{"id":"691fcfc3-321f-43e8-abf1-b44683caec26","arxiv_id":"2509.02395","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A multi-objective reinforcement learning framework that optimizes an emissions-per-bit metric for 6G downlink resource allocation reports roughly 26% lower simulated emissions per bit than energy-efficiency-only baselines.","lead":"This paper proposes an emissions-per-bit metric for 6G networks and a multi-objective reinforcement learning scheme that cuts emissions by allocating energy, computing, and transmission resources. Simulations report about 26% lower emissions per bit than baselines that only optimize energy efficiency.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MORL state in §III-A is only emissions per bit, but the constraints and objectives depend on battery levels, queue backlogs, and channel randomness; the claimed Pareto-optimal global policy is therefore not supported as stated.","rationale":"The central claim has two pillars: (i) the new sustainability metric and MO formulation, and (ii) the MORL solution that supposedly finds a Pareto-optimal global policy and yields a ~26% emissions reduction. The reader's weakest assumption targets the second pillar through MDP state insufficiency, and I agree that this is the most load-bearing concern. The paper provides no proof of Lemma 1 ('we omit the details due to space constraints') and explicitly skips the variational derivation, but the deeper issue is that the state excludes variables required by the constraints and rewards. RL optimality is defined with respect to an MDP; if the state is not Markov, the Q-functions, KL objectives, and convergence claims in Section III-A do not follow. This concern is logically prior to the metric-novelty and reproducibility issues: even a fully reproduced simulation with error bars would not establish Pareto optimality if the MDP is misspecified. The proposed concrete test directly checks sufficiency by comparing feasible sets and delay consequences for two equal states with different hidden variables. If the test confirms the insufficiency, the authors would need to augment the state or prove that b_k and Ω_k are irrelevant under their assumptions, and would need to soften the global Pareto-optimality claim. I therefore agree with the reader's CONDITIONAL verdict; the concern reinforces the condition rather than changing the verdict.","tokens_in":10180,"tokens_out":4191,"duration_ms":48222,"concrete_test":"Analytical check: construct two system configurations at time t with the same state s_t = (Cb_1(t), ..., Cb_M(t)) but different battery levels b_k(t−1) and queue backlogs Ω_k(t). Check whether the feasible action sets from (13b)–(13g) and the delay objective (10) are identical. If the feasible sets or delay values differ, s_t is not a sufficient statistic and the MDP in §III-A is misspecified. A stronger computational version: run Algorithm 1 with state s_t and with augmented state s'_t = (Cb_k, b_k, Ω_k, h_k) under the Table I parameters; if the Pareto front in Fig. 4 shifts or constraints (13d) are violated under s_t, the original state insufficiency is confirmed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section III-A defines the state as s_t = (Cb_1(t), ..., Cb_M(t)) and asserts that \"the state transitions are deterministic and follow (11).\" This is the load-bearing step for the paper's central claim that the global policy π(a_t | s_t, V) achieves a Pareto-optimal tradeoff for problem (13). The assertion fails. From (11), Cb_k(t) = c_k(t) / (T Σ_{u∈M_k} R_u(t)); R_u(t) depends on the stochastic channel h_u(t) and on the chosen transmit powers, so the transition is not deterministic. More importantly, constraint (13d) requires d_k(t) ≤ b_k(t−1), and the delay objective (10) depends on queue backlog Ω_k(t) evolving via the stochastic recursion (9); neither b_k nor Ω_k appears in s_t. Two configurations with identical Cb_k(t) can have different feasible action sets and different delay consequences, so the policy cannot be optimal for the stated MO problem. The Q-functions in (14), the variational derivation in (16)–(18), and Lemma 1 all presuppose a well-posed MDP; with a non-Markov state, the Bellman optimality argument does not go through. The omitted proof of Lemma 1 and the skipped variational steps are explicitly flagged in the text, but even supplying them would not fix the missing-state problem. This concern is independent of the metric novelty or the missing error bars: the central optimality guarantee is unsupported as written.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a carbon-emissions-per-bit sustainability metric S(t) (Eq. (12)) for a multi-server downlink system with renewable and grid energy sources, and formulates a multi-objective resource allocation problem P1 (Eq. (13)) that jointly maximizes long-term sustainability, minimizes energy per bit, and minimizes average delay. The solution method is an MORL framework: the state is defined as the vector of per-server emissions per bit, an LSTM is used as an environment model, and a variational inference procedure over three Q-functions produces a global policy π(a_t|s_t,V). The paper claims this policy achieves a Pareto-optimal tradeoff and reports about 26% lower average emissions per bit than baselines that do not include carbon emissions in their control objective.","tokens_in":10483,"tokens_out":7158,"duration_ms":90423,"significance":"If the formal claims are supported, the paper would provide a concrete, operationalizable way to include carbon footprint in 6G resource allocation and would demonstrate a tradeoff between sustainability, energy, and delay. The emissions-per-bit metric itself, normalized by an all-non-renewable baseline, is a reasonable and useful construct, and the paper explicitly targets joint computing, communication, and energy control rather than energy efficiency alone. However, the theoretical guarantees are not currently backed by a well-posed MDP, the central policy-composition proof is omitted, and the headline 26% improvement is a direct consequence of training on S(t). The significance of the work therefore depends on whether the MDP can be reformulated and whether the matched-QoS and statistical claims can be substantiated.","major_comments":[{"comment":"The MORL state is not sufficient and the transition is not deterministic. Section III-A defines s_t = (Cb_1(t), ..., Cb_M(t)) and states that \"state transitions are deterministic and follow (11)\". This is load-bearing and incorrect. From (11), c_k(t) is a linear combination of energy decisions, and Cb_k(t) = c_k(t)/(T Σ_{u∈M_k} R_u(t)); R_u(t) depends on the stochastic channel h_u(t) and the transmit powers, so the next Cb is not a deterministic function of (s_t, a_t). Moreover, constraint (13d) and battery dynamics (2) require b_k(t−1) in the state, and the delay objective (10) depends on the queue backlog Ω_k(t) evolving stochastically through (9). Two states with identical Cb_k(t) can have different feasible action sets and different delay consequences. Consequently, the Q-functions in (14), the variational updates (16)–(18), and the Bellman-style argument underlying Lemma 1 are not j","section":"§III-A, Eqs. (9)-(13)"},{"comment":"The proof of the central policy-composition result is omitted, and the \"local optimum\" claim is unsupported. Lemma 1 states that the global policy is π = Σ_r ζ_r q_r, and the text concludes that alternating optimization \"leads to a local optimum\". The proof is explicitly omitted (\"We omit the details due to space constraints\"), and the variational derivation after (16) is also skipped (\"we skip the details here\"). The result is not immediate because q_r in (19) itself depends on the current π via p(Z_r,t | s_t, a_t), so the alternating scheme is a fixed-point iteration; no convergence or local-optimality proof is provided. Since this is the mechanism behind the claimed Pareto-optimal global policy, the gap is material.","section":"§III-A, Eq. (16) and Lemma 1"},{"comment":"The headline 26% reduction in emissions per bit is partly by construction. The reward Q1 in (14) maximizes S(t), and S(t) in (12) is a monotone transform of the ratio Σ_k Cb_k(t)/Cb_ψ(t); hence the agent is trained specifically to minimize emissions per bit. Comparing this agent with baselines whose reward does not contain that term will naturally produce a reduction in that same metric. The non-circular content of the claim is that energy and delay are matched and that the reduction is statistically meaningful. However, Figs. 2 and 3 show no error bars or multiple-seed variance, and the paper does not report whether constraints (13d)–(13g) are actually satisfied. Please report the achieved energy and delay values, confidence intervals, and constraint-violation rates.","section":"Abstract, §IV, Eq. (12) and Fig. 2"}],"minor_comments":[{"comment":"Q2 contains a double negative \"--lim\" before the first term; this is presumably a typo and should be a single minus sign.","section":"Eq. (14)"},{"comment":"The interference index is written as \"i≠k\" in one place and \"i≠u\" in another; the notation for channel coefficients also switches between h_u and h_k. Please make the indices consistent.","section":"Eq. (6) and preceding rate expression"},{"comment":"The horizontal-axis label appears corrupted (\"10-4\" appears in the axis title), and the delay units are unclear. Please fix the axis labels and state whether delay is in milliseconds or seconds.","section":"Fig. 4"},{"comment":"The LSTM architecture, prediction horizon T, number of epochs, and the update rule for the Lagrange multipliers λ_i are not specified. These details are needed for reproducibility and for judging whether the predictive model actually tracks (11) faithfully.","section":"Algorithm 1 and §III-A"},{"comment":"Reference [8] already defines network carbon intensity as the ratio of total carbon emissions to total data traffic; the distinction between Cb_k and NCI should be stated more explicitly, since the novel content appears to be the joint optimization enabled by the metric rather than the metric itself.","section":"Related work, §I-A"}],"recommendation":"major_revision","confidential_remarks":"The paper's central theoretical apparatus is not in a publishable state: the MDP is ill-posed as written, the key proof is omitted, and the main empirical claim is partly a consequence of optimizing the reported metric. These issues are fixable in principle—by expanding the state to include battery, queue, and channel information, supplying the missing derivations, and reporting a statistically grounded comparison—but they require substantial rework rather than minor edits. I would advise the editor to require a complete revision addressing the major comments before reconsidering the manuscript."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper tackles a real problem—making emissions per bit an explicit control objective for 6G downlink allocation across CPU/GPU servers and hybrid energy sources—and the specific combination of an emissions-aware metric with MORL and an LSTM predictive model is not in the prior art I know. The system model is detailed, covering battery dynamics, queues, and compute/communication energy, and the intent to optimize carbon rather than just energy is sensible. That is the genuinely new part, and it's worth engaging with.\n\nThe soft spots are not minor. First, the 'novel sustainability metric' in Eq. (12) is a squared, per-server re-encoding of the network carbon intensity ratio from ITU-T L.1333 (their ref [8]); the square is a monotone transform and cannot change the optimum. The overstatement is annoying but not fatal. Second, the abstract's Pareto-optimality claim rests on Lemma 1, whose proof is omitted, and a weighted average of per-objective specialist policies is not generally Pareto-optimal. The text later admits only a 'local optimum,' which is a weaker and more defensible claim. Third—and this is the load-bearing issue—the MORL state is defined as only emissions per bit, with 'deterministic' transitions following (11). That cannot be right: rates depend on stochastic channels and actions, and two of the three objectives (energy, delay) depend on battery levels and queue backlogs that are not in the state. The Bellman optimality argument, the Q-functions, and the variational derivation all presuppose a well-posed MDP. This is not a matter of missing proof details; it is a structural gap.\n\nThe simulations reinforce the concern: the 26% reduction has no error bars, no seeds, unspecified emission factors and energy model constants, and the equal-delay/equal-energy claim is asserted without numbers. Since the baseline by construction does not optimize carbon, part of the gain is circular; the non-circular content would be matched QoS, which is not demonstrated. The paper would be conditionally salvageable if the authors expanded the state to include battery and queue information (or explicitly formulated it as a POMDP), proved or softened the Pareto-optimality claim, and reported proper statistics and parameter tables.\n\nI'd send this to peer review—the problem is current, the approach is plausible, and the flaws are fixable in revision. I wouldn't cite the metric as novel, but I might cite the framework if the experiment story improves. For a reading group, it's a maybe: useful as a case study of how easy it is to overclaim on RL theory.","headline":"Worth a serious referee because the carbon-per-bit control problem is real, but the MDP formulation and Pareto-optimality claims don't hold as written; expect major revisions.","tokens_in":11084,"tokens_out":2655,"would_cite":false,"duration_ms":33639,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C29","68T05","94A05"],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that putting carbon emissions per bit directly into a 6G resource-allocation objective, solved with multi-objective reinforcement learning, reduces average emissions per bit by about 26% at matched delay and energy efficie","keywords":["sustainability metric","carbon emissions per bit","multi-objective reinforcement learning","6G resource allocation","renewable energy","Pareto-optimal policy","energy efficiency","queuing delay"],"falsifier":"Retrain the same multi-objective reinforcement learning controller with the state expanded to include battery level, queue backlog, and current channel coefficients for each server, keeping the same objectives and simulator. If the emissions-per-bit versus delay Pareto frontier does not shift, the emissions-only state suffices; if it shifts measurably, the claimed optimality rests on an incomplete state and the 26% reduction is not attributable to the stated mechanism.","tokens_in":9960,"feed_emoji":"🌱","tokens_out":7406,"duration_ms":79679,"temperature":0.7,"pith_summary":"This paper tries to show that wireless resource allocation in 6G can be controlled directly by an environmental objective—grams of CO2 emitted per delivered bit—rather than by energy efficiency alone, which does not guarantee low carbon output. It defines a closed-form sustainability metric from the energy mix of renewable, grid, and battery sources and folds it into a multi-objective optimization over computing, communication, and energy resources. To solve the non-convex problem, the paper proposes a multi-objective reinforcement learning scheme whose global policy is a weighted mixture of per-objective policies. If the argument holds, operators gain a principled knob for trading emission reduction against delay and energy use, with simulations showing roughly 26% lower average emissions per bit than baselines that ignore carbon. The reader should care because it moves sustainability from a side constraint into the objective being optimized.","feed_headline":"Carbon-aware 6G scheduling cuts emissions per bit by 26%","feed_subtitle":"A per-bit carbon metric puts CO2 into the control loop, balancing emissions, energy, and delay in one policy.","key_machinery":"The load-bearing object is the emissions-per-bit sustainability metric, which compares actual carbon emitted per delivered bit against an all-non-renewable baseline, making it dimensionless and directly optimizable. The solver is a variational multi-objective reinforcement learning update in which each objective gets a local policy, and the global policy is the weighted average of those local policies, with preference weights selecting the operating point on the Pareto frontier. An LSTM predictive model of emissions evolution is used to turn the per-slot metric into a long-term objective.","core_discovery":"The central claim is that carbon emissions per bit can serve as a sustainable and actionable control objective for a downlink 6G system with heterogeneous CPU/GPU servers and mixed renewable/grid/battery energy. The paper defines the per-server emissions per bit and a network sustainability measure equal to the square of one minus the ratio of actual emissions per bit to the emissions per bit that would occur if all energy came from non-renewable sources. It then formulates a multi-objective problem maximizing long-term sustainability, minimizing energy per bit, and minimizing average delay, and solves it with multi-objective reinforcement learning. The resulting global policy is a convex co","pith_inferences":["Editorial extension: the metric's relative form—actual carbon per bit divided by an all-fossil baseline—is not tied to 6G specifics, so it could be reused for edge data centers, IoT gateways, or any system that can attribute energy to delivered bits.","Editorial extension: the 26% figure comes from a simulator with relatively stable emission factors and channel statistics; in the field, emission factors vary hourly, so a deployment would need online estimates, and the realized reduction could differ.","Editorial extension: the most direct test of the modeling assumptions is to add battery levels, queue backlogs, and channel states to the reinforcement learning state; whether the gains survive a richer state is an open question.","Editorial extension: because the preference weights are exposed parameters, the same framework could be adapted to per-user carbon budgets or regulatory emissions caps by adding constraints on cumulative emissions per bit."],"forward_implications":["Operators can optimize a concrete, measurable target—CO2 per delivered bit—instead of treating carbon as a side effect of energy use.","At matched QoS, the reported simulation results indicate roughly 26% lower average emissions per bit under varying packet arrival rates than methods that ignore carbon in the control objective.","Changing the preference weights moves the operating point along an emissions-versus-delay Pareto frontier, from about 33% emission reduction at low delay to about 9% at high delay.","Because the metric is defined from server-level emissions per bit, it applies across heterogeneous CPU/GPU servers and mixed renewable/grid/battery energy, so the same controller can govern different hardware configurations.","Using a predictive model of emissions evolution favors reductions averaged over many time slots rather than instantaneous dips, which matters when renewable supply and battery levels change over time."],"supporting_citations":[{"why":"supplies the variational multi-objective policy-optimization derivation that the paper extends to the wireless setting","marker":"[18]"},{"why":"provides the predictive scheduling idea that motivates tracking emissions over a time horizon instead of per slot","marker":"[2]"},{"why":"defines the emission factors that convert consumed energy into CO2 in the proposed metric","marker":"[16]"},{"why":"gives Little's law used to express average queuing delay in the QoS objective","marker":"[15]"},{"why":"provides the battery evolution model used in the energy constraints","marker":"[11]"},{"why":"supplies the CPU/GPU power consumption model used to compute computing energy and its carbon cost","marker":"[14]"},{"why":"defines the renewable energy ratio used as one of the simulation baselines","marker":"[9]"},{"why":"defines the network carbon intensity metric that the proposed emissions-per-bit measure extends and contrasts with","marker":"[8]"}],"fun_headline_variants":["6G carbon-per-bit metric cuts emissions 26%","Per-bit carbon objective trims 6G emissions 26%","Emissions-per-bit control cuts 6G carbon 26%","Multi-objective RL balances 6G carbon, energy, delay"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The approach assumes that knowing each server's current emissions per bit is enough to decide the best actions, even though delays and energy use also depend on battery levels, queue backlogs, and random channel and traffic conditions that are not part of that state.","fun_headline_variants_meta":{"raw":{"variants":["6G carbon-per-bit metric cuts emissions 26%","Per-bit carbon objective trims 6G emissions 26%","Emissions-per-bit control cuts 6G carbon 26%","Multi-objective RL balances 6G carbon, energy, delay"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000697,"raw_usage":{"total_tokens":2999,"prompt_tokens":768,"completion_tokens":2231,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2170}},"tokens_in":512,"tokens_out":2231,"duration_ms":19491,"temperature":1.0,"reasoning_tokens":2170,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T11:37:11.997517+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Retrain the same multi-objective reinforcement learning controller with the state expanded to include battery level, queue backlog, and current channel coefficients for each server, keeping the same objectives and simulator. If the emissions-per-bit versus delay Pareto frontier does not shift, the emissions-only state suffices; if it shifts measurably, the claimed optimality rests on an incomplete state and the 26% reduction is not attributable to the stated mechanism.","supporting_citations":[{"cited_title":"Towards sus tainable edge computing through renewable energy resources and onli ne, dis- tributed and predictive scheduling,","cited_arxiv_id":null,"evidence_quote":"provides the predictive scheduling idea that motivates tracking emissions over a time horizon instead of per slot"},{"cited_title":"Electric efficie ncy indicators and carbon dioxide emission factors for power generation by fossil and renewable energy sources on hourly basis,","cited_arxiv_id":null,"evidence_quote":"defines the emission factors that convert consumed energy into CO2 in the proposed metric"},{"cited_title":"A Proof for the Queuing Formula: L= λ w ,","cited_arxiv_id":null,"evidence_quote":"gives Little's law used to express average queuing delay in the QoS objective"},{"cited_title":"All ocation of computing tasks in distributed MEC servers co-powered by re newable sources and the power grid,","cited_arxiv_id":null,"evidence_quote":"provides the battery evolution model used in the energy constraints"},{"cited_title":"Power capping of cpu-gpu hetero geneous systems using power and performance models,","cited_arxiv_id":null,"evidence_quote":"supplies the CPU/GPU power consumption model used to compute computing energy and its carbon cost"},{"cited_title":"Optimal des ign of hybrid renewable energy systems in buildings with low to h igh renewable energy ratio,","cited_arxiv_id":null,"evidence_quote":"defines the renewable energy ratio used as one of the simulation baselines"},{"cited_title":"Carbon Data Intensity for Network Energy Performa nce Moni- toring,","cited_arxiv_id":null,"evidence_quote":"defines the network carbon intensity metric that the proposed emissions-per-bit measure extends and contrasts with"}],"review_version":1}