Papers, benchmarks, datasets, and experimental advances worth tracking
APWA proposes a distributed architecture designed for parallelizable agent workloads, addressing the inference, coordination, and computational scaling bottlenecks that multi-agent systems face as task scale and complexity increase.
DDC proposes a unified inference-time scaling framework that reduces token consumption by over 10x while maintaining or surpassing baseline accuracy across 5 benchmarks, utilizing a confidence-weighted Bayesian protocol and trend-aware stratified pruning.
MemEye introduces a visual-centric evaluation framework for multimodal Agent memory, born from the collaboration of 17 researchers, filling a gap in the evaluation of Agent memory systems.
NVIDIA releases MemLens, the first benchmark targeting the multimodal long-term memory capabilities of Large Vision-Language Models (LVLMs), filling a gap in LVLM memory evaluation.
Microsoft Research has open-sourced Orchard, a scalable Agent modeling framework. Spanning from Code Agents to GUI Agents and personal assistants, it achieves cross-domain training through a unified, lightweight environment layer. It achieves 67.5% on SWE-bench Verified, while its GUI Agent reaches the strongest open-source performance using only 400 distilled trajectories.
OpenDataLab has released the CiteVQA benchmark, specifically designed to measure the evidence attribution capability of answers in document intelligence systems. Topping the HuggingFace Daily Papers with 143 upvotes, it signals that trustworthy AI is evolving from a slogan into a quantifiable technical metric.
The CLI-Anything project, released by HKU's HKUDS team, topped GitHub Trending with over 36,000 stars. Its core philosophy is to make all software "Agent-Native"—not just a simple tool, but a fundamental shift in software architecture thinking.
Shanghai Jiao Tong University has released the MMSkills framework, decoupling the capabilities of multimodal visual agents into composable and reusable skill units. Garnering 99 upvotes to top Hugging Face's trending papers, this suggests that the "skill-based" evolution of agents may be closer to the future than the "model-based" approach.
DeepCybo has released the PhysBrain 1.0 technical report, building an AI system capable of understanding physical laws. From intuitive physics to video generation verification, this technical route may bring us closer to true "intelligence" than pure language models.
The Tencent Hunyuan team has released a new paper systematically studying the efficiency of On-Policy Distillation in unlocking model potential. The paper reveals the critical impact of distillation strategy selection on model performance, providing empirical evidence for large-scale model training.
TideGS leverages an SSD-CPU-GPU hierarchical storage management system to train 3DGS with over 1 billion Gaussian primitives on a single 24GB GPU. This represents a 10x improvement over previous out-of-core baselines (~100 million) and roughly a 100x increase over in-memory training (~11 million). The paper has been accepted as a Spotlight at ICML 2026.
Anti-SD identifies, via pointwise mutual information (PMI) analysis, that privileged context suppresses models’ reasoning deliberation tokens—and proposes “anti-self-distillation”: intentionally increasing, rather than decreasing, divergence between student and teacher. On mathematical reasoning benchmarks, it achieves the same accuracy as the GRPO baseline in just 2–10× fewer training steps, with final accuracy gains up to +11.5 points.
CogOmniControl proposes a reasoning-driven framework for controllable video generation, decomposing the generation process into two stages: creative intent cognition and video synthesis. CogVLM—trained on professional anime production data—accurately interprets sparse, abstract conditions; combined with CogOmniDiT and RL alignment, it outperforms existing open-source models on two newly established benchmarks.
GoLongRL introduces a fully open-source reinforcement learning (RL) post-training framework for long-context language modeling, releasing a 23K-sample RLVR dataset and complete training code. The Qwen3-30B-A3B model achieves performance on long-context tasks comparable to DeepSeek-R1-0528 and Qwen3-235B-A22B-Thinking-2507.
OpenComputer introduces a verifier-based framework to build a verifiable software environment for computer-use agents. Covering 33 desktop applications and 1,000 tasks, experiments show that its hard-coded verifiers align more closely with human judgment than LLM-as-judge approaches.
A joint AI research roadmap titled *AI for Auto-Research: Roadmap & User Guide*, released by institutions including the National University of Singapore, systematically analyzes the capability boundaries of AI across the full research lifecycle: papers can be auto-generated for just $15, yet LLMs still fabricate results, overlook hidden errors, and cannot reliably assess scientific novelty.
IAAR-Shanghai and Memtensor Research Group propose SkillsVote, a full-lifecycle governance framework for Agent skills. Offline evolution improves GPT-5.2 by 7.9 percentage points on Terminal-Bench 2.0, while online evolution boosts SWE-Bench Pro by 2.6 percentage points.
ByteDance has released Lance—a natively unified multimodal model trained from scratch—that supports understanding, generation, and editing of images and videos. Leveraging a dual-stream Mixture-of-Experts (MoE) architecture, Lance significantly outperforms existing open-source unified models in generation quality while retaining strong understanding capabilities.
Hugging Face’s #1 Paper of the Day—a survey paper authored by 42 researchers (including prominent academics and industry scientists)—systematically introduces the “Code as Agent Harness” framework, positioning code as a unified infrastructure layer for agent reasoning, action, and environment modeling.
The NVIDIA team has released LongLive-2.0, the first full-stack system for long video generation training and inference based on NVFP4 precision. By introducing sequence-parallel autoregressive training and W4A4 inference, it achieves a 2.15x speedup in training and a 1.84x speedup in inference, with the 5B model reaching 45.7 FPS.
The NUS team has released the *AI for Auto-Research* roadmap, systematically analyzing the reliability boundaries of AI across the entire research lifecycle. It clarifies which stages—from idea generation to publication—AI can handle independently and which still require human oversight.
The Tsinghua team proposes KVPO, an ODE-Native online GRPO framework that aligns autoregressive video generation models with human preferences by shifting the exploration source from random noise to historical KV Cache, achieving improvements in visual quality, motion quality, and text-video consistency.
The Tsinghua team introduces ZEDA, a low-cost framework that converts pre-trained static MoE models into dynamic MoE. It eliminates over 50% of expert FLOPs on Qwen3-30B-A3B and GLM-4.7-Flash, achieving approximately 1.2x end-to-end inference speedup.
ByteDance Research has released Lance, a lightweight native unified multimodal model. Through a dual-stream MoE architecture and multi-task synergistic training, it simultaneously achieves understanding, generation, and editing of images and videos without relying on scaling up model capacity.
NVIDIA has released LongLive-2.0, a long video generation infrastructure built on NVFP4 quantization and parallel inference. With 1.22k GitHub Stars, it explores how to generate longer video sequences without sacrificing quality.
The ARIS system, released by Shanghai Jiao Tong University, enables multiple AI agents to autonomously complete research tasks through adversarial collaboration. It has received 116 upvotes on Papers with Code and 9.7k GitHub stars, making it one of the most notable AI for Science projects recently.
Tsinghua ML Group Causal Forcing++ paper proposes a scalable few-step autoregressive diffusion distillation method, transforming interactive video generation from a multi-minute wait into real-time response. What does this mean for gaming, VR, and interactive content creation?
Darwin Family introduces a training-free evolutionary merging framework that composes the latent capabilities of existing models via gradient-free weight-space recombination. Its flagship model, Darwin-27B-Opus, achieves 86.9% on GPQA Diamond—ranking 6th among 1,252 evaluated models—without any gradient-based training.
The new arXiv paper FORGE proposes a method that enables Agent memory to self-evolve without updating model weights. Through a population broadcast mechanism, Agents can share experiences and learn from each other, achieving continuous memory evolution. This approach bypasses the traditional fine-tuning pipeline, offering a lightweight path for continuous Agent learning.
A new paper proves that through a simple and unified scaling strategy, large language models can reach gold-medal-level reasoning at the International Mathematical Olympiad. No flashy new architectures, no complex training tricks—just scaling. The implications of this may be more thought-provoking than the paper itself.
A paper from the KAIST AI lab proposes a strategy-guided exploration method that proactively steps reinforcement learning training "out of its comfort zone," improving learning efficiency without increasing training data volume. The paper has gained attention on Hugging Face Daily Papers.
A team from Harvard University and Massachusetts General Hospital proposes a multi-pathogen disease forecasting method based on autonomous LLM-guided tree search. LLMs are no longer just conversational tools; they transform into autonomous search agents that explore complex hypothesis spaces to find optimal forecasting models. This work demonstrates a new role for LLMs in scientific modeling.
A new paper systematically evaluates the feedback quality of LLM tutoring agents across different scenarios, revealing a counterintuitive finding: AI tutors perform well when confirming correct answers, but tend to provide inaccurate or incomplete responses precisely when students make mistakes and need high-quality feedback the most.
NVIDIA's MemLens benchmark systematically evaluates the multimodal long-term memory capabilities of large vision-language models for the first time. It reveals the true memory level of current multimodal models and how far they are from 'truly remembering'.
Shanghai Jiao Tong University's MMSkills proposes a multimodal skill learning framework for general-purpose visual agents. Unlike existing approaches that rely on rote memorization, MMSkills enables agents to truly grasp the multimodal nature of skills—knowing not just 'what they see,' but 'how to act.' The paper received 39 upvotes on Hugging Face Daily Papers.
OpenDeepThink proposes a population-based test-time reasoning framework grounded in pairwise Bradley-Terry comparisons. Eight rounds of LLM calls (approximately 27 minutes of wall-clock time) boosted Gemini 3.1 Pro's Codeforces Elo by 405 points. It also open-sources the CF-73 dataset—73 Codeforces problems annotated by International Grandmasters.
SANA-WM is an open-source world model with 2.6B parameters, natively supporting one-minute video generation. Trained for 15 days on 64 H100 GPUs using ~213K publicly available video clips, its distilled variant can denoise a 60-second, 720p video in just 34 seconds on a single RTX 5090 GPU using NVFP4 quantization.
SDAR (Self-Distilled Agentic Reinforcement Learning) introduces on-policy self-distillation as a gated auxiliary objective into RL training for LLM agents, achieving +9.4%, +10.2%, and +7.0% improvements over GRPO on ALFWorld, WebShop, and Search-QA respectively—while avoiding the instability inherent in naive GRPO+OPSD combinations.
Self-Distilled Agentic Reinforcement Learning proposes a new paradigm for Agent training: enabling the Agent to learn from its own experiences through self-distillation, rather than relying on human annotations or external reward signals. This could fundamentally change how we train AI Agents.
Solvita, developed by NJU-LINK Lab at Nanjing University, proposes an Agent Evolution paradigm to enhance the competitive programming capabilities of large language models. Unlike traditional supervised fine-tuning, Solvita enables Agents to evolve stronger programming reasoning abilities through self-play and continuous iteration.
SU-01 is a 30B-A3B Mixture-of-Experts (MoE) model that achieves gold-medal-level performance on the IMO 2025, USAMO 2026, and IPhO 2024/2025 using a simple, unified training recipe. Core pipeline: reverse-perplexity SFT curriculum → two-stage RL (verifiable-reward RL → proof-level RL) → test-time scaling. Supports stable reasoning trajectories exceeding 100K tokens.
Kronos is a foundation model for financial markets that treats financial data as a "language" for modeling. The project has garnered 24,946 stars on GitHub, proposing an approach that uses a tokenizer to discretize financial time-series data into token sequences, which are then predicted using a Transformer. Is this path viable?
A recent arXiv paper highlights a fundamental flaw in current mental health AI safety evaluations—they assess isolated responses or final outcomes, while the most clinically dangerous harms often stem from cumulative effects across interaction sequences: escalating dependency, repeatedly reinforced negative patterns, and gradual deterioration across turns. The paper proposes a theoretical framework of "Temporal Safety Non-Identifiability" and the SCOPE-MH evaluation standard.
NVIDIA's newly released AnyFlow paper introduces an "arbitrary-step" video diffusion model—the same model can seamlessly switch between 1 and dozens of steps without requiring separate training for each step count. Its core method, On-Policy Flow Map Distillation, trains the model by randomly sampling step counts and using self-guided distillation, enabling it to maintain stable generation quality across any inference step count.
The latest arXiv paper OpenDeepThink proposes a population-competition reasoning framework—instead of forcing the model down a single reasoning chain, it pits multiple candidate solutions head-to-head, aggregating judgments via the Bradley-Terry model. Leveraging this, Gemini 3.1 Pro achieved a massive +405 Elo surge on Codeforces, with the entire process taking only about 27 minutes.
