Today’s selection curates four arXiv preprints from the last one to two days based on substance and thematic diversity: all four provide a comprehensible method and robust numbers in the abstract, but cover different areas – the security of image generators, the measurement of agent autonomy, the practical usefulness of forecasting agents, and the technical merging of language models. The selection criterion was that each paper empirically answers a control question about today’s AI systems, rather than just demonstrating a new application.
Image generators: jailbreak framework MIND almost completely bypasses security filters
Dongdong Yang, Deyue Zhang, Zhao Liu, and a six-member team investigate in “Dynamic Defense Profiling Enables Cognitive Jailbreak of Text-to-Image Models” how robust the security filters of current image generators really are against targeted attacks. Their framework MIND treats the search for a working attack prompt as a state-estimation problem (a method that infers a hidden internal state from individual observations) over a model’s not directly visible defense mechanisms: a “Multi-modal Judge” breaks the model’s response down into fine-grained feedback instead of a plain success-or-failure signal, a “Defense Profiler” continuously updates a guess about the active filter strategy from that feedback, and a “Meta-Memory” module draws on historically effective attack patterns. Against Stable Diffusion v1.5, MIND reached a 95.62 percent success rate across six common defense settings; among four tested commercial systems, the highest rate was 91.58 percent against Wan-2.5. This matters because it shows that security filters which judge individual requests in isolation hold up poorly against a learning, adapting attacker – a pattern earlier digest issues already found in text models and which here transfers to image generators.
How autonomous are AI agents really? A new measurement scale
Samuel Presgraves proposes in “The Autonomous Agency Scale: A Behavioral Framework for Measuring Self-Directed Behavior in AI Systems” to make a previously fuzzy concept measurable: existing frameworks, the author argues, capture cognitive ability, task automation, or catastrophic risk, but not autonomy in the narrower sense. The Autonomous Agency Scale (AAS) rates systems across seven dimensions – from cognitive autonomy through temporal persistence to goal formation – separately for active usage phases and for idle phases without user interaction. Its centerpiece is the “Idle-Gap Test”: all external triggers are removed to check whether a system still stays active on its own. Across six examined systems, task agents such as Claude Code, Manus, and Hermes reached active composite scores of 2.3 to 2.4 and idle scores of 0.6 to 1.9 – but their idle activity turned out to be fully traceable to user-configured schedules. Only the companion architecture Airi, Presgraves reports, retained part of its activity even after every trigger was removed. This is relevant because it separates genuine self-direction from merely well-disguised automation – a distinction that carries growing weight for judging AI autonomy in practice.
Football forecasting as a practical test for deep-research agents
Zhaokai Wang, Tianlin Gui, Jiayuan Rao, and colleagues present “WorldCupArena: Fine-Grained Evaluation of Language Models and Deep-Research Agents on Football Forecasting”, a practical test that evaluates language models and research agents before kickoff rather than after the fact. Models either receive a shared evidence package or research on their own, then predict the result, exact score, likely players, and match statistics; the predictions are later compared against what actually happened. Across 104 matches and 13 systems, the authors report that models with similar result accuracy diverge much more clearly on detailed predictions; compared with betting-market and human-fan baselines, the best system showed only small gains in result and exact-score accuracy, but a clearer edge on the graded “Scoreline” measure, which gives partial credit for near-miss scores. This matters because it offers an honest read on where deep-research agents actually gain an edge in a practical, continuously refreshing task – and where they do not.
Merging language models of different sizes – with a seesaw effect
Jiahe Fan, Yinghao Hou, Si Chen, and colleagues ask in “Rethinking Heterogeneous LLM Merging: A Weighted Model Averaging Perspective” whether language models with substantially different parameter spaces can be merged by direct weighted averaging alone – without further training or semantic alignment. The authors test two strategies on Qwen-family model pairs: “union-style” merging expands the smaller model into the larger model’s parameter space, while “intersection-style” merging truncates the larger model down to the smaller one’s space, each followed by training-free dimensional adaptation and a controlled interpolation ratio. Small interpolation ratios measurably improved strong source checkpoints, while near-balanced mixing ratios, the authors report, often collapsed; across several capability areas – from mathematical reasoning to instruction following – a seesaw effect showed up consistently, with gains in one capability coinciding with regressions in another. This matters because it shows that simple weighted averaging remains, against some expectations, a surprisingly strong baseline for heterogeneous model merging, without resolving the seesaw effect itself.
All four works are unreviewed arXiv preprints; their numbers come from the respective author teams’ own experiments. Whether MIND’s jailbreak success rates hold up against updated security filters, whether the Autonomous Agency Scale stays discriminating for future generations of agents, and whether the seesaw effect in model merging can be softened by other interpolation methods will only be shown by independent replications.
