Constrained Network Slice Assignment via Large Language Models¶

Conference: NeurIPS 2025 arXiv: 2512.00040 Code: None Area: Optimization Keywords: Network slicing, 5G resource allocation, LLM, integer programming, constrained optimization

TL;DR¶

This paper investigates the use of LLMs (Claude series) for solving constrained optimization problems in 5G network slice resource allocation. Two approaches are proposed: zero-shot LLM direct assignment and LLM-guided integer programming. Empirical findings show that LLMs alone can produce reasonable initial allocations but may violate hard constraints, whereas combining LLMs with an ILP solver achieves 100% completeness and balanced utilization.

Background & Motivation¶

Background: 5G network slicing creates isolated virtual networks over shared infrastructure to support diverse service requirements, including eMBB (high bandwidth), URLLC (low latency), and mMTC (massive IoT connectivity).

Limitations of Prior Work: Traditional approaches rely on large-scale optimization or heuristic algorithms that are computationally intensive and require expert-level parameter tuning. ILP faces combinatorial explosion as problem size grows.

Core Idea: Leverage the semantic comprehension capabilities of LLMs to interpret service request characteristics and generate initial assignments, followed by ILP-based exact solving to guarantee constraint satisfaction.

Method¶

Overall Architecture¶

The system encompasses three slice types (eMBB for high bandwidth, URLLC for low latency, mMTC for massive connectivity) and a set of user service requests, each characterized by resource demands and latency requirements. The problem is formulated as a constrained assignment problem, where the decision variable \(x_{i,m}\) indicates whether request \(i\) is assigned to slice \(m\).

Key Designs¶

Zero-Shot LLM Assignment:
- Function: Directly generates slice assignment solutions.
- Mechanism: Claude models are provided with slice attributes (capacity, latency) and a list of requests (demand, latency requirements), and are prompted to output assignments in CSV format. The "@" symbol is used as a delimiter to avoid comma ambiguity, and request ordering is shuffled to prevent positional bias.
- Design Motivation: The semantic understanding of LLMs can naturally map "high-bandwidth video streams → eMBB slice" and "low-latency control signals → URLLC slice."
- Limitation: At temperature 0.8, outputs exhibit variability and may violate hard capacity or latency constraints.
LLM-Guided ILP Hybrid Method:
- Function: Precisely solves constrained assignments while leveraging semantic information from LLMs.
- Mechanism: Claude first evaluates the pairwise compatibility of all requests to produce a binary similarity matrix. An ILP is then formulated to maximize the total similarity score of co-assigned request pairs within the same slice. The auxiliary variable \(z_{i,j,m}\) links the joint assignment of request pairs, ensuring all hard constraints (capacity, latency, full assignment) are satisfied.
- Design Motivation: LLMs excel at understanding service semantics (determining which requests "belong" in the same slice), while ILP excels at exact constraint enforcement.
Evaluation Metrics:
- Completeness: Whether all requests are assigned.
- Homogeneity: Consistency of request types within each slice.
- Bandwidth/Density Utilization: Degree of balanced resource usage across slices.

Loss & Training¶

No training is involved. LLMs operate via zero-shot inference, and the ILP is solved using the GLPK solver.

Key Experimental Results¶

Main Results¶

Method	Completeness	Homogeneity	Constraint Satisfaction	Notes
Claude-3-haiku	100%	0.35±0.27	Occasional violations	Poor homogeneity
Claude-3-sonnet	99.5%	1.00±0.00	Occasional violations	Perfect homogeneity
ILP+LLM	100%	High	100%	Most balanced solution

Ablation Study¶

Model Variant	Bandwidth Utilization	Density Utilization	Notes
claude-3-haiku	Severely imbalanced	Slice C overloaded	Weaker semantic understanding
claude-3-sonnet	90–100% balanced	Balanced	Best standalone LLM performance
claude-3.5-sonnet	Moderately balanced	Moderate	Intermediate performance

Key Findings¶

Claude-3-sonnet achieves a perfect homogeneity score of 1.0, demonstrating that its semantic understanding can reliably distinguish between different service request types.
Constraint violations in standalone LLM usage primarily arise from slice capacity overflow, as LLMs cannot accurately track cumulative resource consumption.
The ILP+LLM hybrid combines the semantic clustering capability of LLMs with the constraint guarantees of ILP, yielding the optimal solution.

Highlights & Insights¶

The LLM–solver hybrid architecture represents a promising paradigm: LLMs provide semantic understanding to reduce the search space, while solvers guarantee constraint satisfaction. This pattern is generalizable to other constrained optimization problems.
The paper provides clear empirical analysis of LLM capability boundaries in structured decision-making tasks: strong in semantic understanding, weak in tracking numerical constraints.
The substantial performance variation across Claude variants highlights the significant impact of model selection on task quality.

Limitations & Future Work¶

The study relies on synthetic datasets and has not been validated on real 5G network deployments.
Only static, one-shot allocation is addressed; dynamic and time-varying scenarios are not considered.
The problem scale is relatively small (~20 requests), and scalability remains to be verified.
LLM similarity judgments use binary values; continuous-valued similarity scores may yield greater precision.
The potential of other LLMs (e.g., GPT-4) as semantic engines has not been explored.

Rating¶

Novelty: ⭐⭐⭐ The LLM+ILP hybrid approach is relatively novel in the networking domain.
Experimental Thoroughness: ⭐⭐⭐ Synthetic data with limited scale.
Writing Quality: ⭐⭐⭐ Clear but lacking in depth.
Value: ⭐⭐⭐ An exploratory work demonstrating the potential of the hybrid approach.