supervision

Reusable computer vision utilities for detection and tracking

imagebasic pythonFree tier

Overview

supervision is a Python library focusing on practical computer vision workflows. It bundles utilities for bounding-box handling, segmentation masks, tracking, visualization, and integration with popular CV models. Developers can quickly assemble robust pipelines, standardize evaluation, and avoid re-writing glue code around detectors and trackers, making it especially useful for research prototypes and production CV services.

Best For

+ml-engineers
+researchers

News & Updates

New Method Fixes On-Policy Distillation When Student Models Go Off Track

Researchers Kaixiang Zhao, Tianrun Yu, and Shangzhe Li published a paper identifying a core flaw in on-policy distillation (OPD), a technique where a student model learns from its own generated rollouts under token-level teacher supervision. OPD assumes the teacher can recover a correct answer from any prefix the student writes, a property the authors call 'recoverability' — but they show this fails when the student's rollout goes wrong. Their proposed fix, critique-guided on-policy distillation, uses the teacher's critique of the student's incorrect rollout to guide training instead of relying on the teacher to simply continue from a bad prefix. This matters because OPD is widely used to compress large language models into smaller deployable ones, and the recoverability failure means many distilled models are learning from corrupted supervision signals without anyone noticing.

breakthroughs10/15/2026

New RL Method Cuts Language Agent Training Costs by Learning From Every Step

Researchers Haofei Yu, Yixing Chen, and Yining Zhao published a new method called Co-Evolving Acting and Forecasting that trains language agents more efficiently by using intermediate step supervision instead of only terminal rewards. Standard agentic reinforcement learning relies on a single scalar reward at the end of a long rollout, wasting the rich signal available at each step. Their approach co-evolves an acting policy with a forecasting model, allowing the agent to learn from every action it takes. This directly addresses the bottleneck of high-cost long-horizon rollouts that has slowed progress in training capable language agents.

breakthroughs10/15/2026

VASPO Fixes Flawed Reward Signal in Self-Distilled Reasoning Models

Researchers Fei Wu, Zhenrong Zhang, and Jianshu Zhang introduced VASPO, a new training method that replaces the local token-level preferences used in on-policy self-distillation with a privileged value estimate of the expected return from an unguided student continuation. Existing approaches like GRPO combine self-distillation signals with reinforcement learning but suffer from a mismatch: the self-teacher's dense supervision reflects which tokens look good locally, not which full reasoning paths actually lead to correct answers. VASPO corrects this by training a value model that predicts the outcome of the student's own continuation, then using that value to weight the distillation loss. This yields denser and more accurate gradients for reasoning tasks such as math and code generation. The method is evaluated on standard reasoning benchmarks, showing improved performance over GRPO-based self-distillation baselines.

breakthroughs10/15/2026

Loading AI Digest

Overview

Best For

News & Updates

New Method Fixes On-Policy Distillation When Student Models Go Off Track

New RL Method Cuts Language Agent Training Costs by Learning From Every Step

VASPO Fixes Flawed Reward Signal in Self-Distilled Reasoning Models