Research map / Image generation / Fast diffusion sampling
Plug-and-Play Guidance for Discrete Diffusion Models via Gradient-Informed Logit Correction
Hongkun Dou, Zike Chen, Fengji Li, Hongjue Li, Yue Deng
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| guide the clean-prediction logits instead of the noisy state (ablation) | GILC-DB · DNA diffusion | mean reward | Improved |
| use Monte Carlo value estimation (n>1 samples) instead of deterministic estimation (ablation) | GILC value function · DNA diffusion value-estimation error | L1 error | Improved |
| drop the model Jacobian (Jacobian-free logit correction) (ablation) | GILC · DNA diffusion | guidance stability: Jacobian condition number about 10^4-10^5 | Improved |
| use GILC-DB instead of DRAKES fine-tuning | discrete diffusion DNA design · HepG2 enhancer design | Pred-Activity: 5.61 -> 7.04 | Improved |
| use GILC-DB instead of SVDD | discrete diffusion DNA design · HepG2 enhancer design | ATAC-Acc (%): 51.9 -> 95.2 | Improved |
| use GILC-DB instead of DRAKES | inverse protein folding · Success rate | Success Rate (%): 78.6 -> 82.4 | Improved |
| use GILC-DB instead of DRAKES | DNA design · HepG2 enhancer design | App-Log-Lik: -264 -> -267 | Hurt |
| use GILC-PG (policy gradient) instead of GILC-DB (ablation) | GILC · HepG2 enhancer design | Pred-Activity: 7.04 -> 5.21 Pred-Activity; 0.900 -> 0.910 3-mer Corr | Mixed |
| use GILC-PG instead of GILC-DB (ablation) | GILC · QM9 guidance | MAE: Cv 2.23 -> 1.53; ϵLUMO 988 -> 875 | Mixed |
| use GILC-DB (differentiable reward) instead of multi-trajectory baselines | DNA design · reward evaluations per step | reward calls: 20 -> 5 | Improved |
| remove KL regularization from DRAKES (ablation) | DRAKES · inverse protein folding | Success Rate (%): 78.6 -> 34.1 | Hurt |
Closest papers by meaning
- Simple Guidance Mechanisms for Discrete Diffusion Models (ICLR 2025 · similarity 0.89)
- Derivative-Free Guidance in Continuous and Discrete Diffusion Models with Soft Value-based Decoding (NeurIPS 2025 · similarity 0.89)
- D-DOIT: Training-free Adaptation of Discrete Diffusion via Doob's h-Transform (NeurIPS 2026 · similarity 0.89)
- Token Perturbation Guidance for Diffusion Models (NeurIPS 2025 · similarity 0.88)
- Reward-Guided Iterative Refinement in Diffusion Models at Test-Time with Applications to Protein and DNA Design (ICML 2025 · similarity 0.88)
- Training-Free Guidance Beyond Differentiability: Scalable Path Steering with Tree Search in Diffusion and Flow Models (NeurIPS 2025 · similarity 0.88)
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.
