RSIgen: Reciprocal Self-Improvement for Agentic Visual Generation

Yidi Liu; Haozhe Wang; Jie Xiao; Kai Zhu; Xueyang Fu; Zheng-Jun Zha

cs.CVOct 13, 2026

RSIgen: Reciprocal Self-Improvement for Agentic Visual Generation

Authors: Yidi Liu, Haozhe Wang, Jie Xiao, Kai Zhu, Xueyang Fu, Zheng-Jun Zha

Organizations: University of Science and Technology of China · The Hong Kong University of Science and Technology · Wan Team, Alibaba Group

Abstract

Agentic visual generation produces experience that can improve the native understanding (agent) and generation roles of a unified multimodal backbone, but task success alone does not determine which role should learn. Generator updates can also make the agent's action supervision stale. We introduce RSIgen, a task-centric reciprocal self-improvement framework linking learning allocation, supervision refresh, and recursive curriculum construction. Dynamic Responsibility Routing (DRR) uses matched interventions on evidence, inputs, and supported actions to select learning recipients and construct compatible training targets. After generator learning, RSIgen re-evaluates the relative utility of tested actions with branch inputs held fixed to refresh agent supervision; the retained roles then construct the next curriculum. Across BAGEL, Qwen-Image, and Wan-Image-7B, RSIgen improves AgentGen-Bench Full Overall9 by 3.2--3.3 points over a three-cycle SearchGen adaptation with action costs. Wan controls support allocation gains at matched role-level training exposure, refresh gains at a matched adaptation budget, and benefits from updated curriculum sources while recipients continue learning. In blind evaluation on 200 real-user requests, pooled tie-adjusted preference scores against the same baseline reach 70.0% for task completion and 64.9% for visual quality.

Figures & tables

Figure 1: Reciprocal learning in agentic visual generation. (a) Agent orchestration and generator realization are roles within a unified multimodal model (UMM); PE denotes prompt enhancement. (b) Generator learning changes action utility. (c) Agent learning changes inputs and evidence for later curricula. Images and quality levels are illustrative.

Figure 2: RSIgen’s learning procedure. DRR compiles matched probes into typed curricula. Generator learning precedes re-probing and agent learning; retained roles construct the next curriculum. Pool, admission, and sealed evaluation data have separate roles. UMM denotes the unified model’s understanding/orchestration and generation roles. Evaluator feedback and capability memory support training.

Figure 3: Schematic routing decisions. (a) Holding evidence and generator fixed, a corrected layout repairs a chronology error and yields an agent-construction candidate. (b) A module mismatch under checked input yields a generator-realization candidate when a compatible target is available. (c) Learning, externalization, and deferral depend on the dependency’s properties and branch outcomes.

Probe pattern	Recipient/subtype	Compiled training object
Oracle repairs inadequate retrieval	Agent: acquisition	Valid query and evidence-selection trajectory
Corrected input repairs the output	Agent: construction	Executed corrected prompt or condition
Search repairs a stable-knowledge failure	Generator: internalization	Evidence-free prompt and validated assisted image
Failure persists with sufficient valid input	Generator: realization	Compatible rendering, reference, or edit pair
Previously mastered capability regresses	Original recipient/replay	Validated prior target
Volatile dependency or unresolved diagnosis	Externalize or no-update	Retained access or further probes

Table 1: From probe outcomes to typed curricula. Each route pairs validated inputs with a recipient-compatible target. Oracle supports diagnosis; agent targets use executed trajectories.

Commercial Models
	Checklist		Visual Ref.		Image Quality		Overall9
Model	NS	SI	NS	SI	NS	SI	NS	SI	Full
GPT-Image-2	84.8	74.4	70.9	70.4	73.4	77.4	78.9	75.5	76.0
Grok-Imagine-2.0-Low	83.0	73.1	68.7	68.3	72.5	76.9	76.9	74.9	75.2
Qwen-Image-3-Max	79.1	65.8	64.0	59.6	70.3	71.1	73.7	67.6	68.4
Nano Banana Pro	78.4	60.7	63.0	55.4	70.7	68.1	73.3	63.6	65.0

Table 2: AgentGen-Bench overview (0–100). NS/SI: NoSearch/Search-Intensive; Full: Full-slice Overall9. Colors mark best and second-best within each group. Public baseline scores follow the benchmark release; matched training comparisons appear in Table 3 .

	BAGEL			Qwen-Image			Wan-Image-7B
Method	NS	SI	Full	NS	SI	Full	NS	SI	Full
Direct	–	–	–	–	–	–	62.3	27.4	32.1
Prompt enhancement	–	–	–	–	–	–	64.1	31.2	35.6
Agentic initialization	54.2	25.4	29.3	70.6	35.3	40.0	64.3	33.4	37.5
A-only / Gen-Searcher adapt.	56.3	29.4	33.0	71.2	39.4	43.6	65.2	36.3	40.2
SearchGen-3 + costs	58.2	33.4	36.7	72.3	43.2	47.1	66.2	40.8	44.2

Table 3: Matched within-backbone comparisons on AgentGen-Bench. Trained methods share initialization and resource allowances within each backbone; scores aggregate multiple training seeds. NS/SI denote NoSearch/Search-Intensive, and Full is the global Overall9 mean. Direct and prompt enhancement are untrained Wan references; dashes indicate results not reported in this comparison.

A. NoSearch (NS)
Model	Ch	Ru	Pr	IQ	Tx	Nat	Co	Ph	VR
BAGEL	51.5	49.2	37.7	56.2	16.1	48.8	62.8	58.5	32.1
BAGEL + RSIgen	63.4	62.1	58.3	61.4	50.6	60.2	68.4	73.2	45.6
Qwen-Image	74.9	72.0	65.5	69.2	61.1	61.7	77.0	78.6	60.1
Qwen-Image + RSIgen	77.2	76.1	72.4	70.7	68.4	69.2	80.1	81.3	63.4
Wan-Image-7B	65.3	63.2	57.4	65.1	54.3	57.2	73.4	75.1	49.3

Table 4: Component scores for the three backbones and their RSIgen variants. Ch: Checklist; Ru: Rubric; Pr: Prompt Faithfulness; IQ: Image Quality; Tx: Text Rendering; Nat: AI Naturalness; Co: Composition; Ph: Physical Plausibility; VR: Visual Reference. Overall9 appears in Table 2 .

Backbone	W / T / L	Pref.	W / T / L	Pref.
	Task completion		Visual quality
BAGEL	312 / 198 / 90	68.5%	264 / 234 / 102	63.5%
Qwen-Image	324 / 194 / 82	70.2%	276 / 228 / 96	65.0%
Wan-Image-7B	336 / 185 / 79	71.4%	288 / 217 / 95	66.1%
Overall	972 / 577 / 251	70.0%	828 / 679 / 293	64.9%

Table 5: Blind human evaluation against SearchGen-3 + costs on the same 200 real-user requests. Each backbone has 600 judgments per dimension (three ratings per pair); Overall pools 1,800 judgments on those same requests. W/T/L is from RSIgen’s perspective; preference is

100(W+0.5T)/(W+T+L)

Variant	Full $\uparrow$	S	G/E
Full RSIgen	47.5	1.16	1.30
A-only updates	40.2	1.59	1.45
G-only updates	41.7	1.76	1.51
Uniform reciprocal	43.8	1.41	1.42
Failure-locus-only routing	44.1	1.51	1.44
No post-G re-probe	45.0	1.53	1.39

Table 6: Wan ablations on AgentGen-Bench. Full measures quality; S and G/E count inference searches and generations/edits per request.

Strategy	Full $\uparrow$	$\Delta$ Full	Searches $\downarrow$	Gen./edits $\downarrow$
Compatible-uniform	46.0	0.0	1.29	1.37
Failure-locus	46.3	+0.3	1.27	1.36
DRR	47.4	+1.4	1.21	1.32

Table 7: Wan allocation comparison with matched role-level training exposure.

\Delta

is relative to compatible-uniform.

Backbone	Method	Full	Searches	Generations/edits
BAGEL	SearchGen-3 + costs	36.7	1.21	1.32
BAGEL	RSIgen	39.9	1.18	1.31
Qwen-Image	SearchGen-3 + costs	47.1	1.21	1.31
Qwen-Image	RSIgen	50.3	1.13	1.28
Wan-Image-7B	SearchGen-3 + costs	44.2	1.26	1.33
Wan-Image-7B	RSIgen	47.5	1.16	1.30

Table 8: Inference-call averages for SearchGen-3 + costs and RSIgen. Values are mean search and generation/edit calls per request.

A. Action changes after generator updates
Decision	Evaluated per stage	Preference shift (%)	Stale labels (%)
Pre-generation	200	22.2 [18.1, 26.5]	17.0 [13.4, 20.9]
Post-generation	200	20.8 [16.9, 25.1]	15.7 [12.3, 19.3]

Table 9: Wan action refresh on Test-mini. A: branch changes over 200 requests and 200 canonical post-states per stage, with 95% intervals in brackets. B: matched continuation with the updated generator; the final column gives refresh-minus-stale 95% intervals. All estimates average three stages. Intervals use 10,000 paired request-cluster bootstrap replicates.

Phase	Transition	Count	Share (%)
Pre	Search $\to$ Direct	69	51.9
Pre	Direct $\to$ Search	15	11.3
Pre	SearchModeSwitch	24	18.0
Pre	Search $\to$ PromptEnhance	13	9.8
Pre	PromptEnhance $\to$ Search	12	9.0
Post	Regenerate $\to$ LocalEdit	49	39.2

Table 10: Grouped action transitions pooled across stages. Shares divide by 133 pre-generation or 125 post-generation shifts. SearchModeSwitch includes switches involving HybridSearch; Repair groups admissible repair actions.

Table 14

Source	Valid	New valid	Refreshed	Valid	New valid	Refreshed
	Agent curricula			Generator curricula
Current	5,580	2,660	1,950	7,480	4,140	970
Freeze-A	5,280	2,560	1,660	7,060	3,820	790
Freeze-G	4,690	2,270	400	6,960	3,680	680
Freeze-both	4,430	2,120	0	6,740	3,540	570

Table 13: Wan curriculum construction and supervision changes across the four source controls. Counts sum target entries across three attempts. Valid and new-valid counts describe construction; refreshed counts describe updated targets actually used for training.

Figure 4: Wan role performance and curriculum validity across retained states. Evaluation uses 200 requests and four image seeds per request/state; agent evaluation fixes

G_{0}

. Role scores and curriculum units are defined in Section 5.6 .

Figure 5: Current-source training curricula across three Wan update attempts. Each bar counts selected items by recipient subtype, with totals of 1,800 agent and 2,200 generator items per attempt. Counts are selected training records, including replay.

Figure 6: Five-request comparison with complete prompts at left and BAGEL, Qwen-Image, Wan-Image, SearchGen, and RSIgen outputs in columns. Table 3 reports aggregate comparisons.

Figure 7: Additional RSIgen visual examples, including object depiction, sequential character scenes, product presentation, and stylized character designs.

Self-Improving Agents

Self-improving agents use experience and feedback to update their models, memories, skills, or workflows for better future task performance.

See all 169 papers

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses
Sep 21, 2026Peng Xia, Rujun Han, Zifeng Wang +11OOD Generalization Agent Harness Optimization
Google Cloud AI Research · UNC-Chapel Hill · Stanford University +1
Recursive Self-Improvement through Multi-Agent Self-Supervision
Oct 8, 2026Hyunin Lee, Jinglue Xu, Jeffrey Seely +4Language Model Self-Improvement LLM Agent Workflow Optimization
UC Berkeley · SakanaAI · Sakana AI
VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model
Mar 9, 2026Jinxiang Lai, Wenzhe Zhao, Zexin Lu +7Image Generation Agentic Image Generation
Hong Kong University of Science and Technology · Tencent Hunyuan

Sep 21, 2026J/K move · Enter open · S save

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Peng Xia, Rujun Han, Zifeng Wang, Yanfei Chen+10

Google Cloud AI Research · UNC-Chapel Hill · Stanford University+1

OOD Generalization Agent Harness Optimization LLM Agent Self-Improvement Recursive Self-Improvement

Open paper AskPDF arXiv

An LLM agent's capability is largely magnified by its harness, namely the prompts, control flow, tooling, memory, and context management surrounding the frozen backbone model. Recent methods increasingly automate this process by iteratively proposing and selecting component-wise edits of an agent harness, practically establishing a form of recursive self-improvement (RSI) at the agent-system level. However, such recursive evolution may overfit by memorizing the training tasks, showing large in-distribution gains that shrink or even vanish on out-of-distribution benchmarks. We introduce Regularized Recursive Self-Improvement of Agent Harnesses (RRSI), which incorporates the principles of regularizations into harness self-improvement by constraining the evolution candidate proposal and selection. The proposer operates with a temporally annealed budget, limiting how many edits a candidate can bundle, and it encourages unexplored trajectories based on evolution history. The selector is equipped with a critic and a pruner: the critic screens benchmark-specific proposals, while the pruner, removes changes that are too small, too expensive, or no longer useful. Together these constraints favor reusable agent mechanisms over benchmark-specific ones or even noises. Across eight benchmarks spanning coding, agentic workspace and engineering design tasks, RRSI gains up to 6.0 points on the split it evolves against and up to 4.3 points on the five out-of-distribution benchmarks, while producing a harness that runs on 30% fewer policy tokens than the unregularized evolution. Code is available at https://github.com/google-research/rrsi and project page is https://regularized-rsi.com/.

Abstract

Self-Improving Agents

Explore similar work

RRSI: Regularized Recursive Self-Improvement of Agent Harnesses

Recursive Self-Improvement through Multi-Agent Self-Supervision

VisionCreator-R1: A Reflection-Enhanced Native Visual-Generation Agentic Model