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.

  1. 970Machine Learning

    Let's Scale Step by Step: Compute-Efficient Hyperparameter Transfer for Large-Scale Mixture-of-Experts

    • In stage 1, it formalized μP for MoE models using MLA attention and the Muon optimizer, and showed that the optimal learning rate found on a base proxy model (0.6B total, 0.3B active) transfers unchanged to models with 2, 4, and 8 times the width (up to 30.7B total, 3.6B active).
    • In stage 2, it stops training during the stable phase without a decay phase, extracts multiple checkpoints using EMA (exponential moving average) weights, fits optimal learning rates over 255B to 502B tokens using log-log linear regression, and extrapolates to longer horizons.
    • The regression fit was R²=0.95, and it predicted 3.85×10⁻⁴ as the optimal learning rate for training on 10 trillion tokens.
    • Using the predicted learning rate, it trained a foundation model with 155B total and 17B active parameters from scratch on 10 trillion tokens. The target training run required roughly 98 times the total computation of all proxy experiments, so avoiding a search saves substantial cost.
    • However, the authors state that validation is limited to the MLA and Muon combination, batch size was deliberately left out, increasing sparsity together with width prevented isolation of the sparsity axis's effect, and an exhaustive search to verify whether the predicted learning rate is truly optimal is computationally infeasible.
  2. 969Research

    Exact Fluctuation-Response Relations for Underdamped Langevin Dynamics

    • It derived exact finite-time fluctuation-response identities for underdamped Langevin dynamics that hold under arbitrary time-dependent driving and for general accumulated observables.
    • Using a Doob martingale, it exactly decomposed observable variance into a component attributable to the initial state and a component dynamically generated by subsequent noise.
    • Choosing a friction perturbation in the direction of the irreversible probability current yields an uncertainty relation saturable even at finite times and a variational principle for total entropy production.
    • In driven free diffusion, the friction-response-based factor maintained a universal lower bound even when the existing mean-current-based factor collapsed exponentially with dissipation.
    • The results are a theoretical analysis of underdamped Langevin processes and accumulated observables, so applicability to other noise structures, nonadditive observables, and actual experimental data requires separate validation.
  3. 968Machine Learning

    Harness Continual Learning: Continual Adaptation Beyond Model Parameters

    • It formalized the HCL paradigm, which shifts the target of continual learning from model parameters to the harness (prompts, memory, tools, and routing).
    • It introduced 'harness forgetting' and controlled it through guarded evolution, which commits changes only after they pass checks for improvement, historical retention, and validity.
    • Setting the historical-loss budget to 0 yielded average forgetting of 0.39, while leaving it unlimited yielded 3.45, allowing explicit control over the stability-plasticity tradeoff.
    • It achieved relative improvements of at least 10% across 3 domains, text, multimodal, and open-world, with a highest overall multimodal score of 68.92%.
    • Component-removal experiments showed that experience memory and task-interface updates contributed most to performance and retention.
  4. 967AI / General

    EnvHarness: Awakening Static Worlds for Agent Learning

    • It proposed an EnvHarness layer that changes only behavior by wrapping an existing environment at the reset/step interface with Stage, Contract, and Chain plugins, rather than constructing a new environment.
    • By inheriting the original verifier, the ground-truth scorer, unchanged, it inherently prevents the ground-truth reliability problem of LLM-generated environments.
    • It created an automated loop in which EnvRigger diagnoses policy-failure trajectories, automatically synthesizes environments targeting those weaknesses, and validates them with new rollouts.
    • Across 5 benchmarks in 4 domains, including SWE-bench Verified, it achieved gains of up to +9.0 points and a 9.8% reduction in execution steps over learning in the original environments.
    • Through iterative policy-environment co-evolution, performance continued to rise from 47.67 to 54.79 even with 300 environments, going beyond the point where conventional environment scaling falters.
  5. 966Research

    Self-Normalizing Denominators in Rational Causal Estimation

    • It defined a 'self-normalizing' class of causal-estimation denominators whose standardized values are constant across samples, and proved that such denominators cannot have weak-identification first-order limits.
    • It showed that products of powers of nested covariance minors have this property in every dimension and permit exact Wishart pivots.
    • The converse classification is complete in dimension 2, and in dimension 3 it proposed a factor-rank criterion covering instrumental-variable, front-door, and proxy formulas.
    • It proved that Wald inference for front-door adjustment remains asymptotically valid even when the residual variance of the mediating variable vanishes.
    • An audit of real medical data (right-heart catheterization) showed that 'naive diagnostics' and 'denominator-related diagnostics' lead to different conclusions.
  6. 965Computer Vision

    WithEveryone: Unified Planning and Identity Grounding for Group Image Generation

    • It built a unified model that generates group photos of 5–10 reference individuals in one pass. Existing publicly available techniques were limited to a maximum of 4 people.
    • Layout CoT, which plans each person's position, and a location-grounded identity loss (LG-ID) prevent face matching from breaking down as the number of people increases.
    • It surpassed GPT-Image 2 in target-face similarity, 0.499 versus 0.462, and reduced copy-paste artifacts threefold, from 0.169 to 0.055.
    • With 97.3% coverage of requested people and a 2.8% duplication rate, it delivered the most balanced results among the compared systems.
    • It isolated the planning stage, rather than image generation, as the source of remaining errors, indicating the direction for further improvements.
  7. 964Robotics

    What Matters for Latent Actions in Robot Learning

    • It compared 41 latent action model (LAM) design choices under unified conditions across 3 simulation benchmarks + a real Franka robot.
    • A latent-action dimension of 32 was optimal for both single-arm and dual-arm robots, and additional regularization was unnecessary when pretraining regularization was done properly.
    • Fine-tuning the VLM backbone with latent actions beforehand raised success on 4 real-robot tasks from 64.75%→79.25% (+14.5 percentage points).
    • The LAM-tuned model reached 85% success after just 10k steps, surpassing the baseline's performance at 40k steps (76.25%) and demonstrating data efficiency.
    • The simple reconstruction (FDM) metric was a more reliable predictor of latent-action quality than the conventionally used MLP-probe metric.
  8. 963LLMs / NLP

    SWE-bench Science: Can Coding Agents Resolve Engineering Tasks in Science?

    • It assembled 119 tasks from 98 repositories across 20 scientific fields, including chemistry, materials science, and biology, to evaluate scientific-software repair capabilities.
    • Even the best agent achieved only a 47.9% single-attempt success rate, with a large gap between the public-test score of 96.6% and the full success rate.
    • It classified failures into four types: insufficient scientific knowledge or abstraction, superficial repairs, incomplete integration, and failure to generalize to new cases.
    • In one case, adding scientific information reduced GPT-5.6's success rate from 36.3% to 31.9%, showing that knowledge must be validated alongside execution evidence.
    • The number of tasks per field remains small, and analysis of how scientific knowledge is used in actual repairs is still preliminary, so cross-field rankings and generalization require cautious interpretation.
  9. 962Computer Vision

    Swift-Image: Exploring the Performance Frontier of Compact Unified Image Generation Models

    • It built a unified 6B model that handles text-to-image generation, single-image editing, and multi-image editing with one set of weights.
    • When the prompt enhancer (PE) converted requests into visual task instructions, the knowledge-intensive benchmark score jumped from 2.49→4.63. The bottleneck in complex generation was intent interpretation rather than rendering.
    • Structural pruning retained nearly all performance while reducing the model to 3B, and the few-step distilled model (Turbo) actually achieved a higher overall editing score (4.20 vs 4.16).
    • Training required only approximately 243K GPU hours, showing that strong performance is possible without tens of billions of parameters.
    • It summarized practical lessons such as 'evolve the training-data distribution to match model capabilities' and 'train conflicting objectives separately before integrating them.'
  10. 961Computer Vision

    4DAnyone: Create Anyone in 4D from a Casual Monocular Video

    • It built a pipeline that reconstructs 4D humans, moving 3D people, by consistently generating videos from 16 viewpoints using only video from a single camera.
    • Reference-view compression (RCP) and viewpoint-group cycling (TCR) resolved the loss of consistency caused by increasing the number of viewpoints.
    • On the DNA-Rendering and DyMVHumans benchmarks, its 4D-reconstruction PSNR of 24.15/23.28 substantially exceeded the previous best method's 20.55/19.86.
    • It uses only a 3D skeleton as conditioning instead of unreliable depth information, allowing it to work on videos with unknown camera parameters.
    • It built the MVGameHuman training dataset using its own game engine, achieving generalization robust to varied outdoor videos.
  11. 960Agents

    Intelligent AI Delegation

    • The paper distinguishes task decomposition, which divides work into smaller pieces for allocation, from intelligent delegation. Intelligent delegation also transfers authority and responsibility, clarifies roles and boundaries, aligns intent, and establishes mutual trust.
    • The proposed framework has five pillars: dynamic assessment that continuously monitors the other party's current state, adaptive execution that revises plans as circumstances change, structural transparency that makes processes and outcomes inspectable, scalable market coordination that connects many participants at low cost, and system resilience that prevents cascading failures.
    • The central principle is contract-first decomposition, which means first breaking work into verifiable units. Subtasks whose outputs are difficult to verify or overly subjective should be decomposed further until they admit checks such as unit tests or formal proofs.
    • Selecting a delegate requires separating publicly visible reputation from trust criteria defined by the delegator. Authority should be limited to what the task requires and narrowed further under subdelegation, with progress monitoring, completion verification, and an accountability trail recording who delegated what to whom.
    • The paper argues that MCP lacks mechanisms for fine-grained authority delegation, A2A has no field for attaching proof of completion, AP2 lacks conditional settlement, and UCP is designed for commerce and therefore needs extensions for general task delegation. The proposal is conceptual and is not implemented or experimentally validated in the paper.
  12. 959AI / General

    Physics of Agents: Statistical Mechanics Predicts Collective Behavior of AI Agents

    • The study simulated around 10,000 communities of agents across objective mathematics questions and subjective political statements, and released trajectory data at both the individual and group levels.
    • Collective dynamics collapse into three regimes: indifference, polarization, and consensus. Interaction builds conviction and moves communities toward more ordered states.
    • On objective questions, communication improves collective accuracy: initially incorrect majorities switch to the correct answer more often than initially correct majorities switch to an incorrect answer. On subjective questions, systematic ideological drift can emerge, including movement toward the political right.
    • An extension of the Ising model, with intrinsic-field and interaction terms, predicts individual trajectories from initial opinions alone, outperforms all standard baselines, and generalizes to unseen community graphs.
    • The fitted parameters explain the observations: communities operate below the critical temperature, consistent with conviction buildup; attractive ties are stronger than repulsive ties, favoring consensus; and agents holding the correct answer exert the strongest pull, driving truth-seeking.
  13. 958Robotics

    RoboEdit: Turning Human Manipulation Videos into Scalable Robot Experience

    • The key design choice is to edit human videos directly rather than generate them anew, preserving the original scene context and physical consistency.
    • Cross-embodiment adaptation modules allow a single editor to cover the appearance, kinematics, and contact dynamics of seven robot embodiments.
    • A 3D Robot-State Decoder recovers the robot hand state in every frame, providing structured supervision for downstream learning and control.
    • The automatic RoboEdit-ADC pipeline uses depth normalization and physics-based refinement to reduce artifacts such as object interpenetration, floating contacts, and temporal jitter.
    • The authors built RoboEdit-14M, a dataset of 174,000 paired videos comprising 14 million frames across seven robot embodiments, and showed that it supports training real-world manipulation policies.
  14. 957AI / General

    What is Missing from AI Post-Training AI: An Empirical Analysis

    • The paper separates execution-level capability, meaning local refinement within a chosen strategy, from strategy-level capability, meaning revision of high-level decisions as evidence accumulates, and identifies the latter as the bottleneck.
    • The training strategy becomes fixed at the outset, and the entire remaining budget is spent on local adjustments. An agent can spend more than 10 hours iterating efficiently within the wrong strategy.
    • Strategy fixation reflects the agent's prior knowledge rather than the task itself. The same agent chooses similar strategies across different tasks, while different agents choose different strategies for the same task.
    • An experience scaffold combining an experiment log, a skill library, and an evaluator agent improves execution by 12.6 points on GSM8K and 40.8 points on HumanEval, but does not change the selected strategy.
    • Human corrections can redirect the strategy only during a brief window before training begins; once training starts, the agent returns to local adjustment loops. Increasing inference compute by two to eight times yields almost no benefit on the hardest task.
  15. 956Research

    A metric on $S^2 \times S^2$ with positive sectional curvature

    • It proved by construction that the product of two spheres, S²×S², admits a metric with positive sectional curvature everywhere.
    • Starting from the standard metric, it applied a Cheeger deformation to obtain a nonnegatively curved Cheeger-Müter metric, then added a third-order perturbation.
    • Using the structure of the zero-curvature planes remaining at each generic point and the vanishing first variation of the perturbation, it made even the last flat directions positive.
    • It resolved a longstanding existence problem in differential geometry, expanding the concrete examples of positively curved 4-dimensional spaces.
    • Several calculations in the proof rely on MATHEMATICA, and accompanying code is provided, making independent verification of the symbolic calculations and execution of the code important for confirming the result.
  16. 955Software Engineering

    SkillForge: Self-Distilling Agents for Project-Specific Issue Resolution

    • SkillForge takes a proactive approach: it synthesizes issues directly from the repository and accumulates knowledge in advance, without relying on issue history or incurring costly per-issue test-time exploration.
    • The synthetic issues are test-covered bugs that actually execute, so the learning signal is verified through program behavior.
    • The distilled knowledge is organized in a two-tier store of global diagnostic skills and local intervention skills, then used through global initialization and just-in-time injection during issue resolution.
    • On SWE-bench Verified, SkillForge reaches 72.2% Pass@1 with DeepSeek-V3.2, an improvement of 5.8 percentage points, and 60.6% with GPT-5-mini, an improvement of 5.6 percentage points. On SWE-bench Pro, it improves the two models by 5.8 and 4.1 percentage points, respectively.
    • The type distribution of the synthetic issues does not overlap with that of the real evaluation issues. This indicates that the method targets new model-induced vulnerabilities rather than recycling the evaluation set.
  17. 954Machine Learning

    Inference and Uncertainty Quantification for Streaming $r$-PCA

    • It removed a persistent remainder term from Oja's algorithm's error bound, obtaining a sharp operator-norm rate for general rank and sub-Gaussian data.
    • It proved lower bounds matching up to logarithmic factors in both dense-tail and sparse-tail covariance regimes, supporting the optimality of the convergence rate.
    • By linearizing the iterates, it derived a high-dimensional Gaussian approximation for the general-rank error and an explicit limiting covariance.
    • It proved the consistency of an online multiplier bootstrap, enabling estimates and uncertainty intervals to be computed together in a single pass through the data.
    • The current theory is limited to a constant learning rate, leaving extensions to varying learning rates and data with sparsity, Markov dependence, or quantization open.
  18. 953LLMs / NLP

    SPADE: Self-Play in Adaptive Synthetic Executable Environments

    • SPADE makes environment design itself a learnable component. The distribution of environments produced by the designer coevolves with the agent's capabilities.
    • A hint-based regret signal is the key mechanism. A purely adversarial designer proposes unsolvable tasks, while a purely cooperative designer proposes tasks with nothing to teach. The performance gap with and without hints automatically targets environments that are both feasible and worth learning from.
    • Because each environment is written as executable code with a Gym-style interface, one interface can cover both single-turn reasoning problems and multi-turn tool-use tasks.
    • Grounding the designer in documents from a pretraining corpus and maintaining an accumulated memory of previously created environments were both critical to performance.
    • At the 30B scale, SPADE improves over the strongest fixed-environment baseline by an average of 5.3 points across eight mathematics, science, coding, and reasoning benchmarks, by 5.7 points on BFCL v4 multi-turn, and by 13.9 points on ACEBench-Agent.
  19. 952Machine Learning

    Open-MOPD: Diagnosing and Fixing Capability Imbalance in Multi-Teacher On-Policy Distillation

    • Even oracle routing based on ground-truth domain labels, which assigns each prompt to the correct teacher, fails to integrate the capabilities successfully. The bottleneck is capability integration itself, not routing.
    • Teacher disagreement was not the main cause. The problem was a distorted token budget: instruction-following tasks made up 20.3% of input prompts but received only 0.99% of gradient tokens.
    • The study separates the distortion into three causes: differences in response length across domains, different convergence rates across domains, and stale student reward signals caused by rollout reuse.
    • Combining token-share balancing, gap-aware dynamic budget allocation, and student reward refresh raised the improvement recovery rate from 35.6% to 83.4%.
    • The authors released the full training recipe, training trajectories, and evaluation suite, all reproducible on an academic setup with eight A100 80GB GPUs.
  20. 951Computer Vision

    Depth Anything V4: Dynamic 4D Scene Reconstruction via Riemannian Flow Matching on 4D Gaussian Splatting

    • [Withdrawn paper] The latest arXiv version, v2, states: “This paper has been withdrawn by Jiaming Fan”.
    • The official version history dates the withdrawal to 20 August 2026, and the stated reason is “Major errors in research”.
    • Previously described methods and performance figures were pre-withdrawal claims and are not presented as valid research findings or grounds for recommending adoption.
    • Official withdrawal notice and version history: https://arxiv.org/abs/2608.18388 .
  21. 950Research

    Quantum Magic in High Energy Collision

    • To address the dependence of quantum magic on the basis used to measure spin, it computed both QED and electroweak processes in the lab and helicity bases.
    • It found opposite trends: the helicity basis generally produces less magic in the ultrarelativistic regime, whereas the lab basis generally produces less in the nonrelativistic regime.
    • In a quantum-circuit implementation of ultrarelativistic Bhabha scattering, lab-basis circuits at general scattering angles required more T gates, demonstrating the connection between magic and computational cost.
    • In both bases, nature's weak mixing angle lay close to the value minimizing magic, showing that the earlier conclusion was not merely an artifact of basis choice.
    • The analysis is limited to event-by-event fixed bases and selected tree-level lepton-scattering processes, leaving extensions to higher-order corrections, non-Abelian interactions, and the full Standard Model for future work.
  22. 949Robotics

    Embodied-Navigator: Point, Think, Memorize, and Align for Efficient Navigation

    • The Pixel-to-3D Action Formulation lets the VLM operate through familiar 2D visual prompts, avoiding the spatial hallucinations and poor sample efficiency associated with learning 3D geometric transformations.
    • Selective reasoning activates only at critical nodes and compresses the remaining trajectories into space-time indicators, reducing both reasoning latency and memory overflow.
    • Two-level GRPO combines global outcome rewards with fine-grained process rewards to provide dense supervision and teach the model when to trigger reasoning based on the current state.
    • On R2R-CE, it achieved state-of-the-art performance with a 66.2% success rate while requiring only 90,000 training trajectories.
    • When deployed on a physical robot without additional tuning, it achieved a 60.0% success rate, outperforming StreamVLN at 49.0% and DualVLN at 53.0%.
  23. 948AI / General

    Agent Lightning v1.0: Towards Harnessed Agentic RL

    • In harnessed agentic reinforcement learning, the deployment harness, rather than the training engine, owns the environment-interaction loop, while the trainer observes only LLM request-response pairs. OpenClaw, Claude Code, and OpenHands are cited as examples of such harnesses.
    • The paper identifies four main challenges: retokenization, where identical text cannot be merged when token boundaries differ; a variable number of samples per rollout; loss normalization; and backend scheduling.
    • Computing both advantages and normalization at the rollout level rather than the sample level performed best, reaching a verification reward of 38.2%.
    • After the SWE-smith curation pipeline retained just 6,000 examples, reinforcement learning alone improved SWE-bench Verified performance from 41.8% to 56.4%.
    • Safeguards against reward hacking, including blocking Git commands and allowlisting network access, were necessary to stabilize training.
  24. 947Computer Vision

    GenRec: Knowing Where to Reconstruct and Where to Generate

    • Training observed and unobserved pixels under a single loss makes reconstruction, which has one correct target, interfere with generation, which admits multiple plausible outputs.
    • A monocular depth estimator warps source pixels into the target view to construct an observation mask, which then controls the architecture, loss, and gradient flow.
    • A flow-matching backbone jointly reconstructs RGB and scene-coordinate maps, preserving cross-view consistency in a single sampling pass.
    • On two-view interpolation in RealEstate10K, it reached a PSNR of 19.80, well above the strongest baseline at 15.70, and achieved the best result on every metric in Mip-NeRF 360.
    • Inference takes about 11 seconds, more than two orders of magnitude faster than the strongest baseline at roughly 1,200 seconds.
  25. 946Computer Vision

    aDSL: Agentic 3D Creation via Joint Agent-Program Design

    • The work starts from the observation that LLMs often miss precise numerical placements, such as a leg's exact coordinates, but reason well about structure, such as placing four chair legs beneath the corners of a seat.
    • aDSL emphasizes compositionality and spatial-relation operators, so agents do not have to manipulate low-level numerical values directly.
    • Disabling the self-correction loop reduced the visual question-answering score to 61.53 and also lowered the execution success rate, showing that iterative verification is essential.
    • Removing the planning stage increased the average number of revision rounds from 4.25 to 5.58, showing that structural decomposition determines efficiency.
    • An interactive front end, including an OpenClaw demonstration, supports iterative editing, while integration with an external high-fidelity generator preserves structure and improves detail.
  26. 945Machine Learning

    Abra: Scaling Diffusion Image Training

    • A controlled model family spanning 10^19 to 10^22 FLOPs was used to measure the compute-optimal point for text-to-image diffusion models.
    • The optimum is about 200 image tokens per parameter, ten times the Chinchilla rule for language models at 20 tokens per parameter.
    • Diffusion models are robust to overtraining, so when compute remains available, using more data is preferable to increasing model size.
    • Generative-quality metrics such as FID and CLIPScore, representation quality, and optimal CFG settings also vary predictably with compute.
    • The compute-optimal token count rises with resolution, making high-resolution training compute-bound rather than data-bound.
  27. 944Machine Learning

    Recirculation

    • Transformers pass representations forward through their layers, so shallow layers cannot revise their interpretations using context-sensitive representations formed in deeper layers. Recirculation feeds those deeper representations back down, helping the model track state.
    • Without changing the model weights, and with only modest hyperparameter tuning of the adaptive transform, the method reduced perplexity by 23% and increased GSM8K accuracy by 21% across the Gemma 3 family.
    • The GSM8K error rate fell by 8.8% at pass@1 and by 20.9% at pass@128.
    • During generation, the two stacks can run in parallel with almost no additional cost. Prefilling, however, must process each token sequentially and may become expensive for long contexts.
    • The authors present this as an approach that draws clues for structural improvement from the properties of a trained model, rather than imposing an entirely new architecture.
  28. 943AI / General

    ASI-Bench: At the Dawn of Artificial Superintelligence

    • The benchmark evaluates the same task under four levels of guidance: providing the full procedure, naming only the method, allowing autonomous method selection, and adding distracting information. This separates instruction following from autonomous research ability.
    • Across 18 agent and model combinations, the average score fell from 50.91 with the full procedure to 29.10 when only the method was specified and 26.62 when method selection was autonomous.
    • Scores varied substantially by harness, even for the same model. MiMo V2.5 Pro rose from 16.17 with its own harness to 23.25 with the Claude Code harness.
    • Providing only the method name was the most expensive condition, consuming an average of 6.91 million tokens per task. Incomplete guidance increased the cost of exploration.
    • The best-performing combination scored 51.60%, while a lower-cost combination delivered similar performance at one quarter of the cost, revealing a wide spread in cost efficiency.
  29. 942Robotics

    Hydra-0: Action Flow for Generalist World Modeling and Control

    • When robot commands are represented as point trajectories in the image plane and supplied as conditions, the world model can predict the effects of movement without knowing each embodiment's coordinate system.
    • The researchers trained one general-purpose world model by expressing heterogeneous videos in a common format, including human first-person demonstrations, hand-gripper footage, and dual-arm and single-arm robot videos.
    • In forward prediction, the method reduced robot-motion error by 90.4% and object-motion error by 60.2%. Its reproduced success rates for five RoboLab policies had a correlation of 0.96.
    • In reverse mode, the model observes only object motion from a human demonstration and generates a compatible robot movement. A trained action head then converts that movement into executable commands.
    • The study evaluated policies only in an open-loop setting and identified closed-loop evaluation as future work.
  30. 941Machine Learning

    Agentic ESOpt: Fine-Tuning Long-Horizon LLM Agents with Minimal GPU Requirements

    • Using evolution strategies that evaluate random parameter perturbations through forward passes, it adjusted all parameters with only inference-level GPU memory.
    • It avoided the reinforcement-learning credit-assignment problem, which becomes harder as interactions lengthen, by attributing a long trajectory's reward to a single policy perturbation rather than distributing it across tokens.
    • On 15-turn Sudoku, it exceeded a strong GRPO baseline by 12.5%, and on ReAct mathematics and document question answering it improved over GRPO by an average of 8.3%.
    • It fully fine-tuned a Qwen3.5-27B web agent on four H100s, and a heuristic search that changed parameters and prompts together surpassed the baseline in 28 of 36 settings.
    • It requires additional tuning values such as perturbation scale, and the many forward passes become a disadvantage when environment evaluation is expensive; the issue of drifting in irrelevant directions during continual training also remains unclear.

From research to product use

Operating capabilities, pilots, and technologies in development are identified separately.

View technology