Research map / Continual learning / Online and replay-based
CONCUR: A Framework for Continual Constrained and Unconstrained Routing
Peter Baile Chen, Weiyue Li, Dan Roth, Mike Cafarella, Samuel Madden, Jacob Andreas
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 |
|---|---|---|---|
| replace best single strategy with CONCUR routing | Qwen2.5-7B CoT · Unconstrained routing, ID avg (2WikiMultiHop/MMLU/GSM8k) | accuracy / FLOPs: 74.3 -> 75.2 acc; 41.63 -> 36.50 FLOPs | Improved |
| replace best single strategy with CONCUR routing | Qwen2.5-7B CoT · Unconstrained routing, OOD avg (HotpotQA/GPQA/SVAMP) | accuracy / FLOPs: 62.3 -> 62.6 acc; 46.13 -> 43.23 FLOPs | Improved |
| use RouteLLM single-classifier routing instead of best single strategy | RouteLLM · Unconstrained routing, ID avg (2WikiMultiHop/MMLU/GSM8k) | accuracy: 74.3 -> 53.5 | Hurt |
| global DP optimization instead of local per-task budget (CONCUR, low budget) (ablation) | CONCUR · Constrained routing ID avg, FLOPs budget 25 | accuracy: +5.7 average | Improved |
| global instead of local optimization (CONCUR, high budget) (ablation) | CONCUR · Constrained routing ID avg, FLOPs budget 40 | accuracy: +4.0 average | Improved |
| global instead of local optimization (RTR, GSM8k, high budget) (ablation) | RTR · GSM8k, FLOPs budget 40 | accuracy: +0.2 | No clear effect |
| add general-purpose representation to task-specific only (ablation) | CONCUR · Unconstrained routing, ID avg, representation ablation | accuracy: 72.3 -> 75.2 | Improved |
| add task-specific representation to general-purpose only (ablation) | CONCUR · Unconstrained routing, ID avg, representation ablation | accuracy: 73.3 -> 75.2 | Improved |
| modular per-strategy predictors instead of retraining single router when adding small models | CONCUR · Continual routing Setting 2 (large+small models), ID avg | relative training time: 1.00x vs 7.66x (RTR_FS) | Improved |
| fine-tune existing router on 25-100% of new data instead of full retrain | RTR/EmbedLLM baselines · Continual routing Setting 2 (large+small models), ID avg | accuracy: RTR_FT(75%) 74.7 vs RTR_FT(50%) 71.9 | Mixed |
| use larger Setting 2 (add small models) for CONCUR | CONCUR · Continual ID avg FLOPs | FLOPs: 40.84 -> 36.50 | Improved |
Closest papers by meaning
- Dynamic Latent Routing (NeurIPS 2026 · similarity 0.88)
- SkillOrchestra: Learning to Route Agents via Skill Transfer (NeurIPS 2026 · similarity 0.88)
- LINC: Decoupling Local Consequence Scoring from Hidden Matching in Constructive Neural Routing (NeurIPS 2026 · similarity 0.87)
- Memory-Enhanced Neural Solvers for Routing Problems (NeurIPS 2025 · similarity 0.87)
- Learning to Solve Compositional Geometry Routing Problems (NeurIPS 2026 · similarity 0.87)
- Models Under SCOPE: Scalable and Controllable Routing via Pre-hoc Reasoning (ICML 2026 · similarity 0.86)
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.
