early-stopping

StopLoss is a lightweight external monitor that learns failure recognition from completed coding-agent trajectories to abort doomed runs early, cutting cost per retained solve by up to 48% while preserving successful outcomes on SWE-bench benchmarks.

This paper proposes an analytical framework for estimating the optimal early stopping time in neural network training using Rademacher complexity with L1-norm, avoiding probabilistic assumptions and applicable to both linear and nonlinear models.

This paper challenges the assumption that iterative denoising always improves forecasts in diffusion-based time series forecasting, proposing a global stopping criterion and a Bernoulli sampler to enhance accuracy and speed.

AdaStop is a cost-aware early stopping framework for DNN test selection that optimally stops labeling when the marginal fault discovery rate falls below a threshold, achieving 65-84% fault discovery using only 9-31% of the labeling budget.

This paper introduces a meta-learning framework for anytime-valid certified robustness that uses sequential E-processes to adaptively allocate compute, achieving a 20-fold reduction in sample complexity compared to traditional randomized smoothing while maintaining rigorous statistical guarantees.

ASAG uses attention entropy to detect when reasoning is unproductive, stopping early to improve accuracy and reduce token generation. Experiments on Qwen3-8B show a 4.4% accuracy gain and over 40% fewer generated tokens.

This paper proposes ASAG, a training-free method that adaptively stops reasoning in large reasoning models based on attention distributions, reducing token usage by ~40% while improving accuracy by 3.2% on benchmarks using DeepSeek-R1-Distill and Qwen3 models.

This paper introduces MARS, a stopping rule for parallel LLM test-time scaling that probes partial traces to stop early without sacrificing accuracy, saving 25–47% of tokens across reasoning models on competition math benchmarks.

EvalStop is a scheduling primitive for multi-tenant RLHF platforms that detects and corrects reward overoptimization by monitoring downstream evaluation scores and terminating jobs on consecutive declines, achieving 98% precision and 99% recall while improving job completion time by 9% and cutting wasted compute by 22%.

ESPO introduces an early-stopping mechanism for reinforcement learning that detects and terminates failed reasoning trajectories in LLMs, improving mathematical reasoning performance while reducing compute by over 20%.

This paper introduces LEAP, a training-free method to accelerate inference in Diffusion Language Models (dLLMs) by detecting early-converging tokens, reducing denoising steps by 30% without losing accuracy.