From the arXiv new submissions of the past 24 to 48 hours in cs.AI, cs.LG, and cs.CL, this digest selects five papers that together show how fragile judgment, meeting attention, and introspection remain in today’s AI systems on closer inspection - from a wobbly AI judge to a silent meeting representative, a hallucination problem in peer review itself, a philosophically charged consciousness experiment, and a cultural knowledge gap in image recognition. Selection was based on substance and thematic spread: each paper offers a traceable method with solid numbers in the abstract, and no two papers cover the same subfield.
Prompted meeting representatives fall silent - a new system cuts that to 2.5 percent
Muneeb Khan, Frederic Kirstein, Terry Ruas, and Bela Gipp investigate with CAPA (Collaborative Agent Predictive Architecture) how well an AI agent recognizes when it should speak up on behalf of an absent meeting participant. On the AMI conversation corpus, the authors first show that a purely prompt-based stand-in stays silent on 51.4 percent of suitable speaking opportunities because it lacks the judgment to know when a contribution is warranted. CAPA, made up of five components for perception, prediction, control, generation, and recalibration, cuts that silence rate to 2.5 percent, with a hallucination rate of only 0.6 percent. A purpose-built evaluation protocol using schema-constrained AI judges reaches a Cohen’s kappa of 0.71 agreement with human ratings, and ablations show that tracking meeting state - not simply adding more context - makes the decisive difference. Microsoft Teams already automatically summarizes meetings into recap notes - the new study goes a step further and lets an agent actively join the conversation instead of just documenting it afterward. This matters because companies increasingly want AI stand-ins for absent colleagues, where blind silence leaves real information gaps.
AI judge models can be guessed from the rubric alone - and wobble on reversed answers
Anshul Bagaria, Sowmya S. Sundaram, Gokul S. Krishnan, and Balaraman Ravindran show with a study on rubric artifacts in LLM-as-judge pipelines how reliably automated evaluation systems for AI-generated text actually work. Classifiers trained only on the rubric text - with no access to the response being evaluated at all - predict the AI judges’ verdicts with an accuracy of 82.6 to 90.0 percent, depending on the model. According to the authors, this suggests that the wording of an evaluation rubric already carries usable signal about the expected outcome, independent of the actual model output. In a counterfactual test where the evaluated response was reversed in meaning, the judge model correctly adjusted its verdict in only 37.7 percent of cases; when the rubric criterion itself was reversed instead, the expected adjustment occurred in only 16.8 to 32.2 percent of cases depending on the model. An earlier digest entry had already shown that AI judge models can be swayed to flip their verdict under targeted persuasion in up to 91 percent of cases - the new study exposes a different weak spot in the same class of methods: judges sometimes react too strongly to how a prompt is phrased and too weakly to actual content differences. This matters because LLM-as-judge pipelines are now widely used for automated quality control of AI outputs.
Hallucination detection for peer review fails on real reviews
Tzu-Ling Lin, Dong-Ting Yao, Teng-Fang Hsiao, Wei-Chih Chen, and Hong-Han Shuai present HalluPeer, a dataset built specifically for scientific peer review that tests whether unsupported but plausible-sounding claims in AI-assisted reviews can be reliably detected. The authors develop a review-specific taxonomy of false claims and generate artificially corrupted reviews with built-in quality control, matched against the original papers and human-written reviews. Across 12,000 papers and 38,000 associated reviews, current detection methods struggle to tell genuine fabrications apart from legitimate scholarly criticism; an analysis of authentic human reviews further confirms that the error patterns defined in the dataset genuinely occur there too. An earlier digest entry had already shown that AI reviewers catch only 12.1 percent of deliberately inserted errors in academic papers while consistently judging more leniently than humans - the new study adds the mirror-image problem: not only checking academic papers, but checking the reviews themselves, still lacks a reliable yardstick. This matters because AI tools are increasingly moving into the peer-review process itself.
A deterministic agent appears to show felt preferences
Mark Solms, St John Grimbly, Bruce Bassett, and six further co-authors investigate in a case study on affective consciousness in an artificial agent whether “hedonic place preference” - a behavior known from animal research, the preference for locations where previously pleasant but non-nutritive stimuli were experienced, such as cocaine or morphine in animals - can also be engineered artificially. They build a simple agent that instantiates properties of an affective system, processing something like felt uncertainty about its own intrinsic needs relative to available environmental resources. According to the study, this agent displays the same place preference as its biological counterparts - produced through an apparently subjective form of information processing, even though the entire system remains fully deterministic, meaning it follows the same sequence every time with no randomness involved. The authors draw out implications for understanding the physical basis of consciousness and the experience of free will. An earlier digest entry had already shown how a purpose-built method can measure which silent thought processes an AI model runs through before it answers - the new study approaches the question from the opposite direction: it engineers a system that technically reproduces apparently subjective processing from scratch, rather than tracing it inside an existing model. This matters because the question of whether and how consciousness can be technically reproduced increasingly touches practical debates about the moral status of AI systems.
Multimodal models recognize dishes almost perfectly but misclassify them culturally
Bo Zeng, Linfeng Gao, Peiqin Lin, and nine further co-authors show with CulturalMenuBench, a new benchmark of 4,870 items across 10 languages and 18 regions, that high recognition rates from multimodal language models on food say more about visual matching than about genuine cultural knowledge. Its ten tasks pair images of finished dishes and individual cooking steps with ingredients, procedural text, and regional labels, ranging from basic recognition to process-grounded cultural attribution. Evaluating twelve models exposes a substantial gap between knowledge and application: models that score above 94 percent on standard multiple-choice tasks drop to at most 56 percent when attributing dishes to Chinese regional cuisines - despite an identical four-way answer format. Models also classify dishes 7 to 18 percentage points better from text labels alone than from the image itself, which the authors say shows that near-perfect recognition can conceal an inability to actually apply cultural knowledge. This matters because multimodal models are increasingly used for tasks like automated recipe recommendations or cultural mediation, where image recognition alone is not enough.
None of the five papers presented has gone through a regular peer-review process yet; all are unreviewed arXiv preprints, with figures drawn from the respective author teams’ own experiments and not yet externally replicated. Whether the reported effects hold up across further models, datasets, and independent replications remains to be seen.
