Today’s selection from the arXiv submissions of the past days connects four questions that rarely come up together: how reliable agent networks really are under the hood, how much pure execution details can distort measured agent performance, what a new scientific AI agent model from a major lab achieves, and how multiple models can be merged with less capability loss. The curation criterion was substance and diversity: each paper offers a traceable method with solid numbers in the abstract, and no two works cover the same subfield. Since arXiv does not publish new listings over the weekend, this selection draws on the most recent available issue from Friday.
Reliability certificates for agents: the independence assumption does not hold
Varun Pratap Bhardwaj, Garima Singh, and Arun Pratap Bhardwaj examine an assumption rarely questioned in practice in Agent Behavioral Contracts II: reliability calculations for multi-stage agent systems typically multiply the individual reliabilities of the components, which presumes their failures occur statistically independently of each other. In a pre-registered evaluation of 18,000 missions with two instances of the same model in a two-stage handover scenario, the authors find both agents fail together in 90.0 percent of cases as soon as either one fails - switching the model weakens this coupling, while switching only the vendor with the same model does not. The authors also provide a new, finite-sample certificate that assumes no dependence structure, which raises the guaranteed reliability floor in their tests from 0.2455 to 0.4116. This matters because companies often treat redundancy through multiple agent instances as a safety net - the study shows that this redundancy is structurally overestimated when the model is shared. An earlier digest post already showed that common agent safety benchmarks barely beat a simple “always safe” baseline - this new work adds another fragile measurement foundation to the picture: computed reliability certificates for agent networks themselves.
QuoteBench: when a single special character halves the success rate
Shangao Li, Yao Zhang, Volker Tresp, and Yuanyuan Yang show with QuoteBench that identical model responses can achieve completely different success rates depending on the execution path. Their benchmark replays 56 one-shot tasks derived from real incidents across eight configurations, deliberately inserting an inadequately escaping parser that rewrites Bash commands before execution. According to the authors, this single added processing step alone lowers success by 55.4 to 73.2 percentage points compared with the direct execution path, even though the language model gives exactly the same answer; disclosing the break recovers 30.4 to 60.7 points in six of the eight configurations. This matters because common evaluation methods for coding agents usually report only the matched success rate, invisibly mixing execution-environment errors with model performance. An earlier digest post already showed that the agent harness distorts coding benchmarks more than the underlying model does - QuoteBench supplies a precise mechanism for that: a single unescaped parser step in the command chain.
Intern-S2-Preview: a 397-billion model for scientific agent tasks
A large team led by Lei Bai presents Intern-S2-Preview, a new scientific AI agent model pre-trained on multimodal scientific documents, interleaved image-text data, and diverse scientific corpora, then further refined with supervised fine-tuning, multi-task reinforcement learning, and agent-specific training. The core Intern-S2-Preview-397B model extends time-series modeling from efficient long-sequence understanding to numerical forecasting, while a separate Memory Decoder module serves as an independent extension path for rapid scientific specialization without touching the frozen 397-billion-parameter backbone. Across scientific, multimodal, agentic, and general-purpose benchmarks, the authors report competitive to leading results; the separate Intern-MemDec-4B extension lifts the Biology-Instructions average score from 56.92 to 60.32 without modifying the frozen backbone. This matters because it shows how domain expertise can be layered onto a large base model after the fact, without repeating its costly training run - a practical route for fast-moving scientific specialties.
CABS+: five models merge without weakening each other
Yuchen Liu, Zongzhen Yang, Binhang Qi, Hailong Sun, and Xiang Gao present CABS+, a method that merges several fine-tuned models into a single multi-task model without their capabilities interfering with each other. The core idea is conflict-aware sparsification of weight changes, combined with a new Relative Synergy Score metric that estimates how well two models actually get along before they are merged. Across 27 datasets and five models, CABS+ outperforms two comparison methods by 16.97 and 12.93 percent respectively, according to the authors, while needing markedly less compute and running faster. This matters because model merging promises to combine several specialized models into one without costly retraining - but earlier methods often traded away accuracy to do it. An earlier digest post already showed that language models of different sizes can be merged, but with a seesaw effect between the capabilities each contributes - CABS+ targets exactly this trade-off with more targeted conflict handling.
All four papers are not independently peer-reviewed preprints - the reported numbers come from the respective author teams’ own experiments and have not yet been externally replicated. This applies especially to Intern-S2-Preview, whose model card is explicitly labeled a preview with full technical documentation still pending, and to the certification method in Agent Behavioral Contracts II, which so far has only been tested on a single two-agent scenario with a shared base model. Whether these findings hold up across further models, agent architectures, and independent replications remains to be seen.
