This week's three items each try to measure something that a simple accuracy score misses.
BioBasicsBench: can an agent read a gel?
OpenAI published BioBasicsBench on November 6, 2026. It contains 100 routine wet-lab tasks in five areas, built from raw evidence such as gel and blot images, colony plates, microscopy stacks and plate-reader exports. Scientists wrote the tasks from their own experiments with expert grading rubrics, and each went through three rounds of review.
At the xhigh reasoning setting, GPT-6 Astra earned "an average of 66.7% of the available credit, compared with 43.8% for GPT-5.6 Sol." The task design rewards recognising when evidence cannot support a conclusion, and OpenAI says Astra showed better ability to avoid unwarranted conclusions. It also says there is room for improvement in interpreting images, checking controls and applying experimental context. This is an OpenAI benchmark reporting on OpenAI models, so other labs' results are what would make it useful.
What the experts in a Mixture-of-Experts actually do
A paper posted November 5 trains sparse autoencoders on the output of a Mixture-of-Experts layer. Since the layer's output is a router-weighted sum of expert outputs, each latent's activation decomposes exactly across experts, which measures contribution rather than selection. On layers 4, 30 and 46 of Qwen3-30B-A3B, the authors find contribution is far more concentrated than gating, and most latents have only one or two dominant experts. They also see "a division of labor below coarse domains", with separate experts for Java and front-end web code, or for geometry and for arithmetic and polynomials.
The practical value is for interpretability and debugging of the architecture most large open models now use. A model with many experts may fail on a narrow skill because of a small set of them, and this method offers a way to look.
TB-MathModel: an answer must also be possible
TB-MathModel puts 18 real CUMCM and MCM modelling problems in a terminal environment and scores one JSON answer per task. It uses one-sided scoring that does not penalise beating the reference, and a two-layer feasibility check with a programmatic gate plus 624 Lean 4 obligations. Infeasible answers score zero. Across seven multimodal entries, the strongest reach 64.58 percent, while a $2.54 open model holds 59.81 percent at one hundredth of the cost.
What remains uncertain. We read OpenAI's post and the two paper abstracts and reproduced nothing. Eighteen problems is a small set for TB-MathModel. The SAE results cover one model and three layers, and auto-interpretation labels can mislead.
Analysis: A hard validity gate is cheap to copy. If Geetha's team at Belgavi's Book Store has an agent proposing refund amounts, add a check that zeroes any answer over the order total, before any judge scores the wording. A persuasive answer that cannot be true should not earn partial credit.
Sources and evidence
OpenAI's BioBasicsBench measures agents on 100 raw wet-lab tasks, with GPT-6 Astra at 66.7 percent. A sparse autoencoder study finds MoE experts specialise finely below coarse domains. TB-MathModel adds formal feasibility checks that zero impossible answers and finds a low-cost open model close to the top. Each measures something simple accuracy would miss.