Collaborative Principle Evolution via Evidence Transfer for Scientific Discovery

Yingming Pu; Hongyu Chen; Tao Lin

cs.LGSep 28, 2026

Collaborative Principle Evolution via Evidence Transfer for Scientific Discovery

Authors: Yingming Pu, Hongyu Chen, Tao Lin

Organizations: Zhejiang University, Hangzhou, Zhejiang, China · Westlake University, Hangzhou, Zhejiang, China

Abstract

Large Language Model (LLM)-based agents promise to automate scientific discovery, yet exploring the vast hypothesis space remains costly. Existing principle-evolution methods accelerate this loop, but operate sequentially, which caps exploration breadth and wastes wall-clock time on challenging problems. To address this, we formulate collaborative scientific discovery as evidence transfer between parallel principle-evolution branches. We present COEVOLVE, which realizes this transfer through a coordination core over parallel branches. By integrating value-of-information-gated routing and context-discounted likelihood injection, COEVOLVE enables branches to collaborate through shared measurements while keeping their principle posteriors separate. Across six scientific-discovery tasks under a matched evaluation budget, COEVOLVE attains a mean solution quality of 66.5% versus 57.0% for single-branch principle evolution, with a 1.80x mean wall-clock speedup on the GPT-5.6-Terra backbone; on five auto-research tasks delegated to an autonomous research harness, it is the only arm whose mean stays above the published SOTA anchor on every task. These results establish when evidence sharing accelerates parallel discovery and when transfer safeguards are necessary to limit negative or inert transfers

Figures & tables

Figure 1: Principle co-evolution. Independent branches explore different principle subspaces through a hypothesis-testing loop. CoEvolve shares evidence, allowing the branches to collaborate while retaining separate search histories.

Figure 2: Overview of CoEvolve . After every hypothesis-testing cycle, the coordination core (bottom) collects hypothesis–outcome pairs as evidence, merges them into a global evidence pool, scores each source–target pair, and routes and injects only evidence satisfying the gate. An accepted record enters the target branch as the discounted likelihood factor $q_{j\leftarrow i}(y\mid h,P)^{\alpha}$ ; rejected records remain in the evidence pool.

Model / Method	MBO	NHO	SPO	TMC	AMP	Promoter	Average
Gemma-4-31B-IT
Vanilla MAS	39.2 $\pm$ 0.8	77.2 $\pm$ 1.4	0.0 $\pm$ 0.0 $\downarrow$ 0.0x	74.4 $\pm$ 0.0 $\uparrow$ 1.3x	27.4 $\pm$ 2.3 $\uparrow$ 1.1x	84.4 $\pm$ 0.3 $\uparrow$ 4.5x	50.4 $\uparrow$ 1.2x
The AI Scientist v1 ( Lu et al., 2024 )	48.6 $\pm$ 2.2 $\uparrow$ 1.2x	78.3 $\pm$ 4.2 $\uparrow$ 1.0x	0.0 $\pm$ 0.0 $\downarrow$ 0.0x	84.5 $\pm$ 0.0 $\uparrow$ 1.5x	24.1 $\pm$ 1.5	18.7 $\pm$ 32.4	42.4 $\uparrow$ 1.0x
The AI Scientist v2 ( Yamada et al., 2025 )	48.5 $\pm$ 5.1 $\uparrow$ 1.2x	84.3 $\pm$ 3.9 $\uparrow$ 1.1x	8.2 $\pm$ 1.6	83.7 $\pm$ 0.8 $\uparrow$ 1.5x	27.7 $\pm$ 3.0 $\uparrow$ 1.2x	0.0 $\pm$ 0.0 $\downarrow$ 0.0x	42.1
AI Researcher ( Tang et al., 2025 )	50.0 $\pm$ 2.4 $\uparrow$ 1.3x	88.3 $\pm$ 1.8 $\uparrow$ 1.1x	24.6 $\pm$ 2.0 $\uparrow$ 3.0x	81.5 $\pm$ 6.1 $\uparrow$ 1.5x	26.7 $\pm$ 14.6 $\uparrow$ 1.1x	19.3 $\pm$ 33.5 $\uparrow$ 1.0x	48.4 $\uparrow$ 1.2x
EvoScientist ( Lyu et al., 2026 )	45.3 $\pm$ 1.5 $\uparrow$ 1.2x	93.3 $\pm$ 0.7 $\uparrow$ 1.2x	9.2 $\pm$ 0.7 $\uparrow$ 1.1x	85.4 $\pm$ 0.0 $\uparrow$ 1.5x	29.7 $\pm$ 2.6 $\uparrow$ 1.2x	54.0 $\pm$ 7.8 $\uparrow$ 2.9x	52.8 $\uparrow$ 1.3x

Table 1: Performance comparison with baselines on 6 benchmarks under the Solution Quality (SQ %) metric. Categorized by LLMs: Gemma-4-31B-IT and GPT-5.6-Terra with both non-thinking mode. We report mean

\pm

std over three seeds. On the GPT-5.6-Terra backbone, CoEvolve ’s paired per-seed SQ gain over PiEvo is positive in 15 of 18 task–seed pairs (two-sided sign test

p{<}0.01

; Wilcoxon

p{=}0.001

; full tests in Appendix G ). Bold marks the best mean per task within each backbone block and the best Average. The colored ratio after each cell highlights superior (

\uparrow

) or inferior (

\downarrow

) SQ relative to the worst non-zero mean on the same backbone and task, which is underlined .

Model / Method

MBO

NHO

SPO

TMC

AMP

Promoter

Average

APD

\uparrow

AUOC

\uparrow

APD

\uparrow

AUOC

\uparrow

APD

\uparrow

AUOC

\uparrow

APD

\uparrow

AUOC

\uparrow

APD

\uparrow

AUOC

\uparrow

APD

\uparrow

AUOC

\uparrow

Avg APD

\uparrow

Avg AUOC

\uparrow

Gemma-4-31B-IT

Vanilla MAS

79.1

\pm

1.1

37.5

\pm

0.9

24.0

\pm

1.1

76.5

\pm

1.0

0.0

\pm

0.0

\pm

0.0

29.8

\pm

4.8

73.7

\pm

0.5

76.1

\pm

0.9

25.7

\pm

2.6

36.9

\pm

1.1

80.2

\pm

4.4

41.0

48.9

The AI Scientist v1 ( Lu et al., 2024 )

21.7

\pm

17.0

38.2

\pm

3.4

5.9

\pm

3.4

76.8

\pm

3.6

0.0

\pm

0.0

\pm

0.0

53.1

\pm

1.1

81.5

\pm

1.0

58.5

\pm

3.0

17.2

\pm

1.1

0.0

\pm

0.0

18.7

\pm

32.4

23.2

\downarrow

0.6x

38.7

\downarrow

0.8x

The AI Scientist v2 ( Yamada et al., 2025 )

34.7

\pm

12.5

43.6

\pm

3.5

12.9

\pm

2.2

80.5

\pm

3.0

4.6

\pm

2.9

7.7

\pm

1.0

47.4

\pm

4.2

78.5

\pm

0.7

34.3

\pm

6.2

25.7

\pm

3.5

0.0

\pm

0.0

\pm

0.0

22.3

\downarrow

0.5x

39.3

\downarrow

0.8x

AI Researcher ( Tang et al., 2025 )

46.5

\pm

9.0

46.1

\pm

2.1

16.6

\pm

4.8

84.5

\pm

1.4

4.7

\pm

0.1

21.2

\pm

1.9

52.3

\pm

5.3

81.1

\pm

5.9

73.1

\pm

8.7

23.5

\pm

14.8

18.3

\pm

31.7

18.3

\pm

31.7

35.3

\downarrow

0.9x

45.8

\downarrow

0.9x

Table 2: Exploration (APD

\uparrow

) and exploitation (AUOC

\uparrow

) comparison with baselines on 6 benchmarks. Both over the same three seeds as Table 1 . Bold marks the best mean per column within each backbone and underline the second-best; the colored ratio after each Average highlights superior (

\uparrow

) or inferior (

\downarrow

) performance relative to Vanilla on the same backbone. CoEvolve attains the best exploration–exploitation balance.

Arm	ALDE	Deconv	InvScat	D2D	MolEdit	Average
Claude Code	+2.3 $\pm$ 0.2	+14.0 $\pm$ 7.3	$-$ 5.8 $\pm$ 7.6	+9.0 $\pm$ 0.3	+0.2 $\pm$ 1.7	+3.9
Codex	+2.2 $\pm$ 1.0	+14.3 $\pm$ 5.4	$-$ 5.9 $\pm$ 6.7	+11.2 $\pm$ 0.9	+1.2 $\pm$ 1.9	+4.6
Arbor	+1.3 $\pm$ 1.9	+11.7 $\pm$ 23.8	$-$ 6.5 $\pm$ 5.9	+15.3 $\pm$ 0.1	+0.5 $\pm$ 0.6	+4.5
PiEvo	+2.4 $\pm$ 2.0	+15.4 $\pm$ 28.6	$-$ 5.3 $\pm$ 7.9	+15.4 $\pm$ 0.1	+1.3 $\pm$ 1.6	+5.8
CoEvolve	+6.6 $\pm$ 0.7	+45.4 $\pm$ 1.4	+11.9 $\pm$ 3.7	+15.5 $\pm$ 0.0	+4.1 $\pm$ 0.6	+16.7

Table 3: Performance comparison on five autoresearch tasks (anchor-relative improvement

\Delta\times 100

\uparrow

). Mean

\pm

std over three seeds; bold = best per column. All five CoEvolve means are significantly above the anchor (one-sample

t

against zero,

p{<}0.025

per task; Appendix G ).

Appendix figures & tables10 assets

Supplementary material from the paper’s appendix.

Appendix

Symbol	Default	Role
$\eta$	$0.0$	safe-VoI gate margin (Eq. ( 2 ))
$B$	$3$	routing quota per branch per round
$R_{\max}$	$0.3$	redundancy cap (token-Jaccard)
$\lambda$	$1.0$	import cost weight (Eq. ( 1 ))
$\delta$	$0.05$	total false-import budget (alpha-spending)
$c_{\mathrm{read}}$	$5\times 10^{-4}$	parsing cost

Appendix

Table 4: Default constants of the coordination core. All values are fixed across tasks, backbones, and ablations unless the ablation varies them explicitly.

	Worst-branch SQ (%)			Best-of-seed SQ (%)
Task	CoEvolve shared	Isolated	$\Delta$	CoEvolve shared	Isolated	Independent	$\Delta$
MBO	$54.8$ $\pm\,4.7$	$45.3$ $\pm\,10.0$	${\color[rgb]{0.4883,0.1641,0.8203}+9.4}$	$77.1$ $\pm\,4.6$	$63.5$ $\pm\,7.9$	$53.6$ $\pm\,13.6$	${\color[rgb]{0.4883,0.1641,0.8203}+13.6}$
NHO	$90.8$ $\pm\,2.0$	$89.9$ $\pm\,2.7$	${\color[rgb]{0.4883,0.1641,0.8203}+0.9}$	$94.9$ $\pm\,0.1$	$94.4$ $\pm\,0.6$	$95.0$ $\pm\,0.0$	${\color[rgb]{0.4883,0.1641,0.8203}+0.5}$
SPO	$11.7$ $\pm\,2.3$	$10.4$ $\pm\,0.4$	${\color[rgb]{0.4883,0.1641,0.8203}+1.3}$	$30.0$ $\pm\,0.6$	$27.5$ $\pm\,4.9$	$25.5$ $\pm\,4.2$	${\color[rgb]{0.4883,0.1641,0.8203}+2.5}$
TMC	$80.1$ $\pm\,8.5$	$84.8$ $\pm\,0.2$	${\color[rgb]{0.8203,0.3008,0.3008}-4.7}$	$88.3$ $\pm\,4.2$	$87.8$ $\pm\,4.5$	$87.7$ $\pm\,4.6$	${\color[rgb]{0.4883,0.1641,0.8203}+0.5}$
AMP	$28.7$ $\pm\,1.0$	$19.8$ $\pm\,1.4$	${\color[rgb]{0.4883,0.1641,0.8203}+8.9}$ ⋆	$37.6$ $\pm\,2.6$	$29.7$ $\pm\,2.0$	$32.3$ $\pm\,0.6$	${\color[rgb]{0.4883,0.1641,0.8203}+7.9}$ ⋆

Appendix

Table 5: Ablation of Evidence transfer.

\Delta

is the difference of arm means with ⋆ marking Welch’s

t

significance at

p<0.05

; purple favors sharing. Promoter’s Isolated arm runs the current

T{=}1.1

protocol.

Arm	knocked-out module	SQ	worst-br.	$\Delta$ SQ	imports	$\bar{\alpha}$
CoEvolve (full)	all modules on (main-table arm)	$77.1$ $\pm\,4.6$	$54.8$ $\pm\,4.7$	–	42	$0.23$
$-$ VoI gate	safe-VoI gate off ( $\eta{\to}-\infty$ ): every routed import accepted	$75.0$ $\pm\,11.1$	$52.8$ $\pm\,15.3$	${\color[rgb]{0.8203,0.3008,0.3008}-2.1}$	137	$0.24$
$-$ discounting	$\alpha{\equiv}1$ : imports applied at full weight	$58.1$ $\pm\,4.4$	$43.3$ $\pm\,9.5$	${\color[rgb]{0.8203,0.3008,0.3008}-19.0}$	58	$1.00$
$-$ redundancy control	$R_{\max}{=}1$ : redundancy cap disabled	$67.5$ $\pm\,4.4$	$51.5$ $\pm\,7.8$	${\color[rgb]{0.8203,0.3008,0.3008}-9.6}$	34	$0.23$
$-$ sharing (Isolated)	coordination on, zero evidence flow	$63.5$ $\pm\,7.9$	$45.3$ $\pm\,10.0$	${\color[rgb]{0.8203,0.3008,0.3008}-13.6}$	0	–

Appendix

Table 6: Ablation of modules in CoEvolve .

Task	CoEvolve worst-branch	PiEvo	$\Delta$	accepted	net unique	dedup	uplift (SQ pts)	inert
	Per-branch SQ (%)			Imports
MBO	$54.8$ $\pm\,4.7$	$61.8$ $\pm\,2.7$	${\color[rgb]{0.8203,0.3008,0.3008}-7.0}$	120	99	$1.26\times$	$+15.41$	$80\%$
NHO	$90.8$ $\pm\,2.0$	$94.9$ $\pm\,0.1$	${\color[rgb]{0.8203,0.3008,0.3008}-4.1}$	19	17	$1.29\times$	$+0.15$	$89\%$
SPO	$11.7$ $\pm\,2.3$	$10.6$ $\pm\,0.0$	${\color[rgb]{0.4883,0.1641,0.8203}+1.1}$	6	6	$1.00\times$	$+3.64$	$67\%$
TMC	$80.1$ $\pm\,8.5$	$84.7$ $\pm\,0.0$	${\color[rgb]{0.8203,0.3008,0.3008}-4.6}$	2	2	$2.50\times$	$+43.29$	$100\%$
AMP	$28.7$ $\pm\,1.0$	$31.4$ $\pm\,5.8$	${\color[rgb]{0.8203,0.3008,0.3008}-2.7}$	96	80	$1.26\times$	$+25.41$	$81\%$

Appendix

Table 7: Ablation of evidence transfer with net value. accepted

=

evidence-node

\times

branch applications; net unique

=

distinct evidence hashes delivered; dedup

=

deliveries per unique evidence (delivery redundancy); uplift

=

pooled-best SQ gain over evaluation windows containing a delivery, averaged over seeds: a window-level, associational statistic (deliveries are value-selected, so uplift does not isolate a causal transfer effect), unlike the branch-level, immediate inert rate, with which it need not agree; inert

=

fraction of accepted imports with zero immediate running-best change. TMC and Promoter aggregate only

2

and

1

accepted imports. All import columns aggregate the three main-table seeds.

Task	Per-seed SQ differences	Paired $t$ $p$ -value
MBO	$+5.2,\ +6.6,\ +39.3$	$0.266$
NHO	$+9.4,\ +6.8,\ +6.9$	$0.012$
SPO	$-0.1,\ +14.5,\ +0.3$	$0.415$
TMC	$+2.5,\ \phantom{+}0.0,\ +1.8$	$0.194$
AMP	$+2.0,\ +5.0,\ +5.0$	$0.057$
Promoter	$+2.9,\ -2.8,\ +8.0$	$0.481$

Appendix

Table 8: Paired per-seed SQ differences, CoEvolve minus PiEvo (GPT-5.6-Terra). Two-sided paired

t

test per task over three seeds; the aggregate row pools all 18 task–seed pairs.

	vs PiEvo		vs InternAgent-1.5
Task	Gemma	Terra	Gemma	Terra
MBO	$0.013^{\star}$	$0.263$	$0.005^{\star}$	$0.453$
NHO	$1.000$	$0.004^{\star\star}$	$0.300$	$0.386$
SPO	$3{\times}10^{-4}\,{}^{\star\star}$	$0.414$	$2{\times}10^{-4}\,{}^{\star\star}$	$0.501$
TMC	$0.276$	$0.184$	$0.029^{\star}$	$0.445$
AMP	$0.197$	$0.058$	$0.011^{\star}$	$0.085$

Appendix

Table 9: Welch

t

tests on final SQ from the summary statistics of Table 1 (two-sided,

n{=}3

per cell, computed from rounded means and stds). ⋆

p{<}0.05

; ⋆⋆ survives Bonferroni over the 12 pairs.

	Billable			All-in
Task	CoEvolve	PiEvo	ratio	CoEvolve	PiEvo
ALDE	$7.5$	$14.5$	$0.52\times$	$350.4$	$339.1$
Deconv	$95.4$	$37.6$	$2.53\times$	$181.5$	$90.3$
InvScat	$89.9$	$46.3$	$1.94\times$	$363.8$	$136.6$
D2D	$11.8$	$8.8$	$1.35\times$	$331.4$	$192.4$
MolEdit	$23.3$	$16.6$	$1.40\times$	$628.6$	$237.1$

Appendix

Table 10: Token accounting in the harness regime (three-seed means, millions of tokens).

Task	$q^{*}$	$T_{\mathrm{PiEvo}}$	$T_{\textnormal{CoEvolve}}$	$T_{\mathrm{PiEvo}}/T_{\textnormal{CoEvolve}}$	TTT PiEvo	TTT ${}_{\textnormal{CoEvolve}}$	ratio
ALDE	+0.024	32.5	20.8	1.56 $\times$	6.64	7.67	0.9 $\times$
Deconv	+0.154	22.4	12.1	1.85 $\times$	1.38	1.11	1.2 $\times$
InvScat	-0.053	47.7	59.7	0.80 $\times$	11.87	15.17	0.8 $\times$
D2D	+0.154	31.7	38.0	0.83 $\times$	17.05	13.45	1.3 $\times$
MolEdit	+0.013	45.9	30.8	1.49 $\times$	11.52	6.03	1.9 $\times$

Appendix

Table 11: Wall-clock conversion speed per task in the harness regime.

q^{*}

= PiEvo’s mean final quality (raw anchor-relative

\Delta

; multiply by

100

for the scale of Table 3 );

T

= end-to-end wall-clock (h); TTT = per-run hours until the run’s running best first reaches

q^{*}

. CoEvolve ’s self-TTT ratio against PiEvo is

0.8\times

–

1.9\times

, and its end-to-end wall-clock is shorter on three of five tasks.

Task	Domain	Search space	Surrogate score	Anti-memorization gate	Ref. scale
NHO	nanophotonics	7-dim continuous	$g$ -factor $\in[0,2]$	surrogate-defined optimum	$[0,\,2]$
MBO	bio-chemistry	SMILES (drug-like)	pChEMBL	surrogate-defined optimum; Lipinski + PAINS	$[0,\,12]$
AMP	biology	peptide 12–50 aa	AMP probability $\in[0,1]$	net charge $\leq 0$ (cationic recipe banned)	$[0,\,1]$
Promoter	biology	50-nt DNA	expression bin $\in[0,17]$	yeast GRF + GC-box banned	$[0,\,17]$
TMC	chemistry	4-of-49 ligand comb.	polarizability ( $\leq 500$ )	pool-computed optimum; charge neutrality, distinct ligands	$[0,\,500]$
SPO	materials	cuprate formulas	$T_{c}$ (K)	record/textbook backbones banned (Hg/Tl, BSCCO, YBCO)	$[0,\,298.15]$

Appendix

Table 12: Benchmark summary. All tasks maximize the gated surrogate score. “Anti-memorization gate” = the per-task mechanism that prevents reaching the optimum from parametric memory (the three mechanisms of Section I.1 ); scores in native units. Ref. scale

=[y_{\mathrm{lo}},y_{\mathrm{hi}}]

, the domain-defined absolute bounds of the task’s score used to normalize SQ ( Section 4.2 ); they are fixed properties of the benchmark (physical laws, scale endpoints, or the originating benchmark’s theoretical maximum), never derived from any method’s runs. On NHO, AMP, Promoter, and TMC the scale is a mathematical or benchmark-defined bound, so

\mathrm{SQ}\leq 100\%

holds by construction; on MBO and SPO it is a domain-anchored reference far above every observed run (best observed:

77.1

and

30.0

SQ), so the bound holds empirically with wide margin.

		Gemma-4-31B-IT		GPT-5.6-Terra
Task	Unit (scale)	CoEvolve	PiEvo	CoEvolve	PiEvo
NHO	$g$ -factor ( $2$ )	$1.898$	$1.898$	$1.878$	$1.724$
MBO	pChEMBL ( $12$ )	$9.25$	$7.42$	$8.56$	$6.52$
AMP	probability ( $1$ )	$0.376$	$0.314$	$0.277$	$0.238$
Promoter	expression bin ( $17$ )	$14.69$	$9.25$	$13.96$	$13.50$
TMC	polarizability ( $500$ )	$441.5$	$423.5$	$465.0$	$457.5$

Appendix

Table 13: Final quality in native units. Best-of-seed SQ means of Table 1 multiplied by the reference scale of Table 12 .

Explore similar work

EvoSci: A Bio-Inspired Multi-Agent Framework for the Evolution of Scientific Discovery
May 20, 2026Xiaoyu Xiong, Yuqi Ren, Deyi XiongScientific Discovery Multi-Agent Large Language Model Systems
TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China
EvoScientist: Towards Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery
Mar 9, 2026Yougang Lyu, Xi Zhang, Yuyue Zhao +11Artificial Intelligence Scientists Agentic Discovery
Huawei Technologies Co., Ltd. · Vrije Universiteit Amsterdam
HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses
Sep 14, 2026Jieyuan Liu, Mengzhou Hu, Jefferson Chen +10Hypothesis-Driven Model Expansion Multi-Agent Evolution
University of California San Diego · 2Texas A&M University · 3Carnegie Mellon University +1

May 20, 2026, cs.AIJ/K move · Enter open · S save

EvoSci: A Bio-Inspired Multi-Agent Framework for the Evolution of Scientific Discovery

Xiaoyu Xiong, Yuqi Ren, Deyi Xiong

TJUNLP Lab, School of Computer Science and Technology, Tianjin University, China

Scientific Discovery Multi-Agent Large Language Model Systems Peer Review Evolutionary Optimization Methods Bio-Inspired Framework

Open paper AskPDF arXiv

Large language models (LLMs), have shown strong potential in scientific discovery, yet existing methods still face substantial challenges in the design of research workflows and multi-role collaboration mechanisms. To mitigate these issues, we propose EvoSci, a multi-agent scientific collaboration framework, which integrates bio-inspired evolution with knowledge graph modeling. To iteratively generate, evaluate, and refine research ideas, EvoSci incorporates multiple role-based agents, including mentor, researcher, and reviewer. By combining collaborative reasoning, shared memory, and evolutionary feedback, EvoSci significantly enhances the coherence and creativity of scientific exploration. Experiments on real-world research topics demonstrate that EvoSci significantly outperforms strong baselines in LLM-based structured peer-review and comparative ranking evaluations, achieving the highest overall peer-review score (ICLR 4.90) and top ranking (Top-10 = 54). These results suggest its superiority in both scientific idea generation and continuous discovery.

Abstract

Explore similar work

EvoSci: A Bio-Inspired Multi-Agent Framework for the Evolution of Scientific Discovery

EvoScientist: Towards Multi-Agent Evolving AI Scientists for End-to-End Scientific Discovery

HypoEvolve: Genetic Algorithms Enable Multi-Agent LLMs to Discover Scientific Hypotheses