Prism lets multimodal models teach themselves in both directions

A model asked for an illustrated cooking tutorial must do more than create plausible pictures and sentences. The steps must make sense on their own and progress together. In one example in the paper, a base model’s images show chopped tomatoes before a later step depicts cutting them again.

Researchers from Xi’an Jiaotong University and Alibaba Cloud’s Qwen DianJin Team propose Prism, a post-training method that turns such sequences into teaching material. Their arXiv preprint, Prism: Bidirectional Self-Teaching for Unified Multimodal Models, reports improvements on several image-generation and visual-understanding benchmarks for two models, SenseNova-U1 and BAGEL. The work is a preprint, not an established peer-reviewed result.

A sequence becomes its own training signal

Unified multimodal models handle visual understanding and image generation within one system. Some previous approaches use the understanding side to judge and improve image generation; others use external annotations, reward models or specialized tasks. Prism instead chooses a teaching direction for each generated example.

The model first produces an interleaved sequence: a text step, an image, another text step, another image, and so on. Prism separates that sequence into its complete text chain and image chain. Frozen copies of the starting model review each chain against the original prompt and identify errors. The judgments are estimates, not guarantees of correctness.

If both chains pass, the trainable model learns to reconstruct the text from the images. If only the text passes, a frozen model uses it and the reported image error to regenerate the image chain; if only the images pass, it rewrites the text instead. The corrected sequence is checked by another frozen copy before it becomes a positive training target. A failed correction can be retried, up to four attempts in total, and is discarded if it still fails.

When both chains fail, Prism does not train on them as correct examples. Instead, it reduces the likelihood of reproducing the original text tokens. This negative signal applies to the entire text chain, not just identified errors, and has no equivalent image-side penalty in the reported method. Only the rollout model is updated; the judge, rewriter and verifier remain frozen.

The researchers used approximately 2,400 task-prompt seeds generated with GPT-5.5, covering tutorials, stories, spatial reasoning and other tasks. The prompts were external inputs, even though Prism did not use an external verifier or externally annotated corrected trajectories to construct its main training targets. Training took roughly 24 hours on 16 A800 GPUs, including generation and checking.

What improved, and against what

Against the unmodified SenseNova-U1, Prism raised the reported overall GenEval score from 85.8 to 89.2 and DPG-Bench from 86.1 to 87.3. GenEval tests how well generated images match object-related prompts; its counting category rose from 75.6 to 87.2, although its color category fell from 91.7 to 89.6. For visual understanding, total MME rose from 2282.6 to 2311.6 and MMVP from 58.0 to 62.0.

BAGEL’s overall GenEval score rose from 76.5 to 78.3, and its DPG score from 84.0 to 85.0. Its MMBench score rose from 84.6 to 88.4. On OpenING, which evaluates interleaved text-and-image output through model-judged pairwise comparisons, Prism improved both models’ arena-relative win rates under all four reported tie-handling rules. BAGEL’s share of outputs containing no image fell from 30.7% to 14.8%.

The paper also compares post-training methods on BAGEL. UniCorn achieved higher GenEval and DPG scores than Prism, 80.3 versus 78.3 and 86.2 versus 85.0, respectively. But UniCorn’s total MME fell from the base model’s 2361 to 2337, while Prism reached 2371; SRUM also improved those generation scores while lowering MME. These comparisons support a trade-off in the reported setup, not a claim that Prism wins every metric.

A further SenseNova-U1 control trained directly on unmodified self-generated sequences. It reached 87.7 on GenEval but fell to 2276.3 on MME, below the base model. Full Prism exceeded that control on all seven reported metrics; it also exceeded a random-routing control on all seven. Replacing individual objectives showed mixed effects, including a higher GenEval score when the both-pass objective was replaced.

The judge is part of the limitation

In a separate comparison on 200 training rollouts, the frozen SenseNova-U1 judge agreed with human annotations on 79.8% of 400 text- and image-chain judgments. Among chains it accepted as reliable, 14.4% were flawed according to those annotations. Incorrect judgments can preserve bad teaching signals or penalize useful text; checking rewrites does not establish that accepted targets are error-free.

Results outside the main benchmarks were mixed. On TIIF-Bench, BAGEL’s short-prompt score fell from 0.699 to 0.664 and its long-prompt score from 0.699 to 0.681. SenseNova-U1’s reported editing scores also declined on both editing benchmarks tested. The authors identify the focus on interleaved text-image tasks, reliance on external prompt seeds and whole-chain negative supervision as limitations. The reported gains therefore describe these experiments, not a general guarantee that self-teaching improves every multimodal capability.