An AI Biology-Discovery Agent Nailed the Fit Test. Its Plausibility Score Still Collapsed From 0.98 to 0.59. Here's Everything Else That Shipped.

This week's feature is an ablation study with a genuinely uncomfortable twist: a biological ODE-discovery agent nails its hardest fit test, matching the training trajectory closely — and its plausibility score still collapses, from 0.98 to 0.59, once you strip its mechanistic constraints — full breakdown here. Plus NVIDIA open-sourced a 31-billion-parameter model that autonomously calibrates quantum computers, along with the benchmark it built to measure that, and llama.cpp shipped local vision inference for a large open reasoning model.

The feature: a good data fit and a wrong mechanism can look identical

This week's deep dive covers MEDA(arXiv 2607.13608, submitted July 15), an LLM-and-symbolic-regression agent for discovering the differential equations behind biological systems. Its own ablation shows that removing literature grounding and mechanistic constraints can preserve a strong fit to the observed data while the underlying equation becomes biologically wrong — in the worst case, plausibility falls from 0.98 to 0.59 and term-level F1 from 1.00 to 0.50, "despite good trajectory fitting." Two more papers the same week, on a completely unrelated task (historical-analogy forecasting) and from a position paper out of Oxford and the Max Planck Institute, land on the same underlying gap: models that pattern-match on surface structure instead of reasoning about mechanism. Read the full piece for the numbers and what an Oxford/MPI paper says is baked into the predictive-AI-for-science paradigm by design.

NVIDIA open-sources an agent that calibrates quantum computers on its own

NVIDIA Ising Calibration 1.5 (Lubowe & Cao, published July 27) is a 31-billion-parameter vision-language model that autonomously diagnoses and tunes quantum processors — interpreting experimental results, classifying outcomes, assessing fit quality, and recommending next steps without manual intervention, across superconducting, quantum-dot, ion, neutral-atom, and electron-on-helium qubit modalities. It's 11.4% smaller than its predecessor at BF16 precision, scores 10% better on average than the next-best open model zero-shot, and gains 86.68% over its own predecessor when given in-context examples from related experiments, on a new benchmark NVIDIA built for this purpose, QCalEval, where it's evaluated against Gemini 3.5, Claude Fable 5, and GPT 5.6 Sol. NVIDIA open-sourced the model weights, the QCalEval benchmark and dataset, and a Quantum Calibration Agent Blueprint under the OpenMDW License — and released an NVFP4-quantized version that runs on a single consumer GPU or an NVIDIA DGX Spark, the kind of on-prem deployment this blog keeps tracking as a distinct AI-for-science category from cloud-only agents.

Local vision inference lands for a large open reasoning model

llama.cpp b10142 (merged July 27) adds preliminary vision-tower support for MiniMax-M3, reusing the model's MiniMax-M2-style GQA with per-head QK-norm and DeepSeek-V3-style expert routing, plus a new CLIP-based vision graph. Sparse attention isn't supported yet (it falls back to dense), but the release reports decode holding flat at roughly 7.7–7.8 tokens/sec from 5K to 60K+ tokens of context, about 10% faster prefill, and 20% smaller buffer allocation than the prior implementation. The one operational catch: it's a breaking change, so every MiniMax-M3 GGUF built before this release has to be regenerated. Not a science-agent story on its own, but exactly the kind of infrastructure that determines whether a large open multimodal reasoning model — useful for reading figures and charts inside a literature-review agent — can actually run on a researcher's own machine instead of someone else's cloud.