Research map / Probabilistic and scientific modeling / Molecules and proteins
Online Feedback Efficient Active Target Discovery in Partially Observable Environments
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes DiffATD, a training-free active target discovery method that uses diffusion-model particles as a belief distribution to balance entropy-based exploration with reward-model-guided exploitation under a fixed measurement budget in partially observable environments.
Problem
Active target discovery in partially observable environments under a limited sampling budget, where existing learned policies need large pre-labeled supervised datasets.
Approach
Uses reverse diffusion particles with Tweedie estimates to form a belief distribution over unobserved cells; scores measurement locations by an entropy-based exploration term plus a reward-weighted expected log-likelihood exploitation term, where the reward model is trained online from measurements and the mix is budget-dependent.
Key results
- On DOTA with budget 250, DiffATD achieves SR 0.6379 vs GA 0.5961 and Max-Ent 0.5524.
- On skin disease discovery with budget 150, DiffATD reaches SR 0.9061 vs GA 0.8396.
- With budget 300 on DOTA, DiffATD (0.7309) outperforms GPT-4o (0.5678) and Gemini (0.6453), and is comparable to fully observable SAM-based FullSEG on non-rare targets (0.7309 vs 0.8731 on the reported column, which is ambiguous).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use DiffATD's combined exploration–exploitation policy instead of the strongest listed baseline | Active target discovery · DOTA, budgets 200–300 | SR: 7.01% to 17.23% | Helped |
| Use DiffATD's combined exploration–exploitation policy instead of Greedy-Adaptive | Active target discovery · Coccinella Septempunctata, budgets 100–200 | SR: 8.48% to 13.86% | Helped |
| Use DiffATD's combined exploration–exploitation policy instead of Greedy-Adaptive | Active target discovery · Skin disease, budgets 150–250 | SR: 7.92% to 10.18% | Helped |
| Use DiffATD's combined exploration–exploitation policy instead of the best baseline | Active target discovery · Bone suppression, budgets 200–300 | SR: 41.08% to 59.22% | Helped |
| Use DiffATD instead of fully observable FullSEG | Active target discovery · Rare targets, budget 300 | SR: 0.6221 -> 0.8642 | Helped |
| Use DiffATD instead of fully observable FullSEG | Active target discovery · Vehicles/planes, budget 300 | SR: 0.8731 -> 0.7309 | Hurt |
| Set α=0.2 rather than α=1.0 (stronger exploitation weighting) (ablation) | DiffATD · DOTA, budget 200 | SR: 0.5422 -> 0.5052 | Hurt |
| Set α=5.0 rather than α=1.0 (stronger exploration weighting) (ablation) | DiffATD · DOTA, budget 200 | SR: 0.5422 -> 0.4823 | Hurt |
| Set α=0.2 rather than α=1.0 (stronger exploitation weighting) (ablation) | DiffATD · Skin, budget 200 | SR: 0.8974 -> 0.8465 | Hurt |
| Set α=5.0 rather than α=1.0 (stronger exploration weighting) (ablation) | DiffATD · Skin, budget 200 | SR: 0.8974 -> 0.8782 | Hurt |
Surprising findings
- Greedy-Adaptive exploitation alone overfits noise early in the search, motivating the exploration-first schedule.
- DiffATD outperforms the fully supervised SAM-based FullSEG on rare skin-disease targets while not on common vehicle/plane targets.
Limitations
- Exploitation score unreliable early when reward model is poorly trained; method relies on exploration first.
- The kappa(B) schedule is a design choice and sensitive to alpha (performance drops at alpha=0.2 and 5.0).
Experimental setup
Datasets: DOTA iNaturalist (Coccinella septempunctata) Skin imaging dataset Chest X-ray dataset CelebA
Benchmarks: DOTA iNaturalist Coccinella septempunctata Skin imaging Chest X-ray CelebA
Metrics: success rate (SR)
Baselines: Random Search Max-Ent Greedy-Adaptive FullSEG (SAM) GPT-4o Gemini
Builds on: Tweedie's formula Max-Ent active sampling denoising diffusion models
Closest papers by meaning
- Active Target Discovery under Uninformative Priors: The Power of Permanent and Transient Memory (NeurIPS 2025 · similarity 0.92)
- Tempered Guided Diffusion (NeurIPS 2026 · similarity 0.85)
- Provable Maximum Entropy Manifold Exploration via Diffusion Models (ICML 2025 · similarity 0.85)
- Fast Direct: Query-Efficient Online Black-box Guidance for Diffusion-model Target Generation (ICLR 2025 · similarity 0.84)
- Adapting Actively on the Fly: Relevance-Guided Online Meta-Learning with Latent Concepts for Geospatial Discovery (NeurIPS 2026 · similarity 0.83)
- ALINE: Joint Amortization for Bayesian Inference and Active Data Acquisition (NeurIPS 2025 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.