In plain words
An image request can involve finding information, shaping instructions, generating an image, and fixing it. A wrong result alone does not reveal whether the decision-making role or image-making role needs to learn.
The challenge continues after the image maker improves, because old advice about which action to take may stop working.
The paper introduces RSIgen, a framework that sends checked examples to a suitable role and retests action choices after generator updates. On AgentGen-Bench, a test scoring request fulfillment and image quality, it scored above a repeated-adaptation comparison on all three tested models.
The research
An image request may require finding facts, shaping a prompt, generating an image, or repairing it. If the result fails, how can a learning system tell whether its decision-making agent or image generator needs training, and keep the agent’s choices useful after the generator changes?
RSIgen, a reciprocal-learning framework, routes validated experience to a suitable role and rechecks action choices after generator updates to keep agent guidance current.
AgentGen-Bench tests request fulfillment and image quality; its Full score is a 0–100 global mean across nine evaluator components for each model’s scored coverage. RSIgen scored 3.2–3.3 points above SearchGen-3 + costs, a three-cycle SearchGen adaptation with action costs, on BAGEL, Qwen-Image, and Wan-Image-7B.
The trained arms shared starting checkpoints, tools, resources, and three attempts but used different learning procedures.
Gains were larger on Search-Intensive than NoSearch requests, consistent with the framework’s focus on refreshing action guidance as the generator changes.
The mechanism studies were conducted on Wan, and action comparisons cover only a finite set of tested branches. Routing diagnoses are operational hypotheses; the curriculum-source controls bundle changes in curriculum content with target validity and yield.
What the evidence shows
- Two roles, different learning needs
RSIgen (Reciprocal Self-Improvement for Agentic Visual Generation) is a framework for learning between the understanding agent and image generator inside one multimodal model. The agent may acquire evidence, construct a prompt, choose generation or repair actions, while the generator makes an image from those inputs.
Task success alone does not identify which role should learn or provide a compatible training target.
- Route examples to compatible learners
Dynamic Responsibility Routing (DRR) uses matched interventions to choose a learning recipient, a skill subtype, and whether to learn, replay, keep an external dependency, or make no update. It also compiles a training example suited to that recipient.
For example, a corrected prompt that fixes an image can train the agent, while a validated assisted image paired with a request that leaves out retrieved facts can train the generator.
- Recheck action advice
After generator updates, RSIgen re-tests action choices under matched conditions. A change in the preferred action set is distinct from an old label becoming invalid or materially worse; some old choices may remain near-optimal.
On Wan’s Test-mini requests, preferred sets changed in 22.2% of pre-generation decisions and 20.8% of post-generation states, while stale labels occurred in 17.0% and 15.7%. Each rate is a mean across three update stages, using its own tested decisions or states as the denominator.
- Refresh improves measured outcomes
In a Wan continuation study, the complete-workflow Mini O9 score rose from 47.37 with stale action labels to 50.20 with refreshed labels. A separate Stage-2 comparison, with 100 total cost units for preparation, rollout, and optimization, scored refresh at 50.0, current-generator on-policy learning at 49.1, and stale supervision at 47.2.
The reported passages do not identify the Mini O9 evaluator; these are separate comparisons and should not be combined.
- System scores across three models
AgentGen-Bench tests request fulfillment and image quality; its Full score is a 0–100 global mean across nine evaluator components for each model’s scored coverage. RSIgen scored 3.2–3.3 points above SearchGen-3 + costs, a three-cycle adaptation with action costs, on BAGEL, Qwen-Image, and Wan-Image-7B.
Both trained arms shared starting checkpoints, tools, resources, and three update attempts but used different learning procedures. Gains were larger on the test’s Search-Intensive requests than its NoSearch requests.
- Human ratings add another measure
Three independent blind raters compared outputs for 200 distinct real-user requests reused across the three backbones. They judged task completion and visual quality; researchers pooled 1,800 judgments per dimension and counted ties as half a win.
The resulting tie-adjusted RSIgen preferences were 70.0% for task completion and 64.9% for visual quality against SearchGen-3 + costs. These descriptive results come from the evaluated requests, not a larger sample created by reuse across backbones.
Deep
- Original paper authors
- Yidi Liu, Haozhe Wang, Jie Xiao, Kai Zhu, Xueyang Fu
- Submitted to arXiv
- License
- arXiv non-exclusive