The paper introduces FateMultiplicity, a label‑free framework that builds a Rashomon set of statistically admissible models for single‑cell trajectory inference by cross‑fitting held‑out genes and non‑inferiority testing. It shows that model multiplicity is large and depends more on the diversity of the model space than its size, and that per‑cell certified fate margins do not improve assignment reliability compared to baseline models. The study also identifies two robust constructions— a margin‑erosion ratio for branch‑point detection and a certification filter that reduces disagreement with clonal outcomes— that merit further investigation.
arXiv:2607. 19426v1 Announce Type: cross Abstract: Single-cell datasets are increasingly costly to store, audit, and reuse for model training.
arXiv:2608.00152v3 Announce Type: replace-cross Abstract: AI evaluation can support the wrong inference when an in-domain benchmark success does not survive distribution shift, or when the benchmark...
The paper investigates when post‑training quantization (PTQ) calibration helps or hurts text embedders, focusing on the amount of calibration (dose) and whether to calibrate at all. Across three models and 161 calibration runs at 4‑bit weight quantization, the effect ranges from significant gains to catastrophic degradation, driven mainly by the functional displacement of the quantized model. A validation‑gated abstention rule is proposed that avoids harmful calibrations while capturing most benefits, highlighting the trade‑off between safety and opportunity.
The study evaluates a frozen Geneformer representation for predicting CRISPRi perturbation effects under a tightly controlled, pre‑registered protocol. While the representation shows significant predictive power within the Virtual Cell Challenge dataset, it fails to transfer to external screens, with negative zero‑shot Spearman correlations. The analysis also reveals that the VCC endpoint is heavily influenced by sampling depth, as cell count alone explains most of the variance, indicating a sampling‑depth entanglement that could mask transfer failures in less controlled settings.
arXiv:2607. 19426v2 Announce Type: replace-cross Abstract: Large single-cell datasets are expensive to store, curate, and repeatedly reuse for model training.
PRISM is a tile‑level benchmark for pathology foundation models that evaluates reliability across four dimensions: model choice, label fraction, calibration, and out‑of‑distribution transfer. It tests eight models on six datasets at six label fractions (1%–100%) with 864 linear probe runs, plus a five‑hospital Camelyon17 transfer study. The benchmark reveals four key phenomena, including the changing relationship between AUROC and calibration at low labels, collapse of linear probes on harder datasets, the impact of source‑domain labels on target calibration, and limitations of temperature scaling.
arXiv:2610. 00445v1 Announce Type: new Abstract: Pairwise guide--transcript scores do not enforce conservation of a finite guide-loaded RISC pool when they are interpreted independently as occupancies.
The paper introduces a selection‑safe conformal early‑exit method that calibrates a single split‑conformal quantile using a terminal‑path score derived from prefix maxima. This approach guarantees coverage for any data‑dependent stopping policy that preserves terminal‑set containment, without requiring an α/L correction. Empirical results on eight encoder backbones and CIFAR‑100 show that pathwise selection achieves higher macro coverage (0.901 vs. 0.814) with smaller set sizes and lower latency fractions compared to naive layer‑wise selection, and remains robust to deployment‑time budget changes.
arXiv:2606. 10154v1 Announce Type: new Abstract: Quantized checkpoints are often screened first with quality metrics and only later, if at all, with direct safety tests.
arXiv:2609.36569v1 Announce Type: cross Abstract: Checkpoint selection is a routine decision in supervised fine-tuning (SFT): training produces multiple checkpoints, but only one is retained. Yet fix...
The study investigates how the composition of data during the mid‑training phase of language models affects performance across multiple domains. Experiments with Qwen3‑8B‑Base on five distinct KOR‑Bench domains show that moderate coverage (10%‑40%) yields the best per‑domain results, and that alignment passes cannot fully close the performance gaps created by mid‑training data choices. Additionally, zero coverage in mid‑training severely degrades accuracy, while a carefully tuned allocation can provide the largest overall pipeline improvement.
The paper evaluates additive activation steering in chat and agent contexts, showing that the commonly used gain ratio (Δ_agent/Δ_chat) fails to reliably indicate potency and efficacy across multiple models and dose-response cells. By replacing the gain with a location metric, dEC50 (difference in EC50 between agent and chat), the authors demonstrate a more robust, two‑sided measure that consistently captures cross‑context shifts. The study also reports several refuted and unanswered claims, emphasizing that a single operating point cannot distinguish between displacement and gain effects.