acceptodds
Under review as a conference paper at ICLR 2027

Task-Aware Self-Evolution for Resource-Constrained AI Appliances

Abstract

How should a resource-constrained AI appliance learn to serve its tasks better? For multi-step agents, the work to be served is itself an outcome of execution: model responses trigger tools, change context, and determine subsequent requests. Optimizing isolated inference requests therefore leaves a gap between system performance and useful task delivery. We present a task-aware self-evolution framework that addresses this gap through a persistent cycle of execution, outcome modeling, and resource decisions. The framework couples a conditional three-level action space—instance structure, inference configuration, and serving policy—with task-level probabilistic feedback and cost-aware experiment selection. A cohort model separates task heterogeneity, configuration effects, and batch variation, allowing incomplete executions to inform later decisions. We instantiate the framework on a three-node AI appliance and study 54 experimental attempts on 50 Office tasks. The study reveals quality–time trade-offs, substantial queueing exposure, and changing execution work even under the same nominal configuration. Together, the system design and empirical findings establish a task-centered approach to adaptation: learn from what the appliance delivers, reason about what another experiment would teach, and adjust how resources serve the next task cohort.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Task-aware systems, Self-evolution, LLM agents, Inference serving, Resource allocation, Cost-aware Bayesian optimization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_WjAMOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes a "task-aware self-evolution" framework to make resource-constrained AI appliance better serve its tasks, which is composed by three-level action space with task-level probabilistic feedback and cost-aware experiment selection. The author uses 3 node AI appliance to study limited experiments on 50 office tasks and report observations on the trade-offs of quality and time. However, the problem is not formulated coherently: the research question is posed informally, and the components are introduced separately without a unifying objective. More importantly, the proposal is never systematically evaluated.

Critical strengths

None

Critical weaknesses
  1. The problem is not formulated coherently.

    • The research problem is only presented informally in section 2 and 3; the components introduced by equation 1-5 are introduced without a unifying objective
    • The notation is inconsistent, and some symbols (e.g., n_g(i), n_e(t), η) are undefined.
  2. The proposed framework is not evaluated.

    • There are no baselines, no ablations, and no evidence that the outcome model or selection rule works. The only study introduced is an observational analysis of 54 development runs with changing code versions, on a single fixed set of 50 tasks.
Questions for authors
  1. Problem formulation

    • Q1: What is the objective that this self-evolution framework optimizes, i.e., the total budget, stopping rule, and how success of the learning process is measured?
    • Q2: Please define n_g(i), n_e(t), and η, and clarify the reused symbols (e.g., T for task cohort, delivery time, and sampling temperature; H for instance structure and time horizon).
  2. Evaluation

    • Q3: Can the authors provide baselines, component ablations, and results on held-out tasks beyond the fixed 50-task set?
Reviewer_kAehOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a task-aware self-evolution framework for optimizing AI appliance deployment and serving decisions under resource constraints. It combines a three-level configuration space, probabilistic task-outcome modeling, and cost-aware experiment selection to optimize task delivery rather than individual inference requests. A study involving 54 experimental attempts on 50 Office tasks demonstrates quality–latency trade-offs, execution-dependent workloads, and substantial queueing delays.

Critical strengths

N/A

Critical weaknesses
  • My main concern is that the proposed self-evolution loop is not actually validated. Section 4.1 explicitly states that the historical experiments precede the integrated controller and that comparative learning gains are outside the study. Consequently, the experiments demonstrate why task-aware optimization might matter, but not whether the proposed learning and experiment-selection mechanisms improve resource decisions. This is a substantial gap between the paper's central contribution and its empirical validation.
  • The outcome model seems less task-aware than the motivation suggests. In Equation 3, task and family effects are additive offsets rather than interactions with configurations. This means the model has limited ability to learn that different task families benefit differently from particular serving configurations. Given that heterogeneous agent workloads motivate the framework, I think this modeling choice is somewhat restrictive.
Questions for authors
  • Can the authors demonstrate even a small closed-loop experiment in which the proposed controller improves task delivery over a simple configuration-search baseline under matched experimental cost?
  • Why does Equation 3 exclude task-family configuration interactions? How would the predictor distinguish configurations that benefit document-processing tasks but hurt data-processing tasks?
  • Equation 4 assigns zero utility whenever any task in the cohort misses the delivery deadline. Why is this all-or-nothing objective preferred over delivered quality credit, especially when you emphasize the value of partially completed workloads?
Reviewer_a7wnOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper claims to address resource constrained AI agent continual learning, through a sequence of model calls and tool uses. The evaluation stance proposed is based on task completion, and resource consumption.

The continual learning is applied to the model deployment structure and request policies. The LLM itself do not seem to be fine-tuned.

Results presented in the paper concern task completion in 50 WorkBuddy Office tasks through the CodeBuddy Code harness.

Critical strengths

Important problem with respect to agent autonomy.

Critical weaknesses

insufficient validation.

How does the proposed framework compare with other strategies, e.g., optimization-based. Experiments are unclear and there are no ablations about the different part contributions. The paper itself states "Final-controller comparative learning gains are outside this study", thus the study appears to not be able to demonstrate that the framework has a good performance.

Novelty.

The framework combines established techniques and there is no clear novelty in the paper.

Reproducibility.

No code is shared, the descriptions of the experiments are not enough for clear understanding and reproducibility.

Questions for authors

See weaknesses

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.