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.
- 1060Agents / Infrastructure
HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness?
- HarnessDev defines a harness as execution infrastructure reused across tasks rather than a single response, separates creation from evolution, and evaluates both capability and execution token cost.
- It used 6 generation models and 2,207 unique tasks across 4 domains and 5 benchmarks. Generated harnesses were strong in writing and machine-learning experiments but lagged behind human-built baselines in coding, search, and research.
- All 18 coding harnesses implemented execution loops, but state and memory were their weakest areas. Although 11 declared a State class, only 1 provided a storage interface and only 1 provided periodic checkpoints, and no checkpoint event appeared in 26,679 execution traces.
- Analysis of 73 official versions and 64 version transitions across 9 evolution runs found that visible and hidden scores moved in the same direction 53.1% of the time, and the final selected version was optimal on hidden tasks in only 2/9 cases.
- In evolution experiments with Gemini fixed as the execution model, only the Opus family improved on hidden tasks, while the other 3 families worsened, showing that harness improvements can overfit a particular model and public feedback.
- 1059Machine Learning
WHALE: A Simple Recipe for Joint Harness-Weight Optimization
- It alternates between online rejection-sampling fine-tuning with the current harness fixed and Meta-Harness code search with the updated model fixed.
- Evaluations of Qwen3.5-2B and 4B on search-based question answering, mathematical reasoning, and chess puzzles achieved peak accuracies 7.67–24.38 percentage points higher than single-component baselines and 4.15–13.00 percentage points higher than FST, which changes only prompts.
- The harness was the main bottleneck for search tasks, but mathematical tasks required weight updates first before the same harness search became effective, showing that the bottleneck varies by task.
- Interleaving multiple small updates improved accuracy and rollout efficiency over a staged approach that trains all the weights and then searches for a harness once, providing grounds for designing agent-training pipelines as a joint optimization problem.
- Important limitations are that validation covered only Qwen3.5-2B and 4B and three tasks, and that the rollout-cost comparison excluded compute used by the model proposing harnesses.
- 1058Computer Vision
Qwen-Drive-1.0: An Initial Step towards a Vision-Language Foundation Model for Autonomous Driving
- Without changing the architecture of the pretrained VLM (Qwen3.5-4B), it added just two external modules, a BEV perception head and a Planning Expert, to integrate 3D perception, question answering, and path planning.
- It used the BEV head as a '3D probe' to make it possible to test how well the model understands actual 3D scene structure, beyond simply producing fluent text answers.
- Staged training that mixes driving data with general-purpose vision-language data preserved both driving and general capabilities, mitigating catastrophic forgetting (losing old capabilities while learning new ones).
- It achieved 43.95 mAP for 3D perception and 60.99 mIoU for mapping on nuScenes, a NAVSIM driving score (PDMS) of 90.7, and a Waymo E2E evaluator score of 7.91.
- In line with the trend toward vehicle cockpits (multimedia) and driving functions sharing a single computing platform, it aims for a model that handles them together rather than a model dedicated only to driving.
- 1057Computer Vision
H3-World: Turning Language Understanding into World Control
- Rather than attaching a dedicated control module to the 33B video-generation model MiniMax-H3, it directly reused the text pathway the model already understands as a control interface.
- It achieved precise character and camera control using just 8,000 gameplay samples, 10,000 LoRA training steps, and 0.199% trainable parameters.
- Temporal attention routing restricts each command to its own time interval, reducing command leakage into other intervals.
- In camera-direction-change experiments, the global-prompt approach achieved horizontal flow of only +0.0/-17.3, whereas H3-World accurately followed both directions with +52.7/-106.0.
- The same interface worked unchanged for action combinations never seen during training and for scenes outside games, including first-person views, interiors, and fantasy settings.
- 1056Computer Vision
Lucida: Parse, Generate, and Place for Composable Real-to-Sim Scene Modeling
- It redistributed the requirements of a three-stage pipeline that had demanded 'precise inputs,' redesigning each stage to consume only what real captured data can reliably provide.
- A VLM policy called GizmoAct translates object placement into 3D editor GUI operations, deciding the next adjustment and when to stop using rendered views alone.
- It improved scene-level 3D detection mAP by 69% over Boxer, from 0.351 to 0.592.
- It raised pose-estimation ADD-SB@0.05 on CA-1M from 57.8% to 83.4%, a gain of 25.6 percentage points, and increased the scene-reconstruction F-score from 0.794 (SAM 3D) to 0.924.
- A single policy trained with reinforcement learning after SFT handled initialization methods with different error characteristics without retraining.
- 1055LLMs / NLP
PaperGym: Rubric-Centered Evolution for Research-Plan Generation
- Using paper structure to separate the sources of questions, drawn from objectives and background, from scoring criteria, drawn from methods and experiments, reduced criterion leakage from the previous 11.9–34.1% to 3.7%.
- Two-stage training that uses the same rubric twice, first as privileged context for OPSD self-teaching and then as a GRPO reward, outperformed SFT, either stage alone, and the reverse order.
- For Qwen3-1.7B/4B/8B, it raised the average across 5 benchmarks by +5.6, +5.0, and +4.8 points, respectively.
- In a controlled comparison with the training recipe fixed, a model trained on PaperGym-20k dominated RubricHub Science (28.2%) with a 58.1% win rate in a three-way comparison, isolating the effect of data quality.
- The trained Qwen3-8B scored 73.48 on ResearchQA, surpassing the much larger Kimi K2.6 at 73.19.
- 1054Robotics
LightNav-0: Eliciting VLM Spatial Intelligence for Generalist Embodied Navigation
- A single VLM backbone handled instruction following, open-vocabulary object search, and visual tracking in one model without task-specific prediction heads.
- Dual pointing (an affordance point + an object point) raised average success from 54.7% to 63.1%, and removing it reduced performance on every benchmark.
- A residual vector quantization (RVQ) action tokenizer precisely generated a 10-waypoint trajectory using just 3 tokens.
- It achieved the highest monocular success rate across all 10 public navigation simulations.
- It transferred zero-shot to humanoid, quadruped, aerial, and wheeled robots, and expanding environmental diversity was a more reliable driver of performance than increasing model size (the backbone saturated at 4B and above).
- 1053Machine Learning
Does On-Policy Distillation Really Distill? From Noisy Teacher to Self-Improvement
- Noise in OPD's teacher signals, penalizing correct answers and rewarding incorrect ones, reached 30.6% for a 4B teacher and 50.6% for a 235B teacher, yet training after removing this noise produced essentially the same results.
- It showed that OPD's gains concentrate on low-probability tokens and that a single fixed negative advantage can replace the teacher signal.
- OPSA, with no external supervision at all, raised AIME24 Avg@32 by +35.41 points, a relative gain of 263%, and exceeded OPD by 16.77 points.
- It at least doubled Pass@32 on all 3 mathematical benchmarks: AIME24/25 and HMMT.
- At high-entropy branching points where probabilities diverge, it distributes probability evenly among candidate tokens, preserving exploration diversity while sharpening the distribution.
- 1052Machine Learning
Normalized Low-Rank Adaptation
- It identified that LoRA's early learning is governed by the down-projection A and showed that simply normalizing A along the rank dimension improves optimization.
- It raised the supervised fine-tuning average from 37.93 (LoRA) to 43.37, and even NoRA-init, which applies normalization only at initialization, reached 42.38.
- In reinforcement learning for mathematical reasoning (RLVR), it achieved an overall average of 44.4%, exceeding LoRA (42.8%), and remained stable in contrast to SVD-based PiSSA, which collapsed to 0.2% under the same settings.
- Post-adaptation retention of existing knowledge (average change across 3 evaluations including MMLU) was +0.02, showing much less forgetting than LoRA (-0.56) or MiSS (-0.70).
- Applying the same normalization principle to DoRA also raised its score from 38.30 to 41.40, confirming that the idea is not limited to the LoRA family.
- 1051AI / General
The Illusion of $\textit{What If}$: Evaluating the Breakdown of Counterfactual Reasoning in LLMs
- It released a diagnostic benchmark of 220 open-domain what-if questions spanning STEM, humanities and social sciences, and mixed domains.
- It developed the PRISM evaluator, which converts free-form answers into causal graphs, DAGs of events, states, and mechanisms, to score the reasoning process rather than the outcome.
- The best score among 6 current LLMs was only 64.62%, showing that counterfactual reasoning remains far from saturation.
- The most common failure was a causal gap: models listed plausible outcomes but omitted the intermediate links connecting the premise to those outcomes.
- It concluded that fluent narratives conceal weak causal processes, emphasizing the need for structure-based evaluation.
- 1050Machine Learning
Fast Weight Attention for Continual Learning
- It reinterpreted state-update rules as online learning and developed a framework that separately analyzes temporal alignment, plasticity, forgetting, and rehearsal.
- It showed that, in next-token prediction, the pair the state should learn is offset by one step: (previous key, current value).
- It derived normalized first-order update equations for regression objectives (Falcon-1/2/3) and inner-product objectives (Falcon-1A/2A/3A), and provided recurrent, masked-parallel, and chunk-parallel forms.
- It remained competitive with existing recurrent models in language-modeling evaluations.
- In extrapolation to 33–48-digit addition, Falcon-3A.3 achieved 87.2%, outperforming both RetNet and Transformer baselines.
- 1049Computer Vision
Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning
- It encoded objects, physical laws, and appearance as code to create compact world representations that can also be verified numerically.
- An agent loop of proposal, execution, rendering, verification, and revision, inspired by abductive reasoning, automatically discovered world representations from text or video.
- Using verified executable worlds as supervision, it taught vision-language models quantitative physical reasoning, such as answering questions about speed or mass in videos.
- Code-as-World-VL-9B outperformed much larger commercial models such as Gemini-3.1-Flash on QuantiPhy, while the 27B model achieved the best overall score.
- Its central contribution is translating the scientific discovery process of forming hypotheses and testing them experimentally into an agent loop.
- 1048LLMs / NLP
ContextPilot: Teaching Agents for Proactive Context Management via Fine-grained RL
- Adding planning, long-term memory, and offloading to the tool set raised BrowseComp+ accuracy from 63.49% to 80.96%.
- Reinforcement learning delivered a 5.34-point gain on the more difficult BrowseComp+ benchmark. The longer and harder the task, the greater the effect.
- The baseline agent's input length grew linearly with each turn to 30K tokens, whereas ContextPilot stabilized at 8–10K tokens.
- Reinforcement learning also reduced tool-use mistakes. Memory and offloading tool failures, which were frequent early on, decreased substantially during training.
- It showed consistent improvements across different base models, including Qwen3-8B/14B and Gemma4-E4B.
- 1047Research
Agentic Artifact Creation: Systems, Evaluation, Principles, and Opportunities
- Reviewing 259 studies published through August 2026, it defined agentic creation with state persistence and observation-based course correction as a research field.
- It compared shared challenges and difficulty across 6 artifact families: text, 2D images, audio, video, 3D and spatial artifacts, and behavior such as software.
- It identified the trade-off whereby decomposing a task reduces local complexity but increases coordination and reassembly costs.
- It warned that learned judges sharing the generating model's preferences and blind spots add little independent evidence.
- It proposed principles of making contracts and responsibilities explicit, turning feedback into local repairs, and revalidating affected parts after changes.
- 1046LLMs / NLP
Sliding-window beats linear attention
- 0 additional training runs: SWA, which changes only the attention mask, matched or outperformed conversions to linear attention that require retraining on billions of tokens.
- The gap was larger in long contexts. At a 256K context length, SWA recovered 20–25% of the original performance, whereas linear attention (LoLCATs) recovered only 2.2–5%.
- On MMLU, SWA recovered 93.2% of the original performance, outperforming all major linearization methods, including LoLCATs (83.2%) and Liger-GLA (62.2%).
- SWA (window 64) had the fastest decoding speed and lowest memory use among the tested methods.
- Citations to related work also confirmed that training-free sliding windows are a strong baseline for video generation models.
- 1045Computer Vision
Video Generative Models as Geometry Learner
- By directly inheriting a video-generation model's existing knowledge of temporal consistency, it generated depth and normals together with images in a single pass.
- With approximately one hundredth as much training data, it matched existing state-of-the-art discriminative models.
- Unlike previous methods that predicted depth and normals separately, it unified them in a single model, saving checkpoint storage and computation costs.
- A single prediction surpassed all generative baselines, and a lightweight ensemble (5×5, 10 seconds) further improved performance.
- The extracted 3D geometry enabled applications such as image relighting, controllable image generation, and 3D mesh reconstruction.
- 1044Computer Vision
Self-OPD: On-Policy Distillation for Flow Matching Models without Teacher
- It requires no teacher model at all, reaching the target in approximately 44–48 hours with a warm-start configuration, compared with DiffusionOPD's total of 97 hours (85.8h of teacher training + 11.3h of distillation), making it nearly 2 times faster.
- K stochastic branches + a self-baseline provide dense learning signals at every step, avoiding the high-variance problem of reinforcement learning methods (such as Flow-GRPO) that use only terminal rewards.
- Best-of-K, which follows only the good branches, was unstable during training. Pull-push over all branches + KL regularization was essential for stable convergence (an unconstrained repulsion term caused performance to collapse to 0.768).
- It combined multiple objectives (text, composition, and aesthetics) at the reward level so that a single image could satisfy them simultaneously without gradient conflicts (a preference-score drop of 0.48 vs 1.23 with the teacher-based method).
- 1043AI / General
SwarmWorld: Stigmergic technological evolution in societies of language-model agents
- Despite having no roles, recipes, or central control, identical agents spontaneously differentiated into exploration, construction, maintenance, and coordination behaviors, as confirmed by clustering on 15 behavioral features.
- It measured objective functionality judged by a simulator rather than LLM self-reports, using an evaluation setup that removed all agents before applying 8 unannounced perturbations.
- Compared with isolated exploration, the shared world yielded +1.2 to +6.0 more verified inventions and consistently better resilience metrics in paired comparisons across 4 seeds.
- Technology reuse most often began with firsthand observation of artifacts made by others, rather than through chat.
- Explicit cultural mechanisms such as teaching (TEACH) increased collaboration and organization, but their functional gains varied by evaluation metric and time period, making them no universal solution.
- 1042Machine Learning
Understanding Evolution Strategies for LLM Reasoning: Broader Reasoning Coverage than GRPO
- It showed that, unlike GRPO, ES improves Pass@1 and Pass@K together without entropy collapse (loss of exploration diversity), preserving diversity in reasoning answers.
- A sequential combination that first trains with GRPO and then finishes with ES combined the strengths of both methods.
- Performance gains were concentrated in a small subset of large updates (mainly normalization and attention parameters), showing that functional changes are sparse even when parameters move substantially.
- ES largely preserved existing capabilities in held-out evaluations, providing a counterexample to the idea that large parameter movements themselves cause catastrophic forgetting.
- Z-score reward normalization is essential, and as model size increased, a population of just N=16 approached the performance of N=64.
- 1041Computer Vision
DINOcular: Self-Supervised Visuospatial Representations
- It introduced depth as the third axis of rotary positional encoding (RoPE), allowing geometric information to enter feature representations naturally.
- Through self-supervised learning, learning without labels, it learned representations capturing both appearance and spatial structure from RGB-D video.
- It outperformed existing models of the same size on 3D correspondence matching in NAVI and ScanNet, and even surpassed DINOv3 ViT-B on ScanNet.
- In 3D pose estimation, there were conditions under which the large model, DINOcular-L, outperformed every baseline.
- It identified a trade-off in which improved 3D understanding came at a slight cost to semantic performance.
- 1040LLMs / NLP
Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090
- FP8 low-precision training on a consumer RTX 5090 cluster reduced the pretraining cost of a 2B model to approximately $6,900.
- Training on 1.4 trillion tokens over 22,514 GPU hours and 17.6 days surpassed Qwen2-1.5B and approached Qwen2.5-1.5B performance.
- It proposed the 'Puro cost scaling law,' fitted to the relationship between training cost and performance, and estimated that approximately $4,400 would suffice to reach Qwen2-1.5B-level performance.
- Five factors produced the cost efficiency: hardware selection, low-precision training, Hyperball optimization, curriculum model averaging, and the data recipe.
- It released all 10 sets of weights, data manifests, and training and preprocessing code under Apache 2.0.
- 1039Robotics
Pre-training Visual Dexterity in Simulation
- Using VR teleoperation in simulation, it collected 75 hours of hand-manipulation data, approximately 2000 episodes, in one week without a real robot.
- After pretraining, fine-tuning on just 1–2 hours of real demonstrations outperformed training from scratch on all 5 real-world tasks.
- Providing a history of past observations and using short action chunks of 8 steps achieved both responsiveness and consistency.
- The configuration combining history with short action chunks benefited most from pretraining, improving by an average of 18 points over training from scratch.
- It released the dataset (spd-75h), VR collection software (spd-vr), and 6 tuned scenes.
- 1038AI / General
WikiSkill: Compiling Agent Experience into Persistent Knowledge for Skill Evolution
- It separated execution traces, knowledge, and skills into three layers, with a wiki layer that continuously accumulates information and is never reset.
- It enabled stable evolution through gating that accepts only skill changes that improve validation performance and rolls back changes that make it worse.
- It consistently outperformed existing skill-evolution methods across 5 benchmarks (math, search, spreadsheets, document QA, and household-environment tasks) and 5 models.
- Skill evolution had a larger effect on larger models: in the Qwen family, 4B/9B/27B models improved by an average of +12.3/+17.5/+23.9%, respectively.
- The 9B model + WikiSkill (47.4%) beat the 27B model without skills (39.4%), and skills evolved by another model sometimes outperformed a model's own skills.
- Performance declined in an ablation that removed the wiki, confirming that knowledge accumulation is central to the effect.
- 1037Robotics
CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators
- It developed a cross-embodiment world-model framework that jointly trains a single model on videos of different robot platforms and humans.
- It unified different action spaces through end-effector coordinates, natural language, and latent action representations.
- Curriculum learning progressing from unlabeled video with latent actions to end-effector actions enabled zero-shot deployment.
- In challenging environments such as DROID, it approached or surpassed single-embodiment state-of-the-art methods, with few-shot adaptation widening the gap further.
- Planning at inference time within the world model and RL fine-tuning improved the real-world task success rates of π0.5 and MolmoAct-2.
- 1036AI / General
Tunable Tool-Call Rates in LLM Agents via Representation Steering
- It extracted a linear direction governing whether to call tools from the model's own signals, without training.
- Adjusting the direction's strength changes the call rate monotonically from near 0% to 90% or higher (while preserving valid call formatting).
- Increasing the direction makes the model call tools specifically for questions it cannot answer on its own, while decreasing it reduces unnecessary calls.
- It transfers to unseen tools, and which tool to choose (selection) and whether to call a tool (decision) are separated into different directions.
- It raised QA accuracy from 0.29 to 0.56 in a real execution environment, and the same recipe worked for dense, MoE, and multimodal models.
- 1035Research
High-Temperature Superconductivity of the Fe-Se-H compound
- It produced Fe-Se-H powder by thermally diffusing hydrogen at 180 degrees and 5 atmospheres into FeSe single crystals with a critical temperature of 8K for 10 hours.
- Using EPR spectroscopy, it measured magnetic-field-dependent microwave absorption up to 6000Oe at 3.6–25K and 295K.
- It interpreted U-shaped nonlinear absorption and quasiperiodic signals as evidence of a type-II superconductor's mixed state and Abrikosov vortices.
- Based on these signals, the authors claim that superconductivity exists even at 295K under normal pressure, suggesting the possibility of room-temperature superconductivity.
- However, the work relies on interpretation of EPR signals rather than direct transition measurements of electrical resistance or magnetization, so independent reproduction and cross-validation are essential.
- 1034Robotics
Memory Anchors for Continual Robot Learning
- It found that not all data in a replay buffer are equally important: a small subset in conflict regions determines how well forgetting is prevented.
- It developed a three-stage procedure for selecting memory anchors by locating regions in the policy's latent space where observations overlap but actions conflict.
- Removing only the top 10% of anchors increased forgetting 4.5-fold on the LIBERO benchmark.
- ANCHORER, which fills 10% of the buffer with anchors, reduced forgetting on conflicting tasks by 63%.
- In continual learning over a sequence of two real-robot tasks, it achieved 1.7 times the overall success rate of a random buffer.
- 1033Robotics
Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latency
- It identified the cause of failure: VLA inference latency of 100–300ms changes environment dynamics and breaks RL's Markov assumption.
- It restored an approximately Markov structure by augmenting the state with committed actions and observations collected during inference.
- It enabled stable fine-tuning even under latency conditions where standard RL failed completely.
- It matched or exceeded standard RL's performance under ideal conditions with no latency.
- It demonstrated real-world improvements under asynchronous inference on 3 tasks using a real dual-arm UR5e robot.
- 1032Robotics
FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference
- It began by profiling π0.5 and confirming that action decoding consumes 75% of per-step inference time.
- It created a streaming architecture with a buffer of action chunks at staggered noise levels, denoising all of them by one step in a single forward pass.
- Per-step action-decoding latency was reduced by a factor of up to 20, and the full pipeline became up to 2.43 times faster.
- Causal attention between chunks eliminated temporal misalignment in asynchronous execution without a separate future-state predictor.
- On LIBERO, RoboTwin 2.0, and real Franka experiments, it maintained or improved π0.5 performance while achieving control at 30Hz or higher on a single GPU.
- 1031AI / General
Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings
- Using only a natural-language question, it automated a geospatial prediction pipeline of more than 700 steps, from data discovery to model selection.
- It builds predictive features by combining geospatial foundation-model embeddings with covariates collected in real time.
- It raised average R² across 21 US CDC health indicators from 60.0% to 76.8%, substantially outperforming the expert baseline.
- In forecasting the 2026 Ebola outbreak in Congo, forecasts over 5 consecutive weeks anticipated 15 of the 18 newly affected health zones (Recall@10 83.3%).
- It embedded validation mechanisms such as target-leakage prevention and overfitting guards within the agent loop.
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology