Dynamic resource assignment, the real-time allocation of task streams to heterogeneous processing nodes, is the backbone of modern computing infrastructure. While learning-based schedulers excel in research, industrial deployments still rely on hand-written rules that operators can read, audit, and execute within tight latency budgets. LLM-based Automatic Heuristic Design (AHD) promises to automate writing such rules. However, existing AHD frameworks were developed for combinatorial problems fully specified to the LLM, and they learn only from a scalar fitness score. In real systems, the behaviour that determines a good heuristic, such as processor speeds or power consumption, is unknown a priori: the score reveals which heuristic performs better, but not why. This missing information is recorded in the system logs that every evaluation produces. Exploiting it is non-trivial: logs are massive and noisy, the relevant signals depend on the objective, and their content and format vary across hardware and software stacks, so they can neither be fed to an LLM as is nor processed by a fixed parser. We propose TRACE, which couples an evolutionary AHD loop with an agentic knowledge-extraction workflow. A Reasoner agent analyzes the log schema in light of the objective and formulates hypotheses about the system dynamics; a Coder agent writes and executes schema-specific code to test them, producing insights or executable tools for the evolved heuristics. We evaluate TRACE on a synthetic cloud benchmark and a 5G vRAN scenario built from industrial testbed measurements and operational traffic traces. TRACE consistently outperforms state-of-the-art AHD methods in resource assignment problems and yields more auditable heuristics at under 2% overhead.
Figures & tables
Figure 1: Scalar-feedback AHD (left) learns what works from a single fitness value. TRACE (right) distills system logs into knowledge about why it works (e.g., per-processor speed and energy profiles) and feeds it back into heuristic generation.
Figure 2: Overview of TRACE. Right: the generation loop evolves heuristics; each evaluation returns a fitness J(g) , stored in the algorithm database, and an execution log Λ(g) , stored in a log buffer. Left: at the start of each epoch e , the Agentic Workflow distills the log buffer into Diagnostic Knowledge Ke , merged into the System Information of the generation prompt.
Figure 3: Two runs of the Agentic Workflow. Top: the Reasoner hypothesizes that processing time can be predicted from task features; the Coder writes code that trains a regressor on the logs, yielding a callable surrogate predict() . Bottom: characterizing the energy profile of each processor yields a semantic insight (per-processor average power).
Figure 4: Cost J (log scale, lower is better) over 10 independent runs for Scenario 1 ( N=10 , left) and Scenario 2 ( N=50 , right).
Strategy
Scenario 1 ( N=10 )
Scenario 2 ( N=50 )
Loss
Delay (ms)
Cost ( J )
Loss
Delay (ms)
Cost ( J )
Random
0.70
32.15
103.71
0.10
65.22
117.91
Oracle-Time
0.03
1.30
4.92
0.07
4.81
46.74
Oracle-Queue
0.62
15.76
90.23
0.00
17.32
17.32
FunSearch
0.05
2.43
8.50
0.08
49.39
90.98
EoH
0.05
2.43
8.50
0.01
11.51
15.88
Table 1: Cloud-computing use case: loss rate, mean delay (ms), and cost J (lower is better) of the fixed baselines and of the best of 10 runs of each AHD method, over low, medium, and high loads.
Strategy
Loss
Delay (ms)
Energy ( μ J)
Cost
Random
0.10
0.30
670.01
1.58
Oracle-Time
0.00
0.35
884.22
0.89
Oracle-Queue
0.08
0.27
700.76
1.47
CPU-Only
0.27
0.59
170.46
2.79
HA-Only
0.00
0.21
1132.39
1.13
EoH
0.00
0.31
533.33
0.58
Table 2: 5G vRAN use case: TB loss rate, mean processing delay, mean energy per TB, and cost (equation 2 ; lower is better) for fixed policies (top), industry standards (middle), and the best of 10 runs of each AHD method (bottom).
Strategy
Semantic Grounding
Auditability
Simulatability
Mean
FunSearch
1.00
1.47
2.00
1.49
EoH
1.00
1.33
4.47
2.27
ReEvo
1.00
2.00
5.00
2.67
HSEvo
1.27
1.33
2.27
1.62
TRACE
3.33
4.67
1.00
3.00
Table 3: LLM-as-a-Judge auditability scores for the best heuristics in Scenario 1, averaged over a three-judge panel and five queries per judge. Per-judge scores and rubric definitions are in Appendix F .
Appendix figures & tables7 assets
Supplementary material from the paper’s appendix.
Appendix
Component
Hardware / Service
Tokens (in)
Tokens (out)
Time (s)
Reasoner (Hypothesis gen.)
Google API
454–580
21–25
14–31
Coder (Artifact generation)
Local GPU A100
980–995
1050–1350
7.3–7.8
Sandbox execution
Local CPU Intel i7
–
–
0.1–3.85
Heuristic generation
Local GPU A100
479–495
262–738
1.7–3.5
Heuristic evaluation
Local CPU Intel i7
–
–
0.82–1.3
Appendix
Table 4: Per-component computational profile. Token consumption and execution times are reported as the 5th and 95th percentiles over 10 independent runs.
Block
Component
Avg. call (s)
# calls
Total (s)
% total
Reasoner (Hypothesis gen.)
22.5
1
22.5
4.86
Agentic
Coder (Artifact generation)
7.41
1
7.41
1.60
Workflow
Sandbox execution
2.24
1
2.24
0.48
Sub-total
32.15
6.95
Standard AHD
Heuristic generation
1.92
150
288.0
62.25
steps
Heuristic evaluation
0.95
150
142.5
30.80
Appendix
Table 5: Average wall-clock breakdown for a single TRACE epoch in which the agentic workflow is executed ( I=150 generation steps), over 10 independent runs. TRACE’s Agentic Workflow incurs only 6.95% of the total epoch duration; the remaining 93.05% is attributed to the standard AHD steps shared with baselines such as EoH.
Figure 5: LLM choice for TRACE’s Coder/Evolution loop: distribution of the best per-run cost across 10 independent runs on Scenario 1.
Figure 6: EoH cost distribution across 10 independent runs on Scenario 1 under four LLM backbones, including gemini-3-pro-preview (the model used by TRACE’s Reasoner). Even under the strongest backbone, EoH remains 1 – 2 orders of magnitude above TRACE (cf. Fig. 4 , left, where TRACE’s median cost is below 1 ).
Strategy
Semantic Grounding
Auditability
Simulatability
Mean
Gemini
Claude
GPT
Gemini
Claude
GPT
Gemini
Claude
GPT
FunSearch
1.00
1.00
1.00
1.40
1.00
2.00
1.00
4.00
1.00
1.49
EoH
1.00
1.00
1.00
1.00
1.00
2.00
3.40
5.00
5.00
2.27
ReEvo
1.00
1.00
1.00
2.00
2.00
2.00
5.00
5.00
5.00
2.67
HSEvo
1.00
1.00
1.80
1.00
1.00
2.00
1.00
3.80
2.00
1.62
TRACE
4.00
3.00
3.00
5.00
4.00
5.00
1.00
1.00
1.00
3.00
Appendix
Table 6: Multi-judge auditability scores for the best-found heuristics in Scenario 1. Each cell reports the mean over five queries at temperature 0.7 ; the Mean column averages the nine (rubric, judge) means per method.
Figure 7: Latency-energy trade-off in heterogeneous 5G edge clouds.