SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

Qianru Meng; Xi Zhang; Konstantinos Papakostas; Congfeng Cao; Wenxiang Hu; Tianyi Lu; Xiao Zhang; Lun Zhou

cs.AIOct 14, 2026

SwarmRSI: Recursive Self-Improvement through Evolving Scientific Agent Swarm

Authors: Qianru Meng, Xi Zhang, Konstantinos Papakostas, Congfeng Cao, Wenxiang Hu, Tianyi Lu, Xiao Zhang, Lun Zhou,

Organizations: Huawei Technologies Co., Ltd. · Leiden University · University of Groningen

Abstract

Recursive self-improvement (RSI) in scientific research aims to improve agents' research capabilities through their own experimental experience. Scientific agent swarms can combine evidence and implementations from different research directions to guide their agents' subsequent experiments. However, poor experimental results can leave unclear whether to revise a hypothesis or its implementation, while overall program scores can hide components useful to other agents. To address these challenges, we introduce SwarmRSI, a framework that evolves scientific agent swarms for recursive self-improvement. We devise Feedback-Guided Swarm Mutation to refine hypotheses and their implementations using experimental feedback from other agents. We also propose Complementary Swarm Recombination to reuse retained hypotheses and complementary program components across agents, allowing partial results to support new research. Recursive Swarm Evolution integrates both mechanisms into a feedback loop, using new experimental evidence to guide subsequent research. Experiments on OpenProblems, NatureBench and MLE-Bench Lite demonstrate SwarmRSI's state-of-the-art (SOTA) performance. Further evaluations show that accumulated research experience improves performance on held-out tasks. Biologists also reproduce its programs for follow-up biological analyses.

Figures & tables

Figure 1: Two limitations motivating SwarmRSI . In the top example, a poor result does not show whether the hypothesis or its implementation needs revision, and a promising hypothesis is abandoned. In the bottom example, a program without overall gains holds components useful to another agent’s later research. Dashed links mark open revisions and potential reuse.

Figure 2: Overview of SwarmRSI . Scientific agents use research states assembled from a shared graph. Feedback-Guided Swarm Mutation revises a hypothesis and implements the revision, or retains the hypothesis and repairs its program. Highlighted lines show local modifications. Complementary Swarm Recombination preserves a hypothesis while adapting another agent’s valid program base (rebase), or compares results across instances and adapts components from complementary programs into one new program (crossover). Blue and green code lines indicate crossover sources. The comparisons and code layouts are schematic. All candidates undergo execution and evaluation, and all attempts, whether successful, failed or without gain, update the graph and the next agent’s state $s_{t+1}$ .

Method	Backbone	Denoise	Label	Batch	SVG	Decomp	Visual	Avg $\uparrow$
Human Expert	Task-specific	28.0	85.0	71.0	66.0	49.0	44.0	57.2
Biomni	Claude 4 Sonnet	16.0	89.0	82.0	16.0	53.0	35.0	48.5
ChatGPT	GPT-5	3.0	81.0	83.0	60.0	74.0	38.0	56.5
AIDE †	DeepSeek-V4-Pro	27.4	70.6	78.0	72.8	46.1	39.8	55.8
TusoAI †	DeepSeek-V4-Pro	31.5	58.7	79.5	79.0	84.1	69.8	67.1
	DeepSeek-V4-Pro	33.6	74.2	82.5	83.9	86.0	84.8	74.2

Table 1: Main results on the six OpenProblems tasks. Scores are normalized and scaled by 100. Avg is the six-task mean, recomputed from reported task scores for published references.

\dagger

denotes our means over three seeds. Bold and underlined values indicate best and second best results, including ties at the reported precision.

Method	Backbone	Above Median $\uparrow$	Bronze	Silver	Gold $\uparrow$	Any-Medal $\uparrow$
MLEvolve	DeepSeek-V4-Pro †	81.8	9.1	27.3	36.4	72.7
MLEvolve	Gemini 3.1 Pro Preview	–	–	–	–	80.3
Arbor	DeepSeek-V4-Pro †	86.4	4.5	31.8	40.9	77.3
Arbor	GPT-5.5	95.5	0.0	9.1	77.3	86.4
ML-Master 2.0	DeepSeek-V3.2-Speciale	84.9	13.6	31.8	30.3	75.8
Famou-Agent 2.0	Gemini 3 Pro Preview	93.9	12.1	18.2	50.0	80.3

Table 3: Main results on 22 MLE-Bench Lite tasks (%).

\dagger

denotes our means over three seeds. Any-Medal combines exclusive medal rates before rounding, and – denotes unreported values. Bold and underlined values indicate best and second best Above Median, Gold, and Any-Medal rates, including ties.

Figure 3: SwarmRSI with two backbones. Full bars give Match-SOTA and Any-Medal rates (%), and solid portions give Surpass-SOTA and Gold rates.

Variant	Match	Surpass	Mean $g$
SwarmRSI	44.4	15.6	12.8
Feedback-Guided Swarm Mutation
w/o H-mutation	34.4	8.9	7.2
w/o P-mutation	31.1	6.7	5.4
Complementary Swarm Recombination
w/o recombination	37.8	7.8	7.6

Table 4: NatureBench ablations on all 90 tasks with DeepSeek-V4-Pro, averaged over three seeds and grouped by the mechanism each variant removes. Match and Surpass are Match-SOTA and Surpass-SOTA rates (%), and mean

g

is scaled by 100.

Figure 4: Search behaviour on NatureBench with DeepSeek-V4-Pro. (a) Failure recovery, the share of failed implementations whose later refinement yields a new best, over 90 tasks. (b) Nested shares of proposed hypotheses that are implemented (white), evaluated (light) and improved (hatched). (c, d) Best-so-far

g

of one recorded run per method on symmetric log axes.

Figure 5: Biological case studies. (a–c) UMAP of the reproduced SwarmRSI representation for 833 target enterocytes, coloured by source and by MHC-II/Cd74 and antimicrobial RNA scores (mean normalized log RNA). Green and pink denote helminth day-10 and Salmonella sources. (d–f) BALBc-1 spleen with manual annotations, SwarmRSI program partitions, and expert neighbourhood groups E1 (B/follicular-dendritic-cell-rich), E2 (myeloid/erythroid-rich) and E3 (T-cell-rich).

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Parameter	Value
Discount factor $\rho$	0.9
Exploration weight $\beta$	0.5
Request-pressure weight $\gamma$	0.25
Request-pressure saturation $K$	4
Eligible-wait threshold $W$	8
Tokens per cost unit $T$	1,000

Appendix

Table 5: Shared H/P-selection parameters and search limits across both backbones and all three benchmarks.

Figure 6: Performance against mean tokens per task for SwarmRSI and rerun baselines with DeepSeek-V4-Pro. OpenProblems reports normalized average scores scaled by 100, and NatureBench and MLE-Bench Lite report Match-SOTA and Any-Medal rates (%).

Variant	$\Delta g$	95% CI
w/o H-mutation	5.6	[1.2, 10.2]
w/o P-mutation	7.4	[3.0, 11.9]
w/o recombination	5.2	[2.5, 7.9]
w/o shared graph	4.7	[0.3, 9.3]
w/o H/P selection	8.2	[4.2, 12.4]

Appendix

Table 6: Drop in mean

g

(scaled by 100) from SwarmRSI for each ablation, with task-paired bootstrap 95% CIs.

Method	Seed	Denoise	Label	Batch	SVG	Decomp	Visual	Avg
SwarmRSI (DeepSeek)	1	34.1	74.6	83.1	84.7	85.6	85.2	74.55
	2	33.5	73.5	82.8	83.8	86.6	84.6	74.13
	3	33.3	74.4	81.6	83.1	85.8	84.7	73.82
SwarmRSI (Gemini)	1	40.2	81.5	79.0	86.6	82.7	85.5	75.92
	2	40.2	80.7	78.1	88.8	84.0	86.1	76.32
	3	40.3	80.0	79.4	88.1	83.6	87.5	76.48

Appendix

Table 7: Per-seed OpenProblems scores for all

\dagger

entries in Table 2 . Scores are normalized and scaled by 100. Avg is the six-task mean, shown to two decimal places. DeepSeek and Gemini denote DeepSeek-V4-Pro and Gemini 3.1 Pro Preview, respectively.

Method	Denoise	Label	Batch	SVG	Decomp	Visual	Avg
SwarmRSI (DeepSeek)	1.03	1.46	1.97	1.99	1.31	0.80	0.91
SwarmRSI (Gemini)	0.14	1.86	1.65	2.79	1.65	2.55	0.72
AIDE (DeepSeek)	2.62	1.27	1.38	3.80	1.60	1.23	0.63
TusoAI (DeepSeek)	4.74	2.63	1.62	2.73	1.00	0.29	1.03

Appendix

Table 8: 95% CI half-widths for the OpenProblems means over three seeds in Table 2 , computed using the Student’s

t

-distribution from Table 7 . Intervals are the unrounded means

\pm

these half-widths, on the score scale of Table 2 .

Task	Method	Scored/records	Best $g$
00917	SwarmRSI	18/18	0.1807
	Arbor	3/4	-0.2007
	AIDE	49/50	-0.4742
00639	SwarmRSI	19/21	0.0644
	Arbor	2/3	0.0434
	AIDE	8/28	0.0552

Appendix

Table 9: Selected runs with their scored evaluations, recorded events and best observed

g

Candidate	Recorded change	$g$
4	Retained hypothesis. Outputs unchanged from 3	$-0.08323$
8	Hypothesis mutation produces a new program base	$-0.06388$
9	Rebase: target predictor uses source training and checkpoint rules	$-0.06186$
11	Crossover: source 5 supplies a second instance predictor	$-0.05573$
15	Rebase: target predictor implemented without fallback	$-0.05408$

Appendix

Table 10: Recorded process on 00795. Candidate indices are zero-based. Each

g

comes from a valid evaluation of the resulting program.

Stage	Reuse hit	Other hit	Reuse share
Early	32.0	25.8	40.5
Middle	18.2	16.8	41.1
Late	18.1	15.0	42.6

Appendix

Table 11: New-best rates (%) of reuse and update triggered by feedbacks (rebase, crossover and mutations triggered by feedback) and of other updates across search stages on NatureBench. Reuse share is the percentage of new-best candidates from these updates.

Variant	Match	Surpass	Mean $g$	$\Delta g$ [95% CI]	Tokens
Cold start	41.7	33.3	12.01	–	12.17
Gen 1	58.3	41.7	16.15	4.14 [1.82, 6.49]	11.86
Gen 2	86.1	44.4	20.16	8.15 [5.67, 10.59]	11.52
Gen 3	86.1	58.3	21.89	9.88 [6.88, 13.06]	10.94

Appendix

Table 12: Search on 12 held-out NatureBench tasks with DeepSeek-V4-Pro, starting from an empty state (cold start) or from the state inherited after each generation, averaged over three seeds. Match and Surpass are Match-SOTA and Surpass-SOTA rates (%), mean

g

is scaled by 100, and

\Delta g

is the gain over cold start with a paired task-cluster bootstrap 95% CI from 10,000 resamples. Tokens are means per task in millions.

Source	Cells	MHC-II	Interferon
B1	59	0.354	0.115
B2	46	0.126	0.166
B3	98	0.375	0.136
B4	221	0.100	0.162
B5	75	0.377	0.391
B6	126	1.495	0.636

Appendix

Table 13: Measured enterocyte RNA program means by sequencing source. Scores average normalized log RNA across the four MHC-II or six interferon genes. B1–B4 are controls, B5/B6 helminth day 3, B7/B8 helminth day 10 and B9/B10 Salmonella.

Source	Program	Cells	RNA PCA	Reproduced output	Recorded output
B7	MHC-II	71	0.650	0.768	0.766
B8	MHC-II	57	0.396	0.685	0.686
B9	MHC-II	332	0.331	0.326	0.313
B10	MHC-II	373	0.356	0.351	0.347
B7	Antimicrobial	71	0.311	$-0.045$	$-0.056$
B8	Antimicrobial	57	0.689	0.593	0.612

Appendix

Table 14: Within-source local coherence after covariate adjustment. Entries are Spearman correlations between a cell’s residual RNA score and its ten neighbours’ mean residual score. Reproduced/recorded outputs are the expert reproduction and the representation saved from the original SwarmRSI run. Counts are cells per source.

Figure 7: Within-source expression coherence in enterocytes. Within each source, Spearman correlation compares cells’ measured RNA program scores with their ten nearest neighbours’ mean scores, excluding each focal cell. Neighbours are defined in RNA PCA (blue) or the reproduced SwarmRSI representation (pink), and the correlations are not adjusted for covariates.

Figure 8: Source means of individual measured genes in enterocytes. Hpoly.Day3/Day10 denote helminth day 3/day 10 sources. The antimicrobial set is primarily Reg3g/Reg3b. Zero Lcn2 denotes non-detection in this matrix. Values are mean normalized log RNA.

AI Agents for Scientific Discovery

AI agents for scientific discovery use language models, tools, and experiments to generate, test, and refine scientific knowledge.

See all 288 papers

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement
Sep 14, 2026Siwei Wu, Jincheng Ren, Yizhi Li +12Coding Agents Agent Harness Optimization
University of Manchester · IQuest Research · M-A-P +3
AReaL-RSI: Adaptive Scaling and Coordination of Recursive Self-Improvement
Oct 13, 2026Wenhao Zhou, Jiajun Song, Haitao Wang +11Self-Improving Agents Recursive Self-Improvement
Ant Group · Zhejiang University · HKUST
ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents
Sep 15, 2026Shuhan Xue, Jianyuan Zhong, Ziyuan Nan +10Continual Learning for LLM Agents AI-Assisted Scientific Research

Sep 14, 2026J/K move · Enter open · S save

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement

Siwei Wu, Jincheng Ren, Yizhi Li, Haau-Sing Li+11

University of Manchester · IQuest Research · M-A-P+3

Coding Agents Agent Harness Optimization LLM Agent Self-Improvement

Open paper AskPDF arXiv

Recent work extends recursive self-improvement (RSI) to agent harnesses for long-horizon coding and terminal tasks, enabling agents to improve execution mechanisms from experience. However, generalizable harness RSI remains challenging. First, evolving harnesses on evaluation benchmarks or their subsets makes it difficult to distinguish reusable improvements from benchmark-specific adaptation. Second, single-trajectory updates can conflate systematic harness deficiencies with instance-specific reasoning and solution details, producing modifications that transfer poorly to unseen tasks. Third, localizing recurring behavioral deficiencies within monolithic harnesses is difficult, while whole-harness optimization can entangle unrelated mechanisms and complicate attribution and validation. We propose ModularRSI, a benchmark-disjoint, contrastive, and modular framework for generalizable harness evolution. ModularRSI contrasts successful and failed trajectories for the same task and aggregates evidence across tasks to identify recurring behavioral deficiencies. It decomposes the evolvable harness into five functional modules: Agent Loop, Tool Use, Observation Management, Context Management, and Task Completion Detection. Each module evolves independently within a restricted modification scope, followed by an integration stage that combines the evolved modules into a unified harness and resolves potential conflicts. To support benchmark-disjoint evolution, we curate 2,000 executable evolution tasks from external sources that are disjoint from downstream evaluation benchmarks. Experiments on TB2.0 and SWE-Bench Verified show consistent improvements on unseen in-domain and cross-domain tasks, with the evolved harness also transferring across different foundation models.

Abstract

AI Agents for Scientific Discovery

Explore similar work

ModularRSI: Modular and Generalizable Recursive Harness Self-Improvement

AReaL-RSI: Adaptive Scaling and Coordination of Recursive Self-Improvement

ScienceBuddy: Recursive-in-Recursive Self-Improvement for Interactive Scientific Agents