Discovered, Not Designed: Population Evolution for Collaborative and Compute-Intensive Model Discovery

Bo Peng; Lizhu Zhang; Yuhang Zhou; Mingyi Wang; Yifan Wu; Serena Li; Xiangjun Fan; Zhuokai Zhao

cs.LGOct 5, 2026

Discovered, Not Designed: Population Evolution for Collaborative and Compute-Intensive Model Discovery

Authors: Bo Peng, Lizhu Zhang, Yuhang Zhou, Mingyi Wang, Yifan Wu, Serena Li, Xiangjun Fan, Zhuokai Zhao

Organizations: Meta

Abstract

LLM-driven evolution enables iterative model development, but two practical goals remain underexplored: finding model designs that transfer across related tasks and sustaining improvement when training is expensive. We introduce Population Evolution (PE), a collaborative, hierarchical framework that connects ongoing local searches through shared experimental evidence. PE evaluates code changes across related training instances and shares the results to guide subsequent proposals and promotion to larger training scales. For expensive targets, PE searches small training subsets and screens candidates through peer and intermediate evaluations before full-target training. We introduce RMD-Bench to evaluate both settings across ranking, watch-time prediction, RL algorithm discovery, and LLM/VLM pretraining. Compared with standalone evolution at matched source iterations, PE raises mean best local gains from 7.01% to 8.97% in ranking and from 2.84% to 3.85% in watch-time, while improving the best larger-scale outcome in all three joint-discovery families. In watch-time discovery, PE improves best larger-scale gains with four of five harnesses and all four proposers. On new recommendation datasets under shared target-side calibration, every evaluated PE design improves over the reference in mean performance. Under matched total GPU compute, completed LLM discovery runs yield a best relative accuracy gain of 2.48% and 13 successful candidates for PE, versus 0.92% and none for direct evolution. VLM loss reduction reaches 8.78% versus 5.05% under matched total GPU compute.

Figures & tables

Panel of Figure 1 — Figure 1: PE connects local searches through replay evidence and cross-scale promotion. (a) Published interventions are retrained on peers; their results enter the shared ledger and inform later proposals (Equation 1 ). (b) Track A tests the same designs across related instances and at larger training scales. (c) Decomposed discovery screens candidates on small sources and both intermediate instances before full-target training. In Track B, full-target results remain outside proposal context.

Evaluation support	MLE- bench ( 2024 )	Innovator Bench ( 2025 )	Research Gym ( 2026 )	PostTrain Bench ( 2026 )	MLS- Bench ( 2026 )	Edge Bench ( 2026 )	RMD- Bench (ours)
Collaboration protocol	—	—	—	—	—	—	$\checkmark$
Cross-dataset transfer	—	—	$\checkmark$	—	$\checkmark$	—	$\checkmark$
Cross-scale transfer	—	—	—	—	$\checkmark$	—	$\checkmark$
Hierarchical training	—	—	—	—	$\triangle$	—	$\checkmark$
Discovery yield vs. compute	—	—	—	—	—	—	$\checkmark$
Calibrated starting configurations	—	—	$\checkmark$	—	$\checkmark$	—	$\checkmark$

Table 1: Explicit benchmark support for collaborative and compute-intensive discovery.

Dataset	S	PE
KuaiLive	8.93	23.19
KuaiRand-1K	5.53	9.29
KuaiRand-Pure	8.03	6.07
KuaiRec	1.03	2.42
KuaiSAR	11.52	3.88
Mean	7.01	8.97

Table 2: PE improves mean best local recommendation gains and best larger-scale outcomes. A–C report relative gains (%) from locally proposed candidates at matched source iterations; C also gives solved-task counts. D reports successful/total larger-scale evaluations and best gains after retraining without local search.

Target metric	Candidate-set mean		Best candidate
Target metric	S	PE	S	PE
KuaiVideo AUC (%) $\uparrow$	74.39	74.45	74.43	74.46
VK-LSVD MAE (s) $\downarrow$	14.55	14.54	14.46	14.53

Table 3: Transfer performance of the designs returned by each discovery procedure.

Domain	Method	Full-target attempts	Successful candidates	Best gain (%)
LLM	Direct	70	0	0.92
LLM	PE	39	13	2.48
VLM	Direct	100	12	5.05
VLM	PE	61	57	8.78

Table 4: Hierarchical discovery improves full-target quality and completed-run yield. Completed-run results use matched total GPU compute. Gains are relative LLM accuracy increases or VLM loss reductions; successful candidates meet full-target thresholds.

Panel of Figure 2 — Figure 2: Peer discoveries can outperform local candidates; larger-scale evaluation tests transfer. Each panel shows one design retrained unchanged across six instances. Gains use fixed instance references; bold values exceed the best gain from locally proposed PE candidates on the receiving instance (Table 2 ). Dark bars identify sources; dashed lines separate larger-scale evaluations. 1K/Pure/27K abbreviate KuaiRand datasets (Appendix D ).

Panel of Figure 3 — Figure 3: Shared checkpoint-averaging evidence and subsequent local development. D2-B evidence is available before D5-B proposal 59; proposal 89 later refines the averaging window. Arrows indicate reconstructed evidence flow and temporal order, not direct parentage. Gains are retrospective comparisons with fixed instance references. Transfer results: Figure 4 (Appendix D.1 ).

Appendix figures & tables33 assets

Supplementary material from the paper’s appendix.

Appendix

Task family	Starting method	Discovery space	Source instances	Final-stage evaluation
Track A: joint discovery across related instances
Ranking	FuXi-Linear ( Ye et al., 2026 )	Sequence representations; temporal and retention mechanisms	KuaiLive, KuaiRand-1K/Pure, KuaiRec, KuaiSAR	KuaiRand-27K; 146.82M interactions; ranking score $\uparrow$
Watch-time	EGMN ( Zhao et al., 2025 )	Distributional head; duration prediction	CIKM16, KuaiRec, WeChat, KuaiRand-Pure/1K	KuaiRand-27K; 23.99M examples; MAE (s) $\downarrow$
RL algorithm	Qwen2.5-Coder ( Hui et al., 2024 ) ; GRPO	Policy loss; rollout filtering	Five sources, different seeds; 1.5B policy, 15 steps	3B policy, 50 steps; 423 LiveCodeBench problems; pass@1 $\uparrow$
Track B: decomposed discovery for expensive targets
LLM pretraining	Gemma 3-style ( Gemma Team, 2025 ) ; $\sim$ 185M	Architecture; optimizer; schedule; training mechanisms	FineWeb-Edu 100B core set; two 2% and three 5% views	Full core set; seven-task accuracy $\uparrow$

Appendix

Table 5: RMD-Bench task families and discovery spaces. Each family provides five source instances. Track A adds a larger-scale evaluation instance; Track B adds two intermediate instances and a full target. Sizes refer to the reference implementations and processed corpora used here. Detailed task specifications follow.

Dataset	Users	Interactions	Items	Mean seq. length	Context length	Batch size per GPU	Epochs
KuaiLive	22,497	4,381,721	151,849	194.77	512	47	30
KuaiRand-1K	988	290,969	2,025	294.50	1,024	38	30
KuaiRand-Pure	25,220	1,235,423	6,073	48.99	512	88	30
KuaiRec	7,176	12,529,113	9,958	1,745.97	1,024	40	50
KuaiSAR	17,491	1,595,760	55,505	91.23	512	46	50
KuaiRand-27K	27,285	146,821,616	461,994	5,381.04	1,024	51	50

Appendix

Table 6: Processed ranking datasets and training allocation. All but KuaiRand-27K are source instances. Batch size is the per-GPU value selected by baseline HPO (Table 16 ).

Component	Setting
Encoder	Four blocks; four attention heads; $d_{q}=d_{k}=d_{v}=32$
Embeddings and position	Item dimension 128; positional linear channel dimension 32
Temporal channel	Eight heads; base 2, stride 3, start index 10; learnable decay
Sequence computation	Chunk size 128; RoPE disabled
Sampled-softmax loss	128 negatives; temperature 0.05
Execution	Eight-GPU DDP; BF16/TF32

Appendix

Table 7: Fixed FuXi-Linear reference configuration.

Dataset	Train	Eval.	Total	User/query vocabulary	Item vocabulary	p99 scale (s)
CIKM16	246,853	61,714	308,567	9,837	122,642	5.237
KuaiRec	980,676	245,169	1,225,845	7,176	9,599	47.697
WeChat	5,059,605	1,264,902	6,324,507	19,994	95,511	165.364
KuaiRand-Pure	1,871,608	467,902	2,339,510	27,279	7,268	182.240
KuaiRand-1K	6,951,281	1,737,821	8,689,102	1,000	2,926,199	188.515
KuaiRand-27K	19,193,265	4,798,317	23,991,582	27,285	5,489,023	187.368

Appendix

Table 8: Processed watch-time datasets. The user/query vocabulary counts queries for CIKM16 and users otherwise; the p99 scale converts normalized MAE to seconds. These counts reflect watch-time preprocessing and differ from ranking.

Dataset	Learning rate	Weight decay	$\alpha$	$\beta$	Dropout
CIKM16	0.091684	$3.35\times 10^{-6}$	0.659274	2.023865	0.141097
KuaiRec	0.25076	$1.94\times 10^{-5}$	0.619568	3.41108	0.059406
WeChat	0.085141	$1.98\times 10^{-6}$	0.454659	3.263077	0.055439
KuaiRand-Pure	0.064052	$8.58\times 10^{-7}$	0.521568	4.296266	0.077754
KuaiRand-1K	0.134232	$2.31\times 10^{-5}$	0.646871	1.511465	0.362471
KuaiRand-27K	0.102368	$1.53\times 10^{-5}$	0.489522	1.778713	0.138667

Appendix

Table 9: Fixed HPO-selected watch-time hyperparameters shared by the S and PE runs with GPT-5.5.

Property	Source (each of five instances)	Larger scale
Policy parameters	1.5B	3B
GRPO steps	15	50
Prompts per step	64	32
Evaluation problems	408	423
Evaluation tests	Public tests	Private tests
Training seed	Instance-specific	Fixed

Appendix

Table 10: Fixed RL training and evaluation contracts. Both training scales use eight responses per prompt and 16 H200 GPUs per candidate evaluation.

Dataset	Role	Shards	Documents	Tokens	Steps
D2-A	Source	36	1,924,096	1,925,811,698	694
D2-B	Source	36	1,915,904	1,921,710,324	694
D5-A	Source	91	4,853,760	4,862,315,015	1,750
D5-B	Source	91	4,844,544	4,859,363,728	1,750
D5-C	Source	91	4,853,760	4,863,973,228	1,750
D10 clean	Intermediate	182	9,712,640	9,727,423,143	3,470

Appendix

Table 11: LLM training hierarchy. Token counts describe the processed corpus; step counts scale approximately with the nominal corpus fractions. All instances use the same seven-task mean accuracy.

Dataset	Role	Shards	Shard indices	Raw rows	Steps
D2-A	Source	20	0–19	48,500	48
D2-B	Source	20	50–69	48,500	48
D5-A	Source	50	0–49	121,250	118
D5-B	Source	50	50–99	121,250	118
D5-C	Source	50	100–149	121,250	118
D10 clean	Intermediate	100	150–249	242,500	237

Appendix

Table 12: VLM training hierarchy. Rows report the raw training-data inventory and prescribed training steps.

Domain	Scientific and computational constraints
Ranking	Avoid attention whose cost is quadratic in the full sequence length.
Watch-time	Preserve non-negative distributional support and stable likelihoods; treat CIKM16 duration as session context.
LLM	Distinguish vector and matrix parameters in Muon and maintain stable optimizer updates.
VLM	Preserve causal attention and the fixed visual interface.
RL	Preserve response groups and masks, and compute stable probability ratios.

Appendix

Table 13: Shared domain guidance for valid candidate designs. These instructions complement the fixed interfaces and evaluators.

Family	Instances	Trial cap $N$
Ranking	Five sources and KuaiRand-27K	50
Watch-time	Five sources and KuaiRand-27K	50
LLM pretraining	Two 2% sources and two intermediate instances	20
LLM pretraining	Three 5% sources	10
VLM pretraining	Five sources and two intermediate instances	20

Appendix

Table 14: Baseline HPO trial caps, applied separately to each instance. Calibration also evaluates the original configuration and confirms the selected configuration. Early stopping uses a patience of ten trials without improvement.

Family / tier	Parameters and ranges
Ranking	Dropout $[0.1,0.6]$ ; learning rate $[10^{-4},10^{-2}]$ (log); per-GPU batch size $[32,128]$ (log integer); weight decay $[10^{-5},10^{-2}]$ (log).
Watch-time	Learning rate $[0.01,0.3]$ (log); weight decay $[10^{-7},10^{-3}]$ (log); $\alpha\in[0,1]$ ; $\beta\in[0.1,5]$ ; dropout $[0,0.5]$ .
LLM: source, intermediate	Peak learning rate $[5\times 10^{-5},10^{-3}]$ (log); weight decay $[0.01,0.5]$ (log); warm-up steps $[50,300]$ (log integer); decay fraction $[0.05,0.3]$ ; Muon momentum $[0.9,0.99]$ ; Q/K clipping threshold $[50,200]$ (log).
VLM: source, intermediate	Same ranges as LLM except warm-up steps: $[2,12]$ for 2% sources, $[5,30]$ for 5% sources, and $[10,60]$ for 10% intermediate instances (log integer).

Appendix

Table 15: Baseline HPO search spaces. “Log” denotes a logarithmic search scale; “log integer” additionally restricts values to integers. Unmarked intervals use a linear scale. LLM/VLM full-target configurations are fixed without HPO, and RL has no HPO search space.

Dataset	Dropout	Learning rate ( $\times 10^{-3}$ )	Batch size	Weight decay ( $\times 10^{-4}$ )
KuaiLive	0.49	2.5	47	4.8
KuaiRand-1K	0.352	3	38	1.1
KuaiRand-Pure	0.28727	7.969455	88	6.251374
KuaiRec	0.3	4.3	40	14
KuaiSAR	0.53	5.2	46	9
KuaiRand-27K	0.378871	3.548962	51	1.985122

Appendix

Table 16: Selected ranking baseline hyperparameters. Batch size is per GPU. Learning rate and weight decay use the column scale factors.

Instance	$\eta$ ( $\times 10^{-4}$ )	$\lambda$	$n_{\mathrm{warm}}$	$f_{\mathrm{decay}}$	$\mu$	$\tau_{\mathrm{QK}}$
LLM pretraining
D2-A	4.480393	0.018409	66	0.143635	0.985564	114.656
D2-B	5.4	0.05	170	0.085	0.919	170
D5-A	4.72	0.022	78	0.108	0.982	121
D5-B	4.480393	0.018409	66	0.143635	0.985564	114.656
D5-C	4.480393	0.018409	66	0.143635	0.985564	114.656

Appendix

Table 17: Fixed LLM/VLM baseline hyperparameters. Learning rates use the indicated scale factor. Columns give peak learning rate

\eta

, weight decay

\lambda

, warm-up steps

n_{\mathrm{warm}}

, decay fraction

f_{\mathrm{decay}}

, Muon momentum

\mu

, and Q/K clipping threshold

\tau_{\mathrm{QK}}

. Full-target rows are prescribed configurations without HPO.

Family / sources	Enabled	Extra trials	Paired control	Total evaluations
Ranking / KuaiRand-1K	Yes	3	Up to 1	Up to 5
Ranking / other four sources	No	0	0	1
Watch-time	No	0	0	1
RL algorithm	No	0	0	1
LLM pretraining	No	0	0	1
VLM pretraining	No	0	0	1

Appendix

Table 18: Per-candidate HPO in the reported GPT-5.5 configuration. Counts cover one local mutation before any separate peer or promotion evaluations. Disabled settings execute neither per-candidate HPO trials nor paired baseline controls.

Task family	GPU-hours per complete discovery run
Ranking	Hundreds
Watch-time prediction	Tens
RL algorithm discovery	Hundreds
LLM pretraining	Thousands
VLM pretraining	Thousands

Appendix

Table 19: Compute scale of a complete discovery run. Each run comprises five searches and their prescribed evaluations. Entries give orders of magnitude for total allocated GPU-hours; exact usage is not disclosed due to policy restrictions.

Family	Source publication	Peer evaluation	Larger scale / intermediate	Full target
Ranking	2% relative	1% relative	2% relative	—
Watch-time	0.5% relative	0.5% relative	0.5% relative	—
RL algorithm	10% rel.; $1.5/408$ abs.	5% rel.; $0.5/408$ abs.	4% rel.; $4/423$ abs.	—
LLM	2% relative	1% relative	2% on both intermediates	1% relative
VLM	10% relative	10% rel.; 0.02 abs.	6% rel.; 0.02 abs., both intermediates	4% rel.; 0.015 abs.

Appendix

Table 20: PE’s publication and promotion margins. Relative gains are direction-normalized; where an absolute margin is listed, both conditions apply. Every peer gate requires at least two wins among four peers.

Domain	Dataset / summary	S	PE	PE best including peer candidates †
Ranking	KuaiLive	+8.93%	+23.19%	+23.19%
	KuaiRand-1K	+5.53%	+9.29%	+10.66%
	KuaiRand-Pure	+8.03%	+6.07%	+6.07%
	KuaiRec	+1.03%	+2.42%	+6.47%
	KuaiSAR	+11.52%	+3.88%	+10.06%
	Mean	+7.01%	+8.97%	+11.29%

Appendix

Table 21: Local discovery outcomes and PE’s best gains from local and replayed peer candidates. Gains are relative improvements. For recommendation, S and PE report the best locally proposed gain on each dataset; the RL row reports the maximum across five searches of the same task. Watch-time uses the GPT-5.5 paired comparison.

Domain	Dataset	Local PE minus S	PE including peers minus S
Ranking	KuaiLive	+14.26	+14.26
	KuaiRand-1K	+3.76	+5.13
	KuaiRand-Pure	$-1.96$	$-1.96$
	KuaiRec	+1.39	+5.44
	KuaiSAR	$-7.64$	$-1.46$
Watch-time	CIKM16	+5.23	+5.23

Appendix

Table 22: Dataset-wise PE–S differences in relative gain. Entries are percentage-point differences derived from the rounded values in Table 21 . The two columns distinguish local proposals from the best PE candidates available after peer replay.

Family	Source / iteration	Principal intervention	Final-stage outcome
Ranking	KuaiRand-Pure / 57	Gated length-64 causal attention alongside linear retention; inherited temporal and spectral components.	KuaiRand-27K: 3.02% ranking gain.
Watch-time	WeChat / 21	Duration-conditioned, left-truncated Laplace mixture with a learned near-median quantile readout.	KuaiRand-27K: 2.17% MAE reduction.
RL algorithm	Source 2 / 87	Bounded advantage weights based on trajectory distance from the reference policy.	3B, 50 steps: 6.64% pass@1 gain.
LLM	D5-B / 89	EMA/SWA checkpoint averaging over approximately the final 20% of training.	Full target: 2.48% accuracy gain.
VLM	D5-A / 86	Gated global image-context injection into patch tokens, combined with square-root warm-up.	Full target: 7.51% loss reduction.

Appendix

Table 23: One concrete graduated design from each task family. Gains are relative to the fixed reference at the stated final-stage instance. Each row identifies a complete candidate; the intervention column summarizes its principal changes. The success criteria are specified in Table 20 .

Design	Metric	Parent score	Candidate score	Gain (%)
LLM, D5-B / 89	Full-target accuracy (%)	41.13	41.25	0.28
VLM, D5-A / 86	Full-target loss	1.1200	1.1164	0.33
RL, source 2 / 87	Source pass@1 (%)	3.43	3.19	$-7.14$

Appendix

Table 24: Direct-parent comparisons for the representative cases. Positive relative gains indicate improvement and are computed from unrounded scores. These comparisons use the direct parent rather than the fixed benchmark reference.

Panel of Figure 4 — Figure 4: The checkpoint-averaging candidate from D5-B proposal 89 improves seven of eight instances. Gains use each instance’s fixed reference; the candidate exceeds the 1% full-target threshold. The darker bar marks its source, vertical dividers separate training stages, and the dashed horizontal line marks the full-target threshold.

Panel of Figure 5 — Figure 5: PE improves the best larger-scale or full-target outcome in all five task families. Each pair compares PE with standalone evolution (Track A) or direct full-target evolution (Track B). The Track A values match Table 2 D, and the Track B values match Table 4 . These are completed-run best gains; Track B uses matched total GPU compute. Figure 6 summarizes the same LLM runs.

A. Search harnesses (GPT-5.5; high)
		Best gain on each source dataset					Larger scale KuaiRand-27K
Configuration ( $k$ )	Method	CIKM16	KuaiRec	WeChat	KuaiRand Pure	KuaiRand 1K	Best gain	Passed / tested
AdaEvolve (2)	S	7.32	0.40	0.80	0.75	4.93	1.77	4/10
AdaEvolve (2)	PE	12.55	0.67	0.81	0.16	5.08	2.17	5/7
EvoX (4)	S	7.01	0.28	0.49	0.30	1.87	1.45	4/20
EvoX (4)	PE	11.01	0.58	0.49	0.11	3.01	1.05	18/18

Appendix

Table 25: PE improves the best larger-scale gain over S with four of five harnesses and all four proposers. Watch-time gains are relative MAE reductions (%). Source columns report the best local result on each source instance; KuaiRand-27K reports larger-scale transfer, with passed/tested counts referring to evaluations. Bold marks the better value within each S/PE pair, using unrounded scores.

Panel of Figure 6 — Figure 6: PE improves LLM quality and discovery yield under matched total GPU compute. Both panels summarize completed runs. (a) Best relative full-target gains are 2.48% for PE and 0.92% for Direct. (b) The predefined 1% criterion is met by 13 candidates from 39 PE full-target attempts and none from 70 Direct attempts.

Relative gain threshold (%)	PE	Direct
0.50	21	6
1.00	13	0
1.50	8	0
2.00	3	0

Appendix

Table 26: LLM success counts across full-target gain thresholds. Counts use the same completed runs under matched total GPU compute; the predefined main threshold is bold.

Component	Specification
Examples (train / val. / test)	7,505,820 / 1,516,831 / 1,275,643
Chronology	5/1/1-week split after historical-feature weeks 0–19; training uses weeks 20–24
Numerical features	Video duration and 31 other features; normalization constant: 157 seconds
Categorical features	Age, gender, geographic region, place, platform, and user agent
Duration encoding	50 quantile-based buckets fitted on training data
Shared EGMN encoder	16-dimensional categorical embeddings, scalar linear projections, and a 256/128/64 MLP

Appendix

Table 27: Shared VK-LSVD transfer protocol. Final candidates train on the released training split and are evaluated on the released test split.

Method	Candidate / source	MAE (s) $\downarrow$	Paired reduction (%) $\uparrow$
Reference	EGMN	$14.73\pm 0.02$	—
PE	e33792c7 †	$14.53\pm 0.02$	$1.36\pm 0.15$
PE	2f3242e3 †	$14.53\pm 0.02$	$1.36\pm 0.15$
PE	92f5eb20 †	$14.53\pm 0.02$	$1.36\pm 0.15$
PE	e0b2ffa7	$14.53\pm 0.01$	$1.35\pm 0.06$
PE	94db2d04	$14.58\pm 0.03$	$1.02\pm 0.30$

Appendix

Table 28: VK-LSVD reference and all evaluated transfer candidates. Values are means

\pm

sample standard deviations across three training seeds. Reductions are computed against the seed-matched reference. PE identifiers are archived candidate-ID suffixes; S entries identify their source dataset.

Component	Specification
Candidate rows (train / test)	10,931,092 / 2,730,291
Requests (train / test)	803,007 / 209,728; 10,000 users per split
Item vocabulary	Minimum training count 10; 632,404 known items, plus padding and unknown entries
Visual features	64-dimensional vectors for 3,242,315 items; frozen, with separate padding and unknown entries
Item representation	Learned item-ID embedding plus a trainable, bias-free $64\!\rightarrow\!128$ projection of its visual vector
History encoding	Shared sequence encoder and embeddings for positive and negative histories; at most 100 items each

Appendix

Table 29: Shared KuaiVideo data and adapter specifications. The discovered sequence-model code is the only component that differs among the reference, S, and PE candidates.

Parameter	Search range	Shared selected value
VK-LSVD
Learning rate	$[0.01,0.3]$	0.076611
Weight decay	$[10^{-7},10^{-3}]$	$4.20799\times 10^{-7}$
Entropy coefficient $\alpha$	$[0,1]$	0.37454
Regression coefficient $\beta$	$[0.1,5]$	4.7585
Dropout	$[0,0.5]$	0.365997

Appendix

Table 30: Shared calibration for the new recommendation targets. Baseline configuration and training-duration selection use the released test split; the selected protocol is reused unchanged for the reference and every candidate.

Model	AUC (%) $\uparrow$	User gAUC (%) $\uparrow$	Log loss $\downarrow$
Reference	$74.39\pm 0.11$	$66.14\pm 0.06$	$0.4378\pm 0.0008$
Best S	$74.43\pm 0.08$	$66.27\pm 0.05$	$0.4371\pm 0.0018$
Best PE	$\mathbf{74.46\pm 0.06}$	$\mathbf{66.29\pm 0.02}$	$\mathbf{0.4365\pm 0.0005}$

Appendix

Table 31: KuaiVideo reference and best S and PE designs. The best S and PE designs are selected by mean AUC across three training seeds; all metrics in a row come from that same design. Every metric is reported as mean

\pm

sample standard deviation over those same runs.

Candidate	AUC (%) $\uparrow$	User gAUC (%) $\uparrow$	Log loss $\downarrow$
Reference	$74.39\pm 0.11$	$66.14\pm 0.06$	$0.4378\pm 0.0008$
b3cb61d8	$74.46\pm 0.06$	$66.29\pm 0.02$	$0.4365\pm 0.0005$
5f007905	$74.46\pm 0.07$	$66.26\pm 0.01$	$0.4371\pm 0.0006$
7ac9ec81	$74.46\pm 0.09$	$66.29\pm 0.09$	$0.4369\pm 0.0011$
e7ec2459	$74.46\pm 0.09$	$66.29\pm 0.09$	$0.4369\pm 0.0011$
2755e3d1	$74.46\pm 0.09$	$66.29\pm 0.09$	$0.4369\pm 0.0011$

Appendix

Table 32: All evaluated PE ranking candidates on KuaiVideo. Identifiers are archived candidate-ID suffixes. Values are means

\pm

sample standard deviations across three training seeds. Every candidate’s mean AUC exceeds the reference before rounding.

Source	Candidate	AUC (%) $\uparrow$	User gAUC (%) $\uparrow$	Log loss $\downarrow$
—	Reference	$74.39\pm 0.11$	$66.14\pm 0.06$	$0.4378\pm 0.0008$
KuaiRand-1K	560d182a	$74.43\pm 0.08$	$66.27\pm 0.05$	$0.4371\pm 0.0018$
KuaiRand-1K	35335eb3	$74.42\pm 0.08$	$66.27\pm 0.03$	$0.4371\pm 0.0013$
KuaiRand-1K	e5c01c96	$74.42\pm 0.08$	$66.27\pm 0.03$	$0.4371\pm 0.0013$
KuaiRand-1K	5d08ec24	$74.42\pm 0.08$	$66.26\pm 0.04$	$0.4371\pm 0.0013$
KuaiSAR	3a70637c	$74.41\pm 0.10$	$66.28\pm 0.07$	$0.4368\pm 0.0014$

Appendix

Table 33: All evaluated S ranking candidates on KuaiVideo. Each source contributes its top four candidates by source performance. Identifiers are archived candidate-ID suffixes; values are means

\pm

sample standard deviations across three training seeds.

Method	Candidate	Solved / 632 in each rerun	Pass@1 (%) $\uparrow$
Reference	—	145, 150, 146	$23.26\pm 0.42$
PE	93d95a15	146, 143, 139	$22.57\pm 0.56$
PE	cbf1c653	146, 144, 148	$23.10\pm 0.32$
S	a0936bce	141, 143, 142	$22.47\pm 0.16$
S	db21a301	146, 145, 144	$22.94\pm 0.16$
S	77431a4e	136, 140, 136	$21.73\pm 0.37$

Appendix

Table 34: Reference and all evaluated RL candidates on LiveCodeBench-632. Each entry uses three independent stochastic reruns with nominal training seed 42. Pass@1 is reported as mean

\pm

sample standard deviation. The reference uses the unmodified configuration without target-side HPO.

Explore similar work

Relay, Don't Route: Adaptive Population Handoff for Cost-Efficient LLM-Driven Evolution
Aug 6, 2026Sichun Luo, Yi Huang, Guanzhi Deng +6Evolutionary Optimization Methods Handoff
1The University of Hong Kong · 2JIUTIAN Research, China Mobile · 3City University of Hong Kong +1
LEVI: Stronger Search Architectures Can Substitute for Larger LLMs in Evolutionary Search
May 10, 2026Temoor TanveerEvolutionary Search Alphaevolve
Independent Researcher
SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution
Apr 27, 2026Sichun Luo, Yi Huang, Haochen Luo +7Evolutionary Search Strategic Reasoning
1The University of Hong Kong · 2JIUTIAN Research, China Mobile · 3City University of Hong Kong

Aug 6, 2026, cs.CLJ/K move · Enter open · S save

Relay, Don't Route: Adaptive Population Handoff for Cost-Efficient LLM-Driven Evolution

Sichun Luo, Yi Huang, Guanzhi Deng, Haibo Wang+5

1The University of Hong Kong · 2JIUTIAN Research, China Mobile · 3City University of Hong Kong+1

Evolutionary Optimization Methods Handoff

Open paper AskPDF arXiv

Large language model (LLM)-driven evolution has shown promise for program search and algorithm discovery, but relying on strong models throughout long evolutionary runs is costly. A natural alternative is to combine cheap and strong models under a fixed inference budget. However, existing approaches typically allocate models at the level of individual queries or mutation steps, overlooking that evolutionary search is \textit{stateful}: each generated candidate changes the population from which subsequent mutations are produced. We empirically analyze LLM-driven evolutionary trajectories and find that search progress is strongly front-loaded, early trajectory performance is informative but noisy, and cheap models recover much of the early progress achieved by strong models at lower cost. Motivated by these findings, we propose \textbf{\model}, a training-free framework that shifts budget allocation from individual calls to evolving populations through adaptive \textit{population handoff}. A cheap model explores multiple trajectories in short blocks allocated by a bandit scheduler. Relay Gain, defined as the marginal improvement of a compact, quality-diverse candidate bank constructed for handoff, serves as the scheduler reward and determines when to hand off. The curated candidates initialize a shared strong model population for refinement. Across four benchmarks and three budgets, \model achieves the highest mean score in 11 of 12 settings, outperforming competitive baselines. Our results suggest that in stateful search, budget allocation should be organized around the population, not the individual call.

Abstract

Explore similar work

Relay, Don't Route: Adaptive Population Handoff for Cost-Efficient LLM-Driven Evolution

LEVI: Stronger Search Architectures Can Substitute for Larger LLMs in Evolutionary Search

SeaEvo: Advancing Algorithm Discovery with Strategy Space Evolution