AI Coding Live News

Deep search agents struggle with credit assignment across long, noisy retrieved contexts. Zhejiang University and Tencent open-sourced KnowGPO (arXiv:2610.22331, GitHub: ZJU-REAL/KnowGPO), an RL framework that groups search trajectories by atomic knowledge conditions rather than surface observations. Qwen3-8B fine-tuned with KnowGPO achieves 38.8% on BrowseComp-ZH and 40.2% on GAIA, outperforming GRPO/GiGPO baselines and rivaling 32B open-source search systems.

Zoom

Two new 3D vision and generative compression frameworks solve long-standing efficiency and quality bottlenecks: Zhejiang University's InfiniSplat v2 (arXiv:2610.20341, 118 upvotes on HF) decouples geometry and Gaussian attribute prediction with a frozen geometry branch and large-angle supervision, slashing mean LPIPS by 25.8% and DISTS by 32.9% across six 3D datasets. Meanwhile, Technion's H-DDCM (arXiv:2610.18125) replaces dense Gaussian codebooks in training-free diffusion compression with implicit Walsh-Hadamard dictionaries—requiring O(d log d) compute and O(d) memory—to enable near-real-time streaming video compression on a single GPU.

Zoom

Four new studies advance interactive video world models and streaming video agents: NUS's training-free Velox router (arXiv:2610.22311) uses 3D camera frustum-slab correspondences on H200 GPUs to accelerate chunk-48 attention by 4.36x and denoising by 2.43x with negligible quality loss; Nankai University's DeltaStreamer (arXiv:2610.19849, 39 upvotes) combines change-aware memory with targeted frame replay to reach 66.07% on OVO-Bench; HKPolyU's CoWorld Reasoner (arXiv:2610.21910) co-evolves a VLM-VWM tool harness for +13.0% relative gains on ERQA; and Zhejiang University releases the 260-case WorldMemBench (arXiv:2610.19063).

Zoom

Three new open-source embodied AI frameworks overcome observation-action overfitting in Vision-Language-Action (VLA) policies: Zhejiang University's 2B VLA-Δ (arXiv:2610.21306, 54 upvotes) supervises explicit physical state transitions with zero inference overhead to set a 98.4% SOTA on LIBERO; SJTU and HKU's OneWM (arXiv:2610.21095, 40 upvotes) unifies causal world and action modeling to reach 36.68% on RoboCasa365 (over 2x π0.5 at 16.90%); and UIUC/Amazon's CoDA (arXiv:2610.22362) co-evolves simulation data engines and action policies via Text2potential shaping, raising LIBERO-Pro success from 25% to 84% under position perturbations and from 1% to 76% under instruction perturbations.

Zoom

Enabling service robots to execute open-ended human navigation requests across unfamiliar environments (such as locating reading glasses on a living room table and navigating to them) represents a critical challenge in embodied AI. Conventional navigation models rely heavily on domain-specific fine-tuning within specific simulated datasets, resulting in brittle real-world generalization when deployed in unseen layouts. Researchers from Zhejiang University's 3DV Lab introduce SuperNav (arXiv:2610.12126, code: github.com/zju3dv/SuperNav, project: zju3dv.github.io/SuperNav), an open-source agentic navigation framework that achieves universal scene navigation with zero navigation-specific fine-tuning. SuperNav decouples high-level cognitive reasoning—delegated to a frozen, off-the-shelf Multimodal Large Language Model (MLLM)—from low-level physical control executed by specialized robotic navigation tools. Utilizing a Unified Visual-Point Interface where the MLLM marks normalized coordinates directly on observation images, the agent revises decisions continuously based on visual feedback. Validated in the HM3D virtual benchmark and deployed on a real Unitree Go2 quadruped robot across multi-object and demand-driven navigation tasks, SuperNav decisively outperforms fine-tuned baselines.

Zoom

Large language models predominantly rely on human natural language texts, often capturing surface linguistic statistical patterns rather than intrinsic spatial logic and structural topologies. Researchers from Zhejiang University, SJTU, and BioMap propose Fold2Reason (arXiv:2609.38879), investigating whether learning protein folding—where each resolved crystal structure provides thousands of rigorously verifiable 3D spatial and topological constraints—transfers reusable reasoning capabilities to general LLMs. They construct FoldingCorpus and train via Fold2Reason using two complementary signals: discrete structural answers via native language heads and continuous 3D geometry decoded from shared representations. On FoldBench, Fold2Reason scores 2.7x to 3.5x higher than Qwen3.5-9B; crucially, it generalizes across all 10 evaluated benchmarks spanning spatial, graph, scientific, and general reasoning, raising macro-average accuracy from 45.09% to 48.33% (+3.23 pp) with uniform positive gains.

Zoom

While reinforcement learning induces profound reasoning capabilities in LLMs, the scaling laws governing how and how fast this capability transfers across model sizes via on-policy distillation (OPD) have remained unquantified. Researchers from Zhejiang University and collaborating institutions present a definitive empirical analysis (arXiv:2609.32722, 310+ upvotes on Hugging Face) across weak-to-strong, same-base, and strong-to-weak distillation regimes. The study identifies a universal 'useful-transfer regime' where held-out task accuracy scales linearly with the square root of token-level reverse KL divergence from initialization. Remarkably, across all weak-to-strong pairings, student peak performance strictly surpasses the RL teacher itself. The authors formalize empirical power laws for OPD, demonstrating that smaller RL expert teachers paradoxically deliver superior supervisory density compared to bloated teachers of matched accuracy.

Zoom

Existing Computer-Use Agents (CUAs) face a dilemma: relying strictly on graphical user interface (GUI) interactions is slow and error-prone, while building per-application API tools demands unsustainable engineering and lacks generalizability. Researchers from Zhejiang University introduce HybridCUA, combining the UI versatility of GUIs with the execution speed and determinism of the Command Line Interface (CLI). Supported by the newly constructed HybridCUA-8K dataset (5K hybrid trajectories and 3K RLVR tasks), HybridCUA employs a two-stage training paradigm: supervised fine-tuning across interleaved GUI-CLI modalities, followed by CLI-aware reinforcement learning that rewards reliable, selective shell execution. On the OSWorld benchmark, HybridCUA-9B achieves 53.6% accuracy, improving over the base model by 14.8 percentage points, while delivering a 4.0-point gain on WindowsAgentArena.

Zoom

Addressing the fundamental limitation of benchmarks like SWE-bench that assess coding agents strictly within isolated repositories, researchers from Zhejiang University introduced WideSWE. Mining changes across 103 real-world software ecosystems, WideSWE establishes 120 cross-repository tasks (balanced between 60 bug fixes and 60 feature developments) with comprehensive regression test suites. Across seven frontier agent configurations, full task success tops out at just 42.50%, exposing systemic failures in cross-repo dependency tracking and coordinated verification.

Zoom

Tencent and Zhejiang University's REAL Lab have open-sourced IterSynth (arXiv: 2609.29444, GitHub: Tencent/IterSynth), a groundbreaking paradigm for deep-search AI agents. Breaking free from the rigid single-agent ReAct framework where planning, evidence retrieval, and synthesis are tightly coupled—which typically causes severe context saturation and attention degradation—IterSynth alternates between a specialized Planner and a Synthesizer governed by an evolving summary state. Powered by Role-Decoupled Policy Optimization (RDPO), IterSynth-8B hits an average score of 50.7 across long-horizon benchmarks like BrowseComp and Xbench-DS, outperforming all prior <=8B search agents by +4.2%.

Zoom

Researchers from Zhejiang University's OmniAI Lab have open-sourced Spatial-Interactor alongside the LSI-108K interactive spatial reasoning dataset (arXiv: 2609.23038). Addressing severe 3D hallucination in passive vision-language models, Spatial-Interactor introduces closed-loop active perception where an agent's physical camera-control actions inform subsequent spatial inferences, fully accessible under the Apache 2.0 license.

Zoom