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.
- 910AI / General
Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models
- The starting point is the insight that purely observational data cannot identify causal structure, so direct intervention through experiments is necessary to obtain answers.
- The LLM acts as the hypothesis proposer, while Bayesian machinery, including SMC, SBI, and VoI, handles validation and experiment design.
- When no correct hypothesis exists in the current hypothesis set, the system detects the gap through predictive checks and the LLM expands the search space by adding new models.
- Discovery and experiment design reinforce each other, because the designed experiments confirm proposed mechanisms and the confirmed mechanisms improve prediction, which in turn reveals residual errors for the next discovery step.
- The authors create a new single-neuron electrophysiology benchmark under partial observability called NEURONBENCH.
- 909AI / General
Algebraic Decomposition Theory for Transformer Length Generalization
- Two problems that look structurally almost identical can diverge into cases that do and do not generalize in length, and prior theory could not explain the difference; this paper gives the first complete decision criterion.
- The decision algorithm runs in polynomial time, so it is not just theoretical but also practically checkable.
- The key discovery is that unbounded counting is the seed of transformer length generalization, and the paper formalizes this through iterated wreath products of the integer-addition group.
- When GPT-2 is trained up to length 50 and tested up to length 500, languages inside C-RASP keep their accuracy while languages outside it collapse right after the training length.
- The work provides a theoretical foundation for understanding why structured outputs such as JSON and long agent state traces break down as they grow longer.
- 908Research
Blue Noise as a Lattice Gibbs Ensemble
- It formalized repulsion between points as a Gibbs distribution over binary lattice occupancy, controlling density, repulsion strength, interaction range, and kernel stiffness within one model.
- Using finite-range interactions and backward-tracing sampling, it guaranteed that tiles with sufficient surrounding margins are bitwise identical to generation over the full domain.
- Memory use depends on tile size rather than the full output, allowing a 14K stippled image to be generated with 51MiB of memory.
- It extended the same distributional framework to adaptive density and multiple types of point placement, allowing large graphics workloads to be divided without ordering or communication constraints.
- The lattice quantizes point positions, truncating backward history trades accuracy against cost, and stronger ordering increases the dependency radius, potentially eliminating the benefits of tiling.
- 907Machine Learning
Latent On-Policy Self-Distillation
- It changes the design so that the training process, not a human, decides what to show the teacher, meaning the experience representation itself is optimized end to end.
- After training, only the student model remains, and performance is preserved even if the retrieval database, compressor, and latent tokens are all removed, because the experience has been internalized into the policy.
- It beats GRPO and Skill-SD with less than 30 percent of their rollout budget, which greatly reduces the compute cost of self-evolution.
- The key ingredient is a privileged margin constraint that prevents the teacher from collapsing into the student; ablations show that without this constraint the supervision signal disappears.
- The student's behavior also changes: it shifts from calling tools in a bursty, all-at-once way to executing plans step by step, with calls per step falling from 3.50 to 1.11.
- 906Machine Learning
CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution
- When paired with a compiler that identifies what failed and why, the agent gets much closer to expert-level GPU kernels than it would in a black-box environment.
- The harness itself is also evolved. Repeated failures are turned into verification rules, new IR instructions, and calibration of the performance model, so the system accumulates knowledge instead of using one-off workarounds.
- Under the same budget, clean-start optimization with the Cake IR representation reaches 1.144 times the tuning baseline, while direct CUDA and PTX reach 0.928 times, which means the code representation itself matters.
- The agent-generated Kimi Delta Attention kernel achieves a 2.05x geometric-mean speedup over the official implementation and passes real serving validation.
- It completes library-level integration by separating single-target optimization from a roughly 400-target generalization, or dispatch, stage.
- 905Machine Learning
Training AI Scientists to Replicate Research
- Faraday, a 27B small model, uses a 5-trillion-parameter coding agent as a tool to outperform Claude and GPT-class systems on paper replication, which suggests that it is better to instill scientific judgment in a smaller model than to use a large model directly.
- For open-ended tasks that cannot be fully verified, it presents a recipe in which an automatic rubric judge aligned with human evaluation is used as the reward for stable reinforcement learning.
- It achieves a 60 percent win rate and an average gain of 6 percent on held-out AI-for-science tasks, which shows that it transfers scientific reasoning rather than merely memorizing.
- Faraday does not produce results by blindly coding outputs. Instead, it directly implements the paper's mechanism and scales the experiments down to fit the budget, much like a human scientist.
- The fact that a weaker model can supervise and direct a stronger model becomes practical evidence relevant to AI safety and oversight research.
- 904AI / General
Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
- Even when final scores are the same, the bottlenecks differ. GPT-5.5 is strong at execution, while Gemini-3.1-Pro is strong at feedback control, so their process profiles diverge.
- The gap between models is much larger on average performance, avg@3 at 0.237, than on best@3 at 0.122. Reliability is the deciding factor.
- Only 3 of the 252 top-seed solutions had real methodological novelty.
- Experience transfer cuts both ways. DeepSeek-V4-Pro improves by +0.093, while Gemini-3.1-Pro degrades by -0.017.
- The harness, or agent operating system, mainly affects execution stability rather than peak performance.
- 903Computer Vision
Alaya-EVOKE: From Linear-Scaling Supervision to Endless World
- It stores persistent state in an external geometric bank indexed by camera pose, which keeps the context cost fixed regardless of session length.
- The teacher combines chunk-level sparse attention with a linear-attention global state so that supervision cost grows only linearly with length.
- It distills a three-stage, CFG-free student on a 30-second long-horizon distribution-matching objective, which gives both low-latency interaction and long-horizon consistency.
- It demonstrates two hours of continuous generation and achieves top results on WBench and a score of 66.77 on VBench-2.0, which ranks first among 10 compared systems.
- During a session, text instruction replacement can inject 67 percent of new content, but only 4 percent of content already fixed in geometry is overwritten.
- 902Computer Vision
PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives
- It enables fair comparison across models by using shared goals and adaptive adjustments from agent players, such as Keep, Stop, Extend, Correct, and End, instead of fixed action trajectories.
- It evaluates four core dimensions, geometric consistency, interaction fidelity, out-of-view evolution, and insight evolution, along with baseline capability metrics.
- Among nine models, Genie 3 is the overall best and HappyOyster is second.
- All models score lowest on out-of-view state evolution and insight evolution, confirming that long-horizon persistence is the current bottleneck.
- The automatic VQA metric rankings are positively correlated with human preferences.
- 901AI / General
OmniScientist: An Omni-Modal Omni-Discipline AI Scientist
- It solves the problem that data representation constrains research by making raw observations, including spatial, temporal, and channel relationships, accessible throughout the entire research process instead of relying on text and summaries.
- It controls three autonomous agents for ideas, experiments, and writing through a thin deterministic pipeline, and it enforces freshness, leakage, HARKing, and traceability checks in code.
- Across all 36 real-data cases, spanning five subject areas and four evidence types, it completed the full path from raw data to paper and received an average paper score of 6.3.
- Direct perception outperformed the blinded variant that only received precomputed features across all seven evaluation dimensions and won 85 percent of pairwise comparisons.
- As one example, it found that 21.7 percent of 750 earthquake entries labeled as noise actually contained earthquake signals, and the result was statistically validated.
- 900Computer Vision
Fidelity-Constrained Anchoring for Black-Box Denoisers
- It linearly mixes black-box denoising outputs with the input patch by patch to find the largest mixing coefficient that satisfies a target fidelity.
- It solves PSNR constraints in closed form and efficiently computes SSIM constraints through four root-finding operations using properties of inverse SSIM.
- It verified that target PSNR or SSIM can be controlled without retraining for Real-ESRGAN and a non-local means denoiser.
- SSIM-based anchoring showed a more consistent quality-naturalness trade-off than PSNR-based anchoring across three Gaussian-noise levels, making it useful as practical postprocessing.
- Because it relies on linear mixing, performance degrades when the denoised output differs substantially from the input or when noise is strong enough to damage the original structure.
- 899Computer Vision
StreamTTT: Reconciling Real-Time Perception and Long-Term Memory in Streaming VLMs
- It stores long-term history in TTT fast weights outside the attention context, which structurally prevents the past from diluting the understanding of recent scenes.
- It creates a new 112.4K-scale real-time QA dataset that shifts live queries to answerable time points and trains it together with offline long-video QA.
- On OVO-Bench, it improves real-time perception by 1.4 points to 78.9 and future tracking by 3.7 points to 58.3 compared with SimpleStream-4B.
- On StreamingBench RTVU, it scores 80.48, just 0.11 below SimpleStream-8B, which has twice as many parameters and scores 80.59.
- Adding TTT state to a 4K window recovers about 97 percent of the effect of extending the window to 64K on OVO-EPM.
- 898Machine Learning
Synthetic Persona Pretraining: Alignment from Token Zero
- It proposes persona binding, which attaches value reflections to pretraining documents to install a persona early and then binds it to the assistant identity during post-training.
- Injecting from the first token improves constitutional compliance and leads to more aligned choices even on moral dilemmas not covered during training.
- If the same reflection is injected only at the end of pretraining, the effect is weak and value priorities do not change.
- The effect grows as the pretraining budget increases, from 1.7B/100B to 3B/500B.
- Late injection alone was enough for jailbreak defense, but deep value alignment required early injection.
- 897Computer Vision
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
- It adopts the approach of fixing the model parameters and optimizing the harness, so that failure experience accumulates into improvement of the system itself.
- It introduces a safe update gate that changes only one component at a time and requires the development-performance condition to pass before promotion.
- After adding the learned DesignHarness, all seven code-agent configurations improve by 5.0 to 19.6 points, with the average rising from 54.99 to 67.39.
- On PosterBench, which contains 100 papers across 5 fields, it reaches 78.32 points, the best overall score, and beats the commercial Claude Design system by 7.45 points.
- In a fully autonomous loop, it makes 253 tool calls and 11 edits in 40 minutes for less than 3 dollars and produces conference-poster-quality results.
- 896Machine Learning
Intern-S2-Preview: Scientific Agentic Foundation Model
- We pretrain on rendered scientific documents, including text, figures, and tables, so the model absorbs document structure that is lost in plain text extraction.
- The unified post-training pipeline combines scalable multi-task reinforcement learning with verifiable objectives and black- and white-box agentic RL.
- By freezing the 397B backbone and attaching a separately trained 4B memory decoder, we raise the biology score from 56.92 to 60.32 without hurting general ability.
- The harness-by-task abstraction lets different agent runtimes and tasks share the same rollout, verification, and training protocol.
- It delivers competitive results across scientific, multimodal, agentic, and general benchmarks.
- 895Computer Vision
V-RAE: Rethinking Video Latent Spaces for Generation
- Instead of learning the video latent space for pixel reconstruction, it builds it on top of the semantic feature space of a frozen visual backbone.
- On K600 reconstruction, it achieves an rFVD of 2.13, which beats every large video VAE in the evaluation, and it preserves semantic information in the latent space much better, with 89.13 percent classification on UCF101 versus a VAE best of 30.83 percent.
- Under the same training conditions, it also leads in generation quality, measured by gFVD, and converges up to 6 times faster.
- It shows that reconstruction ranking, measured by rFVD, can differ greatly from generation quality ranking, and it proposes tFVD as a metric that better reflects generation suitability.
- On Cityscapes future video prediction, it also outperforms the Wan 2.2 VAE latent space, with gFVD 111.36 versus 144.47.
- 894Computer Vision
DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation
- We inject the 3D motion path of a robot arm geometrically into attention, preventing the model from moving the wrong arm even when the video looks plausible.
- A depth prediction branch and SAM3 masks, a technique that separates object regions, reduce mistakes where grasped objects disappear or get swapped during the video.
- We enable fast deployment by distilling the multi-stage generator into a few stages.
- On WorldArena 2.0, Track 1, it ranks first among 31 teams with EWMScore-P 60.65, and it ranks joint second on Track 2 with 67.19%.
- On the WorldArena 1.0 offline evaluation, it scores 76.88, which is 3.24 points above the previous best of 73.64.
- 893LLMs / NLP
Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus
- The first pattern is a large spike in values immediately before a full attention layer. The second pattern is those values staying high across several linear attention layers between spikes.
- The same patterns are observed in five linear-attention architectures, six mixture ratios, five input domains, and public models ranging from 1.2B to 397B parameters.
- In a controlled 340M and 1.3B parameter 12-to-1 mixture, alignment between the layer right before full attention and the spike position was 99.4% to 100% overall.
- In the 340M control model, both patterns appeared from the first measurement point, after 1B training tokens. The full-attention output gate reduced magnitude but did not eliminate the layer-wise position pattern.
- The authors explain that fast cancellation makes spikes and slow cancellation makes plateaus, but they leave the cause of the cancellation timing and its actual computational role for future work.
- 892Robotics
FACT: Failure-Aware Causal Training for World-Action Models
- It predicts future video and progress conditioned on actions created first, and it uses failed actions only for outcome prediction, not imitation.
- On 50 RoboTwin tasks, average success is 81.8% without video co-training, 85.6% with video co-training, and 87.5% when failure data are added.
- On a real dual-arm robot, tasks seen during training improve from 82% to 89%, and unseen variants rise from 67% to 77%.
- Video prediction PSNR on failure cases improves from 19.51 to 25.92, while success cases stay almost unchanged at 26.12 and 26.08, so normal prediction is preserved.
- Evaluating four action candidates by progress gives 92% on seen tasks and 82% on new variants, but candidate evaluation requires extra computation.
- 891Machine Learning
Dion3: Full-Stack Orthogonal Updates
- Gram Newton-Schulz rewrites existing methods into mathematically equivalent form and reduces compute by iterating on small symmetric matrices instead of large rectangular ones.
- It adds a symmetric GPU kernel and a grouped-communication strategy for matrices of the same shape. In the 8-way setup for a 1-billion-parameter model, grouped communication alone cuts optimizer-step time from 80.7 milliseconds to 52.1 milliseconds.
- At each step, it orthogonalizes only a subset of rows in the momentum matrix to reduce compute and communication, and it improves both speed and loss over the previous compression method, Dion.
- In experiments training 3B to 14B models on 10 billion tokens, all models achieved lower validation loss, and the 14B model had 0.027 lower loss than NorMuon and 0.7 percentage points higher average accuracy.
- The reported up to 6x speedup applies only to the optimizer step, not full training, and the authors analyze that this step accounts for roughly 1 percent to 17 percent of total training time depending on the setup.
- 890Computer Vision
Map-Det3D: Metric Feed-Forward 3D Reconstruction Prior for Multi-view 3D Object Detection from Streaming Inputs
- It processes RGB frames from a short time window in order and applies the scale ratio predicted by the 3D reconstruction model to create object boxes in real-world units.
- It removes the step that converts 2D detections into 3D and directly predicts boxes in 3D space by combining geometric information from multiple viewpoints.
- On CA-1M, the AP15 of the single-frame baseline was 11.7, and using time information, camera information, and learnable multi-view processing raised it to 21.2.
- Without any extra training, it achieved AP15 of 15.2 on ScanNet200, surpassing the previous monocular best of 11.7.
- On CA-1M, it outperformed existing monocular and multi-view image methods, but it remained below methods that use real depth or 3D point data. It is also limited to detection that does not distinguish between indoor scenes and object categories.
- 889Research
On-Policy Self-Distillation for Multi-Dialect ASR: Mastering Dialects, Retaining Mandarin
- The full training set is about 100,000 hours of standard Mandarin and dialect speech, and the final self-distillation stage uses about 5,000 hours of dialect speech with high error rates.
- The student uses its own generated prefix as input, while the teacher also consults the ground-truth sentence and provides probabilities for each next-output candidate; only the student model is used in deployment.
- Evaluation is performed on eight standard-Mandarin public datasets, five public dialect datasets, and 18 internal dialect datasets.
- Compared with the base model, the average character error rate drops from 3.46 percent to 3.27 percent for standard Mandarin, from 15.37 percent to 12.79 percent for public dialects, and from 21.01 percent to 12.42 percent for internal dialects.
- Continuing ordinary supervised learning on the same data raises the standard-Mandarin error rate to 4.43 percent, while self-distillation keeps it at 3.27 percent, and the overall average error rate is 10.12 percent instead of 10.74 percent.
- 888Machine Learning
Small-Scale Experiments: Are We There Yet?
- The authors experiment from about 4 million to 268 million parameters, fitting the scaling law on models from 4 million to 34 million parameters and then validating it on larger models.
- Testing only 4 or 16 size settings failed to reveal the law, 64 settings were inaccurate, and 256 settings recovered the correct law.
- Small models were highly sensitive to training settings, but the sensitivity decreased as the model size grew, and near the optimum the number of influential hyperparameters fell to about one.
- The law derived at small scale extrapolated well to models about 10 times larger, but the uncertainty grew sharply at farther scales.
- With this method, they reproduce the large-scale result that pre-normalization becomes more advantageous as the model grows, although the mapping between loss and capability holds only when the composition of the pretraining data is kept the same.
- 887Robotics
G0.5: One Autoregressive Stream for Robot Reasoning and Action
- Reasoning and action use the same model weights and the same token stream, and there is no separate action-expert model.
- It trains 14 robot embodiments in both real and simulated environments and maps their different action spaces into one 27-dimensional shared space and one token vocabulary.
- The thought process includes subgoals, object locations, 2D movement paths, and action hints, while visual memory uses the most recent few seconds of video.
- The average success rates are 76.7 percent on real robots, 82.5 percent on DROID's new environments and objects, 98.9 percent on LIBERO, 93.3 percent on RoboTwin 2.0, and 87.3 percent on SimplerEnv-Bridge.
- The BEHAVIOR long-horizon score is 31.4 percent, which is higher than the comparison model's 26.3 percent, but visual memory is limited to a few seconds and lower-body actions are not evaluated separately.
- 886Machine Learning
AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
- It combines prompts, problem-specific handling rules, calculation code, and answer-format checks without changing the target model's weights.
- The hidden test set contained 3,900 questions across four theory-of-mind tasks, and 195 separate validation items were provided for improving the harness.
- GPT-5.4-mini's base average score was 0.488, while the overall average with the automatically generated harness was 0.763 and the best score was 0.912.
- Deterministic code, problem-specific path selection, and strict output formatting had a larger effect than long reasoning or extra sampling.
- The improvement was large on weak target models but some performance degradation also appeared on strong target models, and the experiments were limited to theory-of-mind tasks.
- 885Computer Vision
AVA-Encoder: Towards Agent-Native Video Representation Learning
- It organizes people, backgrounds, objects, style, camera, and audio states under story, event, and scene hierarchies, and links the related assets.
- It turns reconstruction errors into natural-language correction instructions, then improves both the shared encoding policy and the video-specific knowledge graph in two stages.
- Policy improvement used 6 videos, and evaluation was done on 18 non-overlapping videos, 129 scenes, and 246 key frames.
- The overall reconstruction score is 49.0 percent, which is 20.7 points higher than the strongest external baseline at 28.3 percent.
- When comparing policies only, it reaches 45.8 percent, beating the human-tuned policy at 44.4 percent, and it reduces the system prompt from 31,336 tokens to 8,052 tokens.
- 884Computer Vision
StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization
- It stores the world as an editable state composed of each object's 3D shape, position and pose, and semantic attributes.
- Front-view images are used for object appearance and top-view images for spatial layout, and a vision-language model resolves conflicts when the two disagree.
- Scene expansion, style changes, and object movement or pose changes are handled by editing parts of the state or swapping assets, which reduces the need to regenerate the whole scene.
- It proposes camera paths first and then checks rendered results to correct occlusion, bad composition, collision risk, and unstable motion.
- The average VBench score is 0.8484, the highest among the compared methods, and in a user study with 30 participants the combined scene and video scores were both 4.5 out of 5.
- 883Machine Learning
Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control
- It builds 3,200 behavioral scenarios, 20 patterns times 4 situations, to overcome the limits of survey-based personality measurement.
- It finds that an LLM's behavioral style is stable across models but still changes meaningfully with situation, such as first person, advice, or task mode.
- It extracts behavioral mode axes from the activation space and demonstrates across seven models that adjusting only one direction can precisely control behavioral style.
- BMA extracted from the thought process provides much more accurate control than BMA extracted from responses, revealing a trait drift phenomenon.
- It shows that the behavioral control layer is concentrated in a specific region of the model rather than spread uniformly across the whole network.
- 882Robotics
Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning
- The authors propose a two-stage learning method that first learns visual dynamics in surgical scenes from unlabeled endoscopic video and then fine-tunes on a small amount of labeled action data.
- Video pretraining raises the average success rate on four SurRoL benchmark tasks from 63.5 percent to 77.8 percent.
- On PegTransfer, it improves by 20 percentage points, from 66 percent to 86 percent, and the effect is strongest on tasks that involve many contacts and two-handed cooperation.
- The video-pretrained model reaches higher performance with only half the fine-tuning budget, which greatly improves training efficiency.
- The same trend appears on real surgical robot data from JIGSAW, which confirms that the effect is real and not just a simulator artifact.
- 881Robotics
Flex-$π$: A Multi-Stream World-Action Model with Compute Flexibility
- It discovers a free lunch in which a video-generation model's VAE can encode not only RGB but also 3D point maps with almost no loss.
- It builds a multi-stream world-action model that predicts RGB, 3D depth, and DINO object features together in a shared latent space.
- During training it randomly drops modalities, or stream dropout, so that one checkpoint works with any combination at inference time.
- On robot tasks in RoboTwin, with only 50 demos, it is 1.9x better than existing WAMs, and on a real dual-arm robot it achieves 2x to 7x higher success.
- In action-only mode it is faster than π0.5 and still achieves higher success, at about 60 ms.
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology