AI Paper Atlas

Research map / Learning theory and algorithms / Combinatorial and search optimization

LMask: Learn to Solve Constrained Routing Problems with Lazy Masking

Tianyou Li, Haijun Zou, JIAYUAN WU, Zaiwen Wen

ICLR 2026 · Learning theory and algorithms › Combinatorial and search optimization · approach: Unsupervised neural combinatorial optimization

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 changedWhereMeasuredEffect
add backtracking (LazyMask with TSL) instead of lookahead-only maskingLMask · hard TSPTW-100solution infeasibility rate: 31.74% (PIP) -> 0.00% (LMask)Improved
increase the backtracking budget R (ablation)LMask · hard TSPTW-100solution infeasibility rate and inference time: about 2 seconds extra inference time per 100 additional budgetImproved
backtracking with lightweight lookahead versus deeper lookahead (depth 2 to 3) for PIP (ablation)PIP and PIP-D · hard TSPTW-100solution infeasibility vs inference time: an order of magnitude rise in inference time for marginal gainsNo clear effect
add refinement intensity embedding (RIE) (ablation)LMask · medium TSPTW-50optimality gap and feasibilityImproved
raise entropy coefficient lambda from 0 to 0.01 (ablation)LMask · easy TSPTW-100optimality gapImproved
raise entropy coefficient lambda from 0.01 to 0.05 (ablation)LMask · easy TSPTW-100optimality gapHurt
use two-step lookahead (TSL) rather than single-step lookahead (SSL) for initialization (ablation)LMask · hard TSPTW-100solution infeasibility vs time: zero infeasibility within 30 seconds with TSLImproved
train with l1 penalty for violations once backtracking budget is reached (small R in training)LMask · TSPTW and TSPDLfeasibilityImproved
add destination-aware feasibility (PIP-D decoder) over PIPPIP · synthetic TSPTWfeasibility vs optimality gap: TSPTW-50 Medium gap 2.93% -> 3.31%Mixed
increase problem size on TSPTW benchmarkPIP and PIP-D · TSPTW benchmarkinfeasibility and gap: PIP-D infeasibility 5.0% (n=20) -> 66.7% (n=80)Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy