Research library
Search papers on equipment connectivity, quality inspection, production analysis, and manufacturing work.
Search and filter
New paper summaries
Summaries are editorial notes, not statements by the original authors. Check the original paper. Updated 2026-11-20.
- 700AI / General
AREX: Towards a Recursively Self-Improving Agent for Deep Research
- It exploits the asymmetry between discovery and verification, where finding an answer is hard but checking it is easy, and uses verification as the starting point for the next research round.
- A self-contained context compression tool keeps only the essentials from long conversations, making long-horizon research possible.
- With a 4B model, it achieves 82.5 on BrowseComp and 52.4 on HLE, far outperforming similarly sized competing models.
- It teaches search, tool use, and verification step by step through agent mid-training and long-horizon reinforcement learning.
- It reduces the credit-assignment problem of sparse rewards by weighting the moments when key evidence is obtained.
- 699Computer Vision
SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation
- It mixes linear attention and softmax attention in a 3-to-1 ratio so that it can achieve both speed and quality.
- The AttnRes technique improves information quality in deep layers by about 12 percent.
- A single H100 GPU can generate a 720p video in 13 seconds, which is 120 times faster than Wan 2.2.
- It directly learns the optimal balance point by showing that 25 percent softmax attention is the best trade-off between quality and efficiency.
- Kernel optimization through Sol-Engine delivers an additional 3.58 times speedup.
- 698LLMs / NLP
Solar Open 2 Technical Report
- A 250B-A15B MoE structure scales up from Solar Open 1 (102B) while routing only 2.3% of the parameters.
- Hybrid attention, 75% linear and 25% softmax, plus NoPE, supports a 1M-token context.
- The Korean tokenizer is 24% more efficient than global models, meaning it represents the same Korean text with fewer tokens.
- It ranks first among comparable open models and commercial APIs on average Korean benchmarks, and is nearly on par with DeepSeek-V4-Pro (1.6T) on Ko-GDPval.
- It integrates 12 domain experts into a single model through multi-teacher on-policy distillation (MOPD).
- 697Research
Scalable Keyword Spotting via Modular Network Expansion
- It freezes the entire base network, including batch-normalization statistics, so existing keyword performance is preserved at the pixel level.
- With only 10K additional parameters, 6.7% of the 150K base, it improves FRR for new keywords from 6.46% to 4.37%.
- It achieves better performance with fewer MACs, 16.34M versus 18.45M and 20.52M, than Adapter or LoRA.
- A core-first decision rule preserves the exact behavior of the existing keyword threshold.
- It can scale using only new keyword data, without access to the original training data.
- 696Computer Vision
ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion
- An adaptive token expansion, or ATE, module automatically finds regions with large rendering error and adds Gaussians there.
- Compared with dense baselines, it reduces the number of Gaussians by more than 5.7x while still reaching state-of-the-art quality.
- It reconstructs 12 images at 512×960 in under 1 second and renders at 1,136 FPS.
- Its 3D anchor-offset design decouples Gaussian placement from the input pixel grid.
- Uncertainty predictions align well with the actual rendering errors, as shown visually.
- 695AI / General
Pushing the Frontier of Full-Song Generation: Hierarchical Autoregressive Planning Meets Flow-Matching Rendering
- It supports three tasks in one framework: lyric-to-song generation, instrumental generation, and cover-song generation.
- It reaches an Elo score of 1,129 on the Artificial Analysis Music with Vocals leaderboard, which ties for second to third place and is only 12 points behind first.
- In tests on 500 songs across 8 genres and 5 languages, it gets the top score on 15 of 18 metrics.
- FullDiT performs flow matching on the entire song rather than on chunks, which preserves global consistency.
- Post-training with DPO, GRPO, and OPD improves both musicality and rendering quality.
- 694LLMs / NLP
SLAI T-Rex: Full-Parameter Post-training of the DeepSeek-V4 Family on Ascend SuperPOD
- On Ascend SuperPOD, it raises the MFU for DeepSeek-V4-Pro, a 1.6T model, from 11.67 percent to 34.22 percent, or 2.93x.
- It uses an agent called AuraKernel to automatically optimize complex NPU kernels.
- It builds a 10K high-quality CPT plus SFT dataset specialized for operations research and applies solver verification.
- The specialized model achieves an average zero-shot Pass@1 of 71.81 percent, which is 3.98 percentage points above GPT-5.4-Mini and 11.27 points above the base model.
- It provides a full-stack practical example of training a trillion-parameter model on NPU hardware rather than GPUs.
- 693Machine Learning
ISO: An RLVR-Native Optimization Stack
- It discovers and validates a 'spectrum inheritance' phenomenon in which the weight spectrum changes very little after RLVR training.
- ISO-Optimizer matches AdamW accuracy on Qwen3-8B using 2.7 times fewer steps, from 270 down to 100.
- ISO-Merger merges expert models using checkpoints alone, without data, rollouts, or gradients.
- Freezing the spectrum alone does not improve performance much, which means both frames need to remain trainable.
- It shows consistent efficiency gains on reasoning and coding tasks at scales from 1.5B to 8B.
- 692Robotics
Robots Acquire Manipulation Skills in Seconds from a Single Human Video
- We learn a new skill from a single human video in an average of 29 seconds, which is 507x faster than before.
- It reaches a 62% average success rate, 45 points above zero-shot and even higher than the baseline fine-tuned with 50 robot demonstrations.
- It preserves existing skills at 100% because it does not modify the learned parameters, eliminating catastrophic forgetting.
- It resolves temporal misalignment and state mismatch by aligning video and robot trajectories through a task-progress manifold.
- It implements inference-time adaptation through a causal cascade of localization, future observation prediction, and action generation.
- 691Machine Learning
Intelligence from Learnable Novelty
- It shows that the three forms of intelligence, data compression, universal computation, and adaptive behavior, all come from a single principle: maximizing learnable novelty.
- It recovers the Turing-complete rule 110 as the top one-dimensional cellular automaton, which is consistent with decades of complexity-ranking research.
- It achieves representation learning that separates MNIST digit classes without labels, which is fully unsupervised.
- When used as intrinsic reward in reinforcement learning, it improves performance in 9 of 10 environments.
- It has a mathematical structure that avoids both the noisy-TV trap and the dark-room trap.
- 690Computer Vision
Self Gradient Forcing: Native Long Video Extrapolation
- It defines and resolves the past-context gradient gap, which is the critical weakness of Self Forcing.
- A two-pass strategy, where the first pass records rollouts and the second pass reconstructs gradients in parallel, solves the problem without extra memory overhead.
- Using only 5-second training videos, it can generate coherent long videos up to 240 seconds, improving subject identity, background consistency, and temporal stability.
- A user study found that it was preferred over Self Forcing in every setting, with GSB ranging from 29.6% to 48.7%.
- Peak memory increases by only 8 GB, and training slowdown is about 13%, which makes the method practical.
- 689Computer Vision
IGGT4D: Streaming 4D Instance-Grounded Geometry Transformer
- It is the first streaming transformer that predicts geometry and instance identity at the same time from a video stream.
- It processes frames sequentially with causal masking and reuses past context through a KV cache.
- The authors build InsScene4D-147K, a 4D instance-annotated dataset with 147,000 scenes.
- It shows much stronger performance on object tracking, with T-mIoU of 58.84 compared with 45.38 for SAM2.
- Streaming clustering keeps memory constant at 0.7 GB regardless of the number of frames.
- 688LLMs / NLP
Automated Discovery Has No Universally Superior Harness
- It decomposed the harnesses of OpenEvolve and TTT-Discover into components and compared 30 combinations.
- It performed statistically significant comparisons using more than 3.1 million LLM rollouts.
- It found a generalization problem: no fixed harness is consistently best for every model-task pair.
- It proposed an adaptive budget allocation strategy that exploits the fact that early progress strongly predicts final performance.
- It released all execution logs and baseline distributions as reusable statistical infrastructure.
- 687Computer Vision
Latent Riemannian Flow Matching for Geometry-Grounded 3D Foundation Models
- The authors discover that VGGT tokens live on the product manifold of four hyperspheres and apply Riemannian flow matching.
- Compared with Euclidean flow matching, the Riemannian version improves consistently on every metric, with about a 44 percent gain in LPIPS.
- It works with a single input view and is order invariant.
- It reduces Chamfer distance by more than 40 percent on ScanNet++ and by more than 70 percent on ETH3D.
- The generated 3D scans align with the context frame and enable consistent scene reconstruction.
- 686Computer Vision
OmniReasoner: Thinking with Long Audio-Video via Native Tool Use
- We propose a tool-use learning framework that adaptively invokes zoom-in tools for long audio-video inputs.
- TimeAnchor preserves temporal coordinate consistency across different sampling resolutions.
- Temporal Augmented Data Engine automatically creates training data without manual effort.
- Accuracy improves from 29.3 to 34.8 on OmniVideoBench and from 32.0 to 35.4 on LVOmniBench.
- As videos get longer, the zoom-in tool is called more often, and we confirm that this contributes to improved accuracy.
- 685Robotics
WorldScape Policy 2.0: Empowering Steerable World Action Modeling with Reasoning-Augmented Memory
- It proposes a reasoning-augmented memory structure that combines short-term visual memory and long-term event memory.
- It builds a dataset of 5 million event segments called ManipEvent-5M.
- It can control robots through multiple modalities, including text, goal images, and demonstration videos.
- On a real dual-arm robot, it achieves 75 percent success on long-horizon tasks and 86.7 percent on fine-grained commands.
- Its generalization to new object categories improves sharply compared with prior work, from 43.3 percent to 60 percent OOD.
- 684Robotics
Agentic Real2Sim: Physics-based World Modeling with Vision-Language Agents
- It proposes a four-agent pipeline that automatically converts real robot manipulation recordings into MuJoCo simulation episodes.
- A 31B open VLM achieves a similar conversion success rate, 48 out of 100, at only 3 percent of GPT-5.4's cost.
- It handles deformable objects such as cloth and rope, as well as humanoid motion, in the same framework, not just rigid manipulation.
- It shows that narrowing the VLM's role makes backend model replacement easier and more cost-efficient.
- The converted episode twins can be used for robot policy training and evaluation.
- 683Computer Vision
ABot-World-0: Infinite Interactive World Rollout on a Single Desktop GPU
- It builds a multi-source data pipeline that combines AAA games, simulations, and internet video.
- It uses LongForcing to correct autoregressive drift in long rollouts with a teacher model.
- On a single RTX 5090, it runs in real time at 720P and 16 FPS with 1.2 seconds of latency and about 19 GiB of VRAM.
- It shows competitive control performance and consistent world evolution on WorldRoamBench.
- On 60-second rollout evaluation, it accumulates far less visual error than Causal-Forcing.
- 682Computer Vision
Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing
- At 4B parameters, it matches the performance of competitors in the 6B to 80B range while keeping memory usage down to about 18 GB.
- The Mage-VAE tokenizer reduces compute per pixel by 12 to 22 times compared with conventional VAEs.
- CUDA kernel fusion improves training throughput by about 2.5 times.
- The turbo variant generates a 1024 by 1024 image in 0.59 seconds and performs editing in 1.02 seconds on a single A100.
- Its cost efficiency makes it a practical open baseline for researchers who want to experiment and iterate.
- 681Computer Vision
Masked Visual Actions for Unified World Modeling
- It proposes an interface called Masked Visual Actions, which represents robot actions as pixel-level masks.
- A single checkpoint supports both forward simulation and inverse action generation.
- It works well across different environments and robot embodiments even when it is fine-tuned on only 15 hours of real robot data.
- The imagined rollouts from the video model are strongly correlated with real execution, with r = 0.982.
- Given a desired object motion, the inverse model synthesizes the corresponding robot action and reaches a 90 percent success rate.
- 680Software Engineering
TRIM: Reducing AI-Generated CodeSlop via Agent Trajectory Minimization
- It defines CodeSlop by functional behavior rather than surface style, such as verbosity or duplication. Any edit that can be removed while tests still pass counts as CodeSlop.
- The cause is the agent's search process. Once it finds a correct solution, there is no reason to roll back the attempted edits, so they remain in the final patch.
- Telling the agent to reduce its own patch was unstable. In 3.8% to 44.9% of cases it broke behavior or made the patch larger.
- TRIM uses the task history, not the final patch, as the search space because the temporal order of edits approximates dependency structure.
- It uses a hierarchical approach that tries to remove large chunks first, confirms success, and then splits more finely if needed, so the verification cost is about half that of Delta Debugging.
- On SWE-Bench-Verified, it preserved correctness on 327 of 330 successful patches, or 99.1%, and 18 patches became exactly identical to the human-written reference patch.
- 679Distributed Systems
ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels
- More than 95 percent of the weights in an MoE language model belong to the expert part, so duplicating the full model for both prefill and decode wastes a lot of memory.
- The key idea is a hybrid structure in which the heavy experts are shared across the two stages, while less than 5 percent attention is assigned separate GPUs by stage.
- An adaptive persistent kernel, or APK, schedules at matrix-multiplication tile boundaries so that urgent decode work can take over and free capacity can return to prefill without relaunching kernels or involving the CPU.
- A one-way MoE communication path initiated from the attention side removes MoE polling, avoids deadlock and cross-stage network interference, and overlaps communication in one stage with computation in the other.
- In serving experiments with MiniMax-M2.7 and GLM-5.1-FP8, effective throughput rises by 5.65 times over chunked prefill, by 2.01 times over instance-level separation, and by 1.66 times over Green Context colocation.
- Because the preemption interval depends on tile execution time rather than token count, even a prefill with thousands of input tokens can be interrupted by short decode jobs.
- 678Robotics
Closing the Loop in Humanoid VLA: Persistent 3D Object Tokens for Verifiable Loco-Manipulation
- It points out that robots fail on long tasks not because they cannot understand instructions, but because the object state used for action differs from the object state used for verification.
- POT binds objects to role slots such as target, destination, support surface, and handoff partner, and refreshes the 3D position grounding after each action chunk.
- Object records are turned into fixed-format tokens containing 33 features per slot across 8 slots and inserted into the input sequence of the whole-body action model.
- The same record is used to verify grasp, placement, and release through geometry-based checks, and if it fails, the system triggers re-observation, retry, and replanning.
- On eight task groups with a real Unitree G1, it improves the same-condition GR00T-N1.7 baseline from 39/80 to 71/80, and cup stacking jumps from 1/10 to 8/10.
- In ablation, tokens alone improve from 15/40 to 31/40, verification alone gives 22/40, and using both gives 34/40, showing that token conditioning is the main driver of performance.
- 677Robotics
RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning
- Instead of solving long tasks with a single policy, we reformulate the problem as coordinating heterogeneous policies with different capability boundaries.
- Existing planning methods assume homogeneous skills with fixed, non-overlapping capabilities, but we point out that the capability boundaries of real policies change with context.
- We use three groups of skills, understanding, memory, and self-evolution, to handle context interpretation, capability-boundary estimation, and experience-based correction during execution.
- The memory bridge retrieves trajectories associated with the next policy, estimates that policy's normal input region, and guides the robot into that region to make handoff stable without retraining.
- In ablation tests on 500 long-horizon tasks, the full system reaches 86.0% success, removing the memory bridge drops it to 60.4%, removing understanding skills drops it to 37.6%, and using only a single policy yields 0%.
- In 135 real-robot trials, success drops from 86.7% to 66.7% under block re-obscuration perturbations, but remains 80.0% even after partial dismantling of the structure.
- 676Machine Learning
Distilled Reinforcement Learning for LLM Post-training
- Reinforcement learning gives one reward per response, so it is hard to know which intermediate reasoning step to praise, and it mostly reinforces what the student can already do rather than adding new knowledge.
- On-policy distillation forces the model to follow the teacher distribution through KL divergence, creating a dilemma in which there is little to learn if the teacher is too similar and too much noise if it is from a different family.
- Distilled RL does not add a separate imitation objective; instead, it redistributes the learning signal of RL token by token using the reverse importance ratio between the teacher policy and the old student policy.
- It consists of three parts: reverse importance sampling with ratio clipping, negative-sample reset that restores the weight of negative-advantage responses to 1, and response-level geometric normalization that sets the geometric mean of ratios within a response to 1.
- Without negative-sample reset, mean pass@1 drops by 8.81 points on Qwen3-4B and 6.39 points on DSQW-1.5B, making it the most important of the three parts.
- The gains are especially large when the teacher and student belong to different families. In that setting, on-policy distillation remains below RL, while Distilled RL beats all baselines.
- 675Computer Vision
FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry
- The core claim is that video editing training does not need truly natural video pairs, only pixel correspondences from the first frame that remain temporally consistent.
- It converts image-editing samples into video-editing samples in real time by turning both the input image and the target image into 3D grids and building a 4D temporally warped flow field over pixel pairs through grid inverse sampling.
- It treats an image as a one-frame video and uses two modality-mimic losses so that T2I mimics the realism of T2V and V2V editing mimics I2I editing, which converges faster.
- It introduces sense-related tasks such as referring expression segmentation, together with edit-region-aware latent-level and attention-level losses, so the model can locate the edit region without a separate MLLM or mask input at inference time.
- It continues training from Wan2.1-T2V-1.3B, and when the learning rate is warmed up from 5e-6 to 1e-5, video-editing behavior appears within a few thousand steps.
- Without using any specially made video-editing data, it still handles a wide range of editing tasks and even shows SAM3-like ability to segment and track objects specified by language in video.
- 674Computer Vision
Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement
- It diagnoses the limitations of existing monocular geometry estimation models as a structural mismatch rather than a capacity problem. When 3D information is decoded in a 2D coordinate system, features blend across depth boundaries.
- The solution is to lift geometric modeling from the 2D image plane into 3D space, defining feature neighborhoods by 3D distance instead of screen distance.
- It converts the point map into (X/Z, Y/Z, log Z) so that only log depth needs to be corrected, and quantizes this value to resolution D to build a sparse voxel shell.
- A sparse 3D U-Net predicts log-depth residuals at each iteration, and the result redefines the voxel structure for the next iteration. Because of this self-guided scheme, computation is proportional to the number of occupied voxels.
- In replacement experiments with a parameter-matched 2D U-Net, local metrics were similar to or worse than the base model, showing that the improvement comes from 3D inductive bias rather than capacity.
- It was trained with K=3, but even when K was increased to 7 at inference time, the metrics held steady or improved slightly, indicating stable convergence of the learned correction.
- 673Computer Vision
Pixel-Space Diffusion Transformers
- It points out that latent diffusion has an information bottleneck because the VAE's lossy compression irreversibly removes high-frequency texture and sharp boundaries.
- It identifies the core problem as a mismatch between the objectives of a compressor trained for reconstruction and a diffusion model trained for generation, which makes end-to-end optimization difficult.
- Moving to pixel space enables single-stage training, but it introduces new challenges in noise schedules, loss weighting, token efficiency, and scalable architecture design.
- Recent methods are introduced as x0 prediction approaches that directly predict the clean original image and speed convergence by keeping the training path on the data surface.
- It summarizes the trend toward splitting global structure and fine texture across different components through multi-branch designs, frequency separation, and finer tokenization.
- It suggests that if images, text, and task conditions are placed in a shared token space and trained with a single transformer, the field can move toward unified vision foundation models that do both understanding and generation.
- 672Research
Color superconductors and holon metals from doping a Fractional Chern insulator
- Using the parton construction c(r)=f1(r)f2(r)f3(r), the paper places the metallic and superconducting phases of a doped fractional Chern insulator within a unified framework.
- Each color has three hole pockets across three valleys, so the charge minus e over 3 holons form nine pockets in total.
- The SU(3) gauge field mediates pairing in the color-antisymmetric channel, allowing a gas of charge-e/3 anyons to enter a charge-2e superconducting state directly without an intermediate binding step.
- If all nine pockets are weakly paired, c- becomes minus 3 over 2, which differs from the BCS expectation for an f-if superconductor, so an anyonic superconductor can break the usual relation between central charge and pair angular momentum.
- Adding a particle-hole Higgs field yields two Z3 orthogonal metals with one or three pockets, and reducing valley symmetry to a triangular-lattice space group allows a three-pocket U(1)^2 holon metal that preserves C3.
- In a chemical-potential-tuned FCI, a continuous transition to superconductivity is impossible in Z3 metals but possible in SU(3) or U(1)^2 holon metals, so the authors argue that the nine-fermion theory is a natural starting point.
- 671Computer Vision
HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enchancement
- The problem setting unifies inter-subject cases, where the reference images are of different targets, and intra-subject cases, where they are multiple views of the same target.
- It points out the limits of two existing approaches: aligning multimodal model features to the text-embedding space weakens the control power of the text encoder, while replacing the text encoder entirely with a multimodal model introduces a large realignment cost.
- GMG, the first proposal, adds global representations extracted from the multimodal model to the video tokens during query and key computation, which injects knowledge into self-attention.
- MRE, the second proposal, uses learnable embeddings that distinguish token modality and reference index, which prevents errors such as duplicating the same object or text map as separate entities.
- In performance, text recognition accuracy improves by 21.8 percent relative to SkyReels-V3, DINOrec reaches 0.696 and is best among the compared methods, and 40 human raters prefer the overall quality 64.7 percent of the time.
- Ablation studies show that removing GMG drops face similarity from 0.786 to 0.697, and removing MRE drops text accuracy from 0.452 to 0.376.
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology