DailyPapers

TOME-Math: A structured textbook corpus for math reasoning Curated from 2,118 university textbooks into 380K proof-supervision records, preserving chapter order and proposition-proof links. Fine-tuning on it lifts average math benchmark performance by 16.7 percentage points.

NVIDIA just released a 4-bit quantized Kimi-K3.1-Pro on Hugging Face 2.8T parameters, 104B active, near-lossless.

Meta just released MetaEncoder on Hugging Face A unified multimodal encoder for text, images and video, with SOTA retrieval and decision support.

GTAGym by Alibaba's OpenMOSS team A real AAA game tasks environment for evaluating and training agents on GTA V, with 150 tasks across 11 categories and RL support.

Don't just trust your memory WikiTrace links each lesson to its original trajectory, so agents can check the evidence before acting. It reduces memory-induced failures by up to 9.7% across five benchmarks.

Meta just released XRStereo on Hugging Face A synthetic stereo video dataset with ~18k sequences - rectified stereo pairs, dense disparity, poses & intrinsics Built to train StereoSplat for metric-scale novel view synthesis

Learning from the world, not from weights RoboScript lets a frozen model learn manipulation from one demonstration by rehearsing in a simulator, writing the feedback into its context as a Task Memory — no gradient updates, and it hits 83.4% on 50 RoboTwin 2.0 tasks.

Tree-Ising: repairing LLMs with bit-level Ising edits This method fixes model weights by solving sparse discrete optimization problems on bit patterns, outperforming frontier-agent weight editing with just 4 forward passes and zero extra tokens.

MMDM: Measuring Massive Multitask Decision Making A new benchmark evaluates System-1 decision models with 5,448 questions across 16 categories, testing single-select, multi-select, and soft-label choices under long contexts and up to 1,024 options.

dNaViT: discrete codes as universal visual representation Meituan's new tokenizer maps native-resolution images to a shared discrete vocabulary, delivering competitive performance across understanding, reconstruction, generation, and editing – all in one model. Code and weights on Hugging Face.

Longtail-SWE A benchmark of 202 real issue-resolution tasks across 13 underrepresented languages (Clojure, Dart, Elixir, Erlang, F#, Haskell, Julia, Kotlin, Lua, OCaml, R, Scala, Swift), built with a solver–judge–repairer pipeline to make tests and descriptions reliable.

Flide An exact inference engine for decision models, up to 9.5× faster than vLLM and SGLang by forking states across questions in one forward pass.

Reframe: Teaching visual agents to use what they see A new post-training framework that grounds tool rewards and repairs observations during distillation. Beats prior art on 14 benchmarks without SFT cold start.

MiniMax just released the prompt expansion model for H3-Image on Hugging Face Turns a short text request into a detailed caption and aspect ratio recommendation for text-to-image generation.

LingBot-Touch: A Tactile Foundation Model for Dexterous Manipulation Robbyant releases a new model that learns unified representations across magnetic, piezoresistive, and visuotactile sensors, transferring to unseen hands and boosting real-world manipulation tasks.

Tencent just released a real-world group memory benchmark on Hugging Face It tests whether AI assistants can go beyond recalling past chats to actually using those memories to help group members

Google researchers introduce ReproBridge A workflow that turns papers and raw code into refined reproducible artifacts, raising agent reproduction success from 76.4% to 92.6% while cutting API costs by ~20%.

WorldFoundry Open-source infrastructure for reproducible world model inference and evaluation across video, 3D/4D, embodied action, and interactive systems. Defines what a generative world model truly is and tracks evidence with 270 models, 74 benchmarks, and 62 metrics.

Confucius4-R2T2 A streaming ASR model that produces append-only, stable transcripts without revisions, achieving near-offline accuracy with low latency. Open-sourced by NetEase Youdao.