acceptodds
Under review as a conference paper at ICLR 2027

Prism: Bidirectional Self-Teaching for Unified Multimodal Models

Abstract

Unified Multimodal Models (UMMs) aim to support visual understanding and generation within a single model. Existing post-training methods either use understanding as a fixed teacher to improve generation, or rely on ground-truth trajectories to supervise both understanding and generation. Prism instead derives bidirectional supervision from self-generated interleaved rollouts, constructing training targets according to the reliability of their text and image chains. We propose Prism, a bidirectional self-teaching post-training method for UMMs. Like a prism decomposing mixed light, Prism decouples a coupled interleaved trajectory into a text chain and an image chain, assesses the reliability of each chain, and constructs a corresponding training target from their reliability states. Experiments on both SenseNova-U1 and BAGEL show that Prism improves both understanding and generation. On the generation side, GenEval improves from 85.8 to 89.2 (+3.4), and DPG improves from 86.1 to 87.3 (+1.2) on SenseNova-U1. On the understanding side, the total MME score improves from 2282.6 to 2311.6, and MMVP improves from 58.0 to 62.0. Consistent improvements are observed on BAGEL. These results highlight that Prism enables coordinated improvements in generation and understanding by constructing reliability-guided self-training targets from self-generated trajectories, without relying on an external verifier or externally annotated corrected trajectories during target construction.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Unified Multimodal Model, Post Train, Self-Teaching

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_bwn1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper presents Prism, a post-training framework for unified multimodal models that utilizes self-generated, interleaved text-image trajectories. The core mechanism relies on a frozen copy of the base model, acting as an internal judge to evaluate generated text and image chains independently. Based on their identified reliability, Prism dynamically routes target construction: either reconstructs text from accepted images, rewrites the failing modality using the complementary chain, or applies a text-unlikelihood loss when both modalities are deemed unreliable. All rewritten targets are verified prior to fine-tuning. While external assistance is used to construct ~2.4K initial task-prompt seeds, the core judging and rewriting loop is entirely self-contained within the base model. The authors evaluate Prism on SenseNova-U1 and BAGEL under a brief post-training regime across three primary capabilities: image generation, visual understanding, and interleaved synthesis. The empirical study is supported by controlled ablations (e.g., standard self-training, removal of rewriting, randomized routing, and modified loss components) alongside diagnostic analyses on judge reliability, rewrite acceptance rates, and broader capability trade-offs.

Critical strengths

1.The experimental controls are well-designed and convincingly isolate the core contribution. Rather than just showing baseline gains, the comparisons against random routing and standard self-training prove that improvements stem directly from the target-construction strategy rather than compute or action distribution shifts.

2.The empirical evaluation is solid and backed by unusually insightful diagnostic analyses. Testing across backbones like SenseNova-U1 and BAGEL prevents single-architecture bias, while human audits and detailed rejection rates confirm that verification acts as a genuine quality filter.

Critical weaknesses

1.The verification mechanism evaluates modalities independently, which fails to guarantee joint cross-modal consistency. Because text and image checks do not cross-reference the anchor modality, individually accepted chains could still diverge in intermediate details, leaving a core premise of bidirectional teaching unverified.

2.The performance gains appear selective, and small improvements lack statistical uncertainty estimates. Several benchmarks show noticeable regressions (e.g., TIIF and GEdit), while gains on smaller datasets like MMVP lack paired confidence intervals or repeat runs to rule out random variance.

Questions for authors

1.On a sample of accepted rewrites, what fraction is jointly consistent between modalities, and do actual verification prompts access the anchor modality during implementation?

2.Can you provide paired uncertainty estimates or multi-run results to confirm the statistical significance of Prism over random routing?

3.How do you bound your claims regarding general capabilities given the regressions in Appendix B, and do loss ablations shed light on these trade-offs?

5 comments

W1/Q1: Independent verification and joint cross-modal consistency

A. Thank you for raising this issue. In our implementation, the verification prompts do not access the anchor modality. We adopted this design because, during development, the base model's joint judgments tended to overreport inconsistencies and reject otherwise usable targets. We acknowledge that this choice does not guarantee joint cross-modal consistency.

To address your concern, we conduct a joint audit of 100 t2i and 100 i2t targets that were accepted for training. Three trained reviewers manually examine all 200 sampled trajectories for detailed text-image consistency. In the table below, Base denotes the joint judgment of the frozen base model.

DirectionSamplesBase: inconsistentHuman: inconsistentBase inconsistency rate
t2i100823382%
i2t100711571%

The base model reports substantially more inconsistencies than the human reviewers, suggesting that directly using its joint judgments as a filter may reject some usable targets. Nevertheless, the limitation you identified remains: the human reviewers confirm detailed inconsistencies in 33% of t2i and 15% of i2t samples. These may be related to the base model's limited ability. Despite these residual discrepancies, the downstream experiments and ablations in the paper indicate that the constructed targets can provide effective supervision overall.

Q3: Bounding general capability claims and interpreting loss trade-offs

A. Thank you for encouraging us to clarify the scope of our claims. Our results support joint improvements in understanding, generation, and interleaved generation, while the regressions in Appendix B highlight the limits of these gains. Table 5 further shows that different objectives make different contributions to understanding and generation.Replacing the image-to-text rewriting objective reduces MMVP from 62.0 to 56.7. Replacing the text-to-image rewriting objective reduces GenEval from 89.2 to 88.0. Replacing the both-pass reconstruction objective slightly raises GenEval to 89.4 but lowers four of the five understanding metrics. This supports the different capability emphases of the objectives, rather than suggesting that each objective improves every metric. These results highlight the complementary roles and trade-offs of the objectives. We will clarify these findings and the scope of our capability claims more explicitly in the main text.

W2/Q2: Selective gains and paired uncertainty estimates

A. Thank you for the suggestion. Given the limited time and computational resources available during rebuttal, we are unable to conduct repeated training runs. Instead, we provide paired bootstrap 95% confidence intervals for Prism relative to Random Routing on seven benchmarks to address your concern about the stability of smaller improvements.

We perform 20,000 paired, grouped bootstrap resamples with replacement. Each resample uses the same test units for both methods and recomputes the score difference under the original evaluation protocol. For GenEval, we stratify by its six categories and resample prompts, retaining the four images for each prompt. For DPG, we resample prompts and retain their four-image scores. For MME, we stratify by subtask and resample images, keeping the two questions for each image together. For MMBench, we resample original questions, and report row accuracy. The resampling units are multiple-choice questions for MMMU, individual questions for MMStar, and complete question pairs for MMVP.

BenchmarkPrismRandom RoutingDifferencePaired bootstrap 95% CI
GenEval89.387.61.7[0.44, 2.96]
DPG87.286.40.8[0.60, 1.00]
MME2307.52296.311.2[1.18, 22.22]
MMBench87.286.50.7[0.18, 1.20]
MMMU60.759.41.3[0.24, 2.36]
MMStar66.565.70.8[0.07, 1.53]
MMVP62.059.32.7[0.67, 5.33]

Prism outperforms Random Routing with statistical significance on each of the seven benchmarks under the current evaluation setting, as all individual paired 95% confidence intervals for the score differences exclude zero.

Regarding the regressions on TIIF and GEdit, we acknowledge that Prism involves capability trade-offs, although the absolute decreases are relatively small: 0.011–0.035 on TIIF on a 0–1 scale and 0.06–0.27 on GEdit on a 0–10 scale. As discussed in Appendix B, Prism emphasizes planning and visual state progression conditioned on preceding text and images. In contrast, long prompts and style control in TIIF emphasize satisfying all constraints in a single generation, while GEdit requires preserving input-image pixel details and making local changes. Text rendering also requires character-level accuracy and often benefits from specialized post-training, such as reinforcement learning, to improve stability and correctness. These requirements are not directly covered by the current training objectives, which may limit transfer to these capabilities.

Thanks for the authors’ detailed responses. I have increased the soundness score from 2 to 3 and maintain my initial rating.

Reviewer_ZPsROpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

Prism proposes a bidirectional self-teaching method for unified multimodal models. It separates self-generated interleaved trajectories into text and image chains, assesses their reliability using frozen copies of the base model, and constructs training targets through reconstruction, rewriting, or negative supervision. Experiments on SenseNova-U1 and BAGEL demonstrate improvements in understanding, generation, and interleaved outputs.

Critical strengths
  1. Meaningful sample-dependent supervision. Rather than assuming that understanding is always stronger than generation, Prism allows either modality to supervise the other according to the reliability of each trajectory.

  2. Supportive experimental analysis. Improvements across two backbones, together with self-training, random-routing, rewriting, and loss ablations, provide evidence for the proposed target-construction mechanism.

Critical weaknesses
  1. Shared errors may contaminate self-supervision. Generation, judgment, and correction use the same base model and may share systematic blind spots. Appendix C.4 reports errors in 13.7% of accepted text chains and 15.2% of accepted image chains, but it remains unclear whether the audited samples capture harder cases, where correlated errors could be substantially more severe. An independent audit of the final training targets, stratified by task complexity, would strengthen confidence in their reliability.

  2. The methodological advance over related self-training methods needs clearer isolation. Mutual improvement, self-judgment, and self-generated supervision have been explored in UniRL, UniCorn, and HermesFlow. Prism's main distinction is its reliability-guided selection of the teaching direction. Fixed-direction rewriting baselines and clearer data and compute controls would strengthen the evidence for this specific contribution.

  3. Improvements involve capability trade-offs. While the main benchmarks improve, Appendix B reports regressions in some text-rendering, instruction-following, and editing results. Prism also trails UniCorn on the generation metrics in Table 4. The paper should more clearly characterize which capabilities benefit and verify the stability of smaller improvements through repeated runs.

Questions for authors

Why use a frozen copy of the base model for judgment and verification rather than a stronger, independent VLM? Beyond the Qwen3.7 Plus comparison in Appendix C.4, have you considered stronger closed-source VLMs to reduce shared errors, particularly on complex tasks? What are the quality and cost trade-offs?

4 comments

W1: Shared errors and the reliability of complex training targets

A. We agree with your concern. Our additional audit directly examines the final targets actually used for training, rather than the initial trajectories before rewriting. We sample 900 final positive training targets consumed by Full Prism, with 300 each at simple, moderate, and complex task levels. Complexity is assigned before the quality audit, based only on the original prompts: simple tasks involve a single scene and static constraints, moderate tasks involve short sequences or multiple objects and attribute bindings, with the main required states explicitly specified, complex tasks involve multi-step dependencies, inference of intermediate states, and tracking constraints across steps. An independent GPT Astra judge uses the original modality-specific review prompts to examine the final text and image chains separately.

ComplexitySamplesText judged flawedImage judged flawedBoth chains pass
Simple30016 (5.3%)21 (7.0%)271 (90.3%)
Moderate30035 (11.7%)46 (15.3%)236 (78.7%)
Complex30059 (19.7%)94 (31.3%)185 (61.7%)
All900110 (12.2%)161 (17.9%)692 (76.9%)

As task complexity increases, the proportion of targets judged flawed by the independent judge rises, especially for image chains. This supports your concern about residual errors on harder tasks and shows that verification does not eliminate all errors. We will add these findings to the revised manuscript. Please see the response to Question 1 for an evaluation of Astra's judgment accuracy.

W2: Isolating the contribution of reliability-guided selection of the teaching direction

A. As stated in the paper, Prism's distinction is its sample-dependent, reliability-guided selection of the teaching direction, rather than introducing self-training or mutual improvement. To test this design more directly, we add fixed-direction rewriting baselines. Fixed-i2t allows only a reliable image chain to guide text rewriting, while Fixed-t2i allows only a reliable text chain to guide image rewriting. Samples requiring the other direction are skipped. The both-pass and both-fail cases are handled as in Full Prism. This avoids forcing a modality known to be flawed to serve as the teacher.

Given the limited time and computational resources available during rebuttal, and because the different directional settings execute or skip different online operations, it is difficult to strictly match both the number of training samples and the overall computing budget. We therefore use the same GPU time as a common budget: all settings use 16 A800 GPUs for 24 hours, including online target construction and parameter updates.

MethodGenEvalDPGMMEMMBenchMMMUMMStarMMVP
Base85.886.12282.686.159.065.158.0
Fixed-i2t86.186.42306.587.560.466.361.3
Fixed-t2i89.087.22291.886.359.565.760.7
Full Prism89.387.22307.587.260.766.562.0

Fixed-i2t mainly benefits understanding, whereas Fixed-t2i mainly benefits generation. Full Prism achieves higher scores than Fixed-i2t on six metrics, with a slightly lower MMBench score; it exceeds Fixed-t2i on six metrics and ties on DPG. These results support the combined benefits of allowing both reliable teaching directions.

W3: Capability trade-offs and the stability of smaller improvements

A. Thank you for your feedback. Appendix B already discusses these capability trade-offs, and we acknowledge UniCorn's advantage on the generation metrics in Table 4. Our claim is that Prism improves understanding, generation, and interleaved generation together. Training on trajectories primarily reinforces semantic state progression conditioned on preceding text and images. Single-image generation from long instructions, character-level text rendering, and editing that preserves input pixels have different requirements. Text rendering, in particular, often benefits from specialized post-training, such as reinforcement learning, to improve stability and character-level accuracy. These capabilities are not directly reinforced by the current objectives, which may limit transfer. We will highlight both gains and regressions more clearly in the main text and distinguish proposed explanations from experimentally established observations.

To address your concern about the stability of smaller improvements, we train Prism with 3 different random seeds, using 16 A800 GPUs for 24 hours. We report the mean scores and sample standard deviations across these three independent training runs alongside the base model's scores:

ModelGenEvalDPGMMEMMBenchMMMUMMStarMMVP
Base85.886.12282.686.159.065.158.0
Prism89.20±0.26\mathbf{89.20 \pm 0.26}87.20±0.17\mathbf{87.20 \pm 0.17}2307.50±3.63\mathbf{2307.50 \pm 3.63}87.20±0.17\mathbf{87.20 \pm 0.17}60.70±0.36\mathbf{60.70 \pm 0.36}66.50±0.28\mathbf{66.50 \pm 0.28}62.00±0.70\mathbf{62.00 \pm 0.70}

On all seven benchmarks, Prism's mean score remains above the base model's score even after subtracting one sample standard deviation. This provides additional evidence that the improvements are consistent across the different seeds. We will include these statistics and the corresponding training settings in the revised manuscript.

Q1: Why use a frozen base model, and what are the quality and cost trade-offs of stronger independent VLMs?

A. Using a stronger VLM is an attractive option, but introduces additional computational resources, a distinction illustrated in Figure 2. We use a frozen copy of the base model as the judge to construct supervision from its existing capabilities without introducing an external teacher. To assess the benefits and costs of stronger independent VLMs, we add comparisons with Kimi 3.1 Pro and GPT Astra, replacing only the initial reliability judgment and verification after rewriting. All other settings remain unchanged, and each model is trained for one epoch. We first extend Table 14 as follows:

JudgeChainTPFPTNFNAcc.F1FDR
SenseNova-U1 (frozen)Text11318472280.0%85.0%13.7%
Image9517642479.5%82.3%15.2%
All208351114679.8%83.7%14.4%
Qwen3.7 PlusText1282441784.5%89.2%15.8%
Image877743280.5%81.7%7.4%
All215311153982.5%86.0%12.6%
Kimi 3.1 ProText1312144487.5%91.3%13.8%
Image976752286.0%87.4%5.8%
All228271192686.8%89.6%10.6%
GPT AstraText1301550590.0%92.9%10.3%
Image1035761689.5%90.7%4.6%
All233201262189.8%91.9%7.9%

We then evaluate Astra and Kimi as external judges. Only the initial reliability judgment and verification after rewriting are replaced. Rewriting is still performed by the frozen base model. All other training settings are unchanged, and each model is trained for one epoch:

JudgeGenEvalDPGMMEMMBenchMMMUMMStarMMVP
Frozen base model89.387.22307.587.260.766.562.0
Qwen3.7 Plus88.986.92304.187.460.766.559.3
Kimi3.1 Pro89.587.32315.487.559.767.162.0
GPT Astra90.287.82317.688.261.267.962.7

Kimi outperforms the base-model judge on most benchmarks, while GPT Astra achieves higher scores on all seven metrics. This shows that stronger external judges can further improve Prism, although the gains are not uniform across configurations and metrics. These improvements also come at a higher cost: over one training epoch, the API costs of Qwen3.7 Plus, Kimi 3.1 Pro, and Astra are approximately in a 1:7:37 ratio. The internal judge still requires inference compute but incurs no external API calls. We therefore use the frozen base model as the default setting without an external teacher, while external judges offer an option to improve performance at additional cost.

Thank you for taking the time to review our paper and for recognizing the value of letting either modality teach the other based on its reliability in each trajectory. We also appreciate your noting that the results across two backbones and the ablation studies support Prism’s approach to constructing training targets.

Thank you again for the care and effort you put into the review process.

Reviewer_zT4fOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Prism, a bidirectional self-teaching framework for post-training Unified Multimodal Models (UMMs). The method seperate the text chain and image chain from the interleaved trajectories. By evaluating each branch respectively, they rewrite those false part based on the fail reason and finally come into 4 strategies to update the model. The authors deploy Prism to SenseNova-U1 & BAGEL and get improvements on generation/understanding benchmarks.

Critical strengths

The paper introduces the reasonable framework to guide the rewriting direction of the failed text/image chain, instead of only use the original prompt.

Critical weaknesses
  1. The rewriting method actually results in rejecting those bad samples and generating good samples. Therefore the performance may come from those better training samples instead of the bidirectional rewriting. Could you please provide the ablation that discard false trajectories and resample it from the original prompts with the same max 3 times retry?

  2. From sec 4.6 & app. C, the authors compare Prism with those baselines. They only control the one-epoch optimization budget, but the full Prism will introduce additional generation and verification overhead due to the iterative rewriting. Could you provide the performance comparision under the same computing budget?

Questions for authors

Please see my weakness.

8 comments

Thank you for recognizing the value of using reliability feedback to guide rewriting. In response to your concerns about rejection and resampling and the overall computing budget, we have added experiments and summarize their settings and results below.

Additional experimental settings and results

The original experiments controlled the one-epoch optimization budget, but the overhead of online target construction differs across methods. Given the limited time and computational resources available during rebuttal, as well as differences in the processing and verification operations performed by each method, we use the same GPU time as a common computing budget and rerun the relevant comparisons. All settings use 16 A800 GPUs for 24 hours. This budget includes online rollout generation, reliability judgment, rewriting or resampling, verification, and parameter updates.

MethodGenEvalDPGMMEMMBenchMMMUMMStarMMVP
Base85.886.12282.686.159.065.158.0
Self-Training87.586.82278.485.759.464.558.0
Prism w/o Rewriting88.386.32299.486.860.865.760.7
Random Routing87.686.42296.386.559.465.759.3
Rejection and Resampling89.387.02293.385.959.165.358.0
Full Prism89.387.22307.587.260.766.562.0

W1: Bidirectional rewriting versus rejection and resampling

A. The rejection and resampling baseline accepts a trajectory only when both its text chain and image chain pass judgment. If either chain fails, it discards the entire trajectory and resamples from the original task prompt alone, without using the original trajectory or error feedback. In addition to the initial generation, it allows at most three retries, with a new judgment at each attempt. If all attempts fail, the trajectory is discarded without negative supervision.

The results show that rejection and resampling improve generation over the base model, but yield limited gains in understanding, with some metrics remaining unchanged or decreasing. In contrast, Prism achieves higher scores on all five understanding metrics while maintaining comparable or better generation performance. Thus, under the same computing budget, selecting training trajectories through rejection and resampling alone does not match Prism's combined performance in understanding and generation.

W2: Comparison under the same computing budget

A. After controlling the overall budget, Full Prism still outperforms Self-Training and Random Routing on all seven metrics, and Prism w/o Rewriting on six of seven, with nearly identical performance on MMMU. This shows that reliability-guided routing and rewritten targets retain their advantage when the additional overhead of online target construction is included. We will add these results to the revised manuscript and explicitly distinguish a matched one-epoch optimization budget from a matched overall computing budget.

Thanks for the authors’ detailed responses. But the new results also show that rejection and resampling achieve nearly the same generation performance to Full Prism, while the understanding improvements are partly retained without rewriting. Therefore, I remain unconvinced of the contribution of bidirectional rewriting. I will maintain my original rating.

Thank you for your thoughtful response. We have carefully considered your perspective. We believe our previous response may not have fully explained Prism's design and experimental analysis. Your comment appears to focus on rewriting as the paper's entire contribution. However, our design comprises four components, as described in Section 3.3, and our focus is the complete framework, not rewriting alone. We would therefore like to clarify your concerns through the following three points.

1. Both verification after rewriting and rejection and resampling involve filtering, but they differ in their use of reliable information and their efficiency.

We agree that rejection and resampling can produce effective training targets. However, Prism preserves the reliable modality chain and regenerates only the unreliable side before verification, whereas rejection and resampling regenerate the entire trajectory from the original prompt. The distinction therefore concerns not only whether a sample is ultimately accepted, but also the reuse of reliable information and sample-utilization efficiency.

Our experimental statistics show that, because rejection and resampling discard the entire trajectory whenever either the understanding or generation side fails, the final discard rate reaches 51.52%, leading to low sample-utilization efficiency. In contrast, our bidirectional rewriting approach has an average discard rate of 12.57% across all trajectories, demonstrating higher efficiency. Prism's mechanistic advantage is that it preserves the reliable chain and concentrates regeneration on the side requiring correction, avoiding the discarding and regeneration of information that is already usable.

2. The paper's core contribution is reliability-guided joint improvement of understanding and generation, rather than the rewriting operation alone.

The paper proposes selecting the teaching direction according to the modality-specific reliability of each trajectory and constructing corresponding supervision to improve understanding and generation together. Bidirectional rewriting is an important mechanism through which a reliable modality teaches the other. Therefore, comparable generation performance from rejection and resampling does not negate this contribution.

We understand your further concern to be whether resampling could replace bidirectional rewriting and achieve similar benefits. The additional results indeed show that rejection and resampling achieve generation performance comparable to Prism, and we acknowledge this observation. However, their understanding performance differs: under the same overall computing budget, Prism achieves higher scores on all five understanding metrics. Thus, similar generation scores alone do not establish equivalent joint improvements in understanding and generation.

Mechanistically, resampling searches for a new trajectory that passes verification using the original prompt. Prism instead adaptively selects the appropriate training action according to the reliability of the modalities in the current sample. This reliability-guided bidirectional self-teaching between understanding and generation, rather than generic sample filtering, is the paper's core contribution.

3. Retaining some understanding gains without rewriting is consistent with the overall design.  The w/o Rewriting control retains both-pass reconstruction and both-fail negative supervision, so some understanding gains are expected. Table 5 supports their contributions: replacing either objective lowers four of the five understanding metrics.  Rewriting nevertheless provides additional benefits. Under the same overall computing budget, Full Prism outperforms w/o Rewriting on six of seven metrics, with nearly unchanged performance on MMMU. Table 5 further shows that replacing image-to-text rewriting reduces MMVP from 62.0 to 56.7, while replacing text-to-image rewriting reduces GenEval from 89.2 to 88.0. These results support the complementary contributions of the four components. Our contribution is therefore the complete reliability-guided self-teaching framework that jointly improves understanding and generation, rather than the rewriting operation alone.

We hope these clarifications help you reassess our design and experimental results. If any questions or concerns remain, please feel free to raise them, and we will be happy to address them and provide further clarification.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.