As the burgeoning power requirements of sophisticated neural architectures escalate, the information retrieval community has recognized ecological sustainability as a pivotal priority that necessitates a fundamental paradigm shift in model design. While contemporary neural rankers have attained unprecedented accuracy, the substantial environmental externalities associated with their computational intensity often remain overlooked in large-scale deployments. We present GaiaFlow, an innovative framework engineered to facilitate carbon-frugal search by operationalizing semantic-guided diffusion tuning. Our methodology orchestrates the convergence of retrieval-guided Langevin dynamics and a hardware-independent performance modeling strategy to optimize the trade-off between search precision and environmental preservation. By incorporating adaptive early exit protocols and precision-aware quantized inference, the proposed architecture significantly mitigates operational carbon footprints while maintaining robust retrieval quality across heterogeneous computing infrastructures. Extensive experimental evaluations demonstrate that GaiaFlow achieves a superior equilibrium between effectiveness and energy efficiency, offering a scalable and sustainable pathway for next-generation neural search systems.
Figures & tables
Figure 1: Overview of the GaiaFlow framework for semantic-guided, carbon-frugal search optimization. An incoming query q is encoded by the Input & Retrieval Tower into a semantic embedding uq . In the Latent Manifold , a Retrieval-Guided Langevin Sampler explores the configuration space Z , driven jointly by the Carbon-Frugal Gradient ∇zU from the Multi-Objective Engine (Diff-PEIR) and the Semantic Attraction ∇zV that maintains high-quality retrieval behavior. Real-time efficiency is supported by Quantized Inference and an Early Exit mechanism. The Online Calibration Loop uses EW-RLS and PUE correction to adapt to datacenter conditions, while the Operational Safeguard validates, projects, and finalizes the deployed configuration ω∗ .
Figure 2: Measured versus predicted carbon emissions. Each point corresponds to a query instance. The dashed diagonal denotes perfect prediction, while the shaded region indicates a ±10% relative error band. Results show that the predictor remains well-calibrated across a wide dynamic range.
Figure 3: Latency and computational cost (Mop) distributions across retrieval methods. Boxplots summarize per-query latency (left axis) and Mop cost (right axis). GaiaFlow consistently achieves lower latency and computational overhead compared to baseline methods.
Figure 4: Coefficient distributions under full-query training and a 5% subsample. Solid curves correspond to the full dataset, while dashed curves denote the subsampled setting. The strong alignment across coefficients indicates robust distributional stability under aggressive subsampling.
Figure 5: Pareto frontier over system configurations. Each point represents a deployment configuration, plotting average carbon cost against average latency, with color encoding retrieval recall. The frontier highlights favorable trade-offs achievable by GaiaFlow .
Figure 6: Online calibration behavior over time. The curve reports the mean absolute error (MAE) of latency prediction as new observations are incorporated. The decreasing trend indicates effective self-calibration during continuous deployment.
Figure 7: Representative Langevin trajectory during configuration optimization. The potential energy U(z) decreases rapidly in early iterations, illustrating fast convergence and enabling early stopping in practice.
Retrieval method
Intel
AMD
Mop
Flop
MFlop
LR
PL
Mop
Flop
MFlop
LR
PL
BM25 [ 61 ]
0.801
-0.259
0.969
0.990
0.589
0.463
-1.076
0.820
0.884
0.537
BM25-T5 [ 62 ]
0.968
0.367
0.989
0.997
0.000
0.742
0.308
0.950
0.964
0.001
uniCOIL-TILDE [ 63 ]
0.986
0.930
0.985
0.984
0.300
0.979
0.930
0.977
0.977
0.302
DeepImpact [ 64 ]
0.969
0.469
0.987
0.993
0.444
0.957
0.485
0.970
0.977
0.429
uniCOIL-T5 [ 65 ]
0.971
0.939
0.985
0.977
0.261
0.966
0.941
0.981
0.973
0.255
Table 1: R 2 scores for different performance models and retrieval methods.
Method
L_AMD
L_Intel
Mop [ ×106 ]
L_AMD / BM25
L_Intel / BM25
Mop / BM25
BM25 [ 61 ]
13.90
12.26
5.4116
1.00
1.00
1.00
BM25-T5 [ 62 ]
17.15
18.32
7.3242
1.23
1.49
1.35
DeepImpact [ 64 ]
24.50
26.37
17.9721
2.49
2.60
2.70
SPLADEV2 [ 66 ]
277.97
290.97
161.8180
28.24
28.66
24.28
uniCOIL-T5 [ 65 ]
48.92
50.69
32.4773
4.97
4.99
4.87
uniCOIL-TILDE [ 63 ]
40.88
41.32
26.5466
4.15
4.07
3.98
Table 2: Comparison of retrieval methods: average latency (ms) and average Mop.
Configuration
Latency (ms)
Mop ×106
Recall@1000
Carbon ×10−3 g
Steps
Δ Recall
GaiaFlow
9.0
4.00
0.859
2.18
6.1
0
GaiaFlow γ2=0
9.9(+10%)
4.38(+9.5%)
0.857(-0.2%)
2.39(+9.6%)
7.0(+15%)
− 0.2%
GaiaFlow γ1=0
11.3(+26%)
4.95(+24%)
0.860(+0.1%)
2.70(+24%)
9.5(+56%)
+ 0.1%
Table 3: Impact of removing retrieval-guided attraction on MS-MARCO dev. Changes are measured relative to the complete GaiaFlow model.
Setting
Latency (ms)
Mop ×106
Recall@1000
Δ Carbon
γ1=1.8×10−3
9.6 ± 0.3
4.27 ± 0.05
0.858 ± 0.001
+6.5%
γ1=2.0×10−3
9.0 ± 0.2
4.00 ± 0.03
0.859 ± 0.001
—
γ1=2.2×10−3
8.7 ± 0.2
3.89 ± 0.04
0.859 ± 0.001
− 3.2%
γ2=1.8×10−3
9.3 ± 0.2
4.11 ± 0.04
0.858 ± 0.001
+2.4%
γ2=2.0×10−3
9.0 ± 0.2
4.00 ± 0.03
0.859 ± 0.001
—
γ2=2.2×10−3
9.1 ± 0.2
4.03 ± 0.04
0.859 ± 0.001
+0.8%
Table 4: Sensitivity of individual step coefficients on MS-MARCO dev. Default values are shown in bold.
Train → Test
R2
MAE (ms)
AMD → AMD
0.995
0.38
AMD → Intel
0.987
0.42
Intel → Intel
0.993
0.40
Intel → AMD
0.986
0.44
Table 5: Portability of the learned latency model across CPU architectures. “Train → Test” denotes the source and target processors.
The growing deployment of large language models (LLMs) makes per-request routing essential for balancing response quality and computational cost across heterogeneous model pools. Current routing methods rarely consider sustainable energy use and CO2 emissions as optimization objectives, despite grid carbon intensity varying by time and region, and models differing significantly in energy consumption. To address this gap, we introduce Green-Aware Routing (GAR), a constrained multi-objective optimization framework that minimizes per-request CO2 emissions subject to explicit accuracy floors and p95-latency service-level objectives (SLOs). GAR employs adaptive constraint optimization through per-dataset floor tuning and incorporates lightweight estimators for correctness, tail latency, and carbon emissions, enabling real-time routing decisions without additional inference passes. We present GAR-PD, a practical online primal-dual routing algorithm for rolling carbon budgets, alongside heuristic variants that achieve high feasibility coverage while limiting accuracy degradation. Comprehensive experiments across standard NLP benchmarks with heterogeneous LLM pools (7B-70B) demonstrate that GAR achieves substantial carbon reductions while maintaining competitive accuracy and p95 latency guarantees, providing a practical, theoretically grounded approach to sustainable LLM inference.
Disha Sheshanarayana, Rajat Subhra Pal, Manjira Sinha +1
The sustainability constraints of FLaaS consumers pose significant challenges to maintaining carbon-feasible federated training in FLaaS environments. These constraints often lead to infeasible consumer participation and unstable federated training under hard carbon constraints. We propose a Sustainable Federated Learning as a Service (SFLaaS), a carbon- constrained Neural Architecture Search (NAS) framework for heteroge- neous sustainable constraints. We introduce a requirement-driven search space that transforms consumer sustainability profiles into a feasible architecture region before federated execution. We develop a consumer-level carbon feasibility estimation mechanism to evaluate candidate architectures under dynamic carbon conditions. We propose a sustainable con- sumer scheduling strategy that adaptively selects feasible consumers and allocates local workloads to preserve consumer participation and statistical data coverage. An evolutionary search strategy jointly optimised for predictive performance, consumer feasibility, and participation coverage under hard carbon constraints. Experiments on real-world datasets and a simulated environment demonstrate the effectiveness of the proposed approach.
Keya Patel, Sajib Mistry, Sheik Fattah +4
School of Electrical Engineering, Computing and Mathematical Sciences Curtin University, Perth, Australia · King Fahd University of Petroleum and Minerals, Saudi Arabia · Umeå University, Umeå, SE-90187, Sweden
As the Web transitions from static retrieval to generative interaction, the escalating environmental footprint of Large Language Models (LLMs) presents a critical sustainability challenge. Current paradigms indiscriminately apply computation-intensive strategies like Chain-of-Thought (CoT) to billions of daily queries, causing LLM overthinking, a redundancy that amplifies carbon emissions and operational barriers. This inefficiency directly undermines UN Sustainable Development Goals 13 (Climate Action) and 10 (Reduced Inequalities) by hindering equitable AI access in resource-constrained regions. To address this, we introduce EcoThink, an energy-aware adaptive inference framework designed to reconcile high-performance AI intelligence with environmental responsibility. EcoThink employs a lightweight, distillation-based router to dynamically assess query complexity, skipping unnecessary reasoning for factoid retrieval while reserving deep computation for complex logic. Extensive evaluations across 9 diverse benchmarks demonstrate that EcoThink reduces inference energy by 40.4% on average (up to 81.9% for web knowledge retrieval) without statistically significant performance loss. By mitigating algorithmic waste, EcoThink offers a scalable path toward a sustainable, inclusive, and energy-efficient generative AI Agent.
Linxiao Li, Zhixiang Lu
The University of Sydney Sydney, Australia · University of Liverpool Liverpool, United Kingdom