This research introduces state commitment learning, using Counterfactual Erasure RL (CERL) to train language models to distinguish temporary computation from persistent state, reducing answer dependence on hidden thoughts without sacrificing accuracy.

Researchers localized a neural subgraph responsible for temporal preference in a distilled LLM (Qwen3-4B-Instruct-2507), finding that models discount the future less steeply than humans and that this preference is unstable across contexts, with steering vectors capable of modulating it.

PyCC is a Python library that leverages structural skeletons derived from characteristic curves to discover ordinary differential equations from time-series data, incorporating domain hypotheses to address the ill-conditioned inverse problem of equation discovery.

This paper proposes a staged fractional-factorial screening method to efficiently identify high-penalty hyperparameter directions under tight budgets. Through 613 experiments, it finds that main penalties from total batch, depth, and width are largest at short budgets and relax as budget increases. Short designed screens help confirm promising anchors and refine locally, supporting a bridge-centered recommendation up to 24 hours on two hosts, though rankings are not hardware-invariant.

At matched accuracy, open-weight LLMs differ substantially in the shape of their error severity distribution — a difference invisible to the scalar error rate. The Errorquake-10k benchmark scores each response on a continuous 0-4 severity scale across 8 domains and 5 difficulty tiers, revealing that severity profiles provide information beyond error rate.

A new paper proposes a stereological theory for evaluating LLM benchmark coverage, revealing that effective dimensionality of benchmark suites leads to large blind spots that dwarf score differences, and suggests minimal benchmark sets and resolves Gardner's problem.

Ollama 0.30 is now available with improved performance and GGUF model compatibility through llama.cpp, augmenting MLX on Apple silicon and supporting more models on wider hardware.

Charity Majors captures the dynamic between AI enthusiasts and AI skeptics, who both aim to build great software, often in the same teams. Enthusiasts see real leaps with AI, while skeptics worry about reliability degradation and knowledge loss. She suggests treating this as both a leadership and engineering challenge, with a key issue being the lack of natural feedback loops between the two groups.