane intersections. Trained on 10 layouts with one seed per task and then frozen, the tool library is tested on 10 unseen layouts per task across all 42 RoboDojo tasks using Codex with GPT-6-Astra at medium reasoning effort. The authors report 39.7% success and a 48.0 progress score, which they call a new state of the art on RoboDojo, with pushing at 92.5%, memory tasks at 65.0%, but tight insertion only 8.3%. Their analysis says the bottleneck is physical contact rather than reasoning: the agent often knows what to do but cannot execute the motion reliably. The tools are open-sourced.

  • 127 stars/7d
  • 0 citations
  • 289 upvotes

arXiv · HexaFuture/HexaShell

7

Code as Worlds: an agent that turns text or video into an executable physics simulation, then trains VLMs with it

Instead of describing physics in prose, Code as Worlds represents a scene as executable code that encodes object states, physical parameters, and governing dynamics. An agentic discovery loop, inspired by abductive reasoning, proposes, executes, renders, and iteratively refines world hypotheses from multimodal inputs such as natural-language descriptions or real-world videos. The verified executable worlds then provide scalable quantitative supervision for training Code-as-World-VL, which the authors report achieves state-of-the-art on QuantiPhy and surpasses leading proprietary models.

  • 14 stars/7d
  • 1 citations
  • 53 upvotes

arXiv · mirros-lab/code-as-world

8

Ropedia W0: one world model that predicts locomotion and manipulation for humans, humanoids, and robots

Ropedia W0 is billed as the first egocentric loco-manipulation world foundation model spanning humans and robots with diverse embodiments. To handle heterogeneous action spaces, the authors encode locomotion through camera poses and manipulation through projected skeletons, with robot skeletons built from embodiment-specific forward kinematics plus manual keypoint annotation. They pair it with Xp-Data (~10k hours of egocentric human video plus 3k hours of robot data) and a new long-horizon Xp-Teleop benchmark, and report state-of-the-art results on five cross-embodiment benchmarks and held-out egocentric video in zero-shot, few-shot, and full-data settings. They also show a distilled version running real-time, minute-long rollouts for VR interaction, policy evaluation, and model-based planning.

  • 108 stars/7d
  • 0 citations
  • 187 upvotes

arXiv · Ropedia/ropedia-w0

9

A 4B grounding model that points at objects precisely enough to serve as the eyes of robots

GroundingPI is a 4B foundation model that outputs points and boxes as quantized coordinates in a shared vocabulary, trained through multimodal and spatial pretraining, supervised fine-tuning, and GRPO reinforcement learning. The authors report state-of-the-art results across 34 grounding benchmarks covering 11 perceptual capabilities, averaging 73.68% against 44 baselines and edging out the larger GPT-6 Astra (71.54%). Used as the visual backbone for downstream tasks, they report it improves robotic manipulation, reaching the best results among evaluated backbones on RoboTwin 2.0 in all four out-of-distribution settings, and lowers open-loop L2 error to 0.296 m on nuScenes for autonomous driving.

  • 66 stars/7d
  • 0 citations
  • 260 upvotes

arXiv · groundingpi/GroundingPI

10

StartLux-Decision: 0.8B–27B decision models that answer confident cases in one pass and think only when uncertain

Every decision-model call today pays the full cost of a long reasoning trace even when the answer is obvious. StartLux-Decision trains a family of 0.8B–27B models with an Instant variant that answers in a single pass using bidirectional attention in full-attention layers, and an AutoThink variant that routes uncertain cases to explicit reasoning and answers confident ones directly. Its RLCD² stage freezes the model and rewards improvements in decision-probability quality with proper scoring rules, plus a soft penalty on overlong traces; the authors report this raises the share of reasoning traces that finish within the generation limit by up to 48 percentage points and cuts average reasoning length by up to 65% while largely preserving accuracy. They also release DecisionPivot, a programmatically verifiable benchmark of contrast pairs that differ in a single answer-changing detail.

  • 607 stars/7d
  • 0 citations
  • 10 upvotes

arXiv · StartLuxLabs/StartLux-Decision

11

Explorative Modeling: a third pretraining axis that trains on the best of K candidate generations instead of blurring modes

Generative models can't be trained end-to-end because multi-modal distributions force staged pipelines that blur modes together. Explorative Modeling factors the training loop: each step explores K candidate matches between model generations and data and trains only on the best, so predictions commit to a mode. The authors report monotonic gains across image, video and language as exploration scales, 4.1x better FLOP efficiency and 1.43 FID on ImageNet without guidance, and end-to-end reconstructive modeling that matches diffusion on control tasks with 16-256x fewer inference steps.

  • 0 stars/7d
  • 3 citations
  • 19 upvotes
  • #1 d4rl-maze2d

arXiv · alexiglad/XM

12

A benchmark that grades video models on when, where, and what, with audio-visual tracks and paired text/image prompts

OmniCompass evaluates video understanding along three complementary tasks: temporal grounding (when a target appears), spatial grounding (where it is), and question answering (what happens), in two tracks. OmniCompass-VL targets vision-language models with 820 videos and 2,460 instances, while an audio-rich subset forms OmniCompass-AV with 240 videos and 720 instances for models that must combine audio and visual evidence. Every instance is tested under paired text-only and text+image settings, and all data is manually annotated with multi-round verification. Evaluating 50 MLLMs, the authors report open-source models still trail proprietary ones, especially on audio-visual understanding, and that precise temporal and spatial grounding remains hard even for leading models.

  • 22 stars/7d
  • 0 citations
  • 574 upvotes

arXiv · PolyU-VCLab/OmniCompass

13

VideoJev: one 'System One' model that picks time spans and answers from video fast enough to act on

Video understanding agents usually need slow chain-of-thought decoding, which is too late for tasks that demand timely choices from a video stream. VideoJev instead casts temporal grounding and multiple-choice video QA as structured decision tasks: it swaps the MLLM's autoregressive answer head for a shared scoring head that rates candidate time spans and text options against the video directly, trained with GRPO. A cached- and joint-inference scheme reuses visual computation across sequential queries on the same video. The authors report mean-mIoU gains of 6.7 and 10.8 points on 4B and 35B-A3B backbones with 2.8-3.1× grounding speedups, and over 100× speedups on video QA while matching Thinking-mode accuracy.

  • 43 stars/7d
  • 0 citations
  • 205 upvotes

arXiv · MCG-NJU/VideoJev

15

Realtime-Venus: two open 9B full-duplex models that keep talking while delegating tool work to a background harness

Realtime-Venus is a proactive full-duplex interaction stack built from two separately trained 9B models: Realtime-Venus-Omni handles audio-visual conversation and Realtime-Venus-Audio handles spoken dialogue, each integrating continuous perception, conversational control and native speech generation on one shared causal timeline. A dual-loop runtime lets the foreground conversation continue while Realtime-Venus-Harness executes delegated tasks asynchronously and folds results back into the live dialogue; the authors report Realtime-Venus-Omni leads evaluated online models on six of eight video benchmarks and Realtime-Venus-Audio beats Gemini 3.1 Live and GPT-4o on all three Full-Duplex-Bench continuation metrics.

  • 28 stars/7d
  • 0 citations
  • 195 upvotes

arXiv · inclusionAI/Realtime-Venus

How this list is ranked

Large language model papers from the last 90 days, ranked by three independent signals: how fast the official implementation is gaining stars on GitHub, how often the paper is cited, and how much attention it drew on Hugging Face. A paper that scores on only one of the three does not make the list. Rebuilt every day.

A dash means the signal was unavailable, not zero.