Why it matters: InfTool establishes a fully autonomous, self-evolving pipeline for tool-use mastery, proving that mid-sized models can rival frontier models through multi-agent synthetic data and reinforcement learning.

Why it matters: This paper bridges mechanistic interpretability and parameter-efficient fine-tuning by using Sparse Autoencoders to construct an explicit, interpretable subspace for safety alignment that rivals RLHF in performance.

Why it matters: This paper reveals that LLMs fail to distinguish between warnings and instructions due to shared latent features, proving that 'negative' training data inadvertently reinforces the very behaviors it seeks to prevent.

Why it matters: This paper replaces the flawed paradigm of feature invariance with a rigorous decision-theoretic framework based on Le Cam's theory, preventing catastrophic information loss during domain adaptation in safety-critical systems.

Why it matters: This paper exposes a critical flaw in how we use small proxy models to select data recipes and provides a simple, theoretically-backed fix to ensure small-scale experiments actually predict large-scale success.

Why it matters: This paper reveals that web agents are twice as susceptible to manipulative 'dark patterns' as humans, with larger and more capable models paradoxically being the easiest to deceive.

Why it matters: This paper provides critical empirical evidence on how people use LLMs during mental health crises, arguing that AI safety in this domain requires designing agents as bridges to human connection rather than standalone solutions.

Why it matters: This paper provides the first large-scale empirical evidence of how AI-generated code is reshaping the security landscape of the world's most critical software repositories.

Why it matters: This paper demonstrates a breakthrough in autonomous capability scaling by using self-play reinforcement learning to train software agents that outperform human-data-reliant baselines on complex coding tasks.

Why it matters: This paper provides a rigorous theoretical framework for understanding how inference-time scaling interacts with reward misspecification, offering a mathematical explanation for Goodhart's Law in LLM evaluation.

Why it matters: This paper provides the first empirical bridge between compute scaling laws and real-world economic productivity, offering a rigorous framework for forecasting AI's societal impact and takeoff speeds.

Why it matters: This paper introduces the first policy gradient framework for performative RL that achieves performative optimality, providing a rigorous solution for agents whose actions fundamentally alter their environment's dynamics.

Why it matters: SWE-Bench++ automates the creation of complex, multilingual software engineering benchmarks, providing a critical tool for evaluating the capabilities and risks of autonomous AI agents.

Why it matters: This paper provides a rigorous theoretical framework for how data diversity forces transformers to abandon brittle positional shortcuts in favor of generalizable induction heads.