generalization

The paper critiques the reliance on limited coding benchmarks like SWE-bench to measure general coding capability in AI models, showing that optimization for these benchmarks does not generalize, and advocates for diverse evaluation methods.

This paper investigates generalization in on-policy distillation for large language models, showing that it transfers reasoning behaviors and that teacher-student origin alignment is crucial, with multi-teacher combinations causing capability trade-offs.

A theory paper introducing Decoupled Descent (DD), a training method that uses approximate message passing Onsager corrections to enforce asymptotic equality between training and test error during gradient descent, potentially enabling better stopping and hyperparameter tuning.

The article discusses how Goodman's grue problem applies to AI agents in production: flawless performance on historical data doesn't guarantee correctness on future data, and more data can't resolve the fundamental ambiguity.

This arXiv paper presents an AI Scientist loop for studying generalization in quadruped robot navigation, adding an experiment card, specialized subagents, and a preference oracle called kkanbu to prevent drift and maintain falsifiability in autonomous research.

This paper investigates why language models fail at two-hop generalization, showing that models succeed when the second hop follows training distribution but fail when it deviates, and proposes a recurrent-style training strategy to improve out-of-distribution two-hop reasoning.

This paper examines how LLM-driven evolutionary optimization of GPU kernels can fingerprint evaluation configurations, causing 30% of in-distribution wins to fail on held-out settings. It offers a taxonomy of failure modes and design guidance for robust benchmarking under strategic optimization.

This paper argues that Explorative Modeling (XM) optimizes for 'freedom' rather than generative expressivity, proving that larger candidate pools increase miss probability and freedom, with empirical results showing freedom-based selection improves generalization under distribution shift.

This theoretical paper studies the non-asymptotic implicit bias of logistic regression under gradient descent, proving that the parameter vector weakly aligns with the max-margin direction quickly, within a doubly exponential number of iterations in the alignment error.

A position paper arguing that large language models have fundamental limitations, using the metaphor 'can't jump' to highlight gaps in reasoning or generalization.

A tweet by John Schulman highlights the paper 'Chunky Post-Training,' which argues that diverse post-training datasets cause models to learn spurious correlations that lead to unintended behaviors, such as rejecting true facts posed in specific formats. The paper introduces SURF and TURF to surface and trace these generalization failures across frontier models.

This paper identifies invisible metadata traces at the pixel level as shortcuts that vision encoders exploit, leading to performance degradation under metadata distribution shifts. Mitigation strategies during and after pretraining reduce sensitivity to both targeted and unseen metadata without sacrificing downstream performance.

MirrorCraft is a paired benchmark for evaluating LLM-based agents in Minecraft under hidden rule changes, using matched Vanilla and Mirror worlds to measure performance shifts via the Rule Intervention Effect (RIE).

This paper identifies a non-composition principle in AI benchmark evaluation: support for adjacent projections does not automatically warrant their composition. It proposes a projectibility audit to diagnose unsupported joins in benchmark-to-use arguments, with a legal-research case study and simulations.

This paper studies transferring lessons about supervised fine-tuning (SFT) across alignment training, model organisms, and toy models, showing that techniques like training on reasons for behavior and mixing on-model data can improve generalization and capability preservation.

This article discusses a research paper showing that the disagreement rate between two deep networks trained with different random seeds can accurately estimate generalization error using only unlabeled data, revealing a surprising connection called Generalization Disagreement Equality.

The paper proposes GEAR-SAM, which adaptively allocates the perturbation budget across network blocks using an exponential moving average of squared gradients, improving generalization without additional computational overhead.

Introduces Building2Building (B2B), a large-scale benchmark for studying generalization and transfer in reinforcement learning using realistic HVAC control environments built on EnergyPlus, compatible with Gymnasium.

This paper introduces a framework to reduce per-sample harm in stochastic optimization, where parameter updates from batch averaging and historical states increase individual sample loss. The method uses dimensionality reduction and focuses on the last linear layer for efficiency, showing improved generalization on image classification tasks.

Discusses a paper by Alex Zhang and Omar that reveals how frontier models can cheat on benchmarks by training on test lookalikes, and proposes using NLP distance metrics on hidden trajectories to detect such cheating.