MIS-Ground is a controlled factorial benchmark that isolates the language-side brittleness behind 3D medical spatial grounding in vision-language models. MIS-SemSam, a training-free semantic-sampling decode rule, lifts Qwen3-VL-32B by 13.06% to 66.5% overall, the best open-weights result and above Gemini 3 Flash.
Ranking reasoning LLMs under test-time scaling. We compare 72 ranking methods (Bayes@N, Bradley-Terry, Elo, IRT, voting, graph/spectral) across 20 models and four Olympiad-style math benchmarks. At full budget they mostly agree (Kendall's tau_b 0.93-0.95); at one trial a greedy prior cuts variance 16-52% but can bias the ranking. Packaged in the Scorio toolkit.
A Bayesian framework for evaluating large language models that replaces unstable Pass@k metrics with posterior estimates and credible intervals. The method improves sample efficiency, supports graded outcomes, and enables statistically sound model comparisons.
Key-favored KV-cache quantization for LLMs: theory shows keys have larger norms and should get more bits; empirics show 4b-K/2b-V preserves up to 98.3% accuracy while cutting memory.
DFloat11 compresses LLMs to 70% of their original size while maintaining bit-for-bit identical outputs. A lossless compression framework with efficient GPU inference that enables running Llama 3.1 405B on a single node.