Skip to main content
Back to timeline
arXivSource publication:

Agentic image generation stops guessing who learns: RSIgen lets understanding and generation correct each other, gaining 3.2 to 3.3 Overall9 points on three backbones

Synopsis

In a unified multimodal backbone, the understanding-and-orchestration agent and the image generator previously learned separately, so a generator update could leave the agent's action supervision stale. RSIgen links who should learn, which actions remain useful, and what to learn next: matched interventions pick the learning recipient and compile compatible targets, fixed branch inputs re-rank actions after generator learning, and retained roles build the next curriculum. On BAGEL, Qwen-Image and Wan-Image-7B, AgentGen-Bench Full Overall9 is 3.2 to 3.3 points above a three-cycle SearchGen adaptation with action costs.

Source-provided article image: RSIgen: Reciprocal Self-Improvement for Agentic Visual Generation
arXiv

Interpretation

In a unified multimodal model, the understanding role and the generation role can supply training experience to each other, but task success alone does not say which one should learn. SearchGen already trains the generator and then recalibrates its reasoner, and GenEvolve distills contrasting trajectories into an evolving agent; neither separates learning-recipient allocation from whether action supervision is still valid. On BAGEL, Qwen-Image and Wan-Image-7B, RSIgen reaches AgentGen-Bench Full Overall9 of 39.9, 50.3 and 47.5, respectively 3.2, 3.2 and 3.3 points above the three-cycle SearchGen adaptation with action costs.

RSIgen uses Dynamic Responsibility Routing (DRR) to decide whether an experience goes to the agent or the generator, then compiles a training target compatible with that role. DRR runs matched interventions on evidence, inputs and supported actions: Oracle evidence repairing a retrieval failure points to agent acquisition, a corrected prompt repairing the output points to agent construction, and failure under sufficient valid input points to generator realization. In a Wan control with matched role-level training volume, DRR reaches Full 47.4, 1.4 points above compatible-uniform and 1.1 above failure-locus routing, while searches per request fall from 1.29 to 1.21 and generations/edits from 1.37 to 1.32.

After the generator learns, the agent's existing action supervision is re-tested, and stale labels are replaced by supervision re-ranked under the new generator. Old supervision counts as stale only when its action becomes invalid or materially inferior; if both generators improve actions equally, the pairwise difference is preserved, so a preferred set can change while an old label stays competitive. On 200 held-out requests, generator updates shift preferences in 22.2% of pre-generation and 20.8% of post-generation decisions, with stale labels at 17.0% and 15.7%; refreshed supervision raises three-stage Mini O9 from 47.37 to 50.20 and lowers pre/post regret from 5.37/4.60 to 3.30/3.00.

Retained roles build the next curriculum from updated evidence, inputs and outcomes, so recursive improvement moves together with target validity and supervision freshness. Controls that freeze curriculum sources keep recipient learning active and change only where curriculum content comes from; on Wan, dual freezing gives Full 42.8, freezing only the agent 44.8, and freezing only the generator 44.0. Across three attempts, current sources yield 5,580/7,480 valid agent/generator targets versus 4,430/6,740 under dual freezing, with 1,950 versus zero refreshed agent targets, while every arm trains on 5,400/6,600 entries.

Perspective

The demonstrated scope is AgentGen-Bench evaluation on the BAGEL, Qwen-Image and Wan-Image-7B unified multimodal backbones, with mechanism studies on Wan and human preferences from blind evaluation of 200 real-user requests; action comparisons use finite shared probe banks, and the intervals measure request variation conditional on trained checkpoints and frozen evidence.

Routing yields operational learning hypotheses rather than unique causal identification; source-freezing controls change curriculum content and target yield together, so their effects should be read jointly; the protected-capability tolerance applies locally to each accepted update, so permitted losses can accumulate across successive steps; human preferences are descriptive summaries of the evaluated real-user sample without request-cluster confidence intervals.