Hosts: Lenar Kess, Damra Vol. A revision sweep through the cs.AI batch surfaced five independent papers pushing on the same thing: the instruments we use to decide whether a model is right are themselves unreliable — self-consistency, calibration error, count-based F1, and the model judges grading all of it.Kaihua Ding audits self-consistency across 265,000 samples — agreement predicts correctness weakly, and worst on the most consistent frontier model.Jon-Paul Cacioli separates accuracy from metacognitive sensitivity with Signal Detection Theory — and his v3 note retracts his own headline after 1,830 human adjudications.Dekun Yang shows numeric anchoring inflates count-based F1 without improving span localization.A LoRA adapter that makes a fine-tuned model describe its own hidden behavior, and roughly halves the baseline's hallucination rate.Brett Reynolds turns the instrument on the judges — the first one graded its own outputs with the answer visible and still missed the safety-relevant classes.ANTAP routes multi-agent work by empirical probe instead of textual self-description.ATOD anneals on-policy distillation into reinforcement learning for multi-turn agents.FCGraft grafts cached key-value states from validated code skeletons.Ouroboros-Spatial uses solver confidence as a curriculum difficulty signal.VendorBench-100 puts commercial APIs, vision language models, and open-source detectors under one protocol.G2VD attacks shortcut learning in AI-generated video detection.FirstResearch proposed a Research Question Certificate — and was withdrawn by its author on 28 July.RWGBench scores citation decisions instead of text similarity, and checks the metric against expert judgment.Chen and Xie find that making human design visible shifts moral judgment toward rule-based reasoning.
