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.
- 940Research
Mathematics in the age of AI
- Taking the premise that AI can perform research-level mathematics without debating AI capability itself, the paper argues that the mathematics community should explicitly state the goals and values it has long advanced implicitly.
- In the second batch of the First Proof project, seven of ten problems received an effectively perfect passing assessment from at least one AI system.
- The compute cost of these evaluations ranged from tens to hundreds of dollars per problem.
- The paper warns that AI is adept at finding gaps between measurements and their intended targets, making the risk of Goodhart's law especially acute: once a measure becomes a target, it ceases to be a good measure.
- In an age of abundance in which proofs are plentiful, the authors propose shifting cultural emphasis from generating proofs to digesting them through explanation, review, and incorporation into the canon.
- 939LLMs / NLP
Every Coin Has Two Sides: On the Dual Nature of Generalization in On-Policy Distillation of Large Language Models
- On-policy distillation reduces the difference between the teacher's and student's next-token distributions at each point in an answer generated by the student. The student therefore receives guidance in the contexts it actually visits during training.
- For BigMath, the teacher selected 25,000 problems from each of three groups: problems answered correctly on all four attempts, problems answered incorrectly on all four attempts, and randomly selected problems. Across several teacher and student combinations, the three groups produced nearly identical final mathematics accuracy.
- A teacher derived from the same base model brought the student close to the teacher's level in Chinese, long-horizon mathematics, coding, and science, even when training covered only English and short-horizon mathematics. A stronger teacher with a different base model produced much weaker transfer across those areas.
- Changing the ratio of data from two teachers caused the student's scores across several areas to move together toward the original capability profile of the teacher used more heavily. For the first teacher pair, the average score on BeyondAIME and OlymMATH ranged from 25.1% to 27.1% across mixture conditions.
- The experiments were limited to mathematics, coding, science, and instruction-following models. The multi-teacher study also covered only two teachers with fixed domain-based routing. The authors warn that domain partitioning should not be treated as a safety boundary because undesirable teacher behavior may also spread across domains.
- 938Computer Vision
TRACE-Bench: Decomposing and Diagnosing Multi-Reference Image Generation
- The benchmark represents requests through four operations: object preservation, attribute disentanglement, attribute application, and scene composition. It uses the same formula structure to generate problems, score individual items, and trace the causes of failure.
- The dataset contains approximately 1,600 evaluation cases, 631 formula templates, and about 4,000 reference images. Complexity ranges from one to eight operation slots.
- The evaluation compared four commercial models with five open models. Nano Banana 2 achieved the highest average across the four operations at 0.8205, but scored 0.7384 on attribute disentanglement and 0.7989 on attribute application.
- When 200 Emu3.5 cases were decomposed into simpler subproblems, 38.5% of all diagnostic outcomes were classified as interference arising when multiple references were composed together. Automated analysis and human failure localization agreed in 82.6% of cases.
- Gemini-2.5-Pro scores the checklist used for the overall evaluation. When humans rechecked 400 outputs produced by two models across 200 cases, item-level agreement for the four scoring models ranged from 85.4% to 88.4%.
- 937Robotics
HAF: Adapting Generalist VLAs to Humanoid Whole-Body Loco-manipulation via Hierarchical Action Flow and Spectral Latent RL
- HAF-VLA activates head and locomotion, the waist, and bimanual manipulation sequentially across three stages. Stage indicators and the previous stage's KV cache connect earlier body movements with subsequent manipulation actions.
- HAF-Steer compresses the VLA's temporal noise with a discrete cosine transform and learns only the first eight coefficients. It begins with behavior cloning, then applies SAC reinforcement learning that mixes offline data with real interactions while keeping the VLA backbone frozen.
- The researchers tested seven tasks on TienKung 2.0 and 3.0, including loading laundry, carrying a basket, and throwing a ball. They collected 120 teleoperated demonstrations for each task and ran each method on each task ten times.
- Across the seven tasks, HAF-VLA achieved an average progress score of 70.5%, compared with 53.3% for the strongest baseline, π0.5. It recorded the highest or joint-highest score on every task.
- HAF-Steer improved HAF-VLA's success rate across all four conditions covering the original and new locations for two tasks. For example, toy storage at the new location improved from eight successes in ten trials to ten in ten. Its limitations include additional computational latency and dependence on the action range of the base VLA.
- 936Machine Learning
On the Principles Behind Neural Network Optimizers
- Adam is neither always convergent nor always divergent, and the paper gives a safe region that depends on the problem and the number of mini-batches, with smaller batches requiring a larger second-moment coefficient beta2.
- The Hessian of transformers becomes almost block diagonal during training, and the curvature differs greatly across blocks. The authors explain this as the reason Adam, with coordinate-wise learning rates, can be better than SGD.
- They analyze simple neural networks and random matrix theory to show that consecutive large matrix multiplications create this Hessian structure, and this view also leads to a design that splits Muon correction across attention heads.
- Adam-mini replaces more than 99.9% of the second-moment values with block-wise scalars, which cuts optimizer-state memory by 50%. By comparison, Adam needs about 104 GB just for the two moment states in a 13B-parameter model.
- When training Llama 2 architectures from 39M to 1B parameters on C4, Adam-mini matches AdamW validation loss. However, this is a shortened version with no proofs, and rigorous analysis of attention and Mixture-of-Experts structures is left for future work.
- 935Robotics
GigaBrain-0.7: Scaling Embodied Foundation Models to Emergent Capabilities with a Three-System Architecture
- System 2 uses PaliGemma2 3B to understand the scene and task, System 1 uses PaliGemma2 3B plus a 0.5B action model to produce control values, and System 3 uses a 5B world model to predict future video and progress.
- The trajectory data consists of 20,535.65 hours of real-robot data, 8,251.83 hours of UMI human demonstrations, 2,862.36 hours of first-person human video, 1,453.92 hours of simulation, and 4,153.22 hours of world-model generation.
- In the System 3 ablation, the gift-wrapping success rate on AgileX PiPER-X rises from 0 percent for the base model to 80 percent when future video and progress are used together, and the clothing-folding completion time on AgileX PiPER falls from 107 seconds to 75 seconds.
- Under the Co-Train setting on 50 RoboTwin 2.0 tasks, where 50 clean demonstrations per task and one policy are jointly trained, the results are 66.8 percent for Easy, 67.9 percent for Hard, and 67.35 percent on average, while pi0.5 reaches 70.7 percent, 46.0 percent, and 58.35 percent.
- The real-robot tasks are usually evaluated with only 10 to 20 runs per setting and no confidence intervals, and the final MiMo score of 0.5704 also uses evaluation data that was included in training, so it is not an independent test result.
- 934Information Retrieval
Ask to Be Sure: Informative Interactions for Confident Multi-Turn LLM Recommendation
- It computes uncertainty from the frequency and rank of movies in the recommendation list, and it uses how much the current response and the next user response reduce that uncertainty as the reward.
- INSPIRED has 801 training dialogs and 99 test dialogs, while ReDial has 8,631 training dialogs and 1,036 test dialogs after filtering, and the simulated conversations run for up to five turns.
- Averaged over three runs, turn-entropy DPO reaches 27.94% hit rate on INSPIRED simulated dialogs, which is higher than CollabLLM's reward at 26.60%, and on ReDial conversation-level entropy DPO reaches 32.83% versus 30.03%.
- On ReDial, the average number of turns needed to reach the correct movie is 2.74 versus 2.86 for CollabLLM DPO, but this is not measured in real human conversations.
- The evaluation depends on automatic recommendation metrics and LLM simulated users, sensitivity to the setting of five recommendation lists sampled five times is not validated, and direct recommendation Hit@1 remains low at 3.32% on INSPIRED and 2.24% on ReDial.
- 933Research
Contrastive Learning with Variational Regularization for Multi-Session EEG-to-Speech Decoding
- The SpREAD dataset is EEG from one participant who listened to 1,353 Japanese sentences spoken by 18 speakers. Each sentence was measured three times on different days, for 45 sessions over 9 days.
- The train, development, and test sets contain 3,195, 432, and 432 examples, respectively. Repeated measurements of the same sentence were always placed in the same split.
- On the 432 test examples, the character error rate of 0.948 is significantly lower than the baseline 0.968. A classifier trained on the training set and evaluated on the development set reaches only 2.08%, which is no better than chance at 2.2%.
- Using contrastive learning alone makes the acoustic correlation coefficient worse, dropping from 0.282 to 0.262. Adding variational regularization recovers it to 0.274, but that is still not better than the baseline, and speaker information improves only marginally.
- There is only one participant, and the 18 speakers appear in both training and evaluation. The system handles only Japanese speech, and current EEG-based speech reconstruction remains far from practical.
- 932Research
The New Mathematics of Democracy
- The paper has no original experiments of its own; instead, it interprets theory, real election data, institutional adoption cases, and open research problems across three fields.
- Among about 3,000 IRV and STV election analyses cited in the paper, IRV and Condorcet methods choose the same winner more than 99 percent of the time, although voter behavior could also change if different rules were actually used.
- Budgeted voting has been used in 32 real participatory-budgeting elections, and the equal-share method reduced the share of voters who received no approved project in Wieliczka, Poland, from 28 percent under the prior method to 18 percent.
- Determining whether there exists a citizen panel that satisfies overlapping population quotas is NP-complete, and the related tool Panelot was used on 1,071 citizen assembly lotteries over 294 distinct days in 2024 and 2025.
- The equal-share method can leave budget unused, and proportionality can conflict with total satisfaction, so AI should serve as an aid that translates citizen expression into a verifiable mathematical form rather than replacing decisions.
- 931Computer Vision
HiFi-BRep: High-Fidelity Latent Representation for Robust B-Rep Generation
- It separates faces and edges for processing but lets them exchange information only for pairs that are actually connected, and it creates a compact representation with 48 learnable queries whose input length does not depend on the number of parts.
- After duplicate removal and complexity filtering, it trained on 83,611 DeepCAD shapes and 186,148 ABC shapes, and Valid here means the share that passes STEP export, watertight closure, and manifold connectivity checks.
- For unconditional generation, Valid is 72.20 percent on DeepCAD and 32.66 percent on ABC, which is higher than DTGBrepGen's 43.20 percent and 24.88 percent, but DTGBrepGen is better on ABC distribution metrics such as COV, MMD-CD, and JSD.
- On DeepCAD reconstruction, the full model reaches 95.2 percent Valid; when it first reconstructs the shape and only then predicts the connectivity, the score drops to 69.3 percent, and the non-parallel generation time is 3.83 seconds per shape.
- It supports only closed solids and a fixed maximum number of faces and edges, and it still leaves cut faces, mismatched joints, thin faces, and self-intersections, with exact intersection and trimming left to the CAD kernel.
- 930Robotics
Neurosymbolic Embodied Agents
- The first stage checks object positions and relations through multi-view and in-simulator interactions, such as opening or picking up objects. It does not place unseen objects into the state using the model's own guess.
- The second stage only generates action tokens that satisfy the current state constraints. Search time is up to 250 seconds per task, and the completed plan receives no feedback during execution.
- We evaluate 200 VirtualHome tasks and 134 hidden splits from ALFWorld. The total action budget is 20 and 55, respectively, and the model is not shown scene graphs or ground-truth states.
- For Qwen3.5-27B, ALFWorld success is 32.3% with constraints only, 29.2% with search only, and 95.5% when both are combined. Averaged across both environments at the same model scale, the token cost is up to 4x lower than long-reasoning methods.
- Most remaining failures are concentrated in the first stage. The rate of getting the state wrong is about 2.3% on VirtualHome and about 4.5% on ALFWorld, and the typical visual error is missing small objects.
- 929AI / General
Improving the matrix multiplication exponent with modern optimization and AlphaEvolve
- The new result is an upper bound showing that the number of operations needed for matrix multiplication is slightly below the 2.371177 power of the input size, and it is not a measurement of actual runtime for practical matrix multiplication.
- The previous combinatorial loss analysis used about 25,000 variables at maximum recursion depth 3, while the new approach handles about 7 million variables at depth 4.
- The probability distribution is represented with a softmax, and the maximum-entropy distribution is computed with Sinkhorn-Knopp; the optimization uses automatic differentiation, Adam, and parallelized JAX tensor computation.
- Gradient-based methods alone lowered the previous upper bound by about 0.000097, and after AlphaEvolve changed the optimization code the total improvement reached about 0.000162, with one candidate taking about 5 hours on a single GPU.
- The final solution is verified by rounding to rational numbers and applying safe rational bounds in the log domain, and the authors say that any major further improvement will require new mathematical ideas.
- 928Robotics
HiPHI: A Large-Scale Benchmark for High-Precision Human Motion and Object-Interaction
- The full dataset contains 617.5 hours, split into 371.8 hours of full-body motion and 245.7 hours of object interaction, with about 200 million frames, 22 FrameNet categories, and 214 fine-grained motion tags.
- For object interaction, it uses 40 real objects across 12 types, with weights from 0.45 kg to 6.25 kg, and object trajectories are synchronized with 3D models.
- It samples up to 5,000 one-second clips from each dataset for comparison. HiPHI covers 1,620 motion-space cells versus 1,438 for BONES-SEED, and its rare-region share is 14.1% versus 10.7%.
- When comparing datasets with the same floor reference, the floor penetration depth is 8 mm. The non-overlap rate between the object and the person is 98.1%, and the fraction with the person within 20 cm of the object surface is 95.7%.
- In a Unitree G1 full-body tracking experiment matched to 3 hours and 20 hours of data, it showed the highest success rate. Exact numbers are not given in the figures, the real-robot results are presented mainly as case videos, and the full dataset was still planned for release at the time of writing.
- 927Robotics
Zetta $ζ$: An Efficient Closed-Loop Embodied Harness for Self-Evolving Physical Intelligence
- LIBERO-Pro uses a fixed pi0.5 policy, and RoboCasa uses a fixed GR00T N1.5 policy; the monitoring code proposes anomalies and the fixed recovery agent decides whether to execute recovery.
- RoboCasa uses 50 development conditions and 50 separate test conditions for each task, and LIBERO-Pro uses 50 development conditions and 20 non-overlapping test conditions.
- RoboCasa rises from 73.56 percent to 93.56 percent over four rounds of cumulative improvement, while Goal T and Goal S in LIBERO-Pro rise from 31.0 percent to 92.5 percent and from 38.0 percent to 89.0 percent, respectively.
- When the recovery skill learned on PnP-Stove is applied to three related tasks without any additional improvement, the average rises from 64 percent to 84 percent, and on eight A100s it processes 16 concurrent runs at 22.09 per minute.
- The abstract's LIBERO-Pro score of 90.8 percent does not match the 71.13 percent overall average in Table 3, so this summary uses the table value; the paper shows real-robot videos but reports no quantitative evaluation there, so the simulation gains should not be read as real-world performance.
- 926LLMs / NLP
ClawGym II: Exploring Black-Box RL on Agent Harness
- It isolates the task environment and execution tools in separate temporary environments. Model requests are handled by a middle server that records input tokens, output tokens, and generation probabilities.
- ClawGym-Bench uses either code inspection alone or a weighted sum of code at 0.7 and evaluation criteria at 0.3. GPT-5.4 is used for evaluation-criterion judgments, and PinchBench evaluates 30 tasks after removing multimodal ones.
- PinchBench scores for the Qwen3-30A3B family improve from 75.61 to 87.32 in OpenClaw cold-start initial models. In Claude Code, they rise from 54.14 to 71.42 in the original model.
- Under the same execution budget, PPO runs 256 tasks once each, while GRPO runs 32 tasks eight times each. Both methods show stable gains over about 200 to 400 training steps.
- JobBench-Easy improves from 20.46 to 27.20, and OfficeQA-Full from 8.53 to 21.54. The score of a model trained in WhiteBox AgentLoop and transferred to OpenClaw is 50.33, lower than 62.62 for a model trained directly in OpenClaw.
- 925Computer Vision
GenRouter: Unified Workflow Routing for Agentic Image Generation
- GenCanvas provides a fixed framework ranging from lightweight direct generation to heavy workflows that use retrieval, spatial sketches, and iterative verification.
- GenRouter chooses the appropriate workflow in the order of requirement analysis, candidate pruning, experience matching, and quality-versus-cost filtering.
- The main comparison uses Qwen-Image-2512 and Z-Image-Turbo as generators, with common language and vision tools aligned to Kimi K2.5.
- In experiments with 500 mixed prompts from nine benchmarks, adding experience from three benchmarks raises the score from 73.5 to 75.2, while cost and time fall by 8.7% and 7.9%, respectively.
- It uses 10 calibration prompts separated from evaluation for each new benchmark, and it runs all possible workflow and generator combinations for each prompt. Because cost and time weights are also chosen by humans, it is not fully zero-shot.
- 924Robotics
$τ_0$-VLA: a Hierarchical Robot Foundation Model with World-Model-Guided Test-Time Computation
- The high-level policy looks at the current observation, the full instruction, previous small tasks, and execution history together to propose the next task.
- In the extra compute stage, it expands candidates over several steps, keeps only the good ones using the predicted final scene and cumulative score, and then regenerates the final task.
- Four long-horizon tasks were evaluated with 10 real robot trials per task. The mean success rate of direct execution with the same low-level policy rose from 27.5 percent to 45.0 percent in hierarchical execution without retrieval.
- With extra compute turned on, the number of successful trials across three tasks increased from 5 to 7 for making tea, from 6 to 9 for tidying books, and from 5 to 7 for cleaning a room.
- On predicting the next task for unseen book arrangements, accuracy was 50.0 percent for a single prediction, 57.5 percent for one-step candidate comparison, and 74.0 percent for multi-step extra compute. However, task-specific fine-tuning and baseline adjustment were included.
- 923Computer Vision
An Empirical Study of Training Pixel-Space Text-to-Image Diffusion Models
- Experiments that changed only the prediction space while keeping the same Z-Image architecture and corpus of more than 20 billion image-text pairs showed the slow convergence of pixel-space pretraining.
- When run at 1,024 resolution on a single H800 and only the decoder was used, the DiP decoder used 10.09 million parameters and 6.02 milliseconds, and recorded GenEval 0.7545 and DPG 87.54.
- In the noise-scale experiment, 2 worked best with GenEval 0.7545 and DPG 87.54. The theoretically derived value of 8 reached only 0.7316 and 85.64, respectively.
- At 1,024 resolution on a single H800 with 100 model evaluations, Z-Image generation time dropped from 20.12 seconds to 4.56 seconds, and GenEval rose from 0.7510 to 0.7644.
- In the same Z-Image comparison, OneIG fell slightly from 0.566 to 0.564, and LongText from 0.9332 to 0.9252. The validated model families are Z-Image and FLUX2-klein.
- 922Machine Learning
Proteus: Incremental Memory Activation for Long-Context Sequence Modeling
- The base experiment splits memory into 16 blocks, allows reading and writing only to the currently active blocks, and preserves locked blocks until they are reopened.
- With 1.3 billion parameters and 100 billion training tokens, it reduces Titans perplexity on WikiText from 15.36 to 14.94 and on LAMBADA from 13.18 to 13.03.
- After training on FineWeb with an 8K context and evaluating at 16K on S-NIAH-3, Titans improves from 21.4 to 29.8 accuracy, and on S-NIAH-2 it improves from 69.4 to 74.2.
- Across six LongBench tasks, the average score rises from 15.72 to 16.65 for Hope-Attention, from 13.05 to 13.23 for Comba, and from 13.80 to 14.15 for Titans.
- The optimal activation order is not searched separately, and the extension to MLP blocks is only a proof-of-concept checked in one Hope-Attention structure.
- 921Computer Vision
HarnessEval-W: Agentifying the Evaluation of Visual Worlds
- The evaluation dimensions are eight in total: rendering, physical observation, navigation switching, intent switching, physical switching, suppression of unnecessary scene changes, revisit consistency, and off-screen changes.
- The evaluator breaks a question into small measurable problems and collects scene evidence using only the skills and tools that are needed.
- On the same 330 cases, Seedance 2.0 ranks first with an overall score of 75.5, but because the models use different input formats, this should not be treated as a direct apples-to-apples comparison.
- Compared with 5,000 human pairwise judgments gathered from nine models, the Spearman rank correlation is 0.93 for intent switching and 0.87 for physical switching.
- With the same video and GPT-5.5 evaluator settings, agreement with human judgment is higher than in the previous WBench, but the validation focuses on only two of eight conditions and may not find the right method for new situations that the evaluation techniques are not ready for.
- 920Machine Learning
Forecast Collapse in Time-Series Foundation Models
- Finance1K consists of 28,510 hourly time points for 1,000 U.S. stocks from 2015 through early 2026, and it compares stock returns and volume changes under the same conditions.
- When the target has little predictable signal, as with stock returns, even accurate point forecasts can end up smaller than the actual variation. Training to reduce per-stock error does not directly guarantee the cross-stock ranking at the same time point.
- In the baseline model, MSE training produced a prediction spread that was 0.027 times the actual spread and an information coefficient of 0.046 for stock ranking. CalibRank achieved a prediction spread of 1.836 times and an IC of 0.126.
- Across 12 forecasting models, CalibRank improved the IC of every model. The mean IC increased from 0.062 to 0.123, while MSE changed from 0.00043 to 0.00058.
- The stock-level problem and solution were verified only on Finance1K. CalibRank handles only Pearson correlation, does not model tail risk or the full joint distribution, and cannot be applied directly to a single time series.
- 919AI / General
Scaling Domain Data Repetition in LLM Pretraining
- It tested four domains, code, math, wiki, and medical, using unique domain data ratios of 2.5%, 5%, and 10% of the total training set, and repeated each dataset from 1 to 7 times.
- The Pearson correlation between the optimal number of repeats and the minimum validation loss was -0.944. It was 0.400 with model size and 0.018 with unique data ratio.
- The optimal number of repeats for math data was about 5 to 6, and for medical data about 3 to 4. With the same total amount of domain data, math tolerated up to four repeats with little loss increase, but wiki performance degraded noticeably beyond two repeats.
- If the total ratio of math data to general web data is fixed, validation loss on ArXiv and news data stays mostly stable even as the repeat count changes. Lowering the learning rate earlier caused performance degradation to start with fewer repeats.
- Each training run repeated only one domain. It does not yet provide a formula for interactions when multiple domains are repeated together or for predicting the exact repeat count of a large model from small-model results.
- 918Computer Vision
Can We Defend Against AI-Generated Video Attacks on Real-World Crisis Events? A Systematic Evaluation of Detectors, Generators and Social Dissemination
- RA-Bench contains 17,886 videos, made up of 1,830 real videos and 16,056 generated videos, and the real videos cover 10 social risk categories.
- On public data, the AUC of 7 conventional detectors ranges from 67.6 percent to 98.6 percent. Their averages on RA-Bench across generators range only from 43.9 percent to 57.3 percent.
- Humans identified 68.6 percent of open-source generated videos as synthetic, but only 52.9 percent of closed-generator videos. There were also 633 generated videos that all five people judged to be real.
- When compression, downsampling, lower frame rate, and a news badge are all applied, the mean detection rate of a fine-tuned multimodal detector drops from 46.0 percent to 1.4 percent.
- The experiments are limited to 9 image-based video generators and silent video detection. Real misinformation can also involve voice synthesis, editing, and repeated processing across multiple platforms.
- 917Research
The Okinawa Lectures on Entropy
- The lectures explain, from a historical perspective, how Clausius, Boltzmann, Gibbs, Einstein, von Neumann, Shannon, Rényi, and Kolmogorov expanded the concept of entropy.
- Classical entropy is explained as a quantity that captures how quickly the probability of large fluctuations falls with sample size, with the discussion centered on the Sanov, Cramér, Gartner-Ellis, and Varadhan theorems.
- It explains that in classical and quantum hypothesis testing, if one type of error is constrained, the other error decreases at a rate given by relative entropy.
- It moves from von Neumann and Umegaki entropy for finite quantum states to the construction of Araki and Uhlmann relative entropy in arbitrary von Neumann algebras.
- Finally, it connects classical Kolmogorov-Sinai entropy with quantum dynamical entropy, and the author notes that there is no new material except for the organization of the material, the approach, and some proofs, while thermodynamics and black-hole applications are not covered.
- 916Computer Vision
Self-Supervised Visual On-Policy Distillation
- The teacher model sees clean images, while the student sees corrupted versions of the same images and learns to match the teacher's output distribution.
- The best setting lowers the student's image resolution to 0.3x to 0.6x of the original and adds Gaussian noise with 50% probability.
- Training uses Qwen3.5 4B and 9B, and evaluation covers six visual-recognition tasks and three mathematical-reasoning tasks.
- On 12,000 FineVision questions, the 4B model's average across six visual-recognition tasks rises from 70.68% to 77.44%, a gain of 6.76 points.
- Using the same image for teacher and student reaches only 70.52% on average, while the full method with information differences reaches 76.35%. Some analyses use shorter output limits than the main evaluation.
- 915LLMs / NLP
SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning
- It connects only the tokens that occupy the same string span in both models to the teacher model's score, and uses the student model's own score for the parts that do not match.
- The teacher model is SU-01, which can reason over more than 100K tokens. The student models are from the Qwen3, Qwen3.5, Intern-S2, GLM-4.7, and Gemma-4 families.
- ProofBench was scored by DeepSeek-V4-Flash, and Intern-S2's score rose from 21.70 to 44.50. AnswerBench rose from 76.03 to 80.10, and AIME25 rose from 88.33 to 95.00.
- Although it was trained only on math data, Intern-S2's HiPhO score rose from 38.6 to 41.1, and FrontierScience Research rose from 1.7 to 5.0.
- When the teacher-imitation loss on the end token was removed and the output-distribution gap with the initial student model was constrained, the answer truncation rate dropped almost to zero. However, Gemma's AnswerBench score fell from 68.8 to 67.5, so not all capabilities improved together.
- 914Computer Vision
Marionette: Predicting World States, Rendering Geometry, Painting Appearance
- ActionGPT selects actions and PoseGPT creates poses, while a fixed graphics pipeline computes skeletons and occlusion relationships and then a video generation model paints in the appearance.
- The action model was trained on 1,395 segments at 20 frames per second, and the current action range is limited to one kind of monster.
- On 48 test segments, forcing incorrect actions increased joint error from 0.272 meters to 0.357 meters, a 31 percent rise, which shows that action tokens control pose.
- Applying a ground-collision rule reduced the share of frames where joints passed through the ground from 0.337 to 0.114, and a distance-limit rule lowered the distance between characters from 21.2 meters to 5.1 meters.
- The video quality metric FVD was 831, lower than the pixel-based model's 975, but the confidence intervals overlapped, and long-horizon appearance changes and unseen-pose inputs remain unsolved.
- 913Computer Vision
SPARGen: Unifying Spatial Perception and Reasoning through Native Multimodal Generation
- Existing systems often handle 3D reconstruction and spatial question answering with different models or separate modules.
- SPARGen tokenizes camera poses and text answers, while generating depth, point maps, and optical flow as images.
- It trains on spatial reasoning, visual geometry, and optical flow data, and it is trained for 100k steps on 64 H100 GPUs.
- In public model comparisons, it ranks first on 13 of 15 spatial reasoning subcategories. In the KITTI optical flow evaluation without additional fine-tuning, it records an EPE of 4.09 and an F1-all of 13.34.
- A specialized 3D model, VGGT, performs better on some reconstruction metrics, but it predicts normalized depth and point maps, so it cannot recover true metric scale in meters.
- 912AI / General
Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning
- Traditional transformers tie knowledge storage and reasoning compute together layer by layer, so knowledge from other layers has to flow through multiple layers when it is needed.
- Mobius lets every reasoning layer read the same knowledge store and repeatedly refines the internal continuous representation before outputting a token.
- In the 7B comparison, Mobius reaches the same score as a standard transformer trained on the full dataset while using only 62.6% of the data.
- The 35B model achieves an average general-evaluation score of 67.88 and keeps accuracy close to the original Qwen3.5-35B while making question answering about four times faster.
- In one linear-algebra case, it outputs the same answer in 516 tokens, while the original model uses 2,364 tokens, although the reason for the shorter reasoning and higher data efficiency is still only a hypothesis.
- 911Robotics
PRM-as-a-Judge 1.5: A Toolkit for Robot Process Assessment
- Success rate alone cannot distinguish a near-miss failure from one that is blocked from the beginning, and it also misses hesitation and unnecessary motion during successful runs.
- The method turns execution videos into progress curves and computes multiple metrics for reached level, progress efficiency, stagnation, regression, and recovery.
- The new metrics FNS, DRR, and SQS show how close a failure was to success, how much recovery occurred after a large regression, and how stable the successful process was.
- On RoboDojo, model rankings differ between success rate and process metrics, and the rank correlation between simulation and real-world evaluation stays only between 0.18 and 0.58.
- RoboPulse++ evaluates the scoring model with 700 execution trajectories and 2,244 human-annotated segments, and even the best F1 score for progress-decrease segments is only 0.63, so the results should not be treated as absolute truth.
From research to product use
Operating capabilities, pilots, and technologies in development are identified separately.
View technology