ALFWorld
Household interaction
62.98%
Final average success
+15.86 pp
vs Static Harness
Also +2.30 pp vs MemRL.
Plasticity-HCL10.94% average forgetting
Qwen3.5-9B
Harness Continual LearningContinual Adaptation Beyond Model Parameters
1 State Key Laboratory for Novel Software Technology, Nanjing University, China
2 University of Wollongong, Australia * Corresponding author
Continual learning in Minecraft
Zero-shot
Stops at task 22Video unavailableThe recording could not be loaded. Please try again later.
DeepSeek V4.1 Flash (reasoning effort: Max) fails at task 22, “Mine 3 coal,” and cannot proceed to subsequent tasks.
HCL
Completes all 50 tasksVideo unavailableThe recording could not be loaded. Please try again later.
With Harness Continual Learning (HCL), the same model and reasoning effort successfully complete task 22, where the zero-shot run failed, and continue through the remaining tasks, ultimately completing all 50 tasks.
Both videos are recorded in spectator mode, allowing the camera to see through blocks.
The harness becomes the continual learning state.
HCL makes prompts, memories, skills, and routing a persistent learning state around a frozen foundation model.
Changing that state can improve a new task and disrupt an earlier success. Frozen weights do not eliminate forgetting.
Figure 1. The learning object shifts from model parameters to the harness. Both settings face the acquisition–retention challenge.
Guarded harness evolution
The harness changes through feedback. Deployment changes only after evaluation.
Figure 2. The four harness components support execution; the Optimizer and Evaluator govern persistent updates.
What changes in the harness
Prompts, memories, skills, and routing rules are maintained as one evolving, jointly versioned state.
- Task Interface
- Prompts, task templates, and parsing rules.
- Experience Memory
- Interaction records and reusable guidance.
- Capability Map
- External tools and experience-derived skills.
- Adaptive Router
- Memory selection, capability use, and workflows.
How updates are accepted
The Continual Optimizer uses execution feedback to revise selected harness components. The Continual Evaluator checks the candidate before it can enter the deployed state.
Current improvementMeet the validation-gain threshold.
Historical retentionKeep earlier-behavior regressions within budget.
Task validityRespect output, tool-use, and environment constraints.
Commit an admissible candidate; otherwise, keep the current harness. The next interaction uses the committed state.
Continual learning in practice
Four settings. The same frozen model within each comparison.
Compare with a static harness in open-world tasks and with zero-shot inference in the controlled streams.
Minecraft
Long-horizon skill reuse
50/50
Curriculum tasks completed
+35 tasks
vs Static Harness
Fewer actions than both memory-based baselines.
Plasticity-HCL83 actions · MemRL 88 · MemP 91
deepseek-v4-flash-0731
Textual reasoning
Question answering & reasoning
76.10%
Final average accuracy
+12.45 pp
vs Zero-shot
Also +7.25 pp vs MemRL.
Plasticity-HCL0.40% average forgetting
DeepSeek-V4.1-Flash
Multimodal perception
Visual understanding
68.92
Final mean score
+22.15 pts
vs Zero-shot
Also +3.76 pts vs MemP.
Stability-HCL0.22% average forgetting
Qwen3.6-27B
Gains are absolute: pp = percentage points; pts = score points. Performance and forgetting should be read together. Metrics differ across settings.
Citation
@misc{kang2026harnesscontinuallearning,
title = {Harness Continual Learning: Continual Adaptation Beyond Model Parameters},
author = {Borui Kang and Jinrui Gu and Junhan Lv and Wenbin Li and Lei Wang and Yang Gao},
year = {2026},
eprint = {2608.19013},
archivePrefix = {arXiv},
primaryClass = {cs.LG},
url = {https://arxiv.org/abs/2608.19013}
}