Simulated Tasks Reveal a 4.8-Fold Gap in Agent Recovery Costs
A preprint reports that, in the Habitat-Sim simulation, Reflexion and CoPAL had similar success rates but sharply different recovery costs.
Search
2,111 results
1381–1400
Reinforcement learning finds mine-pump designs with fewer simulations
Preprint: In one stochastic mine-pump model, the method reached the highest-reward design after 130 simulations, while Bayesian optimisation outperformed a genetic algorithm on reward summaries.
In one IV simulation, empirical Bayes had the lowest GMM error
Preprint: The method had the lowest error in a baseline simulation and was also applied to schooling data.
AI tutoring answers win more benchmark comparisons after refinement
Preprint: An agent-led process was associated with higher preference rates on held-out Chinese K–12 tutoring tasks, with human checks favoring later answers.
Math preprint presents BBL inequality with separate outputs
Preprint states the bound for −∞ ≤ p ≤ 1/d, with least multiplicative constant C = 1; for m ≥ 2, its power is optimal when C remains 1.
PARTAB reports benchmark gains with less table context
Preprint: PARTAB uses four stages to map a table question to an answer.
AI model rebuilds resting-state brain scans across datasets
Preprint: MnemoDyn showed strong reconstruction and prediction results in tests, but the evaluation did not establish clinical usefulness.
Diffusion models may overfit long before they memorize data
A mathematical preprint reports a best early stopping scale, followed by covariance loss and, at far longer times, collapse of the reverse process onto the training set.
AI System Turns One Image Into Relightable 3D Assets
Preprint: Luce reports leading scores on several 3D-generation benchmarks, while its authors flag limits with tiny details and complex materials.
Preprint links curvature to inequalities in free Langevin dynamics
A mathematical proof study describes a Bakry–Émery framework and derives conditional Poincaré and modified logarithmic Sobolev inequalities.
Sensorless Gripper Reports Lower Damage on Soft TPU Cubes
Preprint tests a controller using encoder position and motor effort, with high grasp success in firmer simulations but a sharp grasp–damage trade-off in the softest cases.
Three Coding-Agent Systems Converge, but a Trust Gap Remains
Preprint: A qualitative review of deepagents, pi and dsh finds five recurring design elements, but none offers an independently verifiable record.
PayPal preprint reports SCOUT's tool-schema context comparison
The report compares 140,200 pre-SCOUT tool-schema tokens with 1,300 production tokens in one implementation and reports an 84.8% first-result hit rate across 45 evaluable benchmark queries.
RAG tests report lower attack success with a new filter
An arXiv preprint reports lower attack success on three question-answering benchmarks, but its guarantee depends on assumptions about the retrieved documents.
WiFi system reports lower errors estimating 3D human pose
Preprint: A physics-informed system reported lower 3D pose errors on benchmarks and in an indoor zero-shot test.
AI model tops tests across four multi-subject learning datasets
Preprint: LT-MKT combines cognitive-load modeling with cross-domain knowledge transfer in an AI learning model.
Three mathematical views of canonical bases are shown to agree
A preprint reports a theoretical result linking the equality to open Richardson varieties and a type-A polytope criterion.
Live-ad system reports 16.1% revenue lift in randomized test
A preprint reports a multi-week experiment on Kuaishou's production traffic covering more than 40 million users, with TAGR outperforming the DLRM baseline on revenue and engagement rates.
Robot-vision model shows more predictable dynamics with physics data
Preprint: Training with paired visual and privileged physical trajectories was associated with lower forecast drift and higher average control success in simulated robot tasks.
AI alignment study finds welfare costs as harm limits tighten
Preprint examines social-choice mechanisms for AI systems affecting many people and shows that tighter welfare floors can carry rising average-welfare costs.