---
title: "Active Learning: The Architecture of a Label-Budget Loop"
description: "Active learning is a loop that spends a fixed labeling budget on the samples most worth annotating. This diagram maps its stages — pool, score, label, train, evaluate — and the return path that closes the cycle."
---

# Active Learning: The Architecture of a Label-Budget Loop

**Type:** [Architecture](https://yomu.fyi/visual-explainers/category/architecture.md)

![A five-layer active learning loop diagram: an unlabeled token pool flows into a two-stage scorer, then a label budget gate, then a model update with a frozen backbone and trainable head, then an evaluation panel, with a return arrow labeled next batch.](https://yomufyi.s3.us-east-1.amazonaws.com/visual-explainers/715e64fe-1b9b-4512-be7a-e94a1605b7c3/image-16722d78-9031-4aec-96cc-f9fd7629b48b.webp)

> Active learning is a loop that spends a fixed labeling budget on the samples most worth annotating. This diagram maps its stages — pool, score, label, train, evaluate — and the return path that closes the cycle.

## Explanation

Active learning treats labels as a scarce budget rather than something to spend evenly. The architecture is a loop: an unlabeled pool feeds a scorer that picks a diverse batch and then refines it with uncertainty, a gate caps how many labels the round may consume, the head trains over a frozen backbone, and an evaluation layer decides whether the round earned its keep. The loop then returns to the pool for the next batch. This shape matters because label cost, not model capacity, is usually the binding constraint; keeping human effort proportional to what the model cannot infer is the whole point.

## FAQ

### Why does the loop score for diversity before uncertainty?

Selecting only the most uncertain samples tends to pull in near-duplicates and outliers, so the batch stops teaching the model anything new. TEAL's two-level MaxHerding first picks representative candidates from multi-layer token descriptors, then refines them with margin-based uncertainty on the decoder grid, keeping the batch both diverse and targeted at real confusion.

### How small can the label budget get before active learning stops paying?

TEAL reaches 97% of fully supervised segmentation quality with just 10 labeled pixels per image, and beats the strongest baseline by up to 24.36 mIoU with a single labeled pixel. Below that, results depend on how well the scorer's frozen representation already matches the task, so the ceiling is set by the backbone, not the loop.

### How do you keep the loop from trusting its own bad labels?

Separate the evaluation from the training signal. Salesforce's Agentforce team uses layered evaluation with an agent-as-a-judge plus human oversight, deliberately using different models to review work and avoid self-preference bias, while AlpaJudge supplies human-calibrated reward signals for selecting training data.

**Tags:** [Active Learning](https://yomu.fyi/topic/active-learning.md), [Human-in-the-Loop](https://yomu.fyi/topic/human-in-the-loop.md), [Machine Learning](https://yomu.fyi/topic/machine-learning.md), [MLOps](https://yomu.fyi/topic/mlops.md), [Model Evaluation](https://yomu.fyi/topic/model-evaluation.md), [Semantic Segmentation](https://yomu.fyi/topic/semantic-segmentation.md)

## Related reading

### [Your Algorithm: Giving People More Ways to Personalize Facebook Recommendations](https://yomu.fyi/post/your-algorithm-giving-people-more-ways-to-personalize-facebook-recomme.md)
- Company: [Meta](https://yomu.fyi/company/meta.md)
- Author: Jiarui Wang
- Published: Nov 10, 2026
- Type: Explainer

Traditional Facebook recommendation systems learn from implicit engagement signals, which adapt slowly and capture little nuance. Your algorithm is a suite of AI-powered tools that lets people state their preferences in natural language to personalize Feed and Reels. The end-to-end system includes profile generation, preference management, ranking, and delivery optimization, with supplementary tail loads and cache invalidation for responsiveness. For interests added to ‘Interested In,’ four dedicated retrieval sources find matching content and a control layer spaces out recommendations and caps daily contribution, while for ‘Not Interested In,’ layered filtering uses embedding comparison, demotion rules, an LLM reranker, and device cache invalidation. LLMs generate the initial profile and Lately Summary, validate each edit for coherence and duplicates, and provide a final reranking check against ‘Not Interested In’ interests, with a separate LLM acting as a judge for evaluation.


### [Running AV Reasoning Models Within Onboard Compute and Latency Limits](https://yomu.fyi/post/running-av-reasoning-models-within-onboard-compute-and-latency-limits.md)
- Company: [NVIDIA](https://yomu.fyi/company/nvidia-developer-blog.md)
- Author: Marco Pavone
- Published: Nov 10, 2026
- Type: Announcement

NVIDIA Alpamayo 2 Edge is an adaptive reasoning vision-language-action model designed for real-time operation on NVIDIA DRIVE AGX Thor. It uses a 2B Nemotron-based reasoning backbone paired with a diffusion action expert, processing surround-camera observations, temporal context, and navigation inputs to generate Chain-of-Causation reasoning and driving trajectories. The model reasons adaptively, producing tokens during routine driving to conserve inference compute. Post-trained with closed-loop reinforcement learning in NVIDIA AlpaGym, it improves closed-loop driving performance by 48.9% over the best checkpoint without closed-loop RL. NVIDIA AlpaJudge provides human-calibrated evaluation and reward signals for measuring reasoning quality and selecting training data.


### [Mitra-v2: Scaling synthetic pretraining for tabular foundation models](https://yomu.fyi/post/mitra-v2-scaling-synthetic-pretraining-for-tabular-foundation-models.md)
- Company: [AWS](https://yomu.fyi/company/aws.md)
- Author: Chris (Luyang) Kong
- Published: Nov 10, 2026
- Type: Announcement

Mitra-v2 is a second-generation tabular foundation model for classification and regression, pretrained entirely on synthetic tables. It uses a Transformer that attends across both rows and columns, and is pretrained on tables with 10 times more rows and roughly three times more columns than its predecessor. A new hybrid SCM prior combines neural networks, decision trees, Gaussian processes, and autoregressive processes within a single causal graph. On the TabArena benchmark of 51 real-world datasets, Mitra-v2 leads the single-model field and completes all datasets within the one-hour budget, which Mitra-v1 failed to do on more than a third. The model is released under Apache-2.0 on Hugging Face, with fine-tuned versions and inference code.


### [Decision Models in Hugging Face](https://yomu.fyi/post/decision-models-in-hugging-face.md)
- Company: [Hugging Face](https://yomu.fyi/company/hugging-face.md)
- Author: Aritra Roy Gosthipaty, Lysandre, Pedro Cuenca, Tom Aarsen
- Published: Nov 10, 2026
- Type: Explainer

Decision models evaluate a state against typed questions—choice, score, or noul—and return a probability distribution over predefined answers without generating text. Unlike traditional classifiers, they accept runtime-defined labels, and unlike zero-shot classifiers, they handle multiple typed questions per request. The interface is State + Questions + Allowed Answers → Decisions + Probabilities, with models like JEV-9B available on the Hugging Face Hub under a dedicated tag, and the ecosystem includes encoder-based and generative backbone models. Hugging Face’s transformers library supports inference via a pipeline and fine-tuning via the Trainer API, with a toy example fine-tuning JEV-9B to route support requests to billing, shipping, or technical categories using cross-entropy loss. Decision models make decisions and probabilities first-class outputs, optimized for reliable, calibrated decisions in software, and are designed to be more efficient than prompting LLMs for structured tasks.


### [Using AI to Build AI: How Agentforce Engineers Scale Automation Without Sacrificing Reliability](https://yomu.fyi/post/using-ai-to-build-ai-how-agentforce-engineers-scale-automation-without.md)
- Company: [Salesforce](https://yomu.fyi/company/salesforce.md)
- Author: Fereydoon Vafaei
- Published: Nov 9, 2026
- Type: Problem & solution

The Agentforce team at Salesforce faced the challenge of using AI to help build and maintain the Atlas Reasoning Engine without compromising reliability. The central risk was that AI could diagnose problems, modify the engine, test fixes, and review its own work while being confidently wrong. To address this, the team built an integrated engineering knowledge base using semantic retrieval to connect fragmented knowledge across incidents, fixes, and discussions. They implemented layered AI evaluation with an agent-as-a-judge model and human oversight, using different models to review work from different perspectives to avoid self-preference bias. As a result, typical bug-resolution time dropped from two to four weeks to one or two days, and system availability reached 100% in weekly reports during the observed period.


### [Pixels to Tokens: Token Space Efficient Active Learning for Low-Budget Semantic Segmentation](https://yomu.fyi/post/pixels-to-tokens-token-space-efficient-active-learning-for-low-budget.md)
- Company: [Apple](https://yomu.fyi/company/apple.md)
- Published: Oct 21, 2026
- Type: Problem & solution

Current active learning strategies for vision transformers in semantic segmentation require extensive pixel labeling, impractical when only one pixel per image is labeled. TEAL is the first pixel-level active learning framework built around ViT representations, using a frozen backbone and diversity-based selection directly on the native token lattice to avoid interpolation artifacts. It applies a two-level MaxHerding strategy over multi-layer token descriptors for representative candidate selection, followed by margin-based uncertainty refinement on a decoder grid. Across CamVid, Cityscapes, ADE20K, and Pascal Context, TEAL improves over the strongest baseline by up to +24.36 mIoU with a single labeled pixel per image and reaches 97% of fully supervised performance with just 10 pixels. The improvements persist when all baselines use the same DINOv3 backbone, showing that token-space selection with decoder-grid refinement makes ViT token geometry effective for low-budget active segmentation.
