In briefLiteReality-Agent: an agent system that rebuilds indoor rooms as interactive 3D scenes
SummaryLiteReality-Agent reconstructs real indoor rooms from RGB-D scans as articulated, simulation-ready 3D scenes. A coding agent gathers evidence with specialised tools and iteratively edits an executable Python script, Room.py, to produce a digital twin. The authors report better geometric accuracy, visual realism and simulation compatibility than recent models such as Astra and Fable. Source code and the data-capture app are publicly available.
In briefAVSD-Scenes: Audio-visual urban scene description dataset
SummaryThis paper introduces AVSD-Scenes, a paired audio-visual scene description dataset for urban environments, containing 12,291 descriptions built from the TAU Urban Audio-Visual Scenes dataset. Audio and visual descriptions were generated separately and then merged into multimodal descriptions using large language models. Multimodal descriptions improved semantic alignment and cross-modal retrieval over single-modality descriptions, reaching up to 94.5% urban scene classification accuracy and 95.4% when combined with embeddings.
Dhanunjaya Varma Devalraju, Arshdeep Singh, Mark D. Plumbley
In briefUsing an LLM as a verifier to find conflicts in AI model specifications
SummaryThe paper introduces VeriSpec, which audits the natural-language text of model specifications, such as the OpenAI Model Spec, to find principles that prescribe incompatible behavior in the same situation. It extracts 405 rules, clusters them by topic, and uses LLM-as-verifier reasoning to detect inconsistencies. Manual validation confirmed five inconsistencies, which were reported to the developers. Compared with five baselines, VeriSpec identified the most validated inconsistencies, with 38.5% precision and the lowest cost per validated inconsistency at $11.12.
WhyIntroduces direct auditing of model specification text for inconsistencies, validating five defects on one OpenAI specification with baseline comparisons.
In briefBeyond Decodability: Do Acoustic Factors Drive Predictions in Speech-Based Alzheimer's Assessment?
SummaryThis paper examines whether recording factors such as noise and reverberation merely appear in self-supervised speech models or actually change Alzheimer's disease predictions. Using the ADReSSo dataset and three large backbones, the authors apply controlled acoustic interventions to different audio segments. They find that these interventions alter predictions across all models, with noise producing the strongest effects despite no significant diagnostic-group difference. The authors conclude that high performance and non-significant acoustic differences do not ensure robustness, and intervention-based robustness tests should become standard for clinical speech models.
Serli Kopar, Alkis Koudounas, Roshan P. Rane et al.
In briefTemporal-Difference Learning for Dragonchess: Comparing Two Adaptive AI Methods in 3D Chess
SummaryThis paper examines how two adaptive AI methods, evolutionary transfer learning and TD(lambda), perform in Dragonchess, a three-dimensional chess variant. The authors rewrote the game engine from PyGame to C++ to speed up play, running 10,000 games with confidence intervals and significance tests. Both adaptive methods outperformed all other agents in a round-robin tournament, but the evolved and learned evaluations did not differ significantly.
In briefOptimal stochastic bilevel optimization using only first-order gradient queries
SummaryThis paper studies nonconvex, strongly convex bilevel optimization when only stochastic first-order gradient information is available. The authors propose MRT-FD, a single-loop method that tracks the upper-level variable, lower-level solution, and auxiliary response, approximating second-order derivatives with finite differences. For any fixed finite smoothness order p, it finds an epsilon-stationary point with O(epsilon^{-4-2/p}) stochastic gradient queries. A matching lower bound of Omega(epsilon^{-4-2/p}) is also proven, closing the complexity gap in this setting.
WhyProposes a single-loop first-order method and a matching lower bound, closing the complexity gap; strong theory, impact mainly within bilevel optimization.
In briefAccuracy Versus Mechanism Consistency in Time Series World Models
SummaryThe paper formalizes time series world models (TSWMs) and introduces a benchmark that tests whether forecasts respond to changed action plans in the direction declared by known action-state relations. It finds that the lowest-error configurations are at or below chance in mechanism consistency on four of five datasets with declared mechanisms. A directional supervision loss significantly raises consistency on penalized mechanisms without changing MAE.
In briefVarda-single-1.0: A 1 km data-driven weather forecasting system for Switzerland's complex terrain
SummaryThis paper presents Varda-single-1.0, a medium-range data-driven weather forecasting system for the Alpine region, providing hourly 1 km regional forecasts and 31 km global forecasts. It uses two independently trained Graph Transformer models within the Anemoi framework, trained on ERA5, a 20-year regional reanalysis, and operational analyses. Over one year of verification, it is competitive with or improves on MeteoSwiss' operational baselines for most headline scores, but underestimates some local wind maxima and produces overly smooth convective precipitation.
Alberto Pennino, Francesco Zanetta, Michele Cattaneo et al.
In briefCode Handles Simulation, Jev Handles Evaluation
SummaryThe paper tests Jev, a judgment model that outputs option probabilities in one call, as an agent's action-selection layer across reflection tests, matrix games, the ALFWorld text game, and robot control. Jev performs well when the correct choice can be judged from the input, solving 99% of counterintuitive Cognitive Reflection Test questions, but fails when the answer requires predicting information not given in the input. The authors suggest that when code supplies this prediction or simulation, Jev becomes an expert controller through its general evaluation ability.
In briefContinuous process-level evaluation for evolving enterprise AI agent skills
SummaryThis paper presents a continuous evaluation framework that checks both final outputs and process-level behavior of enterprise AI agent skills, applied to Revenue and Productivity variants of a Business Value Determination skill. Across 240 trials, 92.6 percent of runs that passed all applicable final numerical checks still contained another evaluator-detected deviation, showing that outcome-only evaluation can miss behavioral drift. The authors also report that dependency attribution reduced a mean of 6.34 failed checks per run to 2.65 roots, and that reusable template tests provided regression coverage across the evaluated configurations.
In briefPooling Helps, Learned Weighting Hurts In-Context: Decomposing Group Attention
SummaryThis paper studies group attention, a mechanism from the Chronos-2 time series forecasting model that attends across variables in a group at a fixed patch index. The authors edit the attention matrix at inference to separate the value/output pathway, which pools a group summary, from the query/key pathway, which sets the weights. Uniform pooling helps in 18 of 20 sensor-network configurations, while learned Q/K weighting is positive or negligible for multivariate forecasting but degrades 8 of 10 in-context learning configurations.
Michael Fore, James Mason Inder, Mrishika Nair et al.
In briefAsymptotic Closeness of Language Model Alignment Methods with Memory
SummaryThis paper examines two methods for aligning a language model to higher rewards: KL-constrained RL and best-of-n sampling, and how close their output distributions are. Prior work showed asymptotic closeness when output tokens are i.i.d. This paper extends the result to Markovian token sequences, which better reflect practical language models. For finite-length outputs, especially length one, it fully characterizes when the KL divergence between the two methods' distributions is zero.
In briefScientific discovery when validation is the bottleneck: ranking candidate designs from experts of varying fidelity
SummaryComputational tools can propose far more candidate molecules and materials than experiments can test, so a small set must be chosen for costly validation. The paper introduces PRISMS, which selects promising candidates using pairwise rankings from experts rather than absolute scores from data-hungry regression models. When experts differ in fidelity and cost, it escalates queries from lower- to higher-fidelity rankers using a Fisher-information criterion. The authors report 50% top-10 recall in about 42% fewer rounds than regression-only active learning on drug discovery libraries, and about 18.8% higher hypervolume than a Bayesian optimization baseline in library-free design.
In briefToken-based communication for teams of embodied AI agents: a framework explained
SummaryThis paper studies token communication, where tokens act both as compact message carriers and as inference units for generative foundation models, to help multiple physical AI agents cooperate. The authors propose a framework with a task-adaptive protocol built from a compact codebook, syntax rules, and contextual examples, guiding transceivers to compress and reconstruct messages. A collaborative object-transport case study reports substantially lower source payload bit consumption while preserving task efficiency and robustness under noisy channels. The paper also outlines future research directions.
