An NVIDIA paper proposes compiling an LLM's reasoning trace into small parameterized Python functions that run on CPU, invoking a fine-tuned 1B model only for lightweight text tasks such as value extraction. The cache matched or beat the source model on 7 of 8 math and science benchmarks, scoring 93.3% versus 83.3% on AIME '25, and covered at least 90% of rephrased or re-numbered variants at 16x to 291x lower cost.
Reasoning Traces
推理轨迹
Records of the logical steps a model takes from input to final conclusion, used to explain decision-making processes.
Curated Intelligence
Artificial Analysis has released VIALS-AA, its implementation of Handshake AI's VIALS benchmark, running 161 scientific image tasks three times each via its open-source Stirrup harness. Models frequently answer incorrectly even after extensive image inspection, and reach correct answers through markedly different reasoning paths.
Researchers from Virginia Tech and others introduced Epi2Diff, a framework that predicts human exam difficulty by analyzing cognitive segments and transition features in LLM reasoning traces. It outperforms baselines on SAT and USMLE datasets, offering a new paradigm for educational measurement using AI process signals.
NVIDIA released Nemotron 3.5 Content Safety, a customizable multimodal safety model for enterprise AI. Building on Nemotron 3, it adds reasoning capabilities and policy generation tools to enable auditable, domain-specific safety decisions across languages and modalities.