{"id":"eded0b48-a867-46e8-8320-f486acd1c636","arxiv_id":"2607.03748","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":7.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"BRAID jointly optimizes text and image generation in interleaved multi-modal reasoning by casting the full trajectory as a unified MDP with shared advantages and a VLM process reward.","lead":"BRAID turns multi-turn text-image-text reasoning into one Markov decision process so a single RL objective can train both the language tokens and the image denoising steps of a unified multi-modal model. The result is a practical way to let visual intermediates actually receive credit for helping the final answer, which matters for spatial and perception tasks that pure text struggles with.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The central credit-assignment claim rests on an unvalidated proprietary VLM judge whose scores are never shown to track actual reasoning utility.","rationale":"The reader correctly isolates the external VLM judge as the weakest link in an otherwise coherent empirical construction. The unified-MDP + modality-native gradient story is technically sound on its own terms (shared Â, GRPO for text, DiffusionNFT soft-label mapping for images), the ablations cleanly separate L_NFT from r_vis, and the Maj@n curves give independent evidence that the policy’s solution space expands. Those elements survive scrutiny. The load-bearing gap is precisely the one the reader named: without any validation that the four hand-chosen GPT-5.2 scores track genuine reasoning utility, the “essential” vision-thinking component remains an unexamined black box. That keeps the paper in CONDITIONAL territory until an open, reproducible process reward (or at least a correlation study) appears; it does not warrant a harsher verdict. No stronger internal inconsistency or missing derivation was found.","tokens_in":22424,"tokens_out":537,"duration_ms":18143,"concrete_test":"On a held-out set of 300 intermediate images drawn from the RL rollouts, collect independent human ratings on the same four criteria plus a binary “did this image help reach the correct answer?” label; compute Spearman ρ between GPT-5.2 scores and human labels (especially r_ru and the binary utility label). If ρ < 0.4 on reasoning utility, the dense reward is unreliable and the claim that it sharpens critical visual branches is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 4.4 and Eq. (9)–(10) treat the four GPT-5.2 scores (visual correctness, fidelity, reasoning utility, trustworthiness) as reliable dense advantages that correctly identify which intermediate images matter for the final answer. The shared trajectory-level advantage is then simply added to these scores for image turns. No correlation, human agreement, or consistency check is reported between the judge’s r^(ru)/r^(tr) and either terminal correctness or human judgments of utility. If the judge is noisy, systematically biased toward aesthetic fidelity, or itself hallucinating, the “vision-thinking guidance” that the abstract and conclusion call essential becomes an opaque distillation signal rather than a principled credit-assignment mechanism. The ablation drop when r_vis is removed shows the signal is used, but does not establish that it is the right signal.","agreement_with_reader":"agree"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes BRAID, a framework that casts multi-turn interleaved text-image-text (TIT-CoT) reasoning in unified multi-modal models as a two-level Markov decision process, so that text and image turns can be optimized jointly under one RL objective. A shared trajectory-level advantage is propagated through modality-native surrogates—GRPO-style clipped policy gradients for autoregressive text tokens and a DiffusionNFT-style reward-aware flow-matching loss for image denoising—while a VLM judge supplies a four-dimensional vision-thinking process reward to densify credit assignment on intermediate images. Built on BAGEL-7B and trained with an SFT-then-RL pipeline, BRAID reports +5.73 average points over BAGEL across seven spatial-reasoning and visual-perception benchmarks, with large gains on SAT and V*Bench, ablations isolating LNFT and r_vis, Maj@n scaling curves, data-allocation and rollout-ratio studies, and qualitative transfer to standalone text-to-image generation.","tokens_in":22668,"tokens_out":950,"duration_ms":18399,"significance":"If the results hold, this is a meaningful step for post-training of hybrid AR-diffusion UMMs: prior work either freezes image generation during RL (COOPER, DeepEyes) or jointly optimizes only single-turn text-to-image (UniGRPO). The two-level MDP factorization and the explicit channeling of one advantage into modality-native gradients are clean and reusable design choices. The experimental package is stronger than average for this area—component ablations, sampling-budget scaling, data-regime and batch-ratio analyses, and case studies that separate correct answers from faithful visual evidence. Limitations are stated honestly (hybrid backbone only; external VLM judge; fixed interleaving). The main significance risk is that part of the claimed “vision-thinking guidance” may be opaque distillation from a proprietary judge rather than validated credit assignment; even so, the joint image-branch optimization result would remain of independent interest.","major_comments":[{"comment":"Section 4.4 and Eqs. (9)–(10): the vision-thinking process reward is load-bearing for the abstract’s and conclusion’s claim that “vision-thinking guidance is essential,” yet no validation is reported that the four GPT-5.2 scores (especially r^(ru) and r^(tr)) correlate with terminal correctness, human judgments of intermediate utility, or even inter-run consistency of the judge. The ablation drop when r_vis is removed (Table 4: 65.19 → 62.68) shows the signal is used, not that it is the right signal. Please add at least one of: (i) correlation of per-turn r_vis components with final answer correctness within groups, (ii) human agreement on a scored subset, or (iii) a controlled swap/noise study of the judge scores. Without this, the process-reward half of the central narrative remains an untested axiom rather than a demonstrated credit-assignment mechanism.","section":null},{"comment":"Table 1 and Section 5.1: main results and the +5.73 avg. claim are reported from single runs with no error bars, confidence intervals, or multi-seed statistics. Several absolute margins (e.g., CV-Bench 3D −1.24, BLINK +3.00) are small enough that seed variance could reorder methods. Given that the paper’s empirical case for “consistently outperforms” rests on these numbers, please report mean±std over at least 3 seeds for BRAID, SFT, and the two ablations on the seven-benchmark average (and ideally on SAT and V*Bench).","section":null},{"comment":"Section 4.3, Eqs. (7)–(8): the mapping from shared advantage  to the soft NFT label r_k = σ(Â/υ) and the scalar balances λ, λ_vis are introduced without sensitivity analysis. Table 3 lists many free parameters (β=0.4, clip ratios, CFG scales, equal 0.25 component weights, etc.), yet only rollout-to-update ratio and data regime are ablated. At minimum, report sensitivity of the seven-benchmark average to λ_vis and to the temperature υ (or to equal vs. learned component weights in Eq. (9)), so that the joint objective is shown to be robust rather than tuned to a single operating point.","section":null}],"minor_comments":[],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The real news is a clean joint RL loop for multi-turn interleaved text-image reasoning on a hybrid UMM. Prior work either freezes the image branch (COOPER, DeepEyes) or does joint RL only for single-turn T2I (UniGRPO). BRAID puts GRPO on text and DiffusionNFT on flow-matching images under one shared trajectory advantage, plus a four-criterion VLM process reward for image turns. That combination is new and the paper executes it carefully.\n\nWhat works: the two-level MDP is a sensible abstraction, the modality-native surrogates avoid scale collapse, and the ablations actually isolate the claims. Dropping LNFT hurts more than dropping r_vis; Maj@n keeps climbing while SFT/BAGEL plateau; the data-allocation and rollout-ratio studies are honest. Gains are real for a 7B model (+5.73 avg over BAGEL, big jumps on SAT and V*Bench) and the qualitative cases show the image branch producing usable intermediates rather than decorative noise. Math and citations look fine; no circular reward definition.\n\nSoft spots, in proportion: the load-bearing process reward is GPT-5.2 scores with no correlation, agreement, or consistency check against terminal correctness or human utility. The ablation shows the signal is used, not that it is the right signal. That is a real gap for the “vision-thinking is essential” claim, but it does not sink the joint-optimization result. Other ordinary limits: single backbone, fixed interleaving, no seeds/error bars, proprietary judge, no code. CV-Bench 3D still dips a bit.\n\nThis is for people doing multi-modal post-training and agentic visual reasoning. Worth a serious referee. I would engage with it and expect the judge to be replaced by an open PRM in follow-ups.","headline":"Solid joint-RL recipe for multi-turn text-image CoT; the VLM-judge credit signal is useful but unvalidated, not fatal.","tokens_in":23342,"tokens_out":475,"would_cite":true,"duration_ms":4756,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Interleaved text-image reasoning works best when both modalities are trained together as one decision process under a single RL objective.","keywords":["unified multi-modal models","interleaved text-image reasoning","reinforcement learning","Markov decision process","flow matching","vision-thinking process reward","spatial reasoning","policy gradient"],"falsifier":"Train the identical joint-MDP setup with the image branch frozen or with the VLM process reward replaced by pure terminal reward only, and check whether the large gains on SAT and V*Bench disappear and whether majority-vote accuracy still scales with more samples; if joint image optimization and the dense judge scores add nothing under that controlled ablation, the central claim fails.","tokens_in":23274,"feed_emoji":"🔗","tokens_out":965,"duration_ms":10889,"temperature":0.7,"pith_summary":"Unified multi-modal models can alternate verbal thinking with intermediate image generation, but most RL methods only train the text steps and leave image generation outside the loop. This paper claims that treating the full multi-turn text-image-text trajectory as one Markov decision process lets a shared reward signal improve both branches at once. A common trajectory-level advantage is fed into text via ordinary token policy gradients and into images via a flow-matching style update, while a vision-language judge scores each intermediate picture for how much it actually helps the final answer. On spatial reasoning and visual perception tests, the resulting 7B model gains several points over its base and beats a strong proprietary system on several hard spatial tasks. The practical point is that half-optimized RL leaves the visual half of interleaved reasoning undertrained, and closing that loop measurably raises the ceiling of what the model can solve.","feed_headline":"Joint RL on text and images lifts multi-modal reasoning","feed_subtitle":"A shared trajectory advantage plus visual process scores lets a 7B model beat stronger systems on spatial tasks","key_machinery":"BRAID: a unified MDP over interleaved turns that shares one trajectory-level advantage and back-propagates it separately through GRPO-style token updates for text and DiffusionNFT-style flow-matching losses for images, augmented by a four-criterion vision-thinking process reward from an external VLM judge.","core_discovery":"The paper establishes that casting multi-turn text-image-text chain-of-thought as a single two-level MDP, computing one shared trajectory advantage, and routing it through modality-native gradients (clipped policy gradient on text tokens and reward-weighted flow-matching on image denoising) plus dense turn-level scores from a VLM judge is necessary for effective interleaved multi-modal reasoning; without joint image-branch optimization the gains collapse, and the method lifts a 7B hybrid model by +5.73 average points across seven benchmarks, with large jumps on SAT and V*Bench.","pith_inferences":["If the external judge is the main source of dense signal, replacing it with a cheap on-policy process reward model would remove a costly dependency and make the method more reproducible.","The same shared-advantage pattern should apply to fully autoregressive unified models that emit discrete image tokens, not only hybrid AR-diffusion backbones.","Adaptive rather than fixed interleaving patterns could further raise sample efficiency once credit can already flow through both modalities.","The approach implies that any multi-modal CoT system whose visual steps are only supervised is leaving a large fraction of its RL potential unused."],"forward_implications":["Policy gradients can cross the text-image boundary end-to-end, so intermediate images become trainable reasoning steps rather than frozen decorations.","Dense turn-level visual scores reduce long-horizon credit assignment failure on multi-step spatial and perception tasks.","A 7B hybrid model can surpass larger understanding-only systems on viewpoint-heavy and fine-grained visual benchmarks once both modalities are jointly optimized.","Improvements in the image branch transfer to standalone text-to-image generation, improving spatial composition and count fidelity.","Half-optimized RL that freezes image generation is strictly weaker than full-trajectory optimization on the same backbone."],"fun_headline_variants":["Shared trajectory RL unifies text-image multi-modal reasoning","Joint text-image MDP with VLM scores lifts 7B interleaved reasoning","One shared advantage + visual scores enable full multi-modal RL","BRAID routes shared advantage through text tokens and image denoising","Unified MDP lets joint policy gradients raise spatial multi-modal gains"],"cache_read_input_tokens":16512,"weakest_assumption_plain":"An external vision-language model scoring intermediate images on four fixed criteria will reliably mark which visual steps actually help the final answer rather than reward superficial or misleading pictures.","fun_headline_variants_meta":{"raw":{"variants":["Shared trajectory RL unifies text-image multi-modal reasoning","Joint text-image MDP with VLM scores lifts 7B interleaved reasoning","One shared advantage + visual scores enable full multi-modal RL","BRAID routes shared advantage through text tokens and image denoising","Unified MDP lets joint policy gradients raise spatial multi-modal gains"]},"model":"grok-4.5","effort":"low","cost_usd":0.001772,"raw_usage":{"total_tokens":934,"prompt_tokens":862,"num_sources_used":0,"completion_tokens":72,"cost_in_usd_ticks":17720000,"prompt_tokens_details":{"text_tokens":862,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":0,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":862,"tokens_out":72,"duration_ms":1247,"temperature":1.0,"reasoning_tokens":0,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-12T00:13:16.693241+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Train the identical joint-MDP setup with the image branch frozen or with the VLM process reward replaced by pure terminal reward only, and check whether the large gains on SAT and V*Bench disappear and whether majority-vote accuracy still scales with more samples; if joint image optimization and the dense judge scores add nothing under that controlled ablation, the central claim fails.","supporting_citations":[],"review_version":1}