Full text of Medical Image Spatial Grounding with Semantic Sampling: MIS-Ground, a factorial benchmark that isolates the language-side brittleness behind 3D medical spatial grounding, and MIS-SemSam, a training-free decode-time rule that lifts Qwen3-VL-32B by 13% overall.
Full text of Ranking Reasoning LLMs under Test-Time Scaling: dense repeated-trial evaluation, Bayes@N as a practical default, and categorical rankings packaged in the Scorio toolkit.
Don’t Pass@k introduces a Bayesian approach to language model evaluation, estimating Bayes@k with posterior uncertainty, credible intervals, and rubric-aware scoring.
Full text of Quantize What Counts: a geometry-driven, mixed-precision KV-cache quantization method that allocates more bits to keys than values to preserve accuracy while cutting memory.
Full text of DFloat11: a lossless compression framework that shrinks LLMs to ~70% of their size with bit-for-bit identical outputs and a GPU kernel for efficient compressed inference.