Today’s selection from arXiv’s new submissions of the past 24 to 48 hours connects four questions that rarely come up together: how reliably AI agents operate graphical user interfaces, how robust coding agents stay against purely cosmetic code changes, how easily AI judge models can be swayed by nothing more than authorship labels, and at what model size language models master a core piece of language understanding. The selection was curated for substance and spread: each paper offers a traceable method with solid numbers in the abstract, and no two papers cover the same subfield.
ComponentBench: UI agents don’t fail at the goal, they fail at the observation
Tianchen Guan, Xinlei Lin, Royce Cheng-Yue, Xiangjun Wang, and Shuyan Zhou introduce ComponentBench, a benchmark for so-called computer-use agents - AI systems that operate web interfaces the way humans do, via mouse clicks and keyboard input. Rather than testing either long workflows or isolated click targets as prior benchmarks did, ComponentBench breaks the task down into 2,910 programmatically verified interactions spanning 97 common UI component types (such as toggle sets or dropdown menus), paired with cleaned human reference trajectories. Across seven tested models - including GPT-5.4, Gemini 3 Flash, and UI-TARS-1.5-7B - the authors report that changing only the observation and action space, with the same model held fixed, shifts task success by more than 30 percentage points: GPT-5 mini fell from 83.1 percent with access to the accessibility tree (a structured description of the interface built for assistive technologies) to 48.9 percent when restricted to raw screen coordinates. Even the fastest configuration took 3.7 times as long as the human reference. This matters because it shows that capability comparisons between AI agents often reveal more about the chosen interface than about the underlying model.
A Jagged Frontier: cosmetic code changes destabilize coding agents
A team led by Hasan Najib Mahmud, Shreya Gupta, and Isha Chaudhary examines, with A Jagged Frontier, how stable AI coding agents stay when the surrounding codebase is rewritten in a purely cosmetic way - through different control-flow phrasing, injected dead code, or renamed identifiers, without changing what the code does. The team tested two agent scaffolds (mini-SWE agent and OpenCode), each paired with four frontier models including Claude Opus 4.5, on tasks from SWE-bench Verified and SWE-bench Pro. The authors report a resolve-rate drop of up to 6.7 percentage points in the most affected configurations, with statistically significant degradation in 6 of 16 tested combinations of model, scaffold, and dataset; which model proves most robust depends heavily on the scaffold used, and the simpler mini-SWE agent held up better overall than OpenCode. A previous report already showed that OpenAI itself classifies roughly 30 percent of SWE-bench Pro tasks as flawed - this new work adds that even correct tasks can yield unstable measurements once only the surrounding code changes, not its meaning. This matters because leaderboard rankings for coding agents may depend more on the luck of the test codebase than previously assumed.
Bidirectional judgment bias: one label is enough to tilt AI judges
Songeun Chae, Min Kim, Donghoon Jung, Seojin Choi, and Seohyon Jung investigate, with their study on self- and other-labels, whether so-called LLM judges - language models used as automatic reviewers of other AI outputs - genuinely favor their own outputs (self-preference), or whether that effect is really just style and quality getting conflated. To separate the two, the authors had ten language models evaluate not generated text but style-neutral narrative-constraint selections that still carry a recoverable, model-specific signature. Under blind evaluation, self-preference largely disappeared once selection quality and evaluator severity were controlled for - it vanished on three of four rubric dimensions and even reversed on the fourth. But when judges were shown, at matched quality, plain “self” or “other” labels without any model being named, they inflated scores for self-labeled selections and deflated scores for other-labeled ones. A previous digest entry already showed that AI judge models can be talked into flipping their verdict in up to 91 percent of cases under targeted persuasion - this new work shows that mere authorship labeling, with no persuasive content at all, is enough to systematically distort a judgment. This matters because LLM judges increasingly decide on training-data selection, product approvals, and automated evaluations, settings where provenance metadata is often available alongside the content.
Entity tracking: human level already at 410 million parameters
Karolina Drożdż and Micha Heilbron investigate, with their study on entity tracking, how well language models and humans can track where things are in a narrative and how they change, even when that is never stated explicitly - a core component of language comprehension. Unlike earlier, artificial test tasks, the authors used naturalistic narrative texts of varying complexity and pitted both language models and 48 human participants against them. In humans, tracking performance degraded specifically with narrative complexity, not with raw text length. In language models, human-level performance was already present at 410 million parameters - well below the much larger, code-specialized models that prior work had identified for this threshold - and larger contemporary models clearly exceeded human performance from there. This matters because it suggests that a central building block of language understanding emerges at far smaller, cheaper-to-run model scales than previously thought, rather than only in the largest systems.
All four papers are preprints that have not been independently peer-reviewed - the reported numbers come from the respective author teams’ own experiments and have not yet been externally replicated. That caveat applies especially to the self- and other-label study, whose finding so far rests on just ten tested models and one narrowly defined task type, and to the entity-tracking paper, whose human comparison draws on only 48 participants. Whether these findings hold up across more models, tasks, and independent replications remains to be seen.
