Papers.Data.Code (@papersdatacode) header
Papers.Data.Code (@papersdatacode) profile picture

Only meaningful ML signals. Papers, repos & datasets. Selected, not collected. Home of The Weekly ML Signal → https://nitter.gq/t.co/wUFGgFytuv

Joined April 2026
📋 ML Weekly Recap · Nov 16 – Nov 22 ⚡ Trends ▸ Inference efficiency targets hidden bottlenecks with sparse routing, caching, and quantization ▸ Long-horizon interactive agents expand across games, software, web, and infrastructure benchmarks ▸ Embodied robotics increasingly leverages frozen foundation models, simulation RL, and richer datasets 🧭 TL;DR 📄 ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters Ruixi Yu, Dairu Liu, Xuchuan Chen et al. 552B frozen VLM backbone dramatically boosts robot control and data efficiency 📄 arxiv.org/abs/2611.21607 ⭐ MACaQ Practical cross-modal quantization cuts memory, time, and supports major runtimes ⭐ github.com/huawei-csl/MACaQ 💡 Frozen foundation backbones are becoming general controllers through efficient adaptation. → t.me/papersdatacode_digests
72
📄 Late Entry in Looped Language Models by Ziqi Jin, Jiaxi Li et al. Uses late entry instead of early exit: tokens run only final recurrent passes over deep context, while skipped passes are completed and verified asynchronously. Key points: • 5.2x exact speedup on Ouro-2.6B; +26% vs K=4 block on GSM8K • 8.4x exact speedup at T=64 on Huginn-3.5B; +153% vs K=64 block • 6.8x lossy speedup on Ouro-2.6B; +16.9 pts over trained early exit In looped LMs, deep cached context is more valuable than spending the same pass budget on token-local early passes.
2
35
📄 Game2Player: Scaling Gameplay Experience for Visual Agents through Environment-Player Co-Development by Xirui Jiang, Yuxi Chen et al. Builds verifiable game tasks from existing web games via builder-player co-development, then trains a pixel agent from code and VLM play with progress-weighted SFT. Key points: • PlayerBench progress improves 14.3% to 31.2% with Game2Player-9B • GameWorld progress rises 30.0% to 53.1% on an external benchmark • Player reports raise correct tasks 65%→72% and solved tasks 34→46 in 20 rebuilds Shows that jointly repairing environments through play and weighting supervision by verified progress can scale useful gameplay data for visual agents.
1
35
💻 ZhouTimeMachine/JVT Generates videos directly in pixel space with a DiT-style transformer for unconditional, class-conditioned, and camera-controlled synthesis. Key points: • Supports 3 modes: unconditional, class-conditioned, camera-controlled • Includes RGB-only and RGB-disparity joint generation on RealEstate10K • Releases training, inference, and evaluation scripts; full training uses 8×80GB GPUs Shows that a simple DiT-style transformer can handle pixel-space video diffusion across multiple conditioning setups. ⭐ 41 stars
1
103
📄 π^4: A Foundation Model for 4D Scene Geometry and Motion by Yifan Wang (@yyfz321021), Weiguang Zhao et al. A shared-backbone model that reconstructs geometry, poses, scale, tracks, and motion masks together; motion is predicted in 2D and converted to 3D via reconstructed geometry. Key points: • Best overall average rank across geometry benchmarks • Competitive single- and multiview 2D/3D tracking and motion segmentation • Uses ~0.5B params, less than half the size of most baselines Shows that conditioning tracking on a strong frozen geometry model can unify reconstruction and motion without direct 3D motion regression.
1
39
📄 ForceHarness: Turning Force into Action and Physics for Frozen Frontier Models by Ke Sun, Kevin Yuchen Ma et al. Closes force control under an API-only model with self-terminating skills, then translates force traces into physical quantities the model can reason over. Key points: • Real robot: 90.0% vs 3.3% zero-shot success for GPT-6 Astra • ForceBench: 87.5% vs 70.0% without force; beats 35.3% best trained • HiddenPhys: 96.4% vs 62.1% without force; beats 59.3% best trained Frozen frontier models can handle contact-rich manipulation well when force is exposed as closed-loop skills plus physics-level feedback, not raw sensor values.
1
31
📄 ScaleVLA: Scaling Vision-Language-Action Models to 552 Billion Parameters by Ruixi Yu, Dairu Liu et al. Freezes a 552B frontier VLM and adds a trainable action DiT, using one encoder pass per action chunk instead of text reasoning at control time. Key points: • LIBERO-PRO: 45.1% vs 31.3% best smaller backbone • LIBERO-Goal: 87.8% with 10 demos vs 85.6% for 2B with 50 • Real charger insertion: 9/10 vs 4/10 successes over 2B Frontier-scale pretrained VLM representations can materially improve robot policy data efficiency, robustness, and contact-sensitive control without per-step language reasoning.
1
57
📊 swe-cowork by Tokenwave-AI (@TokenWaveai) Contains 101 software engineering assignments on real open-source repos with workplace records, chat, and hidden-test evaluation for coding agents. Key points: • 101 tasks across 11 primary languages • 2,712 total requirements; 2,217 held only by colleagues • 2,701 F2P + 561 P2P hidden test nodes; 221 container images Useful for benchmarking coding agents on long-horizon repo work with requirement discovery through teammate interaction. 📥 214 downloads
2
46
📄 K2-Horizon: A Frontier Fleet of Open Source Language Models by K2 Team, Zhengzhong Liu et al. A fully open LLM fleet plus training data, logs, checkpoints, and code; it pairs MoVA sparse value routing with softplus gating and Uno decoding adapters. Key points: • 375B-A23B gets 65.3 on Toolathlon-Verified, beating GPT-5.6 Terra • 375B-A23B scores 48.4 on SWE-Atlas-QnA, leading open-weight peers • Uno speeds decoding by up to 2.2x without a separate draft model Shows sparse value routing can retain dense-model quality at far lower active parameter count while a fully open release enables controlled scaling and sparsity studies.
1
2
90
💻 huawei-csl/MACaQ Quantizes multimodal models for local inference and exports GGUF or compact checkpoints for llama.cpp, vLLM, Diffusers, and PyTorch. Key points: • Quantizes 122B params in 64 min using 32.5 GB device memory • vLLM W4A4 gets 2.33× prefill and 2.58× decode vs BF16 • MACaQ-R cuts KL on Qwen3.6-35B-A3B from 0.015731 to 0.013367 Shows calibrated low-bit quantization can extend beyond LLMs to a shared workflow for text, image, video, and audio models. ⭐ 337 stars
2
3
525