Research map / Probabilistic and scientific modeling / Molecules and proteins
Trajectory-Aware Heuristic Learning for Combinatorial Search
Mustafa Seddiqi, Marta Kersten-Oertel, Tiberiu Popa
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 |
|---|---|---|---|
| use anchored HMM instead of step-based supervision with anchors (ablation) | anchored HMM refinement · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.413 vs. 0.399 (+1.4 pp) | Improved |
| use learned emissions instead of uniform emissions in the HMM (ablation) | anchored HMM refinement · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.413 vs. 0.404 (+0.9 pp) | Improved |
| add HMM forward-backward inference over prior-sampled pseudo-labeling (ablation) | anchored HMM refinement · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.413 vs. 0.381 (+3.2 pp) | Improved |
| add neighbor consistency to goal-rooted restarts (ablation) | CayleyPy-style restarts · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.414 vs. 0.400 (+1.4 pp) | Improved |
| add neighbor consistency to goal-rooted restarts (ablation) | CayleyPy-style restarts · DeepCubeA HTM beam 2^18 | search optimality: 62.9% vs. 59.6% (+3.3 pp) | Improved |
| add neighbor consistency to single-step trajectory pipeline (ablation) | single-step pipeline · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.422 vs. 0.382 (+4.0 pp) | Improved |
| add neighbor consistency to single-step trajectory pipeline (ablation) | single-step pipeline · DeepCubeA HTM beam 2^18 | search optimality: 63.3% vs. 58.1% (+5.2 pp) | Improved |
| add neighbor consistency to scalar regression (ablation) | regression objective · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.418 vs. 0.399 (+1.9 pp) | Improved |
| add classification and KL terms to regression plus neighbor consistency (ablation) | training objective · CubeLocalRank level-1 (HTM) | Aggregate Rank@1: 0.419 vs. 0.418 (+0.1 pp) | No clear effect |
| scale training data from 10B to 52B | Ours · DeepCubeA HTM beam 2^18 | optimality: 63.3% -> 68.1% | Improved |
| scale training data from 10B to 52B | CayleyPy-style · DeepCubeA HTM beam 2^18 | optimality: 59.6 -> 62.3 | Improved |
| scale training data from 10B to 52B | Ours · DeepCubeA QTM beam 2^18 | optimality: 77.8% -> 80.5% | Improved |
Closest papers by meaning
- Combinatorial Optimization with Policy Adaptation using Latent Space Search (NeurIPS 2023 · similarity 0.86)
- RefineEvo: Planning-Guided Heuristic Evolution with Bidirectional Experience (ICML 2026 · similarity 0.86)
- Ballad: Bandit-Based LLM Routing for Automated Heuristic Discovery (NeurIPS 2026 · similarity 0.85)
- Look-Ahead with Mini-Bucket Heuristics for MPE (AAAI 2016 · similarity 0.84)
- PathWise: Planning through World Model for Automated Heuristic Design via Self-Evolving LLMs (ICML 2026 · similarity 0.84)
- Interpreting Neural Combinatorial Optimization via Evolving Programmatic Bottlenecks (NeurIPS 2026 · similarity 0.84)
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.
