generalization

This paper introduces Filtered Posterior Mean Collections (FPMCs), a unified framework for analytical models of diffusion generalization, consolidating existing approaches and demonstrating improvements through soft relaxations and source distribution augmentations.

This paper introduces ASAM (Adversarial Subspace Alignment for Multimodal Knowledge Editing), a method that improves the generality of intrinsic multimodal knowledge editing by generating adversarial yet semantically coherent variants and enforcing low-rank alignment of representations, addressing the limited generalization of existing methods.

A free Stanford lecture by Percy Liang on AI generalization explains why models excel on benchmarks but fail on real codebases, covering benchmark memorization, bias-variance tradeoff, and hallucination.

This paper introduces the Representation Gap, a metric for neural network generalization error with better asymptotic dynamics. Using a geometric perspective and optimal quantization theory, the authors show it is governed by the intrinsic dimension of the task, and verify this empirically on synthetic and realistic datasets.

Introduces Equilibrium Reasoners (EqR), a model that learns task-conditioned neural attractors, achieving high accuracy on hard reasoning tasks like Sudoku-Extreme (99.8%) and Maze (93%).

Geo-Align presents a reinforcement learning framework for camera-controlled video re-rendering that improves generalization through scale-aware perceptual rewards and metric 3D estimation for camera trajectory extraction.

This discussion examines whether liveness detection models trained on historical deepfake samples can generalize to new synthetic media generation techniques, questioning the update cycle for vendors claiming deepfake detection capabilities.

Weasel is a trajectory selection method for offline training of web agents that improves out-of-domain generalization by balancing importance and diversity. It achieves up to 12.5x training speedups and improved performance across several benchmarks.

This paper presents the first theoretical model for out-of-distribution generalization in reinforcement learning, showing that smaller abstract state spaces enable cross-scale generalization in POMDPs.

This paper presents the first quantitative prediction of the grokking delay under AdamW, deriving a closed-form law and validating it on algorithmic tasks with high accuracy.

This paper introduces projection agents for graph combinatorial optimization using reinforcement learning and graph neural networks, operating in a continuous action embedding space to improve generalization and scalability, and releases the LaGCO-RL library.

This paper proposes Emo-Boost, a multimodal deepfake detection framework that leverages emotion cues (audio-visual emotion recognition) as high-level semantic signals to improve generalization to unseen manipulation types, achieving a 2.1% average AUC improvement on the FakeAVCeleb dataset.

GRLO introduces a novel reinforcement learning post-training method that achieves strong generalization across multiple domains (math, code, etc.) from only 5K prompts and 22.7 GPU hours, significantly outperforming in-domain RLVR baselines in efficiency and data requirements.

This paper introduces a margin-based confidence ranking method for LLM-as-a-judge systems, learning a dedicated estimator to ensure monotonicity between confidence and human-disagreement risk, with generalization guarantees and improved ranking accuracy across datasets.

This paper studies generalization in learning through the lens of bounded-rational decision theory, where the learner's response law induces a tradeoff between training loss and sample dependence. The authors show that this tradeoff is governed by an f-divergence regularizer and that generalization can be certified from the learner's hedging behavior.

AgentKernelArena is an open-source benchmark for evaluating AI coding agents on GPU kernel optimization, assessing full agent workflows and generalization to unseen configurations across 196 tasks.

This paper introduces a framework for distribution-aware algorithm design where LLM agents learn to generate solver code specialized to target distributions, achieving high solution quality and significant speedups over standard solvers.

This paper identifies autonomous exploration as a critical capability for LLM agents and proposes the Explore-then-Act paradigm, which decouples information gathering from task execution to improve adaptability and real-world performance. It also introduces Exploration Checkpoint Coverage as a verifiable metric for evaluating exploration breadth.

This paper argues that monolithic scaling of a single model is insufficient for achieving AGI and proposes Agentic AI with multi-agent collaboration as a necessary paradigm, demonstrating theoretically that agentic systems achieve exponentially superior generalization and sample efficiency.

This paper presents a systematic study of long-context continued pre-training for vision-language models, achieving generalization beyond 128K context with an efficient data mixture design and introducing the MMProLong model.