Research map / Continual learning / Continual fine-tuning of LLMs
Incremental Learning of Retrievable Skills For Efficient Continual Task Adaptation
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 IsCiL, an adapter-based continual imitation learning framework that incrementally learns retrievable skill prototypes with LoRA adapters on a diffusion base policy, enabling knowledge sharing, sample-efficient adaptation, and task unlearning on Franka-Kitchen and Meta-World streams.
Problem
Continual imitation learning must accumulate task knowledge from staged, incomplete expert demonstrations without rehearsal, while avoiding catastrophic forgetting and supporting privacy-driven unlearning; existing adapter-based methods isolate parameters and limit knowledge sharing across demonstra
Approach
Map (observation, sub-goal) states into a fixed embedding space and retrieve skills via multi-basis prototypes (KMeans centroids) that index LoRA adapters on a diffusion base policy; new skills are initialized from the most-used existing adapter and incrementally updated, so skills are shared across stages while unlearning is done by deleting a prototype and its adapter.
Key results
- On Evolving Kitchen-Complete, IsCiL achieves AUC 89.8% vs. 86.2% for TAIL-tau and 92.3% for the multi-task oracle.
- On Evolving World-Semi, IsCiL achieves AUC 68.9% vs. 48.2% for the best adapter baseline L2M-g.
- In unseen-task adaptation on Evolving Kitchen, IsCiL reaches AUC-A 72.8% vs. 30.1% for TAIL-g.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use IsCiL instead of TAIL-$\tau$ | Evolving Kitchen Complete · Evolving Kitchen Complete | AUC (%): 86.2{\color{gray} $\pm$5.6} -> 89.8{\color{gray} $\pm$0.5} | Helped |
| Use IsCiL instead of TAIL-$\tau$ | Evolving Kitchen Semi · Evolving Kitchen Semi | AUC (%): 41.2{\color{gray} $\pm$2.5} -> 75.8{\color{gray} $\pm$1.8} | Helped |
| Use IsCiL instead of TAIL-$g$ | Evolving Kitchen Incomplete · Evolving Kitchen Incomplete | AUC (%): 35.4{\color{gray} $\pm$0.7} -> 74.0{\color{gray} $\pm$1.9} | Helped |
| Use IsCiL instead of TAIL-$\tau$ | Evolving World Complete · Evolving World Complete | AUC (%): 85.7{\color{gray} $\pm$5.9} -> 84.3{\color{gray} $\pm$1.1} | Hurt |
| Use IsCiL instead of L2M | Evolving World Semi · Evolving World Semi | AUC (%): 47.0{\color{gray} $\pm$0.7} -> 68.9{\color{gray} $\pm$0.5} | Helped |
| Use IsCiL instead of TAIL-$\tau$ | Evolving World Incomplete · Evolving World Incomplete | AUC (%): 39.7{\color{gray} $\pm$1.0} -> 71.2{\color{gray} $\pm$4.2} | Helped |
| use 1 prototype base instead of 20 (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 30.6{\color{gray} $\pm$12.2} | Hurt |
| use 5 prototype bases instead of 20 (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 66.5{\color{gray} $\pm$7.9} | Hurt |
| use 10 prototype bases instead of 20 (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 83.9{\color{gray} $\pm$2.7} | Hurt |
| use 25 prototype bases instead of 20 (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 88.2{\color{gray} $\pm$1.1} | Hurt |
| use 50 prototype bases instead of 20 (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 89.4{\color{gray} $\pm$1.2} | No clear effect |
| construct skills from full task trajectories rather than sub-goals; use 20 bases (ablation) | IsCiL · Evolving Kitchen Complete | AUC (%): 89.8{\color{gray} $\pm$0.5} -> 67.2{\color{gray} $\pm$17.2} | Hurt |
Surprising findings
- Seq-FT achieves the highest FWT in the Complete scenario yet the lowest BWT, leading to lower overall AUC.
- A single prototype basis (|chi_z|=1) performs similarly to L2M, showing that multi-basis prototypes are essential.
Limitations
- Adapter-based approaches add evaluation-time overhead from skill retrieval.
- Requires sub-goal sequences for training and evaluation.
- Adapter size and prototype size must be tuned to pre-trained model quality and task shift.
Experimental setup
Datasets: D4RL Franka-Kitchen mixed Meta-World heuristic expert demonstrations
Benchmarks: Evolving Kitchen Evolving World
Metrics: FWT BWT AUC goal-conditioned success rate
Baselines: Seq-FT Seq-LoRA EWC L2M L2M-g TAIL-g TAIL-tau Multi-task ER CLPU
Builds on: L2M TAIL LoRA EWC DDPM
Closest papers by meaning
- Policy Compatible Skill Incremental Learning via Lazy Learning Interface (NeurIPS 2025 · similarity 0.87)
- Learning from Demonstrations via Capability-Aware Goal Sampling (NeurIPS 2025 · similarity 0.86)
- CL-LoRA: Continual Low-Rank Adaptation for Rehearsal-Free Class-Incremental Learning (CVPR 2025 · similarity 0.86)
- Continual Knowledge Adaptation for Reinforcement Learning (NeurIPS 2025 · similarity 0.85)
- Interactive and Hybrid Imitation Learning: Provably Beating Behavior Cloning (NeurIPS 2025 · similarity 0.85)
- LLM-based Skill Diffusion for Zero-shot Policy Adaptation (NeurIPS 2024 · similarity 0.85)
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.