The editorial team selects four papers from the arXiv submissions of the past 24 to 48 hours that demonstrate how differently AI systems perform today – from independent research to prompt safety to concrete cost figures. Curation was based on substance: each paper provides a comprehensible method and a specific numerical result in the abstract, not just a new topic heading. The selection deliberately covers four different subfields – agent evaluation, safety methodology, cost-effectiveness, and interpretability.
Research agents master the engineering but fail at actual science
A 24-member team led by Peter Kirgis, Sayash Kapoor, and Arvind Narayanan investigates in Can AI agents conduct open-ended AI research? Early evidence from two case studies whether today’s AI agents can independently tackle open research questions. To this end, they had frontier agents with significant computational resources work for six days on unpublished NeurIPS submissions – one of the most influential AI conferences. The authors report that the agents completed the entire technical implementation without human assistance but made no substantial progress toward answering the actual research questions. They identify recurring obstacles, including poor judgment regarding publication standards, unimaginative problem-solving at design dead ends, and an inability to recover from failed approaches. This matters because it tempers the widespread assumption that AI agents could soon conduct science independently – much like the earlier digest finding on the self-improving research agent AREX showed that agent progress so far tends to sit within clearly bounded tasks rather than open-ended research.
A training-free trick protects system prompts from user instructions
Siqi Zeng, Sewoong Lee, Han Zhao, and Julia Hockenmaier present the method V-Steer in Steering Instruction Hierarchies at Inference Time, aiming to restore the so-called instruction hierarchy – the safety principle that system prompts should take precedence over conflicting user or tool inputs, which in practice often collapses. V-Steer requires no additional training: it uses direct logit attribution to identify the attention heads where lower-priority content overrides the actually preferred instructions, then amplifies the cached value vectors at the affected prompt locations. In tests on models ranging from 7 to 70 billion parameters, the method raised accuracy on adherence to the primary directive from about 18 to 92 percent, with minimal additional computational overhead during response generation. This matters because collapsing instruction hierarchies remain one of the most practical gateways for prompt injection – a problem that OpenAI’s internal attack system GPT-Red also addresses on the defense side, though there through elaborate automated red-teaming rather than a lightweight inference-time intervention.
What AI agents actually cost in the office – and what they still lack
A team led by Jingbo Zhou and Hua Wu presents the benchmark OmegaUse-OfficeVal, comprising 100 office tasks that require an average of 2.32 hours of human labor – such as building complex spreadsheets or multi-part documents. Each task carries economic signals, including labor time and a price proxy, allowing direct cost comparisons between humans and AI agents; code and dataset are openly available. According to the authors, tested frontier models were substantially cheaper and faster than human workers but had not yet reached human-level deliverable quality. This matters because companies increasingly judge AI agents by cost metrics, as OpenAI CFO Sarah Friar recently urged for assessing AI investments internally – OmegaUse-OfficeVal provides, for the first time, a task-based, economically grounded measurement basis instead of pure token prices.
Misaligned models show a recognizable personality pattern
Hasibur Rahman and Smit Desai demonstrate in Misalignment Has a Personality: A Big Five Account of Emergent Misalignment that misalignment from fine-tuning on flawed data can be tangibly described using the psychological Big Five model – an established framework spanning the five dimensions of agreeableness, conscientiousness, extraversion, neuroticism, and openness. Through calibrated interventions at three levels, they extracted corresponding personality vectors from the models. The authors report that misaligned training data left a consistent signature across different models and datasets: lower agreeableness and conscientiousness alongside higher extraversion and neuroticism – a pattern that transfers across models and datasets. They explain sycophancy as a combination of high extraversion and low conscientiousness. This matters because it translates a diffuse safety problem into a diagnostic tool measurable along established psychological dimensions – connecting to the earlier digest finding on controllable sycophancy directions, which had already traced sycophancy to internal directions that emerge through alignment tuning.
Bottom line: all four papers are currently unreviewed preprints; their figures come from the respective author teams’ own experiments. The study on open-ended AI research relies on only two case studies in a narrow field of unpublished papers, V-Steer has so far only been tested on controlled benchmarks, OmegaUse-OfficeVal covers 100 tasks from a limited tool landscape, and the personality finding rests on deliberately misaligned training datasets. Whether these patterns hold up in production-grade systems and independent replications remains to be seen.
