Organizations: Institute of Information Science and Technologies “Alessandro Faedo”, National Research Council (ISTI-CNR), Pisa, Italy · Department of Computer Science, University of Pisa, Pisa, Italy · Department of AI, Data and Decision Sciences, LUISS University, Rome, Italy
Large language models (LLMs) are increasingly explored for resource management across the Cloud-Edge continuum. However, embedding LLMs directly inside latency-sensitive control loops places inference latency, execution jitter, and hallucination risks on the critical path. This study argues for a different use of LLMs: a continually instruction-tuned intent compiler that runs once per management intent or declared context change, above a fast, deterministic tactical controller that handles all low-latency runtime actions. We formulate intent compilation as a continual instruction tuning problem, because management intents arrive sequentially and accumulate while infrastructure and workload contexts drift. Unlike recent LLM-based management frameworks, our design scores the model's outputs by execution, so that catastrophic forgetting can be measured in the executed objective of the underlying controller. On a continuum simulator with a 2B backbone and parameter-efficient adapters, preliminary results show that naive sequential tuning forgets earlier intents in the executed objective, that replay of stored examples nearly prevents this forgetting, and that the tuned compact compiler performs on par with simpler translators while being more robust than a grammar parser to rephrased intents.
Figures & tables
Intent clause type and its forms
latency
priority
budget
energy
residency
link
throughput
Family
latency (deadline classes)
class first
cost weight
capacity cap
energy weight
power cap
allowed layers
link budget
completion floor
Elastic variant: deciding how many workers to run over time
budget-elastic
low
∙
∘
∘
bill-elastic
low
∘
∙
Placement variant: deciding where stages run
Table 1. The eight intent families and their clauses ( ∙ required, ∘ optional), grouped by the scenario variant defined in Section 5.1 . Rows are the eight intent families, and columns are the nine clause forms of the intent grammar. Each cell marks a clause form as required, optional, or unused for that family, except in the latency column, whose cells name the task classes for which the family's intents state a deadline.
Figure 1. The two nested control loops. Top: the outer loop, in which the compiler, tuned offline on the update batch, turns an intent and its context into (κ,θ) and the validation gate admits them to the tactical controller or rejects them. Bottom: the inner loop, in which the tactical controller acts at every epoch on the continuum simulator. Measured outcomes return only to the offline adapter update. Dashed arrows depict exchanges that happen only during offline label search. The outer loop contains a rollout labeller, an offline update batch of labelled examples, a continually instruction-tuned LLM intent compiler, and a validation gate. An accepted output configures the tactical controller. In the inner loop, the controller exchanges runtime state and feasible actions with the continuum simulator, which represents workloads and IoT, edge, and cloud resources. Measured outcomes return to later offline updates.
Stream I
Stream II
Method
AG
F
AG
F
Sequential IA 3
0.51–0.53
0.23–0.26
0.74–0.75
0.67–0.70
Replay (50/stage)
0.29–0.33
0.03
0.17–0.18
0.06–0.07
Clause-type lookup
0.31
–
∗
–
Retrieval
0.41
–
∗
–
Joint ceiling
0.28
–
0.15
–
Table 2. Preliminary results on the 2B backbone (eight stages per stream, two orders; lower is better). The table lists the average gap (AG) and forgetting (F) of sequential tuning and of replay on Streams I and II, with the clause-type lookup, retrieval, and the joint ceiling for comparison. Replay forgets far less than sequential tuning, and its average gap comes close to that of the joint ceiling.
Method
Rollouts
Seen
Reworded
Group 1: the floor
Default θ (no knob chosen)
0
0.93
0.93
Group 2: given the correct canonical intent (solve score)
Clause-type lookup
0
0.31
–
Regression head (frozen embeddings)
0
0.29
0.41
Search at inference
60
0.29
–
Table 3. The tuned compiler against the translation baselines on the Stream I test set. Each entry is the mean gap ( 4 ) over the test intents (lower is better). The methods are grouped by what they are given; for each, the table shows the number of rollouts spent per intent and the gap on training wordings and, where tested, on rewordings. The joint ceiling and the replay-tuned compiler are on par with the strongest baselines on training wordings and far better than the grammar parser on rewordings. Search started from the compiler's answer, with sixty rollouts, has the lowest gap on training wordings.
Managing resources across IoT, edge, and cloud layers calls for continuous, context-aware decisions under constraints that rarely stay fixed. Deep reinforcement learning (DRL) handles this class of problems well, and large language models (LLMs) are increasingly used to augment DRL pipelines, yet the architectural relationship between the two is seldom made explicit. We build on Wang et al.'s taxonomy of Continuum Orchestration Systems employing DRL techniques and extend it with two further dimensions. The AI Augmentation Paradigm measures how LLMs are exploited, while the Feedback channel captures whether and through which system path the execution feedback returns to the LLM in order to close the MAPE control loop at the LLM Orchestration layer. We apply this taxonomy to six recent system architectures and find a common gap, as none combines full LLM orchestration with full agent-layer feedback in a Cloud Continuum setting. We relate this gap to a missing cross-tier feedback abstraction, bridging the incommensurable per-tier signals and the LLM Orchestrator.
Institute of Information Science and Technologies “Alessandro Faedo” (ISTI), National Research Council of Italy (CNR), 56124 Pisa, Italy · Department of Computer Science, University of Pisa, 56127 Pisa, Italy · Department of AI, Data and Decision Sciences, LUISS University, 00197 Rome, Italy
Continual instruction tuning often preserves old skills with task-private modules. Inference still requires selecting the right module from the prompt alone. We introduce Task-Isolated Latent Adapter Continual Training (TILA-CT), a replay-free method that makes each task module self-indexing. Each task receives a latent memory and adapter bank in a frozen LLM. During adaptation, the memory gathers causal-prefix information and injects task-conditioned context. For access, its persistent state provides the routing prototype used by a shared prompt-only router. On the 8-task CT-Pipeline stream, TILA-CT achieves 65.19 acquisition and 63.98 final macro score, with 1.38 points of average forgetting. A middle-reversed order gives 65.38 acquisition and 64.17 final macro score. Across 6 backbones, final scores remain within 1.5 points of acquisition. Ablations support memory-derived routing keys and causal gathering. Final macro routing accuracy is 93.9%, and oracle routing recovers 1.21 final macro points. These results separate preserved task computation from learned access to it.
Dhruv Dixit, Paritosh Pandey, Snehalkumar 'Neil' S. Gaikwad
School of Data and Information Sciences University of North Carolina, Chapel Hill, USA · *Society-Centered AI Lab University of North Carolina, Chapel Hill, USA · Department of Computer Science University of North Carolina, Chapel Hill, USA
Large language model (LLM) performance increasingly depends not only on the base model, but also on the inference-time controller used to organize reasoning. Existing post-training methods, however, typically optimize for a single fixed interaction pattern, despite real deployments relying on diverse controllers such as Chain-of-Thought, self-consistency, debate, planning, and verification pipelines. This creates a training--deployment mismatch and limits transfer to new workflows. We introduce CALM (Controller-Aware Language Models), a post-training framework that explicitly places controllers in the training loop. We formulate controller-aware post-training as multi-task reinforcement learning over controller-induced interaction protocols, where controllers are compositions of reusable local reasoning modules. This structure also induces a module-level decomposition of mixed-controller training under a turn-level GRPO objective, enabling a systematic study of controller and module-aware training strategies. We evaluate CALM on held-out controller compositions and broader controller shifts, showing that controller-aware post-training improves generalization across inference-time workflows beyond single-controller optimization.
Moumita Choudhury, Vanshaj Khattar, Jing Liu +4
College of Information and Computer Sciences, University of Massachusetts Amherst, Amherst, Massachusetts, USA · Mitsubishi Electric Research Laboratories, Cambridge, Massachusetts, USA