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. 1463Agents

    From Delegation to Coordination: Learning Efficient Multi-Agent Deep Research

    • The authors target the problem that training only the main agent or broadcasting the same team advantage to all agents can reward misleading subagent reports and penalize useful ones.
    • LEAD-RL uses a training-time rubric to evaluate the research utility of each subagent report and gates its inherited team advantage.
    • Helpful reports retain only positive team advantages, and harmful reports retain only negative ones.
    • Final-answer correctness remains the only team reward, with no token, step, or speedup reward used.
    • On BrowseComp, LEAD-RL improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit.
    • It reaches 73.5% accuracy on BrowseComp with a 124B-A5.1B model, outperforming the more than twice larger DeepSeek V4 Flash.
  2. 1462RAG / Agents

    RMIT-ADM+S at TREC RAG 2026: Research Agents, Retriever Ablations, and Open-Weight Replication

    • The team submitted seven research-agent runs using proprietary OpenAI and open-weight models to the TREC RAG track, comparing retrieval strategies, writer models, and agent workflows.
    • Pre-submission ablations identified writer selection and adjacent-chunk expansion as the main factors affecting answer quality.
    • In the official evaluation, six of seven runs exceeded the track median across all three rubric/nugget collections, and the best run brief-hybrid tied the track maximum of 0.892 for Batch-1 vital coverage.
    • Hybrid retrieval used 14.4% fewer searches and 8.2% fewer processed tokens than brief-base, providing the clearest efficiency improvement.
    • Human side-by-side preferences differed across evaluation batches: the open-weight owa-glm5 led Batch 1 despite lower automated coverage, while sol-research-first led Batch 2.
    • The estimated generation cost of owa-glm5 was $0.524 per topic, far below $2.823 for sol-research-first.
  3. 1461Research Agents

    ORCHID-SAGE: Fully Reproducible CALPHAD Assessment from Evidence to Phase Diagram

    • The authors propose the ORCHID-SAGE standard, targeting the problem that expert judgments in CALPHAD assessments are rarely recorded, so no one can regenerate or update the databases.
    • The standard organizes the record in three levels: a general protocol, a system-specific protocol with its decisions, and run cards identifying inputs, code, and outputs through hashes.
    • In the Cr–Ni assessment, AI agents carried out most steps from the literature search and evidence verification to the one-shot calculations, model optimization, and protocol assembly.
    • Decisions requiring judgment were reserved to expert humans, and the decision log documents each with its date, responsible person, and justification.
    • A script on a second machine and an AI agent working only through Model Context Protocol tools each regenerated the database from its record alone, passing all 13 checks.
    • The AI agent regeneration required 46 turns and 2 h 38 min, at a cost of US$1.64 in model usage.
  4. 1460Reasoning / RL

    Reward-Driven Collider Physics Challenges for Self-Improving Agentic AI via Reinforcement Learning

    • The authors formalize six collider physics tasks as open-ended optimization environments with continuous, automatically evaluated rewards.
    • On well-established tasks, all evaluated models converged on human-level performance using approaches like boosted trees and OmniFold.
    • On the newly designed diphoton resonance search task, final rewards ranged widely from 0.487 to 0.782, with only one model exceeding the reference program score of 0.689.
    • Over 20 generations of genetic evolution, the 27B model's mean reward on the twelve hidden datasets rose to 0.626 from 0.382 at generation 0.
    • After 30 GRPO reinforcement learning steps, the 27B model's single-shot mean reward rose from a base of 0.116 to 0.457.
    • Reward-based selection eliminated template use of the development background from generation 2 and raised parametric background fitting to 1.00, inducing physically sound strategies without human guidance.
  5. 1459Research Agents

    Science Prediction Universe

    • The authors release the Science Prediction Universe benchmark, which evaluates agents on hidden experimental outcomes across chemistry, materials science, biology, and medicine.
    • The 100 public tasks come from 85 recent papers and the 16 private tasks from unpublished wet-lab experiments, with measured values held only by an isolated grader.
    • Given only the experimental definition, the agent ranks candidates by the measured outcome and may compute and search freely for eight hours.
    • On the public benchmark, configurations separate into two distinct tiers, with task-adjusted Spearman correlations ranging from 0.227 for the lowest-performing configuration to 0.478 for the best.
    • More computation did not yield detectably better predictions, and score showed no detectable correlation with run time, tool calls, or GPU use.
    • The bottleneck lies in scientific judgment rather than tool execution: agents agree with one another at 0.52 but with experimental reality at only 0.33.
  6. 1458Research Agents

    RSI-PDEbench: A Benchmark for Self-Improving Agents in PDE Solving

    • The authors release RSI-PDEbench with 1,001 cases across 42 equation templates for the task of identifying unknown PDEs and coefficients from sparse observations and building numerical solvers.
    • The three evaluation settings are noise-free Full, Noisy with Gaussian noise at 50% of the clean-signal RMS, and Hard-OOD with unseen templates and shifted coefficient ranges.
    • The RSI-PDE harness shortlists five candidate equations by residual loss and uses public relative L2 feedback to refine equation choice, coefficients, and numerical solvers.
    • Only proposals passing the validation gate update the solver, instruction policy, and memory, while the foundation model and executable supervisor stay fixed.
    • Across the four models, the mean pass rate improved by 106.5% to 181.9% relative to single-run baselines and by 3.0% to 19.6% relative to native harnesses across the three settings.
    • Total recorded model calls fell by 46.9% across the 12 model-setting pairs.
  7. 1457Software Engineering

    Mach2: Scientific Discovery Platform with Deterministic Execution Constraints

    • The authors propose the Mach2 platform, which converts natural-language rules into deterministic, lint-like executable predicates, targeting the problem that prompting alone cannot guarantee compliance with operating rules.
    • The Eigen orchestrator decomposes user objectives into bounded tasks, generates execution workflows, commissions specialist agents, and integrates results.
    • Workers propose modifications on logically isolated branches, and the accepted project state is updated only after mandatory rules and integration tests pass.
    • CodeGrapher indexes source-code entities and relationships through abstract syntax trees, supplying structured evidence and executable actions rather than relying exclusively on model-generated descriptions.
    • Mach2 generates task-specific agent packages for independent execution at external endpoints and validates returned artifacts against declared interfaces and evidence requirements.
    • Assertions without recorded sources are flagged as potentially dependent on model-internal knowledge, keeping knowledge provenance traceable.
  8. 1456AI Agents

    Towards an Autonomous AI Steward for Large-Scale Physics Experiments: Integrating Causal Diagnostics, RAG, and Conformal Prediction

    • The authors propose the Helmholtz Co-Pilot, a hybrid Multi-Agent System, to address monitoring of machine data streams and preservation of tacit expert knowledge in long-term physics experiments.
    • An asynchronous causal diagnostics engine continuously monitors hardware streams, performs dynamic causal discovery to trace cascading alarms back to their root causes, and hands a structured diagnostic report to the MAS upon detecting an anomaly.
    • The institutional memory pillar retrieves wikis, logs, and databases through a RAG backbone, and proactively triggers an interview mode upon detecting knowledge gaps to write operator insights back to the vector stores.
    • Critical actions are intercepted by a stateful Human-in-the-Loop guardrail, which serializes the agent state and pauses execution until the shift operator approves the action.
    • Leveraging conformal prediction, the MAS presents operators with evidence-based recommendations bounded by mathematical coverage guarantees that quantify the probability of a specific root cause.
  9. 1455Machine Learning

    COMBIO: Automated Discovery of a Compact Composite Biomarker from Routine Laboratory Panels

    • The authors propose COMBIO, a pipeline that distils laboratory-panel machine-learning models into a compact closed-form biomarker, targeting the difficulty of interpreting and operationalizing such models.
    • Shapley additive explanation values from a gradient-boosted reference model rank five candidate measurements and guide eligible exponent signs.
    • An exhaustive search over small integer exponents from minus two to two, with two to four nonzero terms, selects the formula on an inner validation set before evaluation on the reserved test set.
    • Across the six test sets, COMBIO's AUROC differed by at most 0.03 from logistic regression using the five candidate laboratory measurements.
    • The hospital mortality composite achieved an AUROC of 0.698 in the external eICU database, and a version restricted to widely available measurements achieved an AUROC of 0.786 in a cohort covering 92% of admissions.
    • The authors state that the multiplicative representation provides compactness rather than additional accuracy, and local validation and recalibration are required before clinical use.
  10. 1454AI Agents

    KnowEvo: Knowledge Evolution for Protein Optimization

    • The authors propose KnowEvo, which treats externalized knowledge itself as the optimization target, addressing the problem that acquired expertise stays embedded in model weights or scattered across iterative problem-solving processes.
    • Knowledge is represented as executable expertise blocks whose refinements are organized into strategy mutation trees, aligning the unit of execution with the unit of evidence.
    • The forward phase adaptively distributes evaluation budgets across strategy variants using a Thompson-style posterior allocation strategy.
    • In the backward phase, a code agent operates in an isolated sandbox to translate rollout evidence into localized program updates, and validated revisions are appended as child nodes.
    • On the Rosetta stability task, KnowEvo improved pass@1 from 42.4% to 50.5%, and on the SaProt solubility task it improved pass@1 from 31.6% to 51.5%.
    • It outperformed LLM-guided evolution baselines such as GEPA and AdaEvolve across all six metrics on both tasks.
  11. 1453Software Engineering

    Software Refactoring Research with Large Language Models: A Systematic Literature Review

    • The authors reviewed 50 primary studies that use generative LLMs for refactoring.
    • Java was the most-used language, and datasets were collected from open-source projects on GitHub, Apache, and F-Droid.
    • One-Shot, Few-Shot, context-specific, and Chain-of-Thought prompting were most effective depending on the language; UTRefactor with Chain-of-Thought reduced test smells from 2,375 to 265, an 89% elimination.
    • The definition of accuracy varies significantly across the literature, calling for a standardized measurement.
    • The most often mentioned code smells were Large Class and Long Method, and the most often applied refactoring type was Extract Method.
  12. 1452AI Agents

    Agent Benchmarks Stop Ranking Before They Run Out of Tasks

    • On SWE-bench Verified, strengthened tests rejected 17.7% of the median agent’s passing patches, an error larger than the gaps between neighbouring agents, so the order of neighbours is close to chance.
    • ALE-Bench’s top 10 split into 3 tiers by contest score but became 1 tier when scored pass/fail against the median human, where 5 models pass 39 of the 40 problems, and 47.8% of the runs beat that reference.
    • In a Terminal-Bench re-run, models tied on pass/fail differed widely on the actual quantity: on portfolio-optimization, opus’s median speedup was 10.9× while the others’ was 1.3×.
    • On fixed tasks, a higher bar only moved the problem: the contest-winner bar scored 89% of the earlier models zero on every task.
    • The authors proposed a clipped log-ratio score that maps the baseline to 0 and the reference to 1 with a clip at three reference gaps, and reported cost as a separate axis.
  13. 1451Research Agents

    Sage-Taiyi: A Metacognitive Architecture for Scientific Discovery

    • Sage-Taiyi uses a two-level structure that connects object-level Target Tree, Idea Graph, and Evidence State with meta-level Cognitive State and Cognitive Repertoire through monitoring and control.
    • On Taiyi-Bench, Sage-Taiyi ranked first on all eight metrics of historical rediscovery and open-ended ideation, with an Overall score of 54.0 against 47.9 for the strongest baseline, AI Scientist-v2.
    • In chemistry, reaction-condition optimization raised the yield of cobalt-catalyzed carbonylation from trace amounts to 84%.
    • Photocatalytic allylic amination yield rose from 13% to 69%.
    • The system keeps revising its understanding of evidence, assumptions, and unresolved questions, and uses these assessments to adapt research targets and select reusable reasoning methods.
  14. 1450Machine Learning

    HyPLM Technical Report

    • HyPLM-Flash is a 650M-parameter masked language model trained on 726 billion tokens, 0.73 times the data of ESM-2 650M and a fifth to an eighth of that of E1-600M and ESM C 600M.
    • It reached a ProteinGym zero-shot Spearman correlation of 0.409, statistically tied with ESM-2 650M, ESM C 600M, and E1-600M, and scored highest on organismal fitness among self-supervised single-sequence models.
    • Its long-range contact precision at L/5 on ProteinNet was 0.650 against 0.392 for ESM-2 650M, the best of all evaluated models including ESM-2 15B.
    • With fused kernels and INT8 matrix multiplication, the inference engine embeds about 130,000 residues per second on a single H20 GPU, 1.9 to 2.7 times faster than other 600M-class encoders.
    • The model was trained by Hyra 2.0, the research agent developed by the team, and the weights and inference engine were released.
  15. 1449Research

    Building heterogeneous multistep workflows in automated microscopy: reward design and knowledge transfer for autonomous structure property discovery

    • The framework treats not only where to measure but also the choice between imaging and spectroscopy, the image-window size, and the scan speed as controller actions under a budget constraint.
    • Pre-experiment knowledge, including predictor accuracy and computational cost, empirical learner-maturity functions, and drift characteristics, is transferred into the campaign while kept separate from evaluation references.
    • Two campaigns were run on a real electron energy-loss spectroscopy dataset, showing that response discovery and structure–property learning favor different surrogate models and acquisition strategies.
    • In the primary comparisons, the mean rank association between uncertainty reduction and error improvement was −0.009 for RF, −0.527 for DCNN, and 0.037 for DKL, and the fraction of positive uncertainty-gain actions that worsened error was approximately 50%, 70%, and 48%, respectively.
    • In the drift campaign, explicit registration greatly improved positional accuracy, and the framework measured what this costs in spectroscopy throughput.
  16. 1448Computer Vision

    Breaking survivorship bias in driving world models

    • At only 0.5 s before collision, Vista and GEM imagined the collision in 9% and 12% of the scenes, respectively, whereas the NWM reached 86%.
    • On routine MM-AU clips, the NWM imagined a collision in only 8% of the futures, showing it did not simply associate accident-footage appearance with collisions.
    • On 200 challenging NAVSIM scenes, NWM-guided selection achieved a PDMS of 0.783 against 0.637 for Vista-guided selection, approaching the oracle 0.959.
    • In real-vehicle pedestrian-dummy trials, the NWM chose braking correctly at all decision frames in the Response stage while Vista’s accuracy fell to 30%.
    • The dominant failure modes of both baselines were rewriting the future to avoid the collision, such as making the agent vanish or freezing the scene.
  17. 1447Research

    Amortized Bayesian Experimental Design with In-Context Knowledge Conditioning

    • IMBUE accepts two forms of external knowledge without retraining, user-specified priors and observations from previous instances, encoded as additional input tokens.
    • A learned reliability filter scores each external token against the current observations and excludes tokens with a threshold of 0.5 or below from computation.
    • A naive-concatenation baseline that receives previous instances only at deployment falls below the original policy in both regimes.
    • On CES with K = 5 relevant instances, IMBUE reached an sPCE of 13.93, and with irrelevant instances it stayed near its K = 0 score of 13.46 at 13.14.
    • On Location Finding with a strong prior, IMBUE attained ALINE’s final sPCE several steps earlier in the rollout.
  18. 1446Software Engineering

    Deliberate Pacing for Human Oversight in Agentic Software Development

    • The authors distinguished speed from pace, defining deliberate pacing as intentionally adjusting when and how autonomous development proceeds to preserve opportunities for meaningful oversight.
    • The review of 20 sources found five recurring themes for slowing down: software quality, long-term sustainability, care and ethics, resistance to acceleration, and developing at an appropriate pace.
    • Three factors motivate the vision: changing relative costs of human involvement, the timing of oversight opportunities, and interdependencies between pacing decisions across the SDLC.
    • Retrospective review may not provide the same opportunities to influence autonomous development as intervention before consequential decisions are implemented, so intervention timing determines oversight quality.
    • The authors outlined three research questions and plans to study pacing through developer interviews, quasi-experiments, and simulation-based methods.
  19. 1445LLMs / NLP

    AlienBench: A Benchmark of Structural Novelty in Language Model Outputs

    • AlienBench adapts the structured imagination task to language models, labeling ten biological features as Earth-typical (0) or a departure (1) and summing them into a 0-to-10 score.
    • A panel of three LLM judges agreed closely with Krippendorff’s α = 0.913, and each model generated 180 descriptions, 30 under each of six prompt conditions.
    • Overall scores across ten models ranged from 3.28 (GPT-6 Luna) to 6.55 (Ministral 3 14B 2512).
    • Open-weight models averaged 5.40, higher than the proprietary average of 4.58, and the top-scoring model was open-weight.
    • Under the baseline prompt, all models departed from Earth defaults more often than 100 human participants on every feature.
  20. 1444Research Agents

    Scientist World: Grow a Research World for AI-Scientists

    • A multi-agent pipeline with no human in the loop converts papers into self-contained environments, and the team released the first 1,000 verified environments where each reproduced method serves as the scoring baseline.
    • An AI-Scientist is scored by Frontier-Normalized Improvement (FNI) over the paper’s reproduced method, and frontier AI-Scientists surpassed the human researchers 74% of the time.
    • The environments preserve 6,400 hours of research traces, recording every idea, experiment, failure, and success.
    • Training a small model with only real experimental outcomes as rewards raised its research ideation by 158%, matching the frontier cohort on the median held-out environment.
    • The 1,000 papers were selected from 124,816 indexed papers by quality and compute feasibility, and the world keeps absorbing new literature every day.
  21. 1443Research Agents

    Contrastive Discovery: Open-Ended Scientific Discovery over Competing Explanations

    • The authors propose a contrast set, a bundle of mutually exclusive and collectively exhaustive competing explanations for one scientific question, as the unit of discovery.
    • 89% of the top pointwise candidates were rejections, and re-evaluating the same evidence with contrast sets resolved 73% of these candidates to a specific alternative.
    • Guiding search with contrast sets resolved 67% more questions than pointwise surprisal under the same experiment budget, averaging 161.9 versus 97.2 at K=200.
    • In a blinded study, 38 human experts chose the same explanation as the system on 78% of resolved questions.
    • Most of the gain came from corrective resolutions, where evidence shifted support away from the prior-leading explanation to another one.
    • The authors note the experiments use tabular, observational datasets and a gpt-4o backbone, and on such data a resolution is a statistical rather than a causal finding.
  22. 1442Machine Learning

    Vendi Anomaly Scores for Efficient and Accurate Anomaly Detection

    • The authors propose the Vendi Anomaly Score, which defines anomalies by how much removing a sample changes dataset diversity measured by the Vendi score.
    • They derive a first-order approximation to the exact leave-one-out score so the method scales linearly with dataset size.
    • The method requires no density estimation, partitioning heuristics, or hyperparameter tuning, adapting q automatically to the spectral structure of the data.
    • It achieves state-of-the-art anomaly detection performance across 148 benchmark datasets derived from 10 source datasets and 4 embedding architectures.
    • It remains robust in ImageNet experiments with up to 100k samples as both contamination rate and dataset scale increase.
  23. 1441Machine Learning

    Large Language Models as a Universal Searcher for Open-World Automated Machine Learning

    • The authors encode model architectures, training hyperparameters, and data ratios into unified discrete numerical sequences called NCode, jointly optimizing all three design spaces.
    • The LLM searcher is reframed as a ranking task that picks superior candidates from a pool, performing open-world optimization without task-specific fine-tuning.
    • Trained solely on trajectories from the single vision domain of NAS-Bench-301, it achieves state-of-the-art performance across image classification, image segmentation, speaker verification and text-to-image benchmarks.
    • It scaled to the large-scale training and joint optimization of a billion-parameter Diffusion Transformer, requiring 896 GPU-days and attaining a state-of-the-art GenEval score of 0.91.
    • GRPO-trained Chain-of-Thought reasoning revealed an optimization strategy shifting from early-stage diversity-driven exploration to late-stage anchor exploitation.
  24. 1440AI Agents

    PARIS: A multi-agent LLM framework for end-to-end, PRISMA-guided systematic literature reviews with pre-synthesis corpus quality assessment

    • The authors built PARIS, a framework where a conductor agent running in Claude Code orchestrates 15 specialized large language model (LLM) agents.
    • All agents communicate exclusively through persisted artifacts, and the conductor validates each file before the next step.
    • Three human checkpoints keep study inclusion, corpus adequacy and the synthesis outline under human authority.
    • QualityScore grades the corpus from 0 to 100 before synthesis, and corpus improvements raised it from 74.6 to 81.9 in the first case study.
    • The second case study included 123 of 131 records and recorded a QualityScore of 69.1.
    • The authors' audits exposed failure modes such as incomplete retrieval, counts drifting between agents and unrecorded checkpoint decisions, and PARIS fully supports 8 of the 27 PRISMA 2020 items.
  25. 1439Machine Learning

    HyInv Technical Report

    • The authors trained HyInv, a 1.96-million-parameter backbone-conditioned graph network, with random-order autoregressive pretraining followed by structure-reward reinforcement learning.
    • A sampler that drafts at a higher temperature and rewrites at a lower temperature improved refolding TM-score at matched diversity.
    • An optimized implementation with BF16 numerics designs 128 sequences for a 402-residue backbone in 50.6 ms on a single NVIDIA H20 GPU, about 52 times faster than the reference sampler.
    • Across five benchmark datasets and seven dataset-predictor combinations, HyInv reaches a higher average scTM than all ten external baselines at comparable or higher sequence diversity.
    • The average margin is +0.013 over ProteinMPNN and +0.008 over ProteinZero V2.
    • The same default setting supports de novo backbones, pocket-constrained libraries, enzyme redesign with the catalytic site kept, and loop redesign.
  26. 1438Reasoning / Efficiency

    Late but Valuable: Why Deadlines Change the Measured Value of LLM Reasoning

    • The authors showed that when a deadline makes the system execute a fallback, reporting answer quality and latency separately overstates the delivered value of reasoning.
    • In a replay of 1,200 decisions, thinking answers that missed a deadline were worth about five times as much as on-time ones.
    • Separate summaries predicted about twice the value that thinking delivered and overstated thinking's advantage over fast mode by 0.0241 per decision.
    • Raising the number of simultaneous requests from 8 to 56 on a single-GPU Qwen3-8B reduced the share of thinking answers meeting a 45 s deadline from 79.2% to 1.4%.
    • The authors recommend pairing each answer's value with its arrival time and scoring the action that actually runs.
  27. 1437Distributed Systems

    Advancements in Cloud-Native Systems: A Systematic Mapping Study of Reliability and Scalability across Cloud-Native, Edge-Native and AI-Enabled Systems

    • The authors combined 5 OpenAlex searches with 189 records from an earlier exploratory search, screening 4,611 records in total.
    • The review includes 1,794 primary studies from 2017–2026 and 95 secondary studies.
    • In the 13-category taxonomy, each study was graded as simulation, prototype or testbed, production data, production deployment, or not determinable from the abstract.
    • Of 1,003 studies with a determinable level, 66.2% used prototypes or testbeds, 18.0% simulations and 15.8% production data or deployments.
    • A 93-study synthesis core shows production deployments of cascading-failure prevention, trace- and graph-based root-cause analysis and predictive autoscaling, and production-trace evaluations of cold-start mitigation.
    • Empirical studies in the synthesis show that agent frameworks remain fragile.
  28. 1436Distributed Systems

    DICOMPUTE: An Economic and Privacy-Aware Framework for Distributed Artificial Intelligence Inference

    • The authors present DICOMPUTE, a research framework for distributed inference that allocates requests across privately controlled, regional, and independently operated computing resources under policy-constrained scheduling.
    • The full-cost model accounts for capital amortization, energy, network transfer, orchestration, utilization, failures, and trust-related expenses.
    • An illustrative scenario with hypothetical parameters yields a break-even volume of 5.625 million requests per month.
    • The privacy model separates data residency, transport confidentiality, execution confidentiality, and verifiable accounting.
    • The authors propose falsifiable hypotheses and a reproducible experimental design, without presenting simulated examples as measured performance.
    • They conclude conditionally that centralized clusters retain structural advantages for tightly coupled, memory-intensive, high-throughput inference.
  29. 1435Research Agents

    Scientific Liquidity Agents: An Agent-Based Model of Knowledge Arbitrage in Scientific Communities

    • The authors introduce SLA, an agent-based model that places domain specialists and epistemic arbitrageurs in a shared environment.
    • Researcher decisions use a belief-desire-intention architecture, while a prerequisite-based problem graph, a dynamic interaction network and a publication registry form the shared state.
    • Simulations with 10 and 18 agents registered 68 and 274 papers respectively.
    • In the smaller run, arbitrageurs took part in 16 of the 26 contact pairs that join the community into one connected component, while topical similarity alone splits it into seven.
    • Paper production is abundant, but no recorded replication succeeds, which places the bottleneck in verification rather than production.
    • Solved-problem sets correspond to the prerequisite reachability of the disciplines represented in each community.
  30. 1434AI Agents

    Advances in Agentic Computer Graphics

    • The authors introduce a unified taxonomy that characterizes agentic graphics systems along four dimensions: specification, base model, harness, and graphics interface and tooling.
    • Their collected corpus includes more than 300 papers published through September 2026.
    • They develop an agentic literature survey system that searches for papers, reads their content, and annotates their designs using the taxonomy.
    • The system records taxonomy labels, design insights, and limitations alongside supporting source evidence, supporting comparison across papers and analysis of broader trends.
    • The report discusses open challenges in coordinating agents and tools, designing human-agent collaboration interfaces, sustaining long-horizon planning and refinement, and reliably evaluating interactive graphics systems.

From research to product use

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

View technology