Today’s selection from arXiv’s new submissions of the past 24 to 48 hours is deliberately broad: instead of one common thread, four papers stand side by side, each measuring a different weak point of today’s AI systems - spatial understanding in embodied agents, how reasoning models allocate compute budget, the multilingual reliability of agent benchmarks, and the safety cost of steering vectors. The selection was curated for substance and spread: each paper offers a traceable method with solid numbers in its abstract, and no two papers cover the same subfield.
Spatial cognition: AI agents far behind human level
A team led by Gengyang Xu, Dongwei Xiao, Yiteng Peng, and Shuai Wang introduces MetaSpace, a test framework that checks whether embodied AI agents (systems that interact with their environment through a physical or simulated body) genuinely possess a robust understanding of space, rather than solving tasks only on the surface. Modeled on the principle of metamorphic testing from software engineering, MetaSpace automatically generates test cases from real execution trajectories and encodes physical and logical rules as executable Prolog rules; violations of these rules indicate failures in spatial cognition. Across three embodied scenarios, the method reportedly detects 90,422 such spatial-cognition errors in current state-of-the-art multimodal agents, which average just 0.44 to 0.52 out of 1.0 possible points on the newly introduced Spatial Cognition score - compared with 0.96 for the human benchmark. That matters because previous evaluations have mostly measured only task success, which tends to obscure exactly the kind of hidden weaknesses that could cause safety problems in the real world.
Compute budget: reasoning models don’t allocate thinking time strategically
Chenrui Fan, Yize Cheng, Ming Li, Yongyuan Liang, Tianyi Zhou, and Soheil Feizi show in Thinking Hard, Not Smart that reasoning models (language models that run a longer internal thinking process before producing an answer) handle a shared compute budget across multiple tasks poorly. The authors construct an exam-style evaluation format in which a model must distribute a fixed token budget across several questions of varying difficulty and point value on its own, in order to maximize its total score. Across several open and frontier models, the models reportedly behaved like greedy, sequential solvers: they prioritized questions by their order of presentation, front-loaded effort on early questions, and remained largely insensitive to point value and difficulty - a pattern that grew more pronounced as the number of questions increased and carried over from math to coding tasks. That matters for real-world deployment, because practical applications rarely involve a single isolated request; instead, models often have to handle multiple problems at once under a shared time or cost limit.
Language gap: agent capability collapses outside English
A team led by Andrea Caciolai and colleagues extends OmnilingualGAIA2, the well-known GAIA2 agent benchmark, through machine translation (with partial human validation) into ten target languages spanning five writing systems, complemented by a localized, human-calibrated verifier. Measured across seven frontier and open-weight agents, the authors find a persistent cross-lingual gap of 8.8 to 18.4 pass@3 percentage points that varies in magnitude by agent, concentrates mainly on tool orchestration rather than quantitative reasoning, and does not close with increasing model scale. An error analysis attributes the gap to the model itself in 55 percent of cases, according to the authors, while translation artifacts played a role in only about 6.4 percent of language-scenario pairs; human linguistic experts identify lost morphological cues and amplified ambiguity as the main failure sources in non-Latin scripts. That matters because agent systems deployed worldwide are still tested almost exclusively in English, even though their user base is global and multilingual.
Safety: steering vectors can be separated from their side effects
Yuxiao Li and Gjergji Kasneci show in Safety Cost of Steering Vectors Is Separable and Reducible that so-called steering vectors - a lightweight tool that controls model behavior through targeted interventions on internal activations, without retraining the model - can unintentionally weaken a language model’s built-in safety mechanisms and increase its willingness to comply with harmful requests. The authors show that this safety degradation stems from a separable component within the steering vector that contributes little to the actual steering objective, and they frame removing this component as a constrained optimization problem solved via primal-dual updates. Across multiple models, steering behaviors, and attack suites - including previously unseen attack types - the method reportedly reduces steering-induced safety degradation substantially, with minimal cost in false refusals and without losing the intended steering effect. That matters because steering vectors are gaining popularity as a cheap alternative to costly retraining, while their safety risks have so far received little systematic attention.
All four papers are unreviewed preprints - the reported figures come from each author team’s own experiments and have not yet been independently replicated. The compute-budget study has so far been tested only on math and coding tasks, and the steering-vector correction has only been evaluated on the attack suites chosen within the paper itself. Whether these effects hold up across further models, languages, and independent replications remains to be seen.
