Harness Continual Learning: Continual Adaptation Beyond Model Parameters

Source

• Source name: arXiv
• Source kind: paper
• Authors / organization: Borui Kang, Jinrui Gu, Junhan Lv, Wenbin Li, Lei Wang, Yang Gao
• Venue: arXiv: cs.LG
• Published: 2026-08-19
• Collected: 2026-08-20
• URL: http://arxiv.org/abs/2608.19013v1
• PDF: https://arxiv.org/pdf/2608.19013v1
• DOI: UNKNOWN
• Registry group: academic_apis_and_preprints
• Evidence role: canonical_academic_index
• Language: en
• Country / region: UNKNOWN

Plain-Language Brief

The paper presents a new paradigm for continual learning called Harness Continual Learning (HCL). In this approach, an agent can adapt through a harness of prompts, memories, tools, skills, and routing rules. The harness is designed to evolve around a frozen foundation model. This allows the agent to continually improve its state while retaining previously acquired behavior. The paper introduces four execution-facing components for HCL: Task Interface, Experience Memory, Capability Map, and Adaptive Router. The Continual Optimizer proposes candidate harnesses from post-execution feedback, and the Continual Evaluator commits the resulting candidate harness only after checking current improvement, historical retention, and validity. Experiments show capability accumulation and failure recovery in textual reasoning, multimodal perception, and open-world interaction.

What the Document Says

• The paper introduces Harness Continual Learning (HCL), a new continual learning paradigm that focuses on the evolution of an agent's harness, which includes prompts, memories, tools, skills, and routing rules.
• In this approach, the model parameters are frozen, but the harness can still adapt to improve the agent's state and retain previously acquired behavior.
• The paper presents four execution-facing components for HCL: Task Interface, Experience Memory, Capability Map, and Adaptive Router.
• A Continual Optimizer proposes candidate harnesses based on post-execution feedback, and a Continual Evaluator commits the candidate harness after checking current improvement, historical retention, and validity.
• The experiments demonstrate capability accumulation and failure recovery in textual reasoning, multimodal perception, and open-world interaction.

Key Concepts Mentioned

• Harness Continual Learning
• Continual learning
• Model parameters
• Prompts
• Memories
• Tools
• Skills
• Routing rules
• Task Interface
• Experience Memory
• Capability Map
• Adaptive Router
• Continual Optimizer
• Continual Evaluator

Claims Made by the Document

• HCL is a new paradigm for continual learning that allows an agent to continually improve its state while retaining previously acquired behavior.

Evidence Used

• The paper introduces the concept of HCL and presents its components and evaluation methods.

Methods / Approach

• The paper presents four execution-facing components: Task Interface, Experience Memory, Capability Map, and Adaptive Router. It also introduces a Continual Optimizer and a Continual Evaluator.

Limitations Stated by the Authors

UNKNOWN

Limitations Noted by the Agent

Agent interpretation — not from the source.

• The paper does not discuss the limitations of HCL in detail.

Open Questions

• What are the potential challenges and limitations of implementing HCL in real-world scenarios?
• How scalable is HCL for large datasets?

Notes for Future Q&A

May include interpretation. Labeled where it does.

• Discuss the advantages and disadvantages of using a harness-based approach for continual learning.
• Explain how the Continual Optimizer and Continual Evaluator work in practice.

Raw Abstract / Excerpt

Continual learning has largely been model-centric, treating model parameters as the state that changes with sequential experience. Modern agents can also adapt through a harness of prompts, memories, tools, skills, and routing rules. Because these contents jointly shape later execution, a harness update can disrupt previously reliable behavior even when the model is frozen. This raises a new question: how can an agent continually improve its state outside the model while retaining behavior acquired earlier? We formulate Harness Continual Learning (HCL), a new continual learning paradigm in which the harness evolves around a frozen foundation model, and define the resulting loss of earlier behavior as harness-level forgetting. We instantiate HCL with four execution-facing components: the Task Interface, Experience Memory, Capability Map, and Adaptive Router. We further introduce guarded harness evolution to separate update generation from state commitment. A Continual Optimizer proposes candidate harnesses from post-execution feedback, and a Continual Evaluator commits the resulting candidate harness only after checking current improvement, historical retention, and validity. Experiments on textual reasoning, multimodal perception, and open-world interaction demonstrate capability accumulation and failure recovery, with relative gains exceeding 10% over corresponding baselines in multiple settings. Component ablations assess the contribution of each harness component, while controlled retention sweeps reveal measurable harness-level forgetting and show that the stability--plasticity trade-off can be explicitly adjusted.