{"id":"fd41766c-b0b5-451b-9235-23e4e59f2806","arxiv_id":"2607.17140","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Treating image-editing as a shared target between understanding and generation, with sequential reinforcement learning, improves a unified multimodal model's caption-image consistency and several benchmark scores.","lead":"STBridge is a post-training method that makes a single multimodal model better at both describing and generating the same edited image. It improves scores on several visual-understanding, generation, and editing benchmarks over its starting model.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Uncertainty: alignment-metric gains may partly reflect judge reward hacking; need human or independent-judge replication of Table 7.","rationale":"The reader's weakest assumption — that VLM judges are faithful proxies for human semantic consistency — is precisely the load-bearing concern. I share it because the paper's central evidence (Table 7) is scored by a VLM judge, and the training pipeline uses VLM judges for both target-caption generation (Sec. 3.1) and both RL rewards (Secs. 4.3-4.4). The circularity is not complete: the judge that measures the headline metric (Gemini-3.1) differs from the judges used in training (Qwen-35B and Qwen-122B), which weakens the pure reward-hacking concern. However, the risk does not disappear: all three are large VLM judges of image-text consistency, and the paper reports no human agreement study for any of them. In addition, the evaluation numbers in Tables 1-6 are mostly standard multiple-choice benchmarks with fixed ground truth, so those gains are more robust; the weakest part of the central claim is specifically the alignment analysis and the judge-dependent WISE/ImgEdit/RISE scores. The reader also identified the absence of error bars and significance tests (Tables 7 and 1-6), and the paper's own negative sub-scores (e.g., WISE Cultural -9.00, Time -0.33; ImgEdit Style -0.12, Compose -0.31, Action -0.15) undermine the abstract's 'consistently improves' phrasing. My concrete test — a held-out judge or human subset — directly settles whether the Table 7 gain is genuine. If the gain survives an independent judge, the central claim is much stronger; if not, the paper needs to re-validate its evaluation protocol. Thus I keep a CONDITIONAL verdict rather than upgrading to ACCEPT.","tokens_in":22560,"tokens_out":1847,"duration_ms":15439,"concrete_test":"Re-run the Sec. 5.5 alignment analysis on the same 400 editing pairs with a held-out judge from a different family (e.g., GPT-4o or a human-annotated subset of 100 pairs) and compute agreement/correlation with the Gemini-3.1-Flash-Image-Preview scores. If the held-out judge does not reproduce the 0.5740→0.9000 ordering and rough magnitude, then the headline alignment gain is at least partly an artifact of judge-specific reward hacking. Also report per-condition score distributions (not just means) and a bootstrap confidence interval for the Table 7 gains.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The paper's central claim is that STBridge closes the understanding-generation gap. The most direct evidence is Table 7, where caption-image consistency rises from 0.5740 (BAGEL) to 0.9000 (STBridge-RL), as judged by Gemini-3.1-Flash-Image-Preview. But every stage of STBridge is supervised by the same family of VLM judges: Qwen3.5-35B-A3B writes target captions in Sec. 3.1, Qwen3.5-122B-A10B computes both the Diff Reward (Eq. 6) and the realization reward (Eq. 10) in Secs. 4.3-4.4, and Gemini-3.1-Flash-Image-Preview scores the alignment metric in Sec. 5.5. No human agreement study, no independent judge, and no error bars are reported. If these VLMs share systematic biases—for example, preferring verbose scene-level captions, or rewarding template-like phrasing—then RL can increase Table 7 scores by satisfying the judge rather than by making the model's text and images converge on the intended target semantics. The paper's own ablations (Tables 8-9) show that optimization against judge-based rewards changes benchmark scores without changing held-out semantics in a verified way; the absence of a judge-independent probe leaves the central claim underdetermined. This is a correctness risk, not an internal inconsistency: the training and evaluation signals are not identical (Qwen vs. Gemini), but they are all VLM text-image judges with no demonstrated human correlation.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies the semantic alignment between the textual and visual outputs of unified multimodal models (UMMs) in the image-editing setting. It first diagnoses an 'understanding–generation alignment gap' in a baseline UMM (BAGEL) using a VLM-based consistency protocol, then proposes STBridge, a three-stage post-training procedure: Shared-Target Alignment (SFT on target captions plus target-image flow-matching), Target-Expression Optimization (GRPO with a VLM-based Diff Reward on captions), and Target-Realization Optimization (MixGRPO with a VLM-based editing reward on generated images). The authors report improvements over BAGEL on BLINK, CVBench, OddGridBench, MMVP, VisPuzzle, MM-Vet, WISE, ImgEdit, and RISE, and a large increase in caption–image consistency from 0.5740 to 0.9000 on a custom 400-pair alignment evaluation. They conclude that anchoring understanding and generation to a shared target state narrows the gap between what the model describes and what it generates.","tokens_in":22952,"tokens_out":4366,"duration_ms":44237,"significance":"If the central claim held, STBridge would be a useful and fairly general post-training recipe for UMMs: it converts visual-understanding data into an editing-style interface, uses a shared target caption to condition generation, and applies sequential RL to improve both expression and realization. The paper also ships a diagnostic protocol (Table 7, Fig. 6) that could be reused by the community. The strongest positive evidence is the broad benchmark coverage, the qualitative examples, and the ablation separating source-caption, target-caption, and shared-target SFT. However, the manuscript's own tables contain several sub-score decreases, and all the key rewards and the main alignment metric are produced by VLM judges, with no human agreement study. Those two issues are load-bearing for the abstract's and conclusion's claims of 'consistent improvements' and 'narrowing the gap.' The paper is therefore promising but requires substantial additional analysis before the central claims can be considered established.","major_comments":[{"comment":"The abstract, introduction, and conclusion repeatedly claim 'consistent improvements,' but the benchmark tables show multiple decreases at the sub-score level. For example, Table 4: WISE Cultural drops by -9.00 and Time by -0.33; Table 5: ImgEdit Style -0.12, Compose -0.31, Action -0.15; Table 1: BLINK Count. Rel. -0.84, Rel. Depth -0.66, Multi-view Rel. -0.67; Table 2: CVBench-2D Count -1.27. No error bars, confidence intervals, or significance tests are reported anywhere, so it is impossible to tell which of the aggregate gains (e.g., BLINK +5.15, RISE +21.00) are robust versus noise. Please either provide repeated-run variance and significance testing, or temper the 'consistent improvements' claim to 'improvements on the averages of most benchmarks' and explicitly discuss the sub-score regressions.","section":"Tables 1–6 and Sec. 5.3"},{"comment":"The WISE evaluation is performed with Qwen3.5-35B-A3B as the judge (Sec. 5.2), and the same model is used to generate the SFT target captions in Sec. 3.1. This creates a concrete judge-overlap risk: the gain on WISE (+1.20 overall, +21.50 on Biology) may reflect optimization toward the preferences of the captioning/judging model rather than genuine world-knowledge-informed generation. The manuscript does not provide any independent judge or human agreement analysis for WISE. Please add an evaluation with a different VLM family and/or a human-annotated subset, and report per-dimension agreement.","section":"Sec. 5.2 vs Sec. 3.1 (WISE evaluation)"},{"comment":"The central alignment claim—that STBridge raises caption–image consistency from 0.5740 to 0.9000—is based solely on Gemini-3.1-Flash-Image-Preview as the judge. The training signals (target captions in Sec. 3.1, Diff Reward in Sec. 4.3, realization reward in Sec. 4.4) come from other VLM systems (Qwen3.5-35B and Qwen3.5-122B). All these VLMs are text-image judges, and no human agreement study or independent-judge replication is reported. If these VLM judges share systematic biases—e.g., favoring verbose scene-level descriptions or template-like phrasing—the reported alignment gains could be inflated by satisfying the judge rather than by genuine convergence of textual and visual semantics. Please provide a human evaluation on a subset of the 400 pairs, an independent VLM judge, and per-item error analysis. Also specify how the 400 pairs were sampled.","section":"Sec. 5.5, Table 7"},{"comment":"The Diff Reward depends on the function Align(·,·) and on an 'offline annotated reference difference' Δ*. Neither is defined anywhere in the paper. It is not specified whether Δ* is human-annotated or generated by a VLM, nor how the alignment between the predicted difference and Δ* is computed (exact string match? semantic similarity? another VLM call?). Since this reward is the sole training signal for the target-caption policy, the lack of specification is load-bearing. Please provide the exact construction of Δ*, the Align function, and the prompt/format used for DiffVLM, or release the code/config.","section":"Sec. 4.3, Eq. (6)"},{"comment":"The realization reward RG = b·g combines five VLM rubric scores (Eq. 7) with a gate g (Eq. 9). The gate parameters α=0.1 and β=0.9 are stated but never ablated, and the VLM judge used to produce the five rubric scores is the same Qwen3.5-122B-A10B family that provides the Diff Reward. No evidence is provided that these rubric scores correlate with human preference or with benchmark scores. Given the large RISE jump (+21.00), the reader cannot distinguish a genuine improvement in editing ability from reward overfitting to the VLM rubric. Please ablate the gate components and validate the reward against human judgments or a held-out independent benchmark.","section":"Sec. 4.4, Eq. (10) and Sec. 5.3"}],"minor_comments":[{"comment":"The conclusion that 'the gain comes from coupling target expression with target realization' is based entirely on the BLINK benchmark. This is a single benchmark and the caption-only variants (Source-Caption SFT, Target-Caption SFT) are trained with the same recipe? If they use different numbers of steps or different data, please state it. At minimum, temper the claim or extend the ablation to at least one generation/editing benchmark.","section":"Sec. 5.5, Table 8"},{"comment":"Please clarify the token limits: the first paragraph says 'global and per-sample token limits are both set to 35,000,' while the RL paragraph says 'maximum token budget of 32,768.' Are these different settings for SFT vs RL, or an inconsistency?","section":"Sec. 5.1"},{"comment":"The tables are hard to read because several numbers run together (e.g., '76.9230.77' in Table 1, '71.8386.77' in Table 2). Please add proper spacing or vertical lines, and ensure column headers align with the data.","section":"Tables 1–3"},{"comment":"The alignment protocol in Fig. 6 lists seven metrics, but Table 7 reports only six (Global Scene Alignment, Entity Presence, Entity Attribute, Entity Part Attribute, Spatial Relation, Inter-Attribute Relation) and an Overall. Is 'Fine-grained Avg.' the average of the five fine-grained ones? Please define the relationship clearly and state whether the 'Global Scene Alignment' is the same as the 'Global Scene Alignment' in Fig. 6.","section":"Fig. 6 / Table 7"},{"comment":"The conversion of visual-understanding data into editing-style instructions via τ(q) and ρ(I,q,a) is described at a high level. The examples in Fig. 11 are helpful, but the exact templates for τ and ρ are not given. Please provide the full prompt templates or point to the released code/data.","section":"Sec. 3.2"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically interesting and the proposed framework is well-motivated, but the central 'consistent improvements' and 'alignment narrowing' claims are currently undermined by (i) sub-score decreases without statistical testing and (ii) a pervasive dependence on VLM judges for training rewards and the main evaluation metric. I believe these issues are addressable with substantial additional experiments (error bars/significance tests, human or independent-judge evaluation on a subset, exact specification of the Diff Reward/Align function), so I recommend major revision rather than rejection. The referee should also check whether the WISE judge overlap is disclosed sufficiently; in the current text it is present but the implication is not discussed."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The shared-target idea is genuinely new and the paper does something useful: it treats the target caption and edited image as two modal expressions of the same target state, and trains them sequentially with RL. The ablations are the best part—Table 8 shows that caption supervision alone hurts, and coupling caption with image realization gives the gain; Table 9 shows the Diff Reward matters. That is real evidence the mechanism is doing something, not just extra data.\n\nThe soft spot is the word 'consistently.' The paper's own tables contradict it: WISE Cultural -9.00, Time -0.33, ImgEdit Style -0.12, Compose -0.31, Action -0.15, and several BLINK sub-scores drop. The averages improve, and some gains are large, but 'consistent' is false. Also no error bars or significance tests anywhere.\n\nThe bigger issue is the alignment metric. Table 7 is the direct evidence for the central claim, but the judge is Gemini-3.1-Flash-Image-Preview, the training captions come from Qwen3.5-35B-A3B, and the RL rewards come from Qwen3.5-122B-A10B. No human agreement study, no independent judge. That means the 0.574→0.900 jump could partly reflect the model getting better at satisfying VLM text-image preferences rather than true semantic convergence. It's not fatal—the external benchmarks with fixed answers (BLINK, OddGridBench, MMVP) are independent and improve—but the alignment claim is underdetermined as stated.\n\nI'd also like code/data, but that's not a correctness flaw.\n\nOverall: this is a credible post-training recipe, worth a serious referee. It needs revision, not rejection. The authors should drop 'consistently,' add error bars, and validate Table 7 with a human or at least a different judge family. I'd bring it to reading group and would cite if it comes out in revised form.","headline":"A genuinely useful shared-target post-training recipe, but the consistency claim is overstated and the VLM-judged alignment metric needs external validation.","tokens_in":23443,"tokens_out":2533,"would_cite":true,"duration_ms":25679,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"STBridge's three-stage shared-target training lifts a unified multimodal model's caption-image consistency from 0.574 to 0.900 while also improving understanding, generation, and editing benchmarks.","keywords":["unified multimodal models","understanding-generation alignment","shared-target post-training","image editing","target captioning","diff reward","reinforcement learning","caption-image consistency"],"falsifier":"Ask human raters to score whether the target caption and edited image match on the paper's 400 alignment pairs, and compare human scores with the VLM judge's 0.574-to-0.900 gap. If humans see little or no improvement, the alignment gain is an artifact of the judge.","tokens_in":22451,"feed_emoji":"🎯","tokens_out":4837,"duration_ms":48396,"temperature":0.7,"pith_summary":"Unified multimodal models aim to both understand and generate images in one network, but a model can describe an edit correctly and still draw something else. The paper claims this understanding-generation gap exists because the two outputs are trained as separate tasks rather than as two views of the same target result. It proposes STBridge, a post-training recipe built around a shared target state: given a source image and an instruction, the model must first produce a 'target caption' that spells out the desired result, then generate the edited image that matches that caption. Three training stages—supervised alignment, then reinforcement learning on the caption, then reinforcement learning on the image—progressively close the gap, raising measured caption-image consistency from 0.574 to 0.900 and improving the model across understanding, generation, and editing benchmarks.","feed_headline":"Training around one shared target lifts text-image match to 0.90","feed_subtitle":"A three-step post-training recipe makes unified multimodal models describe exactly what they generate, improving editing and understanding b","key_machinery":"The central mechanism is the shared-target state in the image-editing interface: for a source image and instruction, a target image anchors both modalities. The understanding path produces a target caption that expresses that target; the generation path produces the edited image that realizes it. Training proceeds in three stages: (1) Shared-Target Alignment, supervised fine-tuning on tuples of source, instruction, target caption, and target image, using a cross-entropy caption loss plus a flow-matching image loss; (2) Target-Expression Optimization, which uses the Diff Reward—a VLM infers the edit implied by a candidate caption and compares it with the annotated source-to-target difference—","core_discovery":"The paper's central claim is that architectural unification alone does not make a unified multimodal model semantically consistent: under the same source image and edit instruction, the model's textual description of the intended result and its generated edited image can diverge in fine-grained entities, attributes, spatial relations, and local details. STBridge treats the target image as a shared semantic anchor, so that target captioning and image editing become two modal expressions of the same target state rather than independent tasks. The method uses a three-stage align-then-optimize procedure: supervised fine-tuning first binds the target caption to the target image, then a Diff Rewar","pith_inferences":["Editorial extension: the shared-target interface could be applied to multi-turn or video editing, where each step's target state becomes the next step's source.","Editorial inference: since target captions, RL rewards, and the alignment judge are all VLM outputs, the 0.900 alignment score should be tested against human ratings; otherwise the gain may partly reflect conformity to the judge's preferences.","Editorial extension: a strong robustness test would be to swap in a different judge model at evaluation and see whether the 0.574-to-0.900 improvement persists."],"forward_implications":["If the central claim holds, architecture alone is not enough to make understanding and generation semantically consistent; post-training around a shared target is a viable lever.","Target captions act as a controllable intermediate representation: improving the caption policy first, then freezing it while improving image realization, yields better final edits than joint or independent optimization.","Visual-understanding questions can be converted into editing-style transformations, letting one data interface train both capabilities.","Reward shaping that focuses on the edit-relevant difference, rather than global image-text similarity, is what makes the expression stage work.","Consistent gains on perception, generation, and editing benchmarks imply that shared-target post-training may transfer to other tasks within unified multimodal models."],"fun_headline_variants":["One shared target unifies what models say and generate","STBridge: one shared target for understanding and generation","Single anchor aligns what UMMs describe and generate","Close the gap: model must say and show the same target","STBridge: shared target for text-image consistency"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The whole pipeline presumes that VLM-written captions, VLM-computed rewards, and a VLM judge measure true semantic agreement; if those models share systematic blind spots, the reported alignment gain may measure agreement with the judge rather than with human intent.","fun_headline_variants_meta":{"raw":{"variants":["One shared target unifies what models say and generate","STBridge: one shared target for understanding and generation","Single anchor aligns what UMMs describe and generate","Close the gap: model must say and show the same target","STBridge: shared target for text-image consistency"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000609,"raw_usage":{"total_tokens":2707,"prompt_tokens":812,"completion_tokens":1895,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1818}},"tokens_in":556,"tokens_out":1895,"duration_ms":13972,"temperature":1.0,"reasoning_tokens":1818,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T18:52:50.151738+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Ask human raters to score whether the target caption and edited image match on the paper's 400 alignment pairs, and compare human scores with the VLM judge's 0.574-to-0.900 gap. If humans see little or no improvement, the alignment gain is an artifact of the judge.","supporting_citations":[],"review_version":1}