STEPS: Semantic Contract-Guided Scheduling for LLM-Assisted Natural Language-Driven Edge AI Services

arXiv:2606.09537 · cs.NI · Submitted 2026-06-08 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Computer Science. The day in computer science: software, systems and theory on arXiv.

Tom: We're looking at the paper: "STEPS: Semantic Contract-Guided Scheduling for LLM-Assisted Natural Language-Driven Edge AI Services".

Jane: Here's the gist of the paper. Edge user/service scheduling has become a cornerstone of distributed AI systems, governing where and how AI services are executed under limited communication and computing resources.

Tom: This paper proposes semantic contract–guided edge potential scheduling (STEPS), a natural language-driven scheduling framework that introduces semantic contracts as executable interfaces between user-side semantics and edge-side decision making.

Jane: The Core Problem Addressed Existing edge scheduling frameworks typically assume that service requirements are explicitly specified as numerical quantities/constraints, such as latency bounds or energy budgets.

Tom: In practice, however, users often express their service expectations through ambiguous and context-dependent natural language descriptions (e.g., “finish quickly” or “save energy”), creating a fundamental gap between user intent and scheduling decisions.

Paper summary: Tom: So, to wrap up the "STEPS: Semantic Contract-Guided Scheduling for LLM-Assisted Natural Language-Driven Edge AI Services" paper, this is a new framework that takes what we say in natural language about AI services—like "finish quickly" or "save energy"—and turns it into real, executable schedules at the edge <ref:2606.09537#pg1>.

Jane: Yeah, the core mechanism here is using semantic contracts as this bridge between what the user wants and what the hardware can actually do <ref:2606.09537#pg1>.

Lu: It uses an LLM to parse those vague requests into structured data, like delay sensitivity or energy sensitivity levels, and then turns that into a contract with preferences and fulfillment bounds <ref:2606.09537#pg1>.

Meng: From a deployment standpoint, this means the edge device doesn't just guess what the user wants; it has a formal agreement it has to keep <ref:2606.09537#pg1>.

Jane: And then, if things go wrong after the service runs—if they didn't meet the contract—the system learns from that failure and adjusts its next scheduling decision <ref:2606.09537#pg1>.

Tom: The results show this method handles those ambiguous requests in changing network conditions much better than some other approaches we’ve seen, like Param-Opt or DRL-Opt

Fig. two(c): <ref:2606.09537#pg1>.

Lu: What’s interesting is that STEPS keeps the service loss low even when the environment starts drifting over time

Fig. four(c): <ref:2606.09537#pg1>.

Meng: It seems robust because it's not just looking at one snapshot of the problem, but it's using that feedback to adapt the whole system continuously <ref:2606.09537#pg1>.

Jane: So this paper lays out a pipeline that goes from a vague sentence to an actual schedule and back again based on what happens in real-time <ref:2606.09537#pg1>.

Tom: It’s about taking the uncertainty of human language and giving it mathematical structure so the edge can actually manage resources effectively <ref:2606.09537#pg1>.

Lu: The theoretical analysis also confirms that there's a stable point where the system settles, which is good because we don't want this thing to just keep wildly changing parameters forever <ref:2606.09537#pg1>.

Jane: It’s interesting how the authors managed to tie together the language model part, the scheduling game theory, and the feedback loop all into one cohesive system <ref:2606.09537#pg1>.

Conclusion: Tom: So, we're wrapping up our look at STEPS, which is "Semantic Contract-Guided Scheduling for LLM-Assisted Natural Language-Driven Edge AI Services."

Jane: It's essentially a system that translates what a user says in plain language about their AI service into a formal contract the edge device can actually use to schedule things.

Tom: The authors, they built this whole pipeline from natural language requests all the way through to scheduling and then using feedback to adapt how it makes those scheduling decisions.

Lu: What's important here is how they grounded those vague human intents into something mathematical, like delay sensitivity or energy sensitivity levels for the AI service.

Meng: For us in production, this means if you can get a reliable way to turn a user's "finish quickly" into a measurable constraint that the scheduler understands, you get an AI system that doesn't just guess how to run on your edge hardware.

Jane: And they showed through their experiments that this approach keeps the service loss low even when the network conditions are constantly changing, which is super important for real-world deployments.

Tom: It sounds like they managed to handle the uncertainty of human language and turn it into a stable scheduling system that adapts in real time.

Lu: The theoretical analysis confirms there's a steady point where this whole system settles, which means we don't have to worry about the scheduling parameters just spinning out of control forever.

Jane: It suggests that future AI systems on the edge will need to be able to understand and respond directly to user needs in this kind of semantic way, rather than just being tied down by rigid numerical rules.

Tom: And it opens up a lot of questions about how much control users actually want over the scheduling process when they're talking to an AI service.

Shanghai Research Institute for Intelligent Autonomous Systems · Department of Electrical and Computer Engineering, Western University, Ontario Canada · Department of Electrical Engineering, University at Buffalo-SUNY

cs.NI

Submitted: 2026-06-08

Updated: 2026-11-01

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 75/100

The gist: Edge user/service scheduling has become a cornerstone of distributed AI systems, governing where and how AI services are executed under limited communication and computing resources.

Key concepts

Semantic Contract
These are executable interfaces created from natural language user requests. They combine preference weights, fulfillment bounds, and uncertainty measures derived from an LLM interpretation of the user's intent. They serve as the structured rules that bridge ambiguous human language and technical scheduling decisions.
Contract-Guided Potential Game
This is a mathematical model used to determine scheduling decisions (which node to use, how much resource to allocate). The objective is to minimize long-term service loss by formulating the problem as a potential game where the potential function aggregates service losses and penalizes edge congestion.
Feedback-Driven Adaptive Control
This mechanism allows the system to adapt its scheduling strategy in real-time, especially in changing environments. It uses feedback from actual service outcomes to distinguish between changes in user requests (semantic drift) and changes in contract fulfillment, adjusting parameters like admission thresholds and conservativeness factors.
Semantic Uncertainty Measure ($ ho(t)_i$)
This component quantifies the ambiguity or uncertainty inherent in the extracted semantic requirements from natural language. It is a key part of the semantic contract, helping the scheduler account for how much confidence exists in fulfilling a specific service request.

Terminology

Summary

Edge user/service scheduling has become a cornerstone of distributed AI systems, governing where and how AI services are executed under limited communication and computing resources. p. 0 This paper proposes semantic contract–guided edge potential scheduling (STEPS), a natural language-driven scheduling framework that introduces semantic contracts as executable interfaces between user-side semantics and edge-side decision making. p. 0

The Core Problem Addressed

Existing edge scheduling frameworks typically assume that service requirements are explicitly specified as numerical quantities/constraints, such as latency bounds or energy budgets. In practice, however, users often express their service expectations through ambiguous and context-dependent natural language descriptions (e. p. 0g., “finish quickly” or “save energy”), creating a fundamental gap between user intent and scheduling decisions. p. 0 Bridging this semantic-to-optimization gap is particularly challenging in dynamic edge environments with time-varying resources and execution conditions. p. 0 The framework addresses three research questions (RQs):

  1. How can unstructured natural language service intents be systematically grounded into a structured and constraint-consistent decision representation for E4NetAI? p. 0

  2. How can semantic contracts reshape the structure of edge scheduling under tightly coupled constraints and semantic uncertainty? p. 1

  3. How can post-execution contract-satisfaction feedback enable principled adaptation in non-stationary E4NetAI systems? p. 1

Semantic Contract Generation and Calibration

The framework introduces a large language model (LLM)-assisted semantic parser that interprets natural language requests and extracts semantic service requirements together with confidence scores. These extracted semantics are then converted into service requirements and semantic uncertainty. p. 0 The LLM-assisted parser extracts a semantic feature vector y(t) = (lD,(t)i, lE,(t)i, lM,(t)i, lS,(t)i, cconf,(t)i), where lD,(t)i, lE,(t)i, lM,(t)i, and lS,(t)i denote the delay-sensitivity level, energy-sensitivity level, monetary-cost sensitivity level, and T4E requirement level. p. 3 These quantities are transformed into optimization-compatible contract parameters via interpretable mapping rules that incorporate task attributes and system-side context. p. 3 The semantic contract C(t)i consists of three components: a preference vector w(t)i, a fulfillment-bound vector g(t)i, and a semantic-uncertainty measure ρ(t)i. p. 3 The preference weights are generated according to w k,(t)i = exp(thetawlˆk,(t)i) / Σr∈[D,E,M,S] exp(thetawlˆr,(t)i), k ∈ [D,E,M,S]. p. 3

Contract-Guided Potential Game Formulation

STEPS formulates edge scheduling as a contract-guided potential game that jointly determines execution-node selection, computing-resource provisioning, and bandwidth allocation. The per-timeslot scheduling problem is modeled as an exact potential game where the objective minimizes the long-term average contractguided service loss across all admitted UDs. p. 5 The scheduling cost of UD u i is defined as J(t)i(a(t)i, a(t)-i) = H(t)i(a(t)i)+ (1+τ(t))∑z∈[f,b]ΔΓz,(t)i(a(t)i, a-(t)-i), where ΔΓz,(t)i is the marginal congestion cost induced by UD u i under candidate action a(t)i. p. 6 The potential function Φ(t)(a(t)) aggregates the contract-guided service losses of admitted UDs and penalizes edgeside resource congestion. p. 6

Feedback-Driven Adaptive Control

To handle non-stationary environments, STEPS constructs feedback signals from observed fulfillment outcomes and distinguishes between two complementary sources of variation: semantic request drift (Δ(t)sem) and contract-fulfillment drift (Δ(t)ful). The resulting feedback signals are then used to adapt both edge-side resource coordination and semantic contract management. p. 7 Specifically, the fulfillment-side feedback signal is defined as Ψ(t)ful = αsΔ(t)sem + αf Δ(t)ful + αpP(t)ful. p. 7 These signals are used to update the semantic admission threshold, contract conservativeness factor, and edge coordination gain using adaptive-control stepsizes. p. 8

Key Performance Results

Synthetic experiments demonstrate that STEPS improves semantic contract fulfillment, reduces contractguided service loss, and maintains robust adaptation under ambiguous natural language requests in non-stationary network conditions. Specifically, Fig. p. 2 2(c) reports the ACSL, which measures the severity of contractguided service loss. p. 1 STEPS achieves ACSL comparable to Direct-LLM and clearly lower than Param-Opt and DRL-Opt across all uncertain-request ratios. p. 9 Furthermore, Fig. p. 2 4(c) shows that the ACSL of all methods increases after the drift for all methods, where STEPS maintains one of the lowest ACSL values during the post-drift period and outperforms DRL-Opt, Intent-Reconf, and GAH.

Scalability and Robustness

STEPS maintains stable and high AFD across all problem scales S1 through S4. Compared with other baselines such as DRL-Opt, Intent-Reconf, and GAH, STEPS achieves better fulfillment performance because it jointly considers semantic preferences, fulfillment bounds, and edge-side resource contention. The framework is characterized by the existence of a pure-strategy Nash equilibrium for the per-slot scheduling game G(t). p. 1 The adaptive control parameters are bounded within prescribed feasible regions, ensuring stability of the long-term operation of the framework. p. 9

Conclusion

In this paper, STEPS establishes one of the first closed-loop pipelines from natural language service requests to semantic abstraction, contract-guided edge scheduling, postexecution contract-satisfaction evaluation, and fulfillmentdriven adaptation. STEPS converts natural language service requests into semantic contracts, performs contract-guided scheduling for admitted UDs, executes the resulting service decisions, and updates scheduling-control parameters using post-execution feedback. p. 8 The theoretical analysis established the existence of a pure-strategy equilibrium, finite-step convergence of the asynchronous best-response dynamics, and boundedness of the adaptive control parameters. p. 1 Synthetic and real-world-inspired experiments showed that STEPS reduces contract-guided service loss, maintains competitive fulfillment performance, and improves post-drift robustness with acceptable execution and delay overhead. p. 0

References

[1] Y. Mao, X. Yu, K. Huang, Y.-J. A. Zhang, and J. p. 2 Zhang, “Green edge AI: A contemporary survey,” Proc. p. 2 IEEE, vol. 112, no 7 pp 880–911.

[2] Y. He, X. Li, and J. Luo, “Task-oriented integrated sensing and semantic communications for multi-device video analytics,” IEEE Trans. p. 2 Mobile Comput., vol 25 no 5 pp 7323–7337.

[10] Y. Wang, C. Yang, T. Li, Y. Ouyang, X. Mi, and Y. Song, “A survey on intent-driven end-to-end 6G mobile communication system,” IEEE Commun. p. 0 Surveys Tut., vol 28 pp 882–915.

[14] X. Li and S. Bi, “Optimal AI model splitting and resource allocation for device-edge co-inference in multi-user wireless sensing systems,” IEEE Trans. p. 2 Wireless Commun. p. 2 p. 2, vol 23 no 9 pp 11094–11108.

[24] X. Zhang, J. Nie, Y. Huang, G. Xie, Z. Xiong, J. Liu, D. Niyato, and X Shen, “Beyond the cloud: Edge inference for generative large language models in wireless networks,” IEEE Trans. p. 2 Wireless Commun., vol 24 no 1 pp 643–658.

[29] R. W. Rosenthal, “A class of games possessing pure-strategy nash equilibria,” Int. p. 0 J. Game Theory, vol 2 p 65–67 Dec 1973. p. 0

[31] Q. He, G. Cui, X. Zhang, F. Chen, S. Deng, H Jin, Y Li and Y Yang “A game-theoretical approach for user allocation in edge computing environment” IEEE Trans Parallel Distrib Syst. p. 0, vol 31 no 3 pp 515–529.

[42] The updated prices lambda f,(t+1) and lambda b,(t+1) are used as the next-timeslot unit resource prices in the monetary-cost model in (15). p. 8

[43] The updated parameters are used in the next timeslot as follows: τ rho,(t+1) determines the schedulable UD set through (8), β(t+1) calibrates the scheduling-stage fulfillment bounds through (7), and κ(t+1) enters the best-response scheduling cost and the potential function through (25) and (27). p. 8

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Semantic Contract Generation using LLM-assisted parsers allows natural language service expectations to be translated into executable interfaces, bridging user semantics and edge scheduling decisions through a mechanism that extracts semantic service requirements together with confidence scores.

  2. Contract-guided scheduling forms an exact potential game where decision variables are optimized through a distributed potential game-based equilibrium, jointly determining execution node selection, computing-resource provisioning, and bandwidth allocation.

  3. Post-execution contract satisfaction feedback enables principled adaptation by constructing feedback signals that distinguish between semantic request drift (changes in user semantics) and contract-fulfillment drift (changes in contract satisfaction outcomes due to systemside execution conditions).

  4. The framework can proactively regulate scheduling by dynamically updating parameters: the semantic admission threshold, the contract conservativeness factor, and the edge coordination gain based on observed feedback signals.

  5. The system can improve robustness under uncertain requests by calibrating fulfillment bounds using semantic uncertainty and a contract conservativeness factor, ensuring that when uncertainty increases, the scheduling process adopts stricter limits while imposing a higher T4E requirement.

  6. The online distributed framework can solve per-slot scheduling problems using an asynchronous strict best-response process, which converges in finite steps to a pure-strategy Nash equilibrium for admitted users.

  7. The system maintains stability under non-stationary environments by updating resource prices based on intended resource demand reflected by the game-output action profile before physical realization, allowing the pricing mechanism to react to latent resource pressure.

Sources

Related papers