{"id":"2d8bfad4-3a53-4ba4-9388-9f3413b46714","arxiv_id":"2505.24618","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Active Inference agents managing a three-stage video processing pipeline achieve roughly 80 to 90 percent device-aware service objective fulfillment and adapt when the host device changes.","lead":"This paper places an Active Inference agent next to each service in a distributed stream processing pipeline and tests whether the agents can keep device-specific service objectives fulfilled. It reports over 90 percent fulfillment with hand-built transition models, around 80 percent while learning, and much better sample efficiency than a multi-agent reinforcement learning baseline.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline 90%/80% SLOiD numbers rest on a simulator that samples observations i.i.d. without temporal continuity (Sec. 4.2.3), so the central multi-step adaptation claim is not yet established for real continuum dynamics; one replay-based experiment can settle it.","rationale":"The reader's weakest assumption identifies the i.i.d., non-temporal simulator as the key limitation, and I agree. The center of gravity of the paper is the empirical claim that AIF agents fulfill SLOiDs and adapt to offloading. That claim is only as strong as the simulator's fidelity. Section 4.2.3 is explicit: observations are sampled from a dataset and 'need not follow a temporal sequence.' Section 5.5 repeats this and notes that policy length is not critical. A system with no temporal continuity is effectively a contextual bandit; planning over multiple steps has no object to plan over. The reported high fulfillment rates with expert transition models are further softened by the fact that those models are hand-coded rules in Appendices A–C that specify deterministic responses matching the causal structure of the simulator, with no held-out validation. Thus the 'over 90%' claim is not internally inconsistent—it is supported by the simulation plots, though the unqualified abstract overstates the Worker's lower performance—but its external validity is unestablished. The proposed ordered-trace replay is a direct, feasible check: it keeps all agent code and models identical and only changes the observation generator from i.i.d. to time-ordered. If fulfillment remains high under ordered replay, the concern largely dissolves; if it drops, the central claim must be restated as applying to stationary, memoryless environments. This does not change the reader's CONDITIONAL verdict; it sharpens the condition.","tokens_in":26225,"tokens_out":4718,"duration_ms":62195,"concrete_test":"Re-run the Section 5.1 and 5.2 experiments exactly as described, but replace the i.i.d. observation sampler with a time-ordered replay of the recorded traces: advance through each trace sequentially, preserving within-trace correlations in execution time, latency, energy, smoothness, and device load, while agent actions influence which recorded segment is active or how parameters are drawn, as in the current simulator. Keep the same POMDP definitions, CPTs, policy length, and 200-step/10-repetition protocol. If aggregate SLOiD fulfillment under expert models falls materially below the claimed 90%, or drops by more than a few points relative to the i.i.d. baseline, the central claim is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim—decentralized AIF agents fulfill over 90% of SLOiDs with tested transition models and can recover after environment changes—requires the simulated environment to be a faithful proxy for a dynamic processing pipeline. The manuscript itself undercuts this: Section 4.2.3 says each new observation is sampled from the dataset and 'need not follow a temporal sequence,' and Section 5.5 concedes that the environment has 'limited temporal dependencies' partly because 'new states are sampled to fulfill constraints without necessarily maintaining temporal continuity.' Under this protocol, the environment is an action-conditioned bag of independent samples, not a temporal process with autocorrelated load, network delays, or device-state persistence. Policy length then becomes nearly immaterial, as the authors observe. Additionally, the expert transition models (Appendices A–C) are hand-specified rules that mirror the simulator's own causal structure and are not validated against held-out traces, so the >90% figure may reflect a near-tautological match rather than robust model-based control. If real continuum workloads have temporal coupling, the reported fulfillment rates have no demonstrated transfer to deployment.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a distributed stream-processing pipeline in the Computing Continuum in which each service is supervised by an Active Inference (AIF) agent. The authors define device-aware Service Level Objectives (SLOiDs), specify POMDP-based generative models for a Producer–Worker–Consumer pipeline, and evaluate the agents in five simulation experiments: expert-model SLOiD fulfillment, online learning of the transition model, comparison with multi-agent PPO, adaptation to a sudden hardware change (offloading), and computing cost. The central empirical claim is that AIF agents achieve more than 90% SLOiD fulfillment with tested transition models and roughly 80% when learning the models during deployment, with substantially lower training cost than MARL.","tokens_in":26467,"tokens_out":3633,"duration_ms":39266,"significance":"If the central claim holds, the paper would provide a credible template for decentralized, model-based service management in heterogeneous computing continuum systems, with the notable strengths of publicly available code and dataset, reproducible experiments, and an honest comparison against a standard MARL baseline (PPO). The introduction of SLOiDs as device-aware objectives is a useful conceptual contribution. However, the significance is currently conditional on the simulation environment being a faithful proxy for real continuum dynamics, and on the reported fulfillment numbers being statistically supported; both conditions are only partially met in the present manuscript.","major_comments":[{"comment":"The environment is not a temporal process: Section 4.2.3 states that each new observation is sampled from the dataset and 'this sample need not follow a temporal sequence,' and Section 5.5 concedes that the environment has 'limited temporal dependencies' because 'new states are sampled to fulfill constraints without necessarily maintaining temporal continuity.' Under this protocol, the environment is an action-conditioned bag of independent samples, so the paper's claims about multi-step planning (policy length), adaptation after offloading, and recovery after environmental change are not demonstrated for real continuum workloads, which exhibit autocorrelated load, network delay, and device-state persistence. This is load-bearing for the abstract's central claim, and it needs to be addressed by either replaying the dataset in chronological order, introducing a temporally correlated environment, or clearly restricting the claims to the i.i.d. sampling regime.","section":"Section 4.2.3 and Section 5.5"},{"comment":"The abstract's 'over 90% SLOiD fulfillment' statement conflicts with the visible results: in Figure 6 the Worker's fulfillment curves for Latency and W-consumption stabilize well below 0.9 (approximately 0.7–0.8), and the text itself notes that the 'Worker agent clearly struggles.' The aggregate statement may average over three agents, but as written it overstates the per-agent performance. Please report per-agent aggregate values with standard deviations or confidence intervals, and qualify the abstract's claim (e.g., 'most agents' or 'on average').","section":"Abstract and Figures 6 and 7"},{"comment":"No per-repetition statistics are reported for the ten runs. The text claims that one experiment yielded 'consistently higher' fulfillment for policy length 1 (Section 5.5) and that the Worker exhibits 'greater performance variance' (Section 5.1), but no standard deviations, confidence intervals, or statistical tests are supplied. Given that the environment samples stochastically and that only ten repetitions were run, statements of consistency and relative performance (including the MARL comparison) are not yet supported; please add error bars or per-run summaries and, where relevant, a paired test or effect-size measure.","section":"Section 4.2.3 and Section 5.5"},{"comment":"The expert transition models are hand-specified rules that mirror the simulator's own causal structure (the same DBN structure is also fixed for the learned models, with only CPT parameters learned). Evaluating these expert models on a dataset generated by independently adjusting the same parameters is close to validating a model against a generator built from the same rules; the >90% figure may therefore reflect a near-tautological match rather than robust model-based control. I recommend adding a validation against held-out portions of the dataset or against a differently generated trace, and at minimum discussing this potential circularity explicitly.","section":"Section 4.1 and Appendices A–C"}],"minor_comments":[{"comment":"The heading 'Preferred Ouctomes (C)' contains a typo; it should be 'Preferred Outcomes (C)'.","section":"Section 4.1.1"},{"comment":"The preferred outcome vector for W-consumption is written as {3, 2.5, 0, 5}, but W-consumption has three levels (LOW, MID, HIGH); the vector should have three entries (e.g., {3, 2.5, 0}), and the current four-entry vector is inconsistent with the state definition.","section":"Section 4.1.2"},{"comment":"'state-off-the-art' should be 'state-of-the-art' in the description of the MARL comparison.","section":"Section 4.3.1"},{"comment":"The conclusion says learning agents achieved rates 'approximately 10% lower' than expert-model agents; since the abstract expresses the same difference as 'over 90%' versus 'around 80%', please state whether this is 10 percentage points or 10% relative, and align the wording.","section":"Section 8"},{"comment":"The reference Zahidi et al. (2024) contains 'n/a, n/a' in the volume and issue fields; please complete the bibliographic details.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a reproducible and honest evaluation of a promising application of Active Inference to computing continuum management, but the simulation's i.i.d. sampling protocol is the main risk to the central claim: the recovery/adaptation results may be artifacts of the absence of temporal coupling. I recommend major revision rather than rejection, as the authors could strengthen the paper with a replay-based or autocorrelated environment experiment, per-repetition statistics, and a more careful abstract. The near-tautological expert-model match is a second concern that should be addressed explicitly."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Bottom line: this is a real extension of the Sedlak line—one AIF agent per service instead of a single controller—and the authors ship code, data, and an honest MARL baseline. But the headline 90% figure is not backed by the evidence as stated, and the paper's own Section 4.2.3 tells you why: observations are sampled i.i.d. from a dataset, with no temporal continuity. That makes the central \"adaptation over time\" claim largely untested against real continuum dynamics.\n\nWhat's actually new: a factored POMDP template for three cooperating services (Producer/Worker/Consumer), with SLOiD as a device-aware reframing of SLO. The learning-vs-expert comparison is useful, and the data-efficiency contrast with PPO (400K vs 40M steps) is a fair point even if not a head-to-head benchmark. The offloading experiment is a nice idea, and the learning agent's behavior there is the most interesting result in the paper.\n\nWhere it's soft: first, the over-90% statement in the abstract sits uneasily with the Worker's visibly lower fulfillment in Figure 6; there are no per-agent error bars for the 10 runs. Second, the expert transition models are hand-specified CPTs that mirror the simulator's own structure and are never validated on held-out traces, so the expert-model success is close to a self-consistency check. Third, the i.i.d. sampling means policy length and temporal reasoning are nearly irrelevant—the authors concede this in Section 5.5. The learning case (around 80%) is the more credible number, and even that is measured in the same static environment. None of this is fatal; the paper is careful about its limitations, more than most. But the abstract overstates what the simulation can support.\n\nWho this is for: researchers working on autonomic service management or applying active inference to systems problems. The POMDP design is reusable, the dataset is public, and the paper gives a clear template. A serious referee should engage with it; the fixes are concrete—per-agent statistics, reconcile the abstract, validate or soften the CPT claims, or reframe the contribution as a modeling template rather than a demonstrated adaptation mechanism.\n\nMy recommendation: conditional accept toward a workshop or systems venue, with the temporal-continuity issue addressed by a replay-based experiment or by toning down the claims. It deserves peer review.","headline":"A genuine multi-agent AIF extension with shipped code and data, but the headline 90% figure outruns a simulator that samples observations without temporal continuity.","tokens_in":27017,"tokens_out":2409,"would_cite":true,"duration_ms":29587,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that a decentralized set of Active Inference agents, one per service, can autonomously manage a distributed stream-processing pipeline so that device-aware Service Level Objectives (SLOiDs) are fulfilled most of the…","keywords":["Computing Continuum","Active Inference","Service Level Objectives","SLOiD","multi-agent systems","stream processing pipeline","self-adaptive services","expected free energy"],"falsifier":"Run the same three-agent controller on a physical or emulated continuum testbed with temporally correlated workloads, realistic network delays, and user-driven load changes, and compare agent-level SLOiD fulfillment over the same 200-step horizon; if online-learning fulfillment falls substantially below the reported ~80% or becomes unstable, the dataset-as-environment assumption is where the simulation diverged from deployment.","tokens_in":25994,"feed_emoji":"🤖","tokens_out":7807,"duration_ms":79847,"temperature":0.7,"pith_summary":"The paper is trying to establish that distributed intelligence in the Computing Continuum can be built from Active Inference agents, one per service, rather than a centralized controller. It introduces SLOiDs, service-level objectives that account for the hosting device's capabilities, and models each service agent as a factored partially observable Markov decision process that chooses elasticity actions (changing resolution or FPS, toggling a GPU, enabling or disabling communication) by minimizing expected free energy. In a simulated three-service video processing pipeline, the agents fulfill their SLOiDs over 90% of the time when transition models are given, and around 80% when the models are learned online, stabilizing within about 50 steps. A comparison with multi-agent reinforcement learning shows AIF agents reach comparable performance without millions of training steps, and both kinds of agents recover after the hosting device changes mid-run. If the result holds, decentralized service management could become practical on heterogeneous, resource-constrained devices without large training datasets.","feed_headline":"Three AIF agents keep a video pipeline at 90% SLOiD fulfillment","feed_subtitle":"One agent per service, no training: active inference keeps latency, energy, and quality goals met.","key_machinery":"The load-bearing mechanism is the Active Inference agent defined as a factored POMDP over state modalities (SLOiDs plus system variables such as FPS, resolution, execution time, GPU state, and communication state), where actions are chosen by minimizing the expected free energy of candidate policies. Expected free energy splits into a pragmatic value term, which scores how likely a policy is to produce preferred outcomes encoded in the agent's log-probability preferences, and an information gain term, which scores how much a policy would reduce uncertainty about the hidden states. Because free energy is an additive quantity, if each agent minimizes its own expected free energy, the joint free energy of the whole ensemble is minimized as well, which is what licenses decentralized planning-as-inference. The transition model is either expert-defined through conditional probability tables on a dynamic Bayesian network or learned online starting from uniform distributions, and observations are assumed to be perfect so the likelihood mapping is an identity per modality.","core_discovery":"The central claim is that a pipeline of interdependent streaming services in the Computing Continuum can be managed autonomously by one Active Inference agent per service, with no central coordinator. Each agent supervises its service by maintaining a generative model of the environment and continuously asking which action would both fulfill its preferred outcomes and reduce uncertainty about the world; this is operationalized as selecting policies that minimize expected free energy. The paper reports that with hand-specified conditional probability tables, the three agents achieve over 90% SLOiD fulfillment, and around 80% when those transition models are learned from interaction during deployment. It also reports that, unlike multi-agent reinforcement learning, which needs hundreds of thousands to tens of millions of training steps, AIF agents work effectively from the start, and that when the hosting hardware changes suddenly the agents detect the shift, temporarily raise their expected free energy, and then adapt their policies, with learning agents even deprioritizing a now-impossible energy objective in favor of achievable ones.","pith_inferences":["Because the simulator samples observations independently of time, a real deployment with autocorrelated load and hidden dependencies between devices could degrade the reported fulfillment rates; replaying the same policies over temporally ordered traces would quantify that gap.","The SLOiD encoding is reusable beyond video pipelines: any device-aware objective that can be discretized (energy budgets, thermal limits, latency caps, privacy constraints) can be expressed as preferred outcomes in an agent's POMDP.","The large online-learning slowdown hints that practical edge deployments will need hierarchical control or amortized inference; the paper mentions such hybrid strategies as future work but does not evaluate them.","Sharing learned transition models between agents on similar hardware, which the paper flags as model-transfer future work, would likely close much of the gap between the 90% and 80% fulfillment regimes."],"forward_implications":["A distributed service pipeline can be supervised without a central orchestrator: one AIF agent per service, each minimizing its own expected free energy, keeps device-aware service objectives fulfilled most of the time.","When transition models are unknown, agents can still learn them during deployment and reach about 80% SLOiD fulfillment, stabilizing within roughly 50 time steps rather than needing offline training.","Multi-agent reinforcement learning matches AIF performance only after extensive training, whereas AIF agents are effective from the start, making them a candidate for deployments where training data or compute is scarce.","After an offloading event that changes the host's energy profile, AIF agents detect the shift and adapt, with learning agents able to reprioritize which SLOiDs to pursue.","Policy look-ahead length is not decisive for fulfillment in this near-memoryless simulation, but it strongly drives per-step computation, so short-horizon planning is the practical choice on constrained devices."],"supporting_citations":[{"why":"Supplies the active inference formulation and learning rule that the agents use to select actions and update models.","marker":"Friston et al.(2016)"},{"why":"Provides the preceding equilibrium-in-the-continuum formulation and the recorded processing traces that make up the simulated environment.","marker":"Sedlak et al.(2024a)"},{"why":"Provides the software implementation used to run the active inference agents in all experiments.","marker":"Heins et al.(2022)"},{"why":"Defines the reinforcement learning algorithm used to build the multi-agent baseline that AIF agents are compared against.","marker":"Schulman et al.(2017)"},{"why":"Underpins the factored transition model, representing state dependencies as dynamic Bayesian networks with conditional probability tables.","marker":"Boutilier et al.(1999)"},{"why":"Provides the textbook treatment of active inference and POMDPs that grounds the agent model and the expected free energy objective.","marker":"Parr et al.(2022)"}],"fun_headline_variants":["Active inference agents run pipeline with no central brain","Agents hit 90% SLOiD goals without any training","AIF agents adapt on the fly, beat MARL in computing continuum","Self-organizing agents manage streaming services autonomously"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that the recorded dataset used as the simulation environment is a faithful proxy for real Computing Continuum dynamics, so that sampling an observation conditioned on the chosen action, without preserving any temporal sequence, captures how the actual system would respond.","fun_headline_variants_meta":{"raw":{"variants":["Active inference agents run pipeline with no central brain","Agents hit 90% SLOiD goals without any training","AIF agents adapt on the fly, beat MARL in computing continuum","Self-organizing agents manage streaming services autonomously"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000161,"raw_usage":{"total_tokens":1262,"prompt_tokens":999,"completion_tokens":263,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":615,"completion_tokens_details":{"reasoning_tokens":194}},"tokens_in":615,"tokens_out":263,"duration_ms":4275,"temperature":1.0,"reasoning_tokens":194,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T12:17:40.985951+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same three-agent controller on a physical or emulated continuum testbed with temporally correlated workloads, realistic network delays, and user-driven load changes, and compare agent-level SLOiD fulfillment over the same 200-step horizon; if online-learning fulfillment falls substantially below the reported ~80% or becomes unstable, the dataset-as-environment assumption is where the simulation diverged from deployment.","supporting_citations":[],"review_version":1}